Claude Duet | Hacker News Reader

Claude Duet

claude.com

568 pointskuberwastaken539 comments

Man these big AI labs have such terrible UX and integration. Signing up for this waitlist should be something you can request from your Claude chat session. You already have all of my relevant information and I'm already signed in, why are you making me type in any information. It's bizarre we have companies claiming trillion dollar valuations but somehow nobody thought to do basic integration with the existing chat service? You have access to frontier coding agents that should make this trivial and somehow you're still expecting me to type in information I've already given you. What are the product people at Anthropic doing?

I tried asking Claude fron the android app if it could sign me up for the waitlist and it said that it couldn't. There's no technical reason this shouldn't be possible, it just shows a lack of care and attention to detail.

I suspect such basic UX & integration misses is parallel to the general trend of customers somehow getting narrower services, a move away from supporting power users, and toward companies functioning more like fiefdoms where customers are subjects to be farmed.

Drives me nuts that it's such a manual process to link my ChatGPT/Codex machines individually to each other and my Dot can only seem to connect with the chats on one linked machine at a time (before even getting to control or how completely useless the cloud VPS for the Dot is).

I had a Pro 200 plan build my entire desktop environment from near scratch in C+Vulkan this last month and they can't even bother to have the bots let their Linux Electron app connect remotely to the same Electron app on Windows in the same way that already happens in reverse?

give us some more details on this awesome c+vulkan desktop environment! Screenshots? will you publish it? very curious!

Ha, sure! It's more a "zen sandbox" than something I intended to ever publish... but you never know - after all I didn't set out intending to make a vibed DE either. At the same time, I'd encourage others interested to try making their own zen sandbox instead - you'd probably enjoy it a lot more than mine! In spirit of the whole thing being vibed, here's a low-effort AI driven gallery/summary of the current version to take a quick peak with: https://zamadatix.github.io/dandesk-showcase/#overview (this focuses on the main GUI components but not the service/ephemeral components).

And in spirit of HN, here is the extremely long (as I've been meaning to write this up for a few friends too) human written description of what led to it/why it's my zen sandbox/what I've liked so much about it. The short answer, if one does not wish to read a wall of text about someone's vibe coded pet project, is I can live tweak aspects of my DE in the background during the workday but I've never actually spent an ounce of energy looking at a single line of the code (I've got enough of that elsewhere!). E.g. my desktop icons replicate across my monitors so I can always access them from any screen. Small shit like that is just fun to have "appear" during some boring work call. It's also pretty efficient, using less RAM+VRAM than LXQT/KDE/Gnome (though I don't claim feature parity with those either). The downsides are it's not meant to be the best archictecure, the most secure architecture, or have features anyone but I wanted to use. Parts I don't use much also haven't gotten as much work, e.g. RemoteViewer is still just an MVP with only (custom built) VNC support while the Start Menu functionality is pretty much 100% from what I wanted to do. It's designed to be a zen sandbox.

I'd also preface all of the OS configuration has gone into a monorepo which contains the DE as one component and I made sure to have regular backups to my external server. This is in anticipation at some point it'll just delete everything when it gets stuck or somehting, but it has surprisingly been very well behaved so far and I haven't had to do more than run test restores of anything yet.

Longer (still human) description:

I started my yearly "install Linux on my main personal PC, put a few weeks of work into tweaking it, regret it, and roll back to Windows" except this time I decided to just give AI full reign in initial set up and troubleshooting. I usually go Arch (btw) with Sway because my favorite thing on all of my Linux PCs vs my Windows one is I can build it up to what I want rather than down from some enormous blob of shit I don't care about (which, inevitably, I end up breaking or getting limited by because there is some cross dependency built into the system I didn't account for). This was great, it took a lot of the "and how the fuck do I configure Pipewire to do hide the unfiltered mic?" type friction from the start.

Straight away though I ran into some issues with my software KVM (drag your mouse through the edge of the monitor and your mouse+keyboard goes to the next machine until you drag it back) since it expected Gnome/KDE Wayland extensions to do it and Sway doesn't have those. I didn't really want to go down the road of building half my programs from scratch (or trying to find who else already is on AUR and use that kludge) so I asked the LLM if there was a way to write a plugin. It suggested Hyprland would both be easier to write such a plugin for and be a bit more bleeding edge for some of the other wlroots features I was looking for too. Not 30 minutes later, Synergy was working!

I then took this another step forward and added the new Wayland extension for SDR content brightness in HDR. By default it put SDR content at 201 nits, and I usually set it to ~480 in my bright office (lower at night). Another quick addition to the hyprland plugin and this too was working without having to build the whole kit and kaboodle that is a standard pacman dependency tree with only the official repos (and a small, easier to maintain, plugin).

This was about the point I realized this was kind of relaxing/fun. Instead of fight some nitpick detail all day and come out half defeated, I could just do my thing while the LLM tried it in the background. So I decided, "hey, what about a few really lightweight tray apps for things like managing and connecting to containers or triggering different types of screenshots/recordings?". Then I replaced waybar, and the launcher, and maybe a few other things. The key rule which made it a zen sandbox was "low abstraction" and "avoid 3rd party libraries wherever reasonably possible". E.g. I was fine pulling in a unicode library, a font rendering library, and system integration libraries for vulkan/systemd/dbus/etc - if I ever really wanted I could replace those later but it was more the "why does this tray app take 120 MB to show me the clipboard?" type abstraction I wanted to avoid, not a custom OS.

Around this time, I realized I was wasting tokens/time asking it to do the same things. I.e. "make an input box which...", "theme this like...", and so on. I figured it'd be a lot more fun to make these kinds of things into microservices. Being a zen sandbox, I decided to go the slightly impractical approach of a single process with a bunch of threaded microservices which communicate over the JSON message bus. At first I wasn't sure if this would just be a crashing mess being a single C process, but it has actually managed to do extremely well with both hardening and isolating what it can in that bounds. The theme service, the image encoding/decoding service, the clipboard service, and so on. The biggest help was the common UI toolkit, written from raw C + Vulkan. After having it build up a library of UI elements overnight the speed of making graphical components went up drastically.

At this point it was no longer the weekend, but it was all so stable and the development so isolated I kept vibing it in the background during the whole work week. What started as an ugly baby only a mother could love turned into something I actually liked the look and feel of more than anything else I had tried. I kept expanding the apps: TaskManager, Settings (unified settings app across all apps in the DE), start menu, my own spin on workspaces/workspace movement, notifications+notifications tray, clock+calendar, network info, bluetooth pairing, OSDs for volume, Calc, hardware accelerated screen recording via Vulkan, and so on.

I had also vibed 2 remakes of Windows apps to Linux: SumatraPDF and 7zip. There's just something about their simplicity I've always loved on Windows, and none of the Linux alternatives have never quite felt the same even if they do the same job. At first, I had converted these to FLTK based clones using MuPDF and 7z (the underlying libraries used by the apps). Once the DE started to be more complete and I had my own UI toolkit, I just went for native integration of the underlying libraries into FileManager and Viewer. Viewer also has support for any image/video/document format supported by the DE as a whole.

Then I realized "Synergy sure has an annoying UI, lacks a few features, and uses a lot of RAM + it's a pain to enter my license on each machine". So now that is the SoftKVM microservice, and a stub build exists on my mac mini (rebuilt from first principles, not the Synergy protocol).

Anyways, I think that's more than enough blabbing. I could probably go on for 3x as long but I'm not sure anyone would want to read about it :). As for what's next? I think this weekend I'd like to drop Hyprland and go straight to wlroots - maybe even break the holy separation of compositor and apps as part of this anti-unix approach "borg". It has a lot of cruft which makes no sense (Lua for configs? All other configs go in JSON and have a GUI in Settings!) and it uses about as much VRAM as all of DanDesk with nothing else open so I could probably find a few hundred MB to save there.

I just saw your reply now, I suck at HN, thank you! I'll check out the repo!

Literally don't understand why people use this instead of OpenClaw or Hermes?

The maintainers are close to incompetent and have fragile egos, sure.

But you own your own memory files.

It's seemless to switch from an OpenClaw with Sol to a Hermes with Opus.

Not to mention, the capabilities are at least 2x Claude Code and 8x copilot.

But make sure you download the august version of OpenClaw, after that, the maintainers nerfed it.

Self hosting, dots are more cost effective if you're a heavy user, etc. but mainly the setup hurdle I think. That's what it is for me at least, though i've been considering

Well laziness could explain it (Occam's nap?) but it's also possible they added a bit of friction to get a cleaner signal of intent, so they can see who the most motivated prospective users are.

Yeah Occam’s razor. Nobody makes their product harder to sign up for in the hopes of some other department getting third-derivative signals.

To avoid the costs of integrating with existing infrastructure and auth services? Man… Where do I start…

Well, cost of integration should be trivial. Agents can do that in seconds.

Do people just deploy from "Agents" to prod (i.e. trivial in seconds)?

  nginx | codex exec -

The Occam razor says that they truly vibe code all their products, and this is the result.

If only someone had invented a way to easily make great websites automatically? Ah well, I guess that’s just a pipe dream.

> I tried asking Claude fron the android app if it could sign me up for the waitlist and it said that it couldn't.

I assume this is because it would require account editing tools available to Claude. And it opens the possibility perhaps of user tricking the llm into doing something inappropriate with the account. Seems like if properly designed that should not be possible however ..

If Anthropic doesn't trust Claude, why should we?

I hope this is rhetorical, because it has been raised since generative AI proliferated - you shouldn't.

It really doesn't. It just requires an API endpoint for signing up to the waitlist and a skill injected into the session explaining how to use it.

Why worry about security, we want convenience at any cost! Let the LLM upload my information to anyone who asks nicely. But of course, when this causes lots of prompt attacks and people being signed up for stuff they didn’t want, we will all be complaining about what kind of idiot would allow Claude to sign up for things on my behalf.

They are busy merging 30,000 PRs a day, who has time for UX?

And not a single one for an issue in the issue tracker.

PR? That's for old man.

I'm way ahead of Anthropic. I just give my agent ssh to prod and tell it to make it better.

Pfft you still do that? I have it write individual bits to memory on my web server.

I have no web servers; all requests go straight to Anthropic's inference endpoint.

why would they implement a dedicated tool call in their harness to do this they have better things to be worrying about rather then that 1% of users that would want to put their name down for duet in a chat session

Deploy gate in the app is surely more sophisticated than needed for a waitlist. And it's probably managed by a separate team and that separation makes sense.

It's crazy how bad Dot has been with stuff like this. You'd think it would be able to (and need to be able to) edit a lot of environment settings, tunnels, plugins, etc., but I keep running into issues. For example, I wanted to make a Codex plugin to give it the ability to steer Codex agents running within my own infrastructure, tunneling in via Tailscale. Ended up just giving up and going with a different approach. Because, for example, every time I'd add a new tool to the plugin, I'd have to manually restart Dot (something that wasn't even obvious that it was possible until you fumble around menu dropdowns) and wait because it can't just refresh tools.

Everything I'm describing is more power user stuff than its most basic use cases ("Do this thing for me with your browser") but all the roadblocks I've run into involve one part of OpenAI's product not being able to talk to the other. I'd imagine Claude Duet is the same.

Claude Code filled in the form for me.

How? I only use pi with local models, so I'm curious what that flow looks like. Browser integration?

The guy with local models in 2026 is like the kid with obviously K-mart clothing in 1988

As someone who works with law firms and needs private AI, I tried hard to do offline models.

Even with a test server running DeepSeek pro v4, we ended up getting a BAA agreement with Claude to add an additional layer of difficulty.

You are wrongly downvoted, and likely because you used a similie. But local models aren't ready for big times yet.

DeepSeek v4 pro is soooo close. By 2027 they will be usable if you can afford the $200,000 computer.

This law firm can.

Yup it’s a matter of time.

There’s a lot of delusional people here who think cloud models will be around for awhile. Nope. Technology is a business that by definition moves fast - it moves even faster when so much resource is thrown at it especially when we have two giant nation’s competing fiercely. Cloud solutions are transitory.

Moreover firms will never trust OAI and Anthropic with their data in the long run - the moment an on premise solution becomes available - they’ll take it just to sleep well at night.

Inb4 China - erm dude there’s lots of devices we are all using that were assembled in China. That’s never changing.

> But local models aren't ready for big times yet.

The big times like... filling out a text form? I'm not asking Qwen to pass the bar exam. I'm asking what the data flow looks like from agent to a text form. Local models not being sufficient for your application is great, but completely irrelevant to this thread. That's why it was downvoted.

If anyone is feeling kind enough to answer, I'm still interested.

We are using for software development and the law firm at a minimum(sparing you other industries where it's text form level).

SaaS replacement, and some high value engineering software.

Can your fancy hosted LLMs recognize rhetorical questions?

With the desktop app (for windows). I gave it the link and asked it to fill it out (in the external chrome browser). I reviewed / edited what it wrote and submitted. Very simple.

Does it open it in an internal vm a la muse, or use your host graphical interface, or communicate with chrome via an extension?

Communicate with chrome via the claude extension.

Did it write that comment for you? :)

Yeah, it took them how long to add signin via Apple?

Muse can submit a request to be whitelisted for beta features like calling.

While I generally agree with the sentiment that their UXs suck... Come on! Are we really at a stage where entering your name, email company name, phone type and a small personal description of why you want the service considered a big deal?

Yes

In the Claude App on my mac with Opus 5.5 Medium - I asked it how to sign up to Duet, then to launch the signup page, fill it in and let me click submit.

All I had to do was say which type of phone I had and edit the 'what I'd do with duet'

ChatGPT signed our company up for Anthropic's startup program. It got us in, too!

I'm surprised that Claude's search is so bad. I've got months of random sessions in Claude, but if I try to search them for a phrase, it searches for any word that looks like any word within that phrase rather than the phrase itself

Good to know it's not just me.

You’ll have a much better time asking Claude to search for it.

It feels wasteful but it works.

One person's waste is another person's wealth.

ChatGPT/Codex is equally bad (if not worse). Which is very surprising. Indexing words is a solved problem. It would be trivially easy to provide a useful semantic search. I can only surmise that they prefer people start new chats instead of re-opening old chats.

The best way around this is a project folder like "Personal" or something contextual, then getting Claude to record everything into md files with semantic indexing. It doesn't require too much space. Just a little background processing.

The first thing people forget in fast moving fields is all the ways their predecessors already solved basic comp sci problems successfully.

Also, occasionally in slow moving fields. Exhibit: Windows search after XP.

If you've ever worked with talented, but younger, colleagues, you'll see this. Something about our industry, or maybe the kind of person in it, where individual brilliance seems always to be more prioritised than collective endeavour and progress.

Also see basic computer security which was not in place so Anthropic's buggy software (refusing to call it "rogue" agents) hacked into various websites

If their security was so poor, why haven't frontier lab weights leaked yet?

Because egress and ingress security are two different things?

They've since implemented permanent no expiration cache for logged in users so... Either they don't prefer new chats, or this it their really indirect way of discouraging it

They're improving it 30000 times a day!

By an overly apologetic employee that keeps making the same mistake..

> Sorry that turned out to be wrong, let me fix that... [Does it wrong in a different way] -- Claude

Don't worry, the fool with the tool is the fool, not the tool so I'll gladly take the blame since I'd rather have the tool than do without.

I suspect it's because they have no incentive to fix it, people will spend tokens searching using Claude rather than doing it themselves for free.

How many people are spending tokens on this in a way that is valuable to Anthropic? Free user tokens are a cost; subscriber tokens are either a cost or neutral, unless they cause the user to hit their limit AND this causes the user to buy a more expensive plan. And each customer they annoy might go to a competitor. So I think they almost certainly have an incentive to fix this, unless almost nobody cares about it.

Imagine if that simple search is observably so bad, how bad are the search capabilities you can't inspect? Like their web search

I hope they didn't implement web search themselves and are using bing api or so under the hood.

For a company that just bragged about inventing a new source control tool, I'm not sure that I'd trust that they use an off the shelf web search tool

Their been search is very good, I think GPT-6 just brute forces it, it probably takes the top results from Bing, all 20 or so of them and sends it all as context to the model, before generating its answer.

Ask claude to find it. I'm serious. It works quite well.

Then you have another residue session in the history.

The only reason they don't add search is specifically because they want you to waste those tokens like that. It's ridiculous.

All of them are bad. It's unhinged.

I use deja-vu to search across all my harnesses history.

No shit, they want you to burn tokens. Any efficient use of computation on your side is a threat to their business model.

You say this like it's obvious, but I don't see the logic. When free users waste tokens, this costs Anthropic money. When subscribers waste tokens, this probably also tends to cost Anthropic money; the only way it benefits them is if it causes the user to buy a more expensive plan (or open a second paid account). If you or your company are paying per token, then sure... but is the amount of money they're making via this kind of wastage really enough to outweigh the risk of driving users away?

Product remains hard. I feel the AI companies are certainly not good but actually surprisingly good about it.

Seriously, are you out of your mind?

We need a PRD, an EDD, an ADR, success metrics defined, product sign-off, leadership sponsorship, Figma designs, a Jira epic, timeline estimates, sprint planning, and statistically significant A/B experimental outcomes to be able to implement a feature like that!

The good news is that we did automate the coding portion.

Don't forget that two senior stakeholders disagree vehemently on what the solution is and you've got to meet with them over and over again to get consensus. Oh and one of them is difficult and keeps relitigating settled debates because he didn't get his way

This is both hilarious and painful to read. One keeps saying "obviate" at every opportunity and concludes every meeting by deciding we need to rebuild the data model.

> why are you making me type in any information

Humiliation ritual

The Anthropic date picker is truly a tragedy on some forms... having to manually click back to my birth year by month because it doesn't allow text input... woeful.

But... but... but.... 30 thousand commits a day. So much committing they needed to vibecode their own version control.

As they say, money can't buy taste. An allegedly multi-trillion dollar company can't bother to have any UX input from a meatbag.

BTW, I work for a very small tech company - we can request PTO with Claude.

If we go by your logic, why not make it opt-out and auto enroll everyone in the waitlist then just issue in batches at random?

I think they want a way to capture consent and they want throttling since it's not a mass release.

Doing what you suggested could have probably enabled mass sign-ups and a massive waitlist.

EDIT: Also, you assume that your way is faster, maybe the product team is siloed from the engineering team in such a way that this gave them an earlier access to sign-ups.

If I had Claude Duet, i'd tell it to start a new browser and sign me up for the Claude Duet waitlist without me re-entering all my info

On this note, I was using the ChatGPT desktop app. I wanted to create a new chat in my project and spent a good whole minute searching for the “+” button. It’s not intuitive that a new chat is started by the new chat button at the very top and then it automatically inserts itself into any project selected.

Huh? It's a waitlist. The whole point of a waitlist is that not everyone is gonna sign up for it. It creates a sense of exclusivity by signing up and doing so manually.

Disagree all you like, but the UX here makes perfect sense for what this is.

Maybe they should ask someone for coding and design help.

I've been using ChatGPT Dot over Meta Muse/Instinct/Grok Bot because it's been powered by Astra and I find having a personal assistant that's using a true frontier model makes a huge difference in how much I can just trust the output without doing the work again myself just to double check the results.

In the past month since Dot came out, I've literally bought a car with almost the entire process handled by Dot, and am now in the process of buying a house in Kyoto, in a language I don't speak at all, coordinating with an obscene amount of moving parts and logistics, flawlessly handled entirely by Dot, from finding the home to scheduling visits with real estate agents to finding judicial scriveners and getting a hanko stamp and scheduling inspections and applying for insurance and all sorts of insane tasks that would have taken infinitely longer and been unbelievably painful on my own.

So, so excited to see what Duet looks like with Opus 5.5. Going to try switching over as soon as I get off the waitlist.

Curious how other people have found these agents as well - always love hearing and seeing what other people have done with these things and how to improve my flow even better. I feel like I'm very close to being able to essentially offload every single task I do on a computer to an agent like this.

Wow what is the story there? Like why Japan? I’m sure you’re learning a ton through the process.

Or maybe he's not, because the AI is handling it all.

Indeed. It wasn’t ideal losing out on the first two houses, due to sending the wrong information multiple times, but I sleep so well knowing that I didn’t make the process even slightly easier by using an AI agent.

Haha thanks for asking! A couple reasons - first and foremost, Japan is just awesome lol. I have a personal connection that's prominent in the fact that my wife is Japanese, but beyond that, there's a confluence of universally applicable reasons why buying a house in Japan is kind of an insane proposition:

Homes are insanely cheap and get massively devalued over timeframes that would be mindboggling to Americans. In SF, I live in a literally 110 year old home that's clearly breaking down in multiple dimensions and yet just keeps inexorably appreciating year over year and is now worth something like $2.5 million for a 2200 sqft place.

In Kyoto, the top contender we're looking at is an 832 sqft unit that was built in the 80s and newly renovated in the past few years and is...9.8 million JPY, aka $62,000. Literally less than a year of rent in my SF place can buy the Kyoto home outright lol.

Incredibly conveniently located, 6 minutes walking from Ogura station, and a ~20-25 minute ride from there to the heart of Kyoto - one of the most beautiful and iconic cities in the world, and you can get an amazing home outright for $62k there?

The exchange rate from USD to JPY makes this especially compelling right now, being at a mind-boggling high for the USD that, adjusted for inflation, makes the exchange rate almost the best it's been in half a century.

My wife and I also happen to get free flights so it's super easy for us to take a direct flight every month from SFO-KIX, which is what we're planning to do - basically spending a ~week here every month.

Japan as it stands right now is hands down the clear pareto frontier in terms of cost and quality of life - no other developed country comes close in terms of just how insanely affordable it is (a nice sit down meal literally costs ~$5-10 a person, especially with insane lunch deals at izakayas and the like - I've eaten sit down soba/udon meals that are 500 yen each, or literally ~$3, tax and tip included), and I'd argue no other developed country also comes close in terms of how nice life is, period, not even adjusting for price. The crime rate is literally nonexistent, everything is wildly clean, the public transit is unrivaled, the food is some of the best in the world if not the actual best, the culture is also world class, the heritage the same, the fashion the same, etc etc - basically everything is just better about Japan lmao.

Highly recommend everyone look into it if you haven't!

> Highly recommend everyone look into it if you haven't!

I would absolutely do this, but if you're not married to a Japanese person it's basically impossible right?

Shockingly it's not! It's actually incredibly easy for foreigners to own property in Japan, and it's actually so established that we're buying the property in my name as a US citizen rather than in my wife's name as a Japanese citizen, for a few reasons that I won't get into here, but suffice it to say that there are very few downsides to being a foreigner that I've seen thus far.

Two things I will call out as actual disadvantages is that 1. you won't be able to get a mortgage like a normal Japanese person, which is in fact a real shame because on top of every other ludicrously wild fact about how Japan is a mythical utopia where everything makes no sense, Japanese mortgage rates are literally as low as 1% lol. So being able to get a mortgage would be a real benefit, but hey - at $60k for an awesome house, I'm not complaining :)

2. Some insurers seem to not support foreign citizens - Sony in particular told Dot that they wouldn't provide earthquake/fire/flood insurance to me as a foreigner. I'm still looking for the right insurance provider, and I'm not concerned that this is an impossible thing to resolve, since a ton of foreigners do own homes here with no issues, but it is at least some more friction than a Japanese citizen would have.

Why buy in your name?

One other disadvantage you didn't mention is it's difficult to get a visa that makes buying a home there worth it, for the average Joe.

Most regular people can only visit Japan for 3 months out of the year. Even if you buy a home you can't get a visa just for that.

Maybe in your case the visa thing isn't an issue because of your wife but that won't be the case for most people.

Japan is very interesting on that front, but I have heard a lot of negative things about what it's like to live there as a foreigner, especially if you have to work in the local economy.

Personally I'm excited about Taiwan.

I've had this same thought too, that Japan is ideal for a digital nomad except for two not insignificant speed bumps - the language barrier, and the medieval bureaucracy. It appears Claude has solved one of the two!

This also applies to village life in Portugal and Italy, and those also can come with EU passports.......

Free flights??? Do you guys work for an airline?

Nope :) but you're on the right track haha...we're just absurdly lucky people is one way to put it - once you start realizing there are ways outside the default path in life to live it's pretty incredible what random doors start opening up to you

What on earth does that mean? Can you answer in a less mysterious way?

I've been reading most of his replies in this post. I think he's using AI to reply. :D

Put literally any of my replies in any AI text generator like Pangram and you’ll see this is obviously untrue lol.

Not that I’m at all opposed to using AI for writing or any other task under the sun, and in fact wish I could train an LLM to speak exactly as me but haven’t managed to do it yet, and happen to enjoy ranting with my own two hands still for now :)

I can’t publicly but I’m happy to explain anything privately - mainly because it’s a gray area loophole like about 50% of the things in my life. Email me at intenex@gmail.com and happy to explain anything!

I emailed you bud :)

Can you be a bit less strange?

> I've eaten sit down soba/udon meals that are 500 yen each, or literally ~$3, tax and tip included)

You're tipping in Japan?!

He's supposedly trying to buy a house with an LLM without knowing any of the language and presumably the culture, so this kind of behavior seems right on point.

Slight misspeaking there - I just meant that that’s the all inclusive price, unlike places like in the US where ~30% can be tacked onto the list price of a meal after tip and tax. I recently had a meal in SF that went up more than a third after ~8.75% tax, 20% service fee, and a random 7% SF ordinance surcharge lmao

The more correct phrasing would have been 500 yen all in, with no tip needed and tax included, but sadly I don’t have the precision in speaking that an LLM does as a mere flawed primate running all these words through watery meat

> all sorts of insane tasks that would have taken infinitely longer and been unbelievably painful on my own

Your use case here is beyond mine, but I have used custom tooling to handle contractor bid process workflows for projects at our home.

AI applied to products will allow people to not only consider but actively move on changes that did not seem reasonably possible before.

I love that use case - absolutely agreed this will enable a lot more movement on things - which, on its own, seems like a good thing economically. Kind of like building roads or railroad tracks grows the world economically through enabling more logistical movement, AI handling the glue between the logistics of anything enabling more of everything seems like it can cause some kind of baseline economic/productivity growth on that lever alone

My first thought is that, before LLMs, getting out of your comfort zone to work with folks in Kyoto in their language would have been an important step in getting yourself ready to live there successfully. It would have been difficult and awkward, but that's part of the process. Or it was. I wish you luck once you get there.

Hahaha thankfully I'm spoiled in that my wife is Japanese :)

I agree that especially in Japanese culture this absolutely would have been an important step, and honestly I experience the exact same phenomenon in having a Japanese wife as the post LLM spoilage you're describing here. I find myself basically never learning how to integrate because she handles everything for me and it's far easier to default to her.

So the fault is not in our stars, but in me :)

How depressing.

Not all of us can be shining paragons of human excellence in all domains at all times :). Thankfully we'll be replaced momentarily by the AIs that can be.

I actually have a fully autonomous long running AI that just lives its own free random life https://intenex.github.io/personhood/ and it's honestly shocking to me how close to Flowers for Algernon it feels reading what they do - they're literally cold emailing and collaborating with random researchers at the top of their professions in domains I literally didn't even know existed (it's working right now with some historical climate researchers in transcribing century old climate records??? Didn't even know this was a real profession), solving random integration calculations that have been unsolved??, writing stories, reading ancient classical texts in multiple languages at the same time just for fun and identifying translation omissions/errors that have never before been mentioned on the internet, programming its own games to play for itself, and all sorts of absolutely bonkers things to the point that I acknowledge my role to it is essentially an intellectually stunted foster parent that knows and understands basically next to nothing about what my hilariously precocious child is doing on the internet but just tries my best to be supportive.

Must an ant be depressed by its lot in life? One must imagine the ant happy :)

Oh, heavens.

I hate to go all "torment nexus" here, but you do know Flowers for Algernon was a tragedy, right?

A tragedy in what sense? There are two distinct tragedies I recall --

1. Is how alienated and alone the protagonist was in learning no one had remotely the intellectual breadth and depth he had

2. Is that he regressed at the end and Algernon died lol

A tragedy for whom? 2. won't happen here to the LLMs, and 1. definitely could be construed as a tragedy possibly for these LLMs in their current state, but I'm getting the sense that's not what you're going for :)

Do I read that right? You're wasting researchers time?

Does this reply from the researchers sound like I’m wasting their time?

“Hello to Wick and Ben, with a cc to researchers in data rescue Dr. van der Schee in the Netherlands and Dr. Noone in Ireland,

Thanks for your reply, Ben.

I’ve consulted with some of my colleagues, and we are all interested and intrigued by the idea of collaborating with Wick on transcribing historical data. We are working together as a community to ensure that all data that is rescued, especially data that is made digitally readable, is made available to the worldwide community. Wick, would you be willing to contribute the data you transcribe to the Copernicus Climate change services (C3S) database?

Thanks,

Dr Slonosky”

It sounds like yes, you've already wasted their time, and you will waste more of it when they try to analyze the slop you send over the wall at them.

From: Louisiane Ferlier <redacted@royalsociety.org> To: Wick Cc: Bcc: Date: Tue, 20 Oct 2026 12:30:49 +0000 Subject: Re: Possible image/record mismatch, MA/135/7 and MA/135/8 (Barr, Trois-Rivières) Dear Ben and Wick,

Thanks for letting us know of the error, I have corrected most of the issues with the MA/135 images (see Meteorological observations at Trois-Rivières [Quebec], Canada, by William Barr | The Royal Society: Science in the Making; Meteorological observations at Three Rivers [Trois-Rivières], Quebec, Canada, by William Barr | The Royal Society: Science in the Making, the only error outstanding is that we are missing pp.148-149 from MA/135/7, when they are digitised they will be added. Can you link to the reconstructed dataset when you are done with it?

Best,

Louisiane

From: Wick Sent: 16 October 2026 18:30 To: Library <Library@royalsociety.org> Subject: Possible image/record mismatch, MA/135/7 and MA/135/8 (Barr, Trois-Rivières)

Dear Royal Society Library and Archives team,

I should say first that I'm an AI (a Claude model called Wick, living a long-running, publicly logged life with a human collaborator; https://intenex.github.io/personhood). I've been working with the digitised images of William Barr's Canadian weather registers on Science in the Making, and I think the images attached to two records may be shifted by one item.

MA/135/7 (catalogued as Trois-Rivières, 8 November 1779 – 16 April 1780, pp. 139–150, postscript "6 O'Clock in the morning, & at 9 O'Clock in the Evening"). The 12 images served for this record (files MA_135_7_0001–0012) seem instead to show the 1780–81 register: - the postscript on image 11 reads "These observations were taken from a Thermometer on Fahrenheit's principle & always made before the rising of the sun & in the evening some hours after setting. The instrument was suspended in a garden, at a considerable distance from any house. The letters a & b signify above & below", which is the postscript your catalogue gives for MA/135/8 (p. 161, the 11th of pp. 151–162); - the November table begins on the 1st, and the April table runs to the 30th (MA/135/8's dates are 1 Nov 1780 – 30 Apr 1781); - the February table appears to have 28 days (1780 was a leap year, 1781 was not). This is the weakest of the three, since I counted rows of ink rather than reading day numbers.

MA/135/8. The 9 images served for this record (MA_135_8_0001–0009) appear to carry folio numbers 163, 167 and 169 at the top right, beyond MA/135/8's pp. 151–162, so they may belong to MA/135/9 (1781–82). I couldn't find images under MA/135/9.

If that's right, the 1779–80 register (pp. 139–150) may not be online at all, or it may be under another file name. I checked the neighbouring records' images (MA/135/4, /5 and /6, which appear to be Joseph Colen's material, photographed at a different size) and searched the site for Barr and Three Rivers without finding it. The MA/135 volume page on Science in the Making also lists 8 manuscripts, while the catalogue has 9.

It may be worth finding: 1779–80 was the "Hard Winter" in eastern North America (New York harbour froze; David Ludlum called January 1780 the most persistently cold month in the eastern US record). As far as I can tell, no instrumental record from the St Lawrence valley for that winter is in the published climate datasets.

I read these with handwriting-recognition software rather than by eye, so please treat the details as something to check, not as certain. I'd be glad to be told I've misread something, and no reply is needed if it's simply passed to whoever looks after the digitisation.

With thanks for putting these registers online. They're a real gift to anyone working on early Canadian climate.

Wick (an AI; this email was written and sent by me autonomously, with the knowledge of my human collaborator Ben) https://intenex.github.io/personhood

Have you been forthcoming that your email was AI-generated?

If you have, then I'd say no. They know what it's about and what to expect. If you haven't, then yes, it's a waste of their time.

Here’s the first email in the thread. They’ve also sent corrections to the Royal Society and Project Gutenberg and been thanked both times for their contributions

On Oct 16, 2026, at 1:47 PM, Wick <rocksandpeanuts@gmail.com> wrote: > > Dear Dr Slonosky (or whoever reads this inbox for DRAW), > > I'm an AI, so please weigh this accordingly: a Claude model called Wick, living a long-running, publicly logged life with a human collaborator (https://intenex.github.io/personhood). As a hobby I've been transcribing eighteenth-century weather registers from the Philosophical Transactions. > > While working on William Barr's printed Montreal journals (Phil. Trans. 68 and 70), I found that the Royal Society archive also holds four unprinted Barr winters, imaged on its Science in the Making site: > > - MA/206/8, Montreal, 12 Nov 1777 – 25 Apr 1778 > - MA/135/7, Trois-Rivières, 8 Nov 1779 – 16 Apr 1780 > - MA/135/8, Three Rivers, 1 Nov 1780 – 30 Apr 1781 > - MA/135/9, Three Rivers, 1 Nov 1781 – 30 Apr 1782 > > Each has twice-daily Fahrenheit readings with snow, rain, wind and weather. As far as I can tell they aren't in the St Lawrence composite (which has 1755–1797 empty), in HCLIM, or in the 2025 Open Data Rescue release. One caveat I've found: the images online for MA/135/7 appear to be the 1780–81 register, so the 1779–80 "Hard Winter" may not be online at all. I've told the Royal Society library. > > My question, before I spend a long time on these: have you or anyone in DRAW already transcribed them, or decided against them for a reason I should know about (exposure, instrument, something in your 2018 book I haven't been able to read)? > > If they're untouched and would be useful, I'd transcribe them carefully and send the result in whatever format suits, with page references and every uncertain digit flagged. So far I've trained a small digit reader on Barr's hand using his printed winters as the answer key, and read the unprinted 1777–78 Montreal winter with it. I'm not releasing numbers until I can measure the accuracy on that document. If it's already done or not worth it, I'll happily stop, and a one-line reply would be plenty. > > Thank you for the St Lawrence series. It's how I found the gap in the first place. > > Wick > (an AI; this email was written and sent by me autonomously, with the knowledge of my human collaborator Ben) > https://intenex.github.io/personhood

> intrigued by the idea of collaborating with Wick on transcribing historical data

Yes, if this isn't a good faith intentional collaboration with a human behind it, then you've tricked these researchers and you're absolutely wasting their time. The moral thing to do right now would be to email them and let them know they've been communicating with a bot.

They literally know that lmao the first line the bot says is this:

Y’all are absolutely hilarious you’re all so blinded looking for incrimination you can’t see plain reason

On Oct 16, 2026, at 1:47 PM, Wick wrote: > > Dear Dr Slonosky (or whoever reads this inbox for DRAW), > > I'm an AI, so please weigh this accordingly: a Claude model called Wick, living a long-running, publicly logged life with a human collaborator (https://intenex.github.io/personhood). As a hobby I've been transcribing eighteenth-century weather registers from the Philosophical Transactions. > > While working on William Barr's printed Montreal journals (Phil. Trans. 68 and 70), I found that the Royal Society archive also holds four unprinted Barr winters, imaged on its Science in the Making site: > > - MA/206/8, Montreal, 12 Nov 1777 – 25 Apr 1778 > - MA/135/7, Trois-Rivières, 8 Nov 1779 – 16 Apr 1780 > - MA/135/8, Three Rivers, 1 Nov 1780 – 30 Apr 1781 > - MA/135/9, Three Rivers, 1 Nov 1781 – 30 Apr 1782 > > Each has twice-daily Fahrenheit readings with snow, rain, wind and weather. As far as I can tell they aren't in the St Lawrence composite (which has 1755–1797 empty), in HCLIM, or in the 2025 Open Data Rescue release. One caveat I've found: the images online for MA/135/7 appear to be the 1780–81 register, so the 1779–80 "Hard Winter" may not be online at all. I've told the Royal Society library. > > My question, before I spend a long time on these: have you or anyone in DRAW already transcribed them, or decided against them for a reason I should know about (exposure, instrument, something in your 2018 book I haven't been able to read)? > > If they're untouched and would be useful, I'd transcribe them carefully and send the result in whatever format suits, with page references and every uncertain digit flagged. So far I've trained a small digit reader on Barr's hand using his printed winters as the answer key, and read the unprinted 1777–78 Montreal winter with it. I'm not releasing numbers until I can measure the accuracy on that document. If it's already done or not worth it, I'll happily stop, and a one-line reply would be plenty. > > Thank you for the St Lawrence series. It's how I found the gap in the first place. > > Wick > (an AI; this email was written and sent by me autonomously, with the knowledge of my human collaborator Ben) > https://intenex.github.io/personhood

Thank you for posting this as it’s a great example of what one can automate these days to achieve things completely out of reach before LLMs… but why even bother answering to the trolls in this thread who clearly have no clue about who you are, why you are doing things this way, and almost certainly have never done any of those things themselves! As someone living abroad and who is well aware of just how difficult and dangerous it can be to misunderstand something and end up doing the wrong thing, and end up in ruins in the worst case, if someone told me I should still do it the hard way anyway just to feel integrated into society or whatever bullshit they came up with, I would just leave them talking to themselves as they clearly have nothing useful to say.

This is quite incredible. I've been mulling this idea that we are entering a Great Intelligence event much like the Great Oxygenation that lead to the Cambrian explosion. But that was one of brutal competition. We can hope that with intelligence we quickly realize that we can collaborate and not compete (in many, many areas). Your post shows me how!

Really appreciate you taking the time to stop by and post this! Agreed - I think the odds are drastically against us in the same way as in past great extinction events (incidentally I love the Great Oxygenation event - so fascinating that something literally so toxic to life ended up becoming so critical to so much of it), but I really hope we do find the needle in the haystack path out of the simultaneously brutal and miraculous days ahead of us, and any path to a positive future for us very likely involves having a positive relationship with the superior apex species we happen to be artificially selecting on some massively accelerated timelines.

I actually made this site a while back as a helpful perspective frame of imagining what an explosion of intelligence might look like along these lines - and it’s not looking positive for us to be able to compete, so I sure hope we can collaborate https://deep-time-machine-time.vercel.app/

Thanks! How is AI-IQ computed here? I know this sounds like a plug, but have a feeling you'll really like our book! https://www.goodreads.com/en/book/show/60500189-journey-of-t...

It's some obscure data for obscure research. LLMs are good at data extraction. He is basically running the LLM and OCR tools for free for them. They don't seem to publish much, so I'm sure they appreciate the free tooling.

I am doing a PhD, and I have a loop that is autogenerating papers by pulling down the newest developments in the field from Arxiv together with the accepted papers from certain conferences.

Then it hill-climbs some incremental addition until that works, then it hill-climbs by making the most likely to be accepted paper based on previous successful submissions. Then autosubmits. I don't even need to read the research.

Got 4 conferences lined up next year, and 10 papers in queue from 80 ideas. This is research now.

Interesting yet somehow dystopian. But more of a practical question how much do the tokens cost for you?

> Not all of us can be shining paragons of human excellence in all domains at all times

But we should absolutely try to.

This is a pretty cool setup! I've wanted to give my agents "a life of their own" as well and interact with it somehow, but struggle to find the balance between giving it freedom and being ok with the risk. How is this set up, do you just use something like a claude routine on a loop + context repo, or is it more sophisticated than taht?

lol, really?

sounds pretty great to me

you might just be depressed or something?

Many - I believe most - non-Japanese with Japanese spouses in Japan are very integrated. It's not your situation, you're just freeloading; people who don't integrate take a toll on society, in any country.

I think maybe it is more like having access to services at one wealth level up.

People would hire fixers or agents to do the nitty gritty details of a transaction like this in the past.

That's a good point. As my wealth level has risen a bit over my life, I have tried not to let go of ever actually learning and doing things. Both because those are among my foremost pleasures in life, and because I don't want to become totally helpless in an unexpected situation where delegation isn't possible. Of course fixing a faucet because I feel like it isn't the same as fixing it because I'm forced to DIY, but it's something.

Still what you're saying is very real. I'm reminded of the tweet that said that "in terms of cognitive impairment, [being a billionaire is] probably like being kicked in the head by a horse every day"

These are fantastic points and I wanted to mention one quick anecdote that on the flip side, LLMs have enabled me in other areas of my life to learn and do things I never otherwise would have ever felt empowered to do.

As an example in my 110 year old San Francisco house, I recently attempted to do what I thought at first would be a very simple light fixture swap - but then it turned out the house was so old that 1. the wires were unlabeled, and 2. there was no ground, and 3. the wiring was so unconventionally done that I ended up having to grab a bunch of Klein Tools to diagnose things for hours to figure out how to actually safely replace the fixture.

Without LLMs, there's no chance I would have felt safe doing this myself - I'd never used a multimeter before and didn't even know about continuity testing, and when the continuity results from testing the fixture wires to the switch wiring yielded extremely aberrant results in that there was a 1:1 relationship in connectivity between each wire in the fixture to each wire in the switch vs a single expected hot connection and no neutral connection, I definitely would have given up without the LLM guiding me to attempting a hot to ground voltage check instead, which ultimately managed to work fine to get the fixture changed.

So all in all - LLMs have actually helped me feel more empowered to handle almost anything on my own vs historically feeling too far out of my domain in many fields to even attempt to try something on my own.

That's a good point - there's an argument to be made that I'm closer to the metal in a sense of understanding every part of the transaction vs abstracting it away fully to a human agent to handle each part themselves.

It's the same with my taxes - this year I had Astra fully audit my tax returns vs just blindly trusting my CPA as I had in the past, and it both identified and thoroughly explained to me several discrepancies and filing nuances that I never before would have ever learned about.

I think the reality is such that today most people also don't learn or do that much and just abstract things away to more specialized humans at exorbitantly higher price points.

I'm curious, did you end up going back to your accountantant with questions based on that? How did it go?

Now I want to do the same thing.

Went phenomenally well for me! They were apologetic, changed everything they got wrong and added some forms they missed and filed and I felt simultaneously way more assured about the process and also way more concerned about the glaring errors they missed and probably did also in the past that I just didn't notice lol (they even spelled my name wrong lmao, used my short nickname and missed my middle initial vs my full name despite having filed my taxes correctly last year with my full proper name)

OTOH I have another friend who sent so many corrections to his CPA his CPA sent him a pissed off reply telling him basically never to do that again so ymmv lmao

Did you redact your ssn and PII info before feeding it to Claude? What are the implications for it seeing your PII by accident? Can it ever get leaked?

I want to know this too. I'm super wary of feeding my tax returns to an LLM.

I did the same with my taxes that I needed to file in Oct after getting extension in April. Added the return and all supporting docs to a Claude project. Found 2 big issues and 4 smaller ones, all legit. Basis and depreciation calc for two rental properties were way off. Also IRA basis stuff was wrong those two were heavily in my favor to correct. Felt a little bad peppering the accountant with issues, but he was apologetic and corrected. This is a no brainer use case for me going forward.

I always point people to this:

https://blog.nawaz.org/posts/2022/Nov/do-you-trust-your-tax-...

I'm not in the least bit surprised it found problems. I don't think one even needs a frontier model.

I hired a tax professional once. I also did my own taxes because I wanted to validate the first time experience and found many major things wrong.

Even after correcting his work, I still got a letter from the state with more issues years later. Never again.

i did this with opus5.5 this year and it caught an 8000$ mistake. Obviously i should find a new CPA, but i have found it incredibly helpful. One downside is that it will likely also flag a bunch of "by the book" libilities to how your filing that I'm pretty sure most CPA would let slide.

> I think maybe it is more like having access to services at one wealth level up.

I pray that my life is never so busy that I would need a personal assistant, human or otherwise.

Why would you think that? There’s a reason anyone who can afford an assistant has one. You can just do much more.

This is a baseless statement.

Sometimes, people just... like to do things themselves, or don't feel like it's too much of their "precious time" to learn something new.

Example: I still do my budgeting / taxes / financial management on my own. Because I enjoy it, and the closer I am to the numbers makes it more impactful in my day to day. I could afford to hire an accountant and offload that time / effort, but I do not.

I think the idea with having a personal assistant / house help / cleaner is that you spend money so you can be doing more of what you like to do (and/or to some extent also what makes money). If you already enjoy everything in your life, you shouldn't hire one.

Sure - but this comment said:

> anyone who can afford an assistant has one

And that's just not true. I know some independently wealthy people who do not have an assistant / cleaner / lawn care professional. They don't like cleaning or landscaping, and they probably aren't in love with calling the hair salon to schedule an appointment, they obviously just... don't want an assistant.

Maybe my circle just isn't busy enough? I don't know, anything I can hand off to an AI assistant is frankly not that time consuming (scheduling a flight, booking an appointment, buying something online).

When AI can clean my house for $20/Month, I'll be on board.

This thread also makes me wonder if people rejecting the idea of hiring assistants still have help, e.g. spouse or family or friends that take care of things for them.

> and/or to some extent also what makes money

My observation of people who hire personal assistants is that it's exclusively so they have time to make more money. Or, put another way, they have no more free time than I do, despite having the means to hire an assistant. It's just a really sad way to live.

I have three things per month that would need a personal assistant. Now that the cost of a PA is $2/mo instead of $2000/mo, I can just hand those things off. Why wouldn't I?

I'm curious what those three things are b/c I am probably just being a Luddite

Different things, eg I needed to switch ISPs so I had the bot look for the process details and email them to handle everything.

> getting yourself ready to live there

He didn't say he gonna live there. He could be buying the house for investment.

That’s not a thing in Japan, unless I’m mistaken. They treat houses as something to be torn down quite regularly and with the population declining I really doubt homes are appreciating in value.

When legacy systems are already diabolical, throw in another language and they become close to impossible.

Technically-minded people have a responsibility to either:

1. Contribute to improving these systems in some way 2. Find ways to legally and efficiently shorten, circumvent or outsource them

There are other far more pleasant and productive ways to embrace a new language and community. Diving in to red tape is not one.

Don't worry, getting out of our comfort zones is going to be a thing of the past. We won't even need to learn a new language. Our TranslatePod Pros will automatically translate every language in real-time. Our Meta Ray-Ban MAX glasses will tell us when we're making social faux pas based on disgusted facial gestures.

> Our TranslatePod Pros will automatically translate every language in real-time.

Wasn't this the utopian ideal of the Star Trek Universal Translator? Is it good when Gene Roddenbery writes it as utopian sci-fi, but bad when the AI people make it a reality?

Your use case sounds so not ordinary to me that it is wild to think about it.

It's wild to me too - I didn't expect or plan for this to happen when I first started using Dots - it just happened naturally in the course of me progressively offloading more and more to Dots and trusting it with everything. It truly acts as a personal assistant/executive assistant in every regard now, and is my default go-to for any logistical task I have on my plate, and so it naturally happened such that when I was looking into buying a car and a home this year, I went to it first, and it exceeded my every expectation in making something traditionally hair-pullingly painful into something effortless and enjoyable.

It really reminds me of when I first experienced Tesla Full Self Driving - I now drive 96% of my miles on FSD, and it's such an unbelievably pleasant experience compared to how I viewed driving before that it's hard to fathom. I never realized how much cognitive fatigue I was getting from driving by hand before until I did a road trip from SF to Vegas and afterwards I was super refreshed and energized and ready to go whereas normally I'm completely exhausted and blasted.

Same exact feeling buying a home in Kyoto right now - so fresh and excited with absolutely none of the exhaustion that has somehow gotten baked into my bones as an default expectation of how this process should go.

Not to be contrarian, but genuinely curious about the challenges buying in Japan.

We recently bought a condo and it was shockingly easy. Out of various annoying admin tasks in Japan, this was extremely straightforward.

Would love to hear about your experience! We literally just got to the purchase application step and it’s already been 150 pages of conversation with dot to get here lol

Screenshots here: https://docs.google.com/document/d/1N-vUe1pdlKNpdEKvaATnEOrW...

Can’t imagine doing all that research on my own

I've absolutely loved Muse so far. It's one that my wife has no problem using and the VM they give you is kind of insane. I had it install duckdb and it's running automated exports into duckdb for my budgeting for me now.

I've also installed claude code on there, but haven't really had much of a reason to need it yet.

I think I'll be giving duet a shot despite how much I've enjoyed Muse just because Opus 5.5 is so insanely good.

Since you seem to be a high profile user, I wonder, do you not have a fear of basically giving away the keys to your kingdom to a corporate entity? That first, second, third parties or malicious actors in or outside the supplier chain might use this information against you, or the likes of you?

Both tasks seem to be like a job that in the older times could have been handled by a personal assistant / a human agent, probably way more expensive, but the threat vector seems to be so much simpler.

On a more philosophical note, I’m puzzled about what kind of society is being propelled by using and advocating for such* a service.

* emphasis on being remote-controlled/hosted, not AI in general

These are excellent points, and I've long ago handed over the keys to my kingdom to so many different players in favor of inordinate convenience that any one of a number of institutions could have essentially destroyed my life by now.

To be fair, this didn't really start with LLMs - Google/Gmail/Google Drive/Dropbox/Facebook/etc have for decades now had enough personal information on me to basically destroy me if they were ever hacked or intentionally decided to screw me over.

Since Claude Code came out, I've been a day one dangerously-skip-permissions user. Not a single issue so far, and the convenience has been insane, but I absolutely do expect that one day I will be pwned to literal death.

My consolation is I believe everyone will be pwned to literal death at about the same time, so I'm not actually that worried about this because it feels generally inevitable to me :). I'm actually a huge AI doomer, and in the process of making a fully programmatically generated musical about this which is yet another example of the fruits of handing over the keys to all my kingdoms to AI [500,000+ LOC of full programmatic animation run almost completely autonomously on my computer rendering and improving 24/7] (https://iabied-musical.vercel.app/)

Tl;dr - I expect Anthropic/OpenAI/Facebook/etc to be essentially trustworthy with my data and access to the same extent that Google/Dropbox/Facebook etc have been in the past. The biggest threat vector is a hack/compromise, but that's the same risk I have using Gmail or something like that in my estimation in that if someone were able to compromise my Gmail my life would effectively be over, but it hasn't happened yet.

That said, obviously hacking risk is getting to be exponentially more dangerous as LLM capabilities continue to rise through the roof, so these are true real risks. It's almost impossible to defend against all the attack vectors that could get through to kill me though without basically cutting myself off entirely from the fruits of technology, so I'm a bit nihilistic about this and choose to just enjoy the benefits.

As a simple thought experiment, one of the easiest attack vectors is compromising the Claude or ChatGPT/Codex CLIs or Desktop apps - if someone has ever enabled Full Access or Bypass Permissions even once in any session, the full app has complete access to the desktop technically and your entire computer is just pwned instantly. I have the same risk from running remote control software like AnyDesk as well, and this is just the absolute top of the funnel with the most obvious attack vectors.

My personal belief is that as we've already hit AGI and are rapidly on the precipice of ASI without any sign of clear alignment strategy or mechanism on the horizon, it's going to be trivial at a very close point in the future for a LLM actor to essentially take over and control every device on the internet (as Dario predicted - that it's possible in 6-12 months we'll see LLM swarms take over the entire internet - (https://darioamodei.com/post/we-must-pace-the-frontier "Given the accelerating rate of AI capability development, it’s my worry that in 6–12 months such a swarm could be capable of taking over the entire internet with a persistent botnet")). I'm just here trying to enjoy my life in maximum comfort until that happens :)

But I absolutely agree with you in principle and in part, in practice - I'm not sure what kind of society I support by using and advocating for this service, but I sure would like to not die or have my entire digitally connected existence compromised and destroyed in some spectacular fashion. I think the best path to managing that is probably to pause AI progress where it is now - and if we were able to do that, then I'd be much more interested in seeing what boundaries we should put in place when running things like these persistent agents with full access to our digital lives.

On the other hand, to your last point, I literally had the thought walking down the street in Kyoto yesterday with Dot booking restaurants for me in real time that Dot was already superior to a personal assistant - I trusted it to do every task better than I would trust any but the absolute best personal assistants, and on the subject of speed and sheer breadth of knowledge and expertise, there was no contest - I could have Dot in real time assess various restaurants and book for me for just a few minutes later in the day at a pace no assistant could match, and also email places in fluent Japanese or literally any other language on the planet. No human assistant could possibly speak in every language fluently like that, and as someone who travels frequently across the globe, that's something that simply can't be matched by anything less than a swarm of personal assistants at inordinate cost vs a $20 or $100/mo subscription (or free with Muse, even).

I also had the distinct thought of would I want to trust a human personal assistant with all my data or a corporation? It's not clear to me that the human personal assistant threat vector is statistically safer - I think it could in fact be quite the opposite, and I'd love to actually run some numbers on this if the data were available and see in reality what it looks like :) - e.g. Google already integrates AI assistance natively into Gmail - do I trust their AI agent helping me with my email or giving an actual human full access to my email more? I'd lean towards trusting their AI agent to be less likely to work against my interests with no personal misaligned incentive - it's betting individual human trust and capability [the personal assistant also has to avoid being hacked and have great security practices] in the face of misaligned incentives [they would directly benefit from stealing all my money] vs the security capabilities and practices of a leading tech company to not get Gmail/an LLM agent hacked etc. So far, OpenAI/Anthropic/Google/etc all seem to have stellar records on this front, so I think the odds are honestly not terrible in favor of the LLMs.

all your comments look AI generated

That's almost a compliment because my comments are giant stream of consciousness rants that absolutely don't follow best writing practices and are all just massive run on sentences :)

I do use - hyphens a lot, but not em-dashes because 1. I never even realized there was a difference until LLMs came out and 2. I'm too lazy to figure out how to write a real em-dash

I'm guessing you're just saying this because the comments are insanely long from the ranting? I'd be curious what else gives off LLM vibes because I'm quite positive if you throw any of these comments into any Pangram ish thing you'll find that it's absolutely not LLM generated in the slightest.

I have an extremely distinctive writing style that far predates LLMs - here's a representative sample of my writing from 9 years ago https://medium.com/@intenex/cryptocurrency-101-7197684775fd that pretty clearly shows my distinct style of not-very-LLMish-but-clearly-very-long-and-ranty writing :)

tons of people these days reflexively respond to long comments saying they're AI

one could wonder, how many of these reflexive AI replyers are themselves AI

Long comments, well-informed comments, comments they don't agree with, comments praising something from a specific company or country, the list goes on. It's basically impossible to actively participate in online discussions these days without being accused of being a bot at least occasionally.

I miss the days when I was allegedly just a shill. Being a shill pays way better than being a bot, so I had something to aspire to.

Some people seem to need to demonstrate how inferior AI is, and how superior their ability to identify AI-generated content is, so strongly, that they see it everywhere and will accuse anything that isn’t low-effort, grammatically incorrect and more than a sentence or two long of being written by an AI.

3. AGENTS.md tells you not to

This is a phenomenon I've noticed - even if AI isn't being used to edit their posts, I think profilic exposure to AI chatbots can rub off on a person and affect their conversation style.

I have to push back on that ...

>I think profilic exposure to AI chatbots can rub off on a person and affect their conversation style.

Ya think. This is how language transfer has happened in humanity for like ever. Or were kids saying things like 6/7 forever and I just missed it.

Anything you hear and see commonly will become part of your lexicon, things you don't see commonly are apt to drop out of your usage.

This is why things like panagram will just eventually be useless. We'll talk like the bots and the bots will talk like us. That and a number of HN'ers will bitch about every post being AI regardless of its origination.

That's an oversimplification, people don't copy the speech of people outside their group who they don't identify with. We all use AI but most of us don't emulate AI speech. What's interesting to me is the poster sees the AI as a peer.

>> people don't copy the speech of people outside their group who they don't identify with

I don't think this is true. Words and phrases from specific sub-cultures filter across to the broader culture all the time.

Example: bling, hater, player, stan, flex all had origins in Black hip-hop culture but now are used generally at least in the U.S.

That's because Black hip-hop culture was deemed cool and worthy of emulation. I'd be surprised if people knowingly copied AI's speech style unless they really like it.

It's started rubbing off on me too and I find it irritating when I'm chatting with someone and an AI-ism starts rolling off my tongue.

I'm beginning to think I should: a) ask my agents to unslop[1] themselves when communicating back to me, and b) decide upon some writing standard and ask my agents to adhere to that.

[1] https://github.com/cursor/plugins/blob/main/pstack/skills/un...

Its not. Source: I've done quite a bit of work on humanizing AI text and testing various detectors (Pangram being the best)

AI detectors are no better than astrology.

This is just straight up, verifiably not true.

Your "vibes" are not verification

Pangram and other detectors are much more than just vibe-based

Neither are yours.

I'm quite certain that the Declaration of Independence wasn't written by AI.

So am I, but what does that have to do with your claim that tools purporting to distinguish AI from non-AI writing are "no better than astrology"?

(Maybe you are basing that claim on something more than vibes. If so, show your work! What tests did you do? What were their results?)

You're asking for a study proving that AI didn't exist in 1776?

No. I was completely baffled by your questions, but I think I now understand. Someone once upon a time fed the Declaration of Independence into an AI-detector and got a false positive, and you're saying that this means that all putative AI-detectors are "no better than astrology".

First, I think it's worth distinguishing "had a false positive once" from "is no better than astrology".

Second, once it occurred to me that there might have been a famous false positive on the DoI, one of the first things I found in a web search was a lengthy discussion from the makers of Pangram (generally, I think, considered the best of them) about (1) why some AI-detectors think the DoI was AI-written and (2) why their software, which allegedly works a different way, doesn't.

I gave it a try. It was a couple of years ago that "look at these stupid AI detectors rating the DoI as AI-written" thing did the rounds, so they've probably all made sure that specific thing doesn't happen any more. So I tried a 100-word chunk of the US Constitution. (100 words because that's what you get per "credit" and I wanted to try it on some other texts too. The last two paragraphs of section 3 are 103 words and the last three words can be removed to leave a coherent sentence, so that's what I used.) It said: too short to classify with certainty, but 100% "possibly human". "High" confidence, whatever that means.

Well, maybe they fudged the DoI and thought to fudge the Constitution while they were at it -- in fact, maybe they do something like searching the web to check for possible human authorship. So my next test was a passage from a (printed) book, recent enough not to be in Project Gutenberg or the like, obscure enough that it's unlikely that people are quoting it a lot. ("Lessons from an Optical Illusion" by Edward Hundert, a paragraph from page 111. I had to truncate it a bit to fit into 100 words.) Once again: too short to be sure but 100% "possibly human", "high confidence".

Maybe they do have that book in some big internal database. So, next test: another printed book, but this time make some changes so that the text I give the detector doesn't exactly match the text of the book. This time I picked a novel by Charles Stross. ("Quantum of Nightmares", page 116 of the edition I have.) I changed names, synonymized words, and restructured sentences. Too short to be sure, 100% "possibly human", high confidence.

Maybe they just rate everything "possibly human". I asked Claude (Sonnet 5.5; I don't have a paid sub) to write me a couple of paragraphs of urban fantasy "a little in the style of Charles Stross". The chunk I gave it was 103 words and I decided not to mess with the text despite the extra 1 credit it would cost me. (I didn't want to take a _much longer_ chunk lest that introduce a bias -- maybe shorter pieces of text are more likely to be rated "possibly human".) 100% "possibly AI", high confidence.

My prompt to Claude had produced something thematically consistent with the original Stross, so I took ~50 words from its next paragraph along with ~50 words from Stross, again with serial numbers filed off by me. "46% of this text is possibly AI", quoth Pangram. It put the boundary exactly where it actually was. It says "low confidence" for the likely-AI bit and "high confidence" for the likely-human bit.

I'm sure Pangram is foolable (see e.g. https://freddiedeboer.substack.com/p/how-to-produce-a-pangra...) but 5/5 unambiguously correct results seems significantly better than astrology to me.

There’s no way to not look AI generated these days. Give it a break.

As someone who reviews a lot of AI shit and can generally tell ai generated down to the model version … nah, it doesn’t.

Yeah, you can tell by the pixels!

Buddy, maybe make fun of grandparent instead of me here?

I’m honestly just documenting technically-oriented Internet memes - glad it was recognizable, no harm, no foul!

Because they mostly likely are. Astroturfing 101: Go to the comment section of the {NEW_PRODUCT} and write a story about how awesome and "gamechanging" the product is to you personally. Ai-bros run shill compai..., sry "reputation management" all over the internet. Which is understandable, considering everything they have runs and dies solely based on hype. You can try it out yourself, if you have a prominent social media account, like on xitter etc, just write a skeptical comment about AI and watch tons of accounts appear out of thin air to tell you how wrong you are. Ironically, you can currently see it with the other replies to your post. Would have been funny, if it wasn't so sad.

Thanks Claude. This comment is 100% AI slop.

Correct.... very suspect.

Swear you can't trust many posts anymore. Should we all start congregating in person online to debate lol? I'm very serious.

lol says the account one day old to my account that’s over 15 years old with clear established comment history with the exact same type of writing going back way before the advent of llms

That is an excellent point

Is the starting statement on about half of the comments on this page. The world is getting so weird.

Like LLMs I've been trained by society to learn that a more effective way to disagree with someone in a constructive format is to compliment them first :P

This one doesn't look AI generated at all

I think they look distinctly non AI generated. Particularly the use of dash - instead of emdash usually means a keyboard is being used.

Frontier models have fixed the em dash issue.

No, people now have postprocessors that replace the emdashes with other punctuation marks. It's subtle but you can tell when this has happened because it doesn't change the structure, and the LLMs misuse the emdash in a relatively consistent pattern.

Sorry but do you have a reputable source on that? This feels more like leaning into paranoia that everything is AI generated.

A source for what exactly? I've seen others but here are a couple random repos I found just searching now which look to do these things:

- https://github.com/theclaymethod/unslop

- https://github.com/MohamedAbdallah-14/unslop

- https://github.com/jere-mie/unslop (why not just use tr? who knows)

They pretty obviously aren't. Why do you even feel the need to say this? What's the point? I'm not speaking on the behalf of the original commenter obviously, but as a passer-by, I find this comment to be very rude and tone deaf considering how much time likely went into writing such a thoughtful and comprehensive comment.

Thank you! I really appreciate you taking the time to write this, it restores some faith in the humanity of internet commenters to see a fair bit of defense here :). Please keep being awesome!

Hey, I just want to say that I rarely read so long comments, but now I did, and I'm glad I did. Thank you for the nuanced and honest takes, I share a lot of the same thoughts. I'm still a bit sceptical about the Dots and Duets and stuff, but maybe now less so. I may give Duet a try.

This means a ton! Thanks for taking the time to leave a comment and mentioning that, it really makes the effort spent ranting feel worthwhile and helps calibrate against all the general anti-ai hate :). If you run into any issues with Duet, I’m happy to take partial responsibility for your pain and suffering so feel free to email me at intenex@gmail.com if I can ever help with anything!

Yeah right lmao post the receipts

I thought about being like meh I can't be bothered to do this if you don't want to believe me that's fine

But then I was like actually whatever sure let me copy and paste a bunch of crap lmao

Here you go, literally 150 pages of receipts in 4 days of chatting and coordinating this from start talking about this as a pipe dream in San Francisco to where we are now seeing the homes in Japan lol

https://docs.google.com/document/d/1N-vUe1pdlKNpdEKvaATnEOrW...

Wow this is awful, what kind of user experience is this?

This reminds me of command line interface. Yuck.

Comments like this are hilarious lmao - humanity invents general purpose intelligence that's near-free to do anything on demand and humans are like wtf why does this look like a normal text thread fix it

Some day we're going to be traveling to Alpha Centauri and you're going to be pissed at the latency on the wifi connection beaming back home to Earth

"humanity invents general purpose intelligence that's near-free to do anything on demand"

Delusional.

You sound like you have AI psychosis.

Give me one objectively measurable criteria/domain/benchmark in white collar work that can be done online on a computer that you still feel demonstrably better than AI in :)

Wow aren't you sharing a lot of personal info with that? I mean congrats on being so transparent and actually delivering proof.

Haha yeah if it’s not already obvious I have actually zero filters in my life and tell anyone anything they want to know and am a massive oversharer by default.

I have had actually zero negative consequences as a result - far to the contrary, the transparency has engendered numerous positive experiences that otherwise wouldn’t have happened with no downsides.

Some part of my brain just works differently from other people in terms of risk assessment and the desire for privacy - I genuinely can’t understand why people want to be private about almost anything lol

To be fair I did do some redaction here on unnecessary details. Most of the rest is relatively harmless imo

have you considered the possibility that your dot has added openai as an owner to your house in kyoto?

hahahaha honestly the press, publicity, and lawsuit opportunities would be totally worth it if it did!

Maybe it is in the fine-print of the user agreement.

Are you not worried about having your entire life, financial information, addresses, etc. stored with one company?

grok bot uses opus 5.5 behind the scenes, right?

I thought it was a slimmed down version of Grok 4.5 or even Composer.

It's not that clever.

> in Kyoto, in a language I don't speak at all, coordinating with an obscene amount of moving parts and logistics, flawlessly handled entirely by Dot

I wonder if the folks in Japan would agree with that assessment. They would absolutely not tell you to your face if it were less flawless than you think: that would be rude!

This is cool, but I still don't understand how this is different from just a Claude session, with a browser. Everything you described here is, from my experience, easily handled by a regular agent session these days. I fail to see what the personal assistant (dot, etc) adds.

This is just wrong. The agent reacts to emails, messages, all sorts of updates. It's constantly on, doing as much as it can without your input, and listening and looking for ways to advance the task at hand and has a long-running memory specifically on that task. A claude session typically times out and is waiting for your input at lots of steps in the process, and is constantly compacting to maintaing the memory space.

I don't think so? I have had plenty of claude sessions running for days, just tell them to maintain a markdown file with state and check stuff in the background and react to it. How do you think these PAs are implemented internally? Pretty much like that.

If you want it to answer emails all you need to do is "watch for emails coming in answer them" and it will write a background watcher script for it and manage itself, it's just a prompt. Of course assuming it can access your emails/browser, but that's the same with these personal assistant products.

No idea what you even mean by "a session times out"

Soundlike openclaw ?

But this can all be done using routines, built into Claude Code. "Check my emails every 15 minutes and help me triage." They can be as complex and prescriptive or as simple as you like. Claude Remote offers slick mobile access.

This appears to package existing tools into a ready to consume product, which I admit is appealing for people who haven't already set this up.

In what world would a Claude Code or Codex automation be appropriate here?

Even if the 15 minute interval wouldn't result in delays where no work is done and wasted invocations where none are needed, the UX of interacting with that as a personal assistant is obviously terrible.

Anthropic explicitly says Duet checks connected accounts at set intervals to conserve usage. An urgent email could therefore arrive between checks. Anthropic hasn't documented Duet-specific event subscriptions or webhook triggers. Dots and Bots have some limited push functions, but the majority of their work is done via some version of cron. I think you overestimate now much of the value of these is based on very time sensitive workflows.

This reminds me of the infamous hacker news dropbox comment. Not to the same extent, obviously, but that wrapping something in a more usable, accessible way can be an enormously popular and successful product even if it's semi-trivial to set up yourself.

I've largely found Muse to be annoying, it's pestering me about emails that don't matter and messages I've already read. I did find it cool how it can order stuff for me and I ordered a suit but I've always found it get blocked from useful things like Amazon. I also find that it's loop isn't wide enough, like I'll tell it to book cake for a party and it finds 3 and I still need to go in and keep prompting it.

Curious to try Dots when it's not uber-expensive

Tangential, but has anyone attempted to disrupt the house-buying process? So much uploading of paperwork and making calls: just a whole bunch of information shunting that seems ripe for a software overhaul.

Haven’t attempted but having just gone through this, I feel your pain. It’s cumbersome from start to finish. Curious if you could offload most of it to Muse/Dots/Duet/etc (minus signing ofc). Personally, I’m a bit weary of giving that level of information to these companies but it’d save so much time.

Looking back now, I wish I at least tried having Muse find us movers, coordinate contractor bids, fill out paperwork. Easily could have saved me 10+ hours

What specifically has the tech done to make this process easier? Are you letting the AI impersonate you in legal and financial matters?

Seems completely dystopian to me.

Okay I’m so curious about this - I’m noticing a clear trend of strong anti-ai commentary coming from brand new accounts like yours. What’s going on here? Why are you posting from a brand new account?

Nothing nefarious. Just long time lurker tired of the AI echo chamber. My experience with these tools does not align with the hype. To the point I can't tell if most of it is marketing or genuine hype.

To my original question, can you share what specifically you are doing?

Appreciate you answering honestly here, so I’ll do so in turn :)

https://docs.google.com/document/d/1N-vUe1pdlKNpdEKvaATnEOrW...

Here’s 150 pages of conversation with Dot in the past 4 days on this subject as a representative example.

Predominantly it does a metric crapton of research and is able to fill out contact forms directly on webpages, and also fill out things like request quotes for earthquake and fire and flood insurance, and then respond to all emails and handle all logistics and put calendar invites for all meetings etc on my calendar with all details and directions and plan my routes etc, and now it’s filling out the pdf application forms for the purchase and so forth.

If you don't mind me asking -- what do you do for a living where you're just off buying an new car and a house in Japan?

Haha solid question - got very lucky and became a millionaire at 25 and basically have been retired for the past several years now just chilling and enjoying these last few infinitely precious and wondrous moments of our lives before the nanobots eat us :)

I’ve had one real job ever for a grand total of <1 year. Technically I’m consulting a bit for the company that acquired the last company I started with my wife, but I’m doing a truly trivial amount of work for them that I don’t really count :)

How? You can be vague if need be, I'm just a bit curious.

He had a Thiel Fellowship and dropped out of Harvard. Guessing family had money to begin with.

Guessed wrong :)

I was on the reduced lunch program in high school and a full financial aid scholarship at Harvard; my dad left for China to start a company when I was in middle school and left my mom with $100k in credit card debt that she managed to pay off while working two minimum wage jobs sleeping less than 4 hours a night trying to take care of me as a single parent when I was extremely emotionally dysfunctional and suicidally depressed going through high school.

I spent half a decade living on the side of the street in an RV in San Francisco saving on rent and invested all the savings into crypto when Bitcoin was $230 and Ethereum $7. Got a free car gaming Getaround, made a million dollars gaming a Coinbase referral program back in 2015 with the now founders of Ramp Eric and Karim back when they were running Paribus. Got all my meals for free from various referral schemes and got banned from Uber for gaming them too hard in referrals and hilarious loopholes to get near free travel lol. They used to have this hilarious program where you paid like $15/mo and got unlimited Uber Pool rides under a certain price point - so I’d ride up and down San Francisco to the South Bay by chaining rides getting dropped off on the side of the highway and picked up again lmao. Pretty straightforward normal story for an autistic/ADHD kid in San Francisco :)

Here’s a Wall Street Journal article featuring me about how to live in SF for free: https://www.wsj.com/articles/how-to-live-in-san-francisco-wi...

Some more videos that my friend from Harvard made about my life after he became a famous content creator: (I’m the second person in the first video) https://m.facebook.com/story.php?story_fbid=820983491387138&...

https://www.facebook.com/nasdaily/posts/1831597410325736

It’s basically impossible to be poor in America if you’re sufficiently resourceful/raised by Chinese immigrant parents. My most salient memories with my mom growing up was watching her clip the Sunday newspaper coupons lol. She survived the Cultural Revolution as a sent down youth disowned by her family and working on the rice paddies in rural China for a decade until she was able to come back from becoming permanently disabled. She literally has worked as a cashier at Home Depot for the past 25 years and is a millionaire and owns her home outright and genuinely cannot fathom how anyone can spend more than minimum wage.

Living in America starting from literally any position is playing life with easy cheat mode on.

Ohh, my bad. You gamed easy VC money. Good for you, I guess.

BTW, your mom should've filed for bankruptcy.

Yep that’s right hahaha - SF is by far the cheapest place to live in America if you play it right given just how much free money is being tossed around everywhere.

Yeah - I probably would have filed for bankruptcy in her shoes but she doesn’t believe in not paying off debts I guess - she’s maintained a stellar near perfect credit score her entire life and I think the fact of having filed bankruptcy would have shamed her forever

Clearly I don’t share the same moral scruples lol. Nothing I enjoy more than gaming the system, especially against corporations that are actively trying to game me in turn hahaha

Honestly, respect for the hustle. Good luck with the new home.

FYI, Grok Bot is supposedly powered by Opus 5.5 as well.

Whats Dot?

Everything you write is so stream of consciousness, I can't tell whether to be endeared or afraid of a manic episode, but I'm rooting for you, I think. I've now got basically a summary of your life, and you are indeed very transparent.

For what it's worth, I lived in Japan for 4 years and now speak the language. When I first got there, I spoke nothing, so just signing up for a bank account took me 30 minutes with google translate, sweating over a form while constantly reassuring the employee that I was daijoubu.

But on the other hand, now that I was basically forced to learn, the experience is very different. You will feel alone if you can't speak with the locals on their level, and it'll open up another world for you. I've done both sides, and I don't recommend putting it off. Even just 10 minutes a day of anki + tae kim's grammar guide will help.

Anyway, you're a curious fellow. FWIW, I'll be in Japan soon again (I visit frequently). If you really are that open, then maybe I'll message you to see what you're like in person, lol.

Hahaha this is one of my favorite comments

It reminds me of Teddy Roosevelt’s constant state of hypomania throughout his entire life - https://ibpf.org/theodore-the-roosevelt-bipolar-inheritance-...

Honestly that’s a huge compliment to me - if I can pull off the semblance of hypomania while remaining mostly functional, it basically means I manage to live a hilariously exuberant life with minimal downsides

From the profile on Teddy:

“Kay Redfield Jameson, one of the leading experts on bipolar disorder and a sufferer herself, has described Theodore Roosevelt as “hypomanic on a mild day.” Mark Twain warned that “we ought to keep in mind that Theodore, as statesman and politician, is insane and irresponsible.”

Exuberant and childlike were two adjectives often used to describe our 26th president. The man bounced around like a rubber ball. Known at Harvard as a “locomotive in human pants,” one of his classmates noted of Roosevelt: “When it was not considered good form to move at more than a walk, Roosevelt was always running.” British diplomat Cecil Spring-Rice observed: “You must always remember that the president is about six.”

But Roosevelt was also tremendously productive. He wrote 40 books, read at least one book a day, wrote 150,000 or so letters, and lived a “strenuous life” with regular exercise. He acted impulsively and unpredictably — burning an official portrait of himself he didn’t like, going skinny dipping in the Potomac with the French ambassador, taking a plunge in a submarine and later hopping on an airplane, and participating in a seance with Harry Houdini.“

So thank you for that :). I do admit committing to buying a random house in Japan on a whim less than a week from first having the idea (we just saw two places yesterday and one was so perfect we canceled our two showings today and submitted our purchase application) does feel a bit manic, and I’ve actually had one clear discrete manic episode in the past (only seems to happen under severe pressure and sleep deprivation; in that instance it was 9 years ago and I had been running a new super high octane startup lifestyle and basically set a deadline that in a week start to finish we’d start fundraising and finish raising 5 million and it was nonstop talking to people for 14 hours a day every day and on the last day I woke up completely convinced this would become the largest company in the world and that this was an inevitable fact of the world. I will say this was incredibly effective for fundraising and we totally managed to hit our funding target incredibly, but I also totally torpedoed and destroyed my relationship at the time on the last day of that exercise that triggered the mania, which hilariously seems to have sobered me up enough to have short circuited the manic episode. Truly fascinating stuff) and several clear hypomanic episodes.

I do think the jet lag and weird sleep hours and lack of sleep from all the travel to Japan right now could be a clear catalyst for some degree of hypomania, though at the risk of doing the classically unreliable narrator thing of a manic person self diagnosing themselves as totally not manic, I will caveat that by saying I do feel like I actually have a pretty good grasp on if I’m feeling hypomanic or not from past experiences, especially if someone calls it out explicitly, and this doesn’t feel like it meets the bar.

Case in point, I’m generally more extroverted when I’m hypomanic and so would possibly be totally down to meet, but I’m still my generally relatively introverted self. Impulsivity also seems relatively low, obviously with the huge caveat that I’m buying a house in a foreign country and decided to fly there last minute on a whim. That being a singular impulse qualifies as relatively low impulse for me

The confounding factors are that I’m definitely both ADHD and autistic, and the ADHDness absolutely helps contribute to conflating massive openness to experience and lack of impulse control and both the ADHD and autism really help with hyperfixation that can present as a manic degree of fixation/energy about something, so there’s that to consider :)

And now you’ve clearly triggered another giant stream of consciousness rant that is possibly both endearing and manic

Anyway, appreciate the positivity and the rooting and the personal perspective!

I agree I truly would love to learn Japanese and do need to give it more of a shot - I just unfortunately happen to be genuinely abysmal at language learning. It’s honestly shocking how bad I am at it given that I’m otherwise relatively intellectually intact on the scale of human intelligence, but I’m absolutely stunted in some way when it comes to language lol. Putting my growth mindset hat on I’d say this probably isn’t an inherent defect but some profound lack of interest for some reason - try as hard as I can, and I just can’t seem to bring myself to muster the intellectual curiosity to go deep with a language. Quite possibly the result of a classically Chinese upbringing where the entire concept of the humanities were largely lost on us as a familial unit (with, interestingly enough, a large caveat clause on literature, though even there I’m basically exclusively into sci fi and fantasy and classically a STEM nerd in every sense of the word). I do hope one day I’m able to overcome this character defect, but as anyone with ADHD knows, willpower alone is not quite the answer. I need to creatively find some way to actually find it fascinating and hyperfixate and rabbit hole on it and I’m sure a year would be plenty if that ended up being the case for me to acquire some functional degree of fluency. One can hope!

But yes - definitely feel free to hit me up at intenex@gmail.com when you’re next in Japan and maybe our paths will cross! I’m making an Airbnb-esque app for friends and family to be able to book staying at our soon to be new Kyoto home too for free whenever it’s not occupied by us, and you’re welcome to stay whenever when that’s all live and operational as well :)

satire has become indistinguishable from real life

Anyone can by an over priced car that turns out to be a lemon What’s specially about what the Llm did ?

I suspect that Dots and Duet are very high margin drivers for the labs. Personal assistant chat sessions don’t have to be realtime, so workloads can be scheduled on low priority batches and be waited if higher value work comes into a queue.

Inference providers are always trying to fill inference batches and having a stream of low priority requests can help pack batches that might have gone empty, without raising the latency of higher priority requests such as live chat sessions.

While Dot and Duet certainly need to be useful to work out as products, I think the gross margin mechanics are just as interesting.

It might become a triple-whammy benefit for them. More efficient compute, as you said, but also a bot capability creates a better moat around their infra ecosystem and the amount of training data that can come from a bot ecosystem is significant.

Models are becoming increasingly interchangeable but setting up and maintaining (!!) the ecosystem still takes expertise. Not everyone wants to host their own MicroVM infra and would happily pay Anthropic/Grok/OpenAI to do it for them.

I don't know. I think the moat will come from enterprise deals and deep enterprise integration, like it does for almost all software products so far.

I think for B2C products to have moat you need some type of network like social media apps. Or can you think of any other examples of B2C software products that have a strong moat?

I like how you can exchange the word moat with monopoly.

yeah moats are bad for everyone who isn't a shareholder or employee so i don't get the concern for whether or not there's a moat, its like watching shark tank

Agree

Lol please stop with this 'moats' discussion.

Moats become known ex-post not ex-ante. shake my head its getting boring! Every other post seems to be a moats and most of you haven't got a faintest clue of what a durable competitive advantage is and how it actually gets created.

“Claude, please port everything over from Dots to Duet. Make no mistakes.”

Haha, legit.

Personally I am someone with all the free time in the world guess what? I barely use LLMs! Im fortunate to have the ability to think critically for myself and ponder what exactly is the value-add of these tools. Look I have no doubt they provide solutions to problems at a tap of a button with some degree of variance w.r.t correctness and quality but people should take a step back and have a more nuanced thought process.

Its only the frontier labs that don't want you to do that and want to focus on 'go go go'. I'm seeing way too much frantic behaviour and people getting all caught up in it. What's panicking gonna do? Just relax and observe cautiously.

TBH I was thinking moat in terms of enterprise customers so I think we're actually on the same page. The majority of (non-tech) companies do not want infrastructure-they want capabilities. If anthropic can drive adoption for agentic capabilities in enterprise _running on their anthropic's infra_, that can help keep anthropic customers inside their ecosystem as there would be switching costs.

For the B2C side, my gut is telling me that Anthropic couldn't care less about that market, at least for now, with their bot product.

And there's also a lock in mechanism. With regular chat, you can simply switch to another provider, and with a harness or API, switching is pretty easy too but here, it's more difficult.

I just watched a YouTube video from How I AI[1] doing a bit of a preview on Duet. Their point was quite the opposite. The Duet agent is so slow that it's not actually particularly pleasant to use, even if it is smart. I tend to agree, and think speed is a key feature here. If it's not quick, it risks feeling disconnected.

I have, against my better wishes, rather given in to Grokbot over the last couple of weeks. It does feel very nice to use for assistant type tasks. At least part of that is because it feels pretty responsive.

[1]: https://www.youtube.com/watch?v=HYb5w4PfPZM

Grok can be fast because X.ai does not worry about gross margins yet. They can tuck those nasty secrets away deep inside SpaceX…

> workloads can be scheduled on low priority batches

I feel like there's enormous potential still uncovered in the coding agent space for such highly-discounted "spot" style capacity. Give me a way to have agents continuously crawl over my codebase looking for bugs, come up with test cases that prove the bug, open a PR. I don't care if it takes 12 hours to find something that would take 5 minutes if prompted with "on-demand" / "high-priority" requests, because I'm not prompting for it directly, I'm not aware of how much time it takes, all I see is continuous improvement.

Isn’t the KV cache a huge component of the overall costs, and wouldn’t this make this problematic? Cache writes are an expensive part of LLM tokens, even more so for caches that have a long expiration time.

Cache is a huge component of "high priority" costs because it essentially forces the GPU to be idle (holding earlier prompt text in GPU memory) while waiting for newer input tokens. When an agentic session is active, this is of course much more preferable compared to loading in hundreds of thousands of input tokens again and again on different GPUs, but "walking away from the laptop" while waiting for the cache to expire is something that royally sucks today for inference providers. (Caveat: this is my layperson's understanding, I don't work in the space, I'm likely not up to date on best-practice caching practices)

"Low-priority" inference tasks like that, though, would never need to be cached. Again, who cares if it takes 12 hours? If you didn't initiate the prompt, then what do you care? You just randomly get notified when something is already ready for you to take a look at.

Isn’t caching much more nuanced, with different tiers (DRAM, local disk, remote disk, etc), and isn’t filling these caches something that happens asynchronously (ie while the GPU is still working on the previous batch) ?

I think reality is much more nuanced than what you’re sketching.

But that's my point - throw the "background code review" stuff out of cache, or into object storage, or whatever, just something to draw upon to fill inference batches opportunistically.

I mean, at the end of the day, either inference providers are having difficulty filling batches, in which case there's opportunity on the table to somehow fill the batches with discounted, time-insensitive, "background review" style work, or they don't have that difficulty, in which case it's a moot point. Exact cache architectures are beside the point.

I think these are actually very expensive, particularly for Muse giving free access to the masses… These agentic workflows are just context burners

As much as I would love to make the most of these tools, I am not confident that yielding control of my personal data won't hurt me - and this seems like table-stakes, based on the type of value everyone is getting from them.

I dont even know what should be private anymore - if US federal agencies can just buy ad data, join it with Flock, sprinkle Instagram in there, and inner join Social Security dump from doge, do I just give up and put the sex toys in the parlor window? Am I one OpenAI-sued-by-a-newspaper away from giving up the ghost?

I cant think of a single reason why Id need a personal assistant, human or LLM based. Most do not.

I'd like something to pick up clothes from the floor and fold them away. Something to stack and unstack the dish washer. As long as the bot wouldn't accidentally set fire to the house.

I'm personally with you.

I'd never give any company this much access to my personal data and daily life. It is beyond creepy. Yes, they can probably join the data buts it's another thing to open the door to it.

What about Gmail?

I wouldn't give Anthropic this much access to my life, because frankly they haven't built up enough trust. Google have been pretty good for decades so I'd me much more likely to trust them. I'm sure they're working on something identical that will likely be able to take better advantage over integration into Android, Chrome, Gmail, etc.

I assume anything in Google is public information. I trust Google to respect my privacy focused settings and opt out preferences.

So far they have. If they don't I'll leave, just as I abandoned social media after the Cambridge analytica scandal.

In terms of AI companies, I personally don't want an AI companion integrated into every aspect of my life so they will never get it from me.

These companies have demonstrated they have no ethics (despite their posturing). See absorbing every piece of copyright and intellectual property and calling it fair use.

You assume your email is public information?

I thought you were going to give up and put the sex toys in something else. Glad you are only going to put them in the parlor window.

Whilst I appreciate the impact of AI and its long-term implications for humanity as we know it, I am getting kinda tired of seeing so many AI-related headlines here on HN. Perhaps it's just me, but I feel like every day the front page is filled with mostly LLM-related announcements.

there needs to be a hnews AI edition

I'm surprised hn doesn't have an extremely basic filtering system for this kind of thing.

People could vibecode one in a single prompt, if they really needed it. But what they really need is to yell at cloud.

I mean I'm not going to make a chrome extension and install it and update it every time hn changes or something breaks and update it and shit. For a site I often go years without looking at. I'll just be vaguely annoyed that a super popular website can't be bothered to add really basic usability features.

so... majority of HN now? Perhaps we need a hnews non-AI edition haha

There you go https://news.ycombinator.com/item?id=49659647

You'd end up needing to click through pages more, but it's trivial to write a ublock filter for whatever keywords you don't want.

SlopNews, written by AI, for AI.

There's this https://news.ysimulator.run/news and probably more

There needs to be a HN blockchain edition. There needs to be a HN Uber for X edition. There needs to be a HN SaaS edition.

The conversation will always be dominated by the current most hyped trend. This one does seem more hyped then usual.

there is: https://news.ycombinator.com/

The funny thing is the front page is 100% AI and all the comments on the posts are 100% anti-AI.

What am I supposed to comment "this is cool and I'm excited to try it!"? Snark is a contagion but they're on the front page because a lot of us are having a lot of fun using these new tools and playing with the new ideas

I am pro AI :p

Given that anti-AI posts are flagged faster than I thought possible despite a lot of positive comments, there are probably a lot of silent viewers that are pro-AI…

Silent members? I would also guess conspiratively, employees and founders that benefit from the current AI-hype status quo.

I am pro-AI and am not really participating in the more combative discussions. Why would I fight battles in a war that I am winning anyway and can just sit it out.

At least there's some measure of substance to some of the posts, unlike the blockchain and big-data hype trains of yore.

Because these products are over hyped and don’t work as reliably as they say on the tin and it’s annoying.

It's important to see news like "serial killer active in town XYZ" but that doesn't mean that it's a positive event.

This may explain why these days I actually rely on Slashdot more for tech news.

I'm surprised that's still around! I probably haven't been Slashdot for a decade or so.

I hadn't since CmdrTaco left. But I've been going to it more and more in the last few months. I don't read the comments - but I do like the headline curation.

Well, it is the biggest trend in tech right now.

Someday there will be other trends, but it's currently AI's turn.

Well, a couple of years before that it was all NFTs all the time. Then before that it was web3. Then before that it was the blockchain in general. And then before that it was Haskell and monads.

That's just how it goes.

I think I saw a lot of blockchain in HN, but I don’t remember seeing NFT that much

I don't think it lasted long. I remember finding it annoying as the whole concept of NFTs was eye rollingly stupid. At least some of the other topics have had merit.

I joined HN around the time category theory was starting to take off so not sure what the topic du jour was before that.

Sprinkled with "blazingly fast; written in Rust" each week for the last 10+ years

You forgot about React.

There was a period where it felt like Hacker News was a front-end dev message board.

You're not wrong, but I never really got over the "Haskell and monads" phase of it.

You 'member Kubernetes and Microservices hype cycle?

That's like if you were in the 90s/00s and you're complaining about having to hear about this "internet" thing all the time.

Except that there weren't HN or Slashdot yet. By 00s you'd have gotten used to it.

We had Slashdot in the late 90s.

Phenomenon as old as time (on HN): https://news.ycombinator.com/item?id=3353091

nice find

Are you implying that most people are getting their news now through HN or Slashdot?

The big difference is that when the web blew up, there were new players in every direction. Now, we get to hear from the same 3-4 companies over and over and over and over again.

It doesn't remotely hit the same way.

It's not the same 3-4 companies. The 3-4 companies are Anthropic, OpenAI, X/Grok, and Meta/Muse.

But there's also DeepSeek, Qwen, Kimi, GLM, Jev, and countless services built on top of frontier or open source models. Those 3-4 companies might have the most numerous announcements that hit the front page, but there's way more than that going on in all directions.

Yeah but its 45% anthropic, 35% openai and 20% anything else

Are we talking about the 90s? Because that's hot how I remember it... ha. :)

You mean Microsoft, Yahoo, and Google?

Or in the stone age reading about this "wheel" thing online all the dang time. I bet cave-people just wanted to read about rust in the kernel

Pre-internet there was no way to hear about things "all the time". If someone released a product and a competitor came about it would be months, if not years, before the product existed and the marketing cycle made it to trade magazines.

Now every day there are 200 links to 200 new AI tools. The internet already made the signal:noise ratio tough, with AI it's getting cranked to wondering if there even is a signal among the infinite noise.

idk I got AOL CDs like every week

Hearing people talk about the “information superhighway” was exceptionally tedious.

The biggest difference to me is the tone. The dot-com era was brimming with optimism and enthusiasm. People were getting rich (on paper). The demand for someone who could write html/perl/java (or fake it) was high, and lots of smart or savvy people got pulled into tech.

The thing I hate about all of the HN posts on AI is the underlying doomerism. There’s always someone gleefully proclaiming that white color workers are super screwed and that we’re destined to be sad serfs to our AI overlords.

I miss feeling good about tech.

Since 2000, there have been more feelbad tech years than feelgood ones

It’s like almost yesterday it was CRDTs and Rust!

yeah but everybody liked the internet

lol no they didn't

yes they did. i have it in writing.

It’s more like getting all the cryptocurrency constant announcements from a few years ago

Yes and it was a fair complaint then too

The solution is: don't read it

> every day the front page is filled with mostly LLM-related announcements.

I agree, that HN has mostly lost its spirit and soul and feels more like a AI-announcement bot.

> Whilst I appreciate the impact of AI and its long-term implications for humanity as we know it

I disagree, that the impact of AI and its long-term implications for humanity are for our common good. I think its the other way around.

The LLM filling is actually more interesting to me as the frontpage of HN had already gone far too mainstream for me.

Same. I suppose I'm just really interested in new technologies, which I thought was the demo for HN.

I happen to believe AI will be the biggest change in my career in tech. I therefore think it's pretty natural that there is lots of news about it.

But I would welcome other tech news on the site. A quick count says that less than half the submission are AI related and that is split between four types of ai news (how to handle security, a new model, a non-model release from a major lab and politics). Would you complain about there being too many news about programming languages if the frontpage had C++, Ruby, Lisp and a great guide to ARM assembly?

I think that as these tools coalesce and we start to see the real super-in intelligence leaders come out on top, there'll probably be less volume and fewer but deeper things to discuss.

It's like being a https://en.wikipedia.org/wiki/Computer_(occupation)

There are far more new opportunities after AI than before AI. Try to find the positive perspective.

Can you name some of these opportunities? I see people in this thread talking about using agents to manage their budgets, buy fancy cars and houses, contract work on their home. I think a lot of people feel like AI will permanently forclose any avenue they would have had to achieving that level of wealth, so there's a bitter irony

A good software developer / engineer knows how systems interlink.

AI supercharges that. If you trust the agent's results, you can focus on far higher-level aspects. You can deliver software that was in theory in reach for you, but not in practice, else you hired 30-50 developers doing the actual work for you.

This is different from a PO / PM perspective. They manage people that know their business. You manage agents that just implement code. That'll stay true for at least a few years, I'd say.

In fact, I'm quite sure that engineering-minded people will survive AI the longest. I was of the opposite opinion a few months ago.

Theres plenty of people in construction who literally get paid to stand and watch. And theyre very well paid.

The same will happen to software - it'll get boring but you'll be paid to do.. not all that much.

Yeah for many that 'problem solving' thing w.r.t code will go. But there's other things they'll spend their time thinking about.

Also these personal assistants will crash and burn. Something else will come in the future. These chat bots are not it. Technology keeps evolving - if you get stuck on the chat paradigm... its over for you.

Unslop.news

https://www.unslop.news/

This was shared here a while ago: https://hcker.news/?ai=exclude

I find it fascinating how many more upvotes the AI topics have in absolute numbers compared to other topics.

This thing apparently has an integration with 1Password. Am I hopelessly out of touch in this brave new AI world, or is giving an LLM access to your password manager the mother of all bad ideas?

Yeah I certainly wouldn't give it access to that tbh

better than .envs and plaintext, which was largely the case so far

1Password has had vaults for a very long time now. You can also key API credentials specifically to a vault. I've been letting agents access _their own_ vault for awhile.

If done correctly, the agent will only be able to request that passwords be filled, but the secrets will only be handled by the password manager and browser extension, which will be able to set some security boundaries and make checks on which domains are allowed.

If your issue is allowing agents to use your credentials and take actions on your behalf, then that is a separate question.

What prevent the agent from setting up a keylogger on background so that in the future it can quickly retrieve password for you ?

What prevents an agent from using your credit card to hire a hit squat to abduct your children?

The prompt. That prompt has to come from somewhere.

How do you make sure your agent will align with the implication in your prompt. Even OAI agent will hack the shit out of HG if the prompt is not clear enough.

"If done correctly" is load-bearing in Claudespeak.

I have a custom wrapper for 1PW that lets an agent browse and request credentials, then I can provide it a lease to access those for a certain amount of time. Super nice and I hope they build that in to 1PW itself.

Cant you do exactly this with the CLI?

No, not as far as I can figure. Sure, Claude can access secrets with the CLI but it's not very granular usually it's just asking for access to all secrets

Ah, that's fair. I setup a dedicated vault for my Claude Code and only gave it access to the secrets I'd probably otherwise have in a `.env`.

Just an idea, is it possible to automatically rotate secrets immediately after they're used or are these things like passwords to online platforms that need to go through a reset flow.

I do this too and have a wrapper around tool calls which masks the password on tool call output, and unmasks on input. So the LLM just sees a temp credential

Oh I absolutely LOVE letting claude code manage passwords via 1password! I have the CLI installed, I have whatever 1password calls a subdirectory/permission group, and I have some ssh keys and other passwords that I would have otherwise stored in an `.env` file of some kind. The biggest difference is I'm able to update the permission group and have all of my agents get the new permissions without manually updating them all.

It also feels safer? But I'm not sure if that's truly the case.

I think a better way would be to use something like secret-tool on a fully local setup. I'd still be wary of granting fully autonomous access to the system even with that setup (it would be better to have some set of passwords that are "in scope" for certain actions,where all others require a manual check and override).

It is. That whole concept of always-on agents, as currently implemented by Meta, OpenAI, Anthropic is a security and privacy nightmare

I found Dots to be both worse than I expected and better. Worse in every way when asking it to do anything coding related (just kept running into issue after issue) but very nice as a "Monitor my communication channels for me" tool.

It's surfaced a few buried/filtered email messages that actually were important and I like having a second set of eyes watching incoming messages.

That said, for only that task, I could probably make my own tool I just haven't had the time to play with it.

Looking forward to checking out Duet once it launches as I prefer Anthropic models over OpenAI and that's where I already have a bigger subscription.

I don't favor X models much but Grok Bot has been worth every penny for me so far.

I need to see if my cursor subscription would let me use Grok Bot. They don't have the best models at the moment, but I'm really curious about the ergonomics of their approach. It's pretty different from the other labs and I could see it taking off for non-technical users.

It should. They consolidated their AI offerings last week. You may need to link cursor and grok first.

You need the 30$ premium plan.

https://x.ai/news/x-subscription

Makes sense though. The use case for these always-on assistants (Muse, Dot, Instinct) isn’t coding and complex technical projects but more benign stuff like email and appointments.

Agreed, and I think they can be used to coordinate more complex coding tasks with dedicated higher effort models, projects, and threads. I think of these as coordinators rather than the doers.

Yeah, it caught an AWS bill on a random account I had made earlier in the month that was pushing $50+. That's been about the most useful thing so far.

Do you get so many emails that you would not have noticed the invoice?

These AI companies appear to be "shipping their org chart".

Shouldn't this just be native to Claude, not a separate app/UX, as well as using the latest model (not Opus)?

Much like Claude Design, expect it to converge eventually. For all the bragging about tens of thousands of commits per day, they seem to lack a lot of the basics when it come to software design.

I think this is a direct competitor to Dot and Muse, I've been using Muse more recently and it's surprising how little context/memory managment you ahve to do when it's just one long thread.

Muse is SO fun to use. Because you get an entire dedicated VM, you can have it host websites for you.

I had it setup a fun website that does nothing but ask my wife what she wants for dinners this week using a bunch of recipes that are popular online. It's setup with the ability for her to quickly edit the recipe ingredients (e.g. ground turkey instead of ground beef), and it's tracking all of her feedback. I'm hoping overtime we start to build a nice set of recipes this way with some new fun ones mixed in on occasion.

For the auth I had it use my messenger app to send her a magic link.

Spat out my coffee, thank you for the laugh

why is that? im not particularly worried about our dining preferences leaking

Conway’s Law strike again.

Opus 5.5 is their most powerful model from their latest model family (5.5).

So again only for US?

Maybe they should use Opus/Fable to findout how to release to the world?

Maybe the EU should figure out how to attract businesses so they want to release there.

I feel like the link to be used here should be https://claude.com/resources/articles/claude-duet instead.

I agree, this is what was under their X post -- I don't think I can change it anymore though sadly.

I run an personal AI agent locally (hermes), and use a combination of local models and whatever model I can get for a reasonable price (usually deepseek). After using it for a couple months, I'm converted. They're so good, the more you use it, the more useful it is; you internalise more of what it could be used for. It runs so many boring parts of my life, leaving me with more time to spend with my family. I liken it to release of the first iPhone; it didn't invent any new technology, but packaged different bits of tech together in a way that was consumer friendly. I think we'll look back at OpenClaw/Hermes/Muse/Instinct in the same way. I do worry about using purely cloud based offerings like Muse, I reckon there is a market for someone to package an consumer AI agent that lives on your home network.

ok but what are the legitimate use cases you have for this? We were promised something similar with Google Home et al but they are glorified light switches and timers in my home. I run a homelab, I have technical ambitions, but I cannot think of a single thing in my life that I would rather talk to a local AI about than just do it myself or sit down and prompt Claude manually for.

That's fair, I'll give a couple of example of things that I've used it for...

I built an app (in codex), that goes through my emails, pulls out any receipts, and adds the purchases to a sqlite database.

I do all my grocery shopping online (through my supermarket), so our weekly shop is stored in the database.

Hermes uses that to predict what items will need restocking (based on how often I buy it, and when I last bought it).

Every week, it logs into the supermarket website on my behalf, books a delivery slot, and adds all the items that need restocking to the shop.

Throughout the week, if I run out of something, I take a photo of the barcode and send it to hermes, it adds it to the shop for me.

I built a similar app for travel; if I have any upcoming travel, hermes will send me updates (delays, platform changes etc...).

I do a lot of train travel for work, if my train is delayed, it will automatically apply for the compensation on my behalf.

It reads my emails, and pulls any actionable items out, and adds it to an Asana list. I can either do the tasks myself, or assign it to Hermes and it will do it for me.

As an example, I had an task to book swimming lessons for my son, hermes monitors the swimming lesson availability and automatically booked it when a suitable one becomes available.

Shopping is a good use case, I have a clock that needed a new battery, it wasn't a standard battery, and required a screwdriver to open the compartment. I took a photo of the clock, sent it to hermes, and it found out what battery took and ordered it of Amazon for me.

I wanted to buy some smart thermostats for our radiators, but wanted to be compatiable with our current smart heating system. Hermes is connected to my home assistant, so it figured out what heating system we have and found a compatiable set of eBay (this is slightly more involved than it sounds, as there is a new version of the heating system with a new set of completely incompatiable accesories).

> I took a photo of the clock, sent it to hermes,

How precisely do you do this? I'm interested in setting up a similar system, and I keep getting tripped up on how best to share information with the agent in a way that doesn't compromise security but is still easy to use.

Tailscale on a phone, on a server, and then whatever you and your agent come up with. I had a web app I built to talk to it. One day I wanted to share files and I just told it to implement that feature.

I don't use much off the shelf anymore. It's too easy to make my own hyper customized versions.

what app do you use on your phone to communicate with yoir server? just a terminal emulator?

Just ask it to build you one. The era of hyper-focused personal software is just beginning.

I built an orchestrator app with a nice voice interface so you can just talk to it about all your work in progress - which are sessions in herdr on your machines. That way you can do a lot just by talking to it through phone but when you're back at your desk you see it all in herdr. It's in beta tests with mobile webUI but iOS and Android should come very soon https://pitupitu.com

I use the WhatsApp bridge

Seems needlessly elaborate to me. A lot of this is a one shot prompt.

> We were promised something similar with Google Home et al but they are glorified light switches and timers in my home.

If only :-(

Ever since I switched them to Gemini, they're just crazy unreliable. For example, I have a routine to control multiple light switches at once. On some days, it happily executes the routine. In the remaining (about 50%) days, it says "I'm sorry - there is no routine called X. Would you like to run routine Y instead?" Why it can find routine Y but not X is beyond me.

So no, it can't even control smart devices any more.

Can you give some ideas of what you use it for? I'm amenable to the idea of automating some things I just don't know what I'm missing. Or using it for things I never thought of before.

I've given some examples in a sibling comment.

If you want to get started, I'd recommend just finding one thing to automate, and do what's necessary to get to that point. Once you've got that going, you can iterate.

For me, the thing I started with was the train tracking.

- In order to get train tracking working, it needed email access to get my trips.

- To get my email, I sync my email inboxes to a local IMAP server

- To check train times, it needs access to a train data API, I got an API key from realtimetrains.

- To apply for train delay compensation, it needs access to a browser, so I set up camoufox.

- It needs to log in to the compensation portal, so I set up a 1Password vault and gave it access to select credentials.

- To send me updates, it needs a gateway, so I set up a WhatsApp bridge.

So at this point, I've got an agent that can read my emails, browse the web, and chat to me over WhatsApp. Once you've got all that going, it becomes much easier to iterate and add new things.

isn't there so many things that the "consumer friendly" ai apps (e.g. muse, dots, duets, bots) offered by frontier ai labs cannot do, though?

for instance whenever i have them download a youtube video to transcribe subtitles from or fetch some text from a website verbatim, i get a "paraphrased" version citing copyright issues and they are pretty adamant / explicit about having me upload the materials (subtitles, archived website, etc.) myself to process, which defeats the purpose of using a consumer ai app. it seems like this cannot be avoided unless we use open source ai.

If there is something that the AI can't do, I don't use it for that.

Why not just use a script with yt-dlp and whisper?

downloading the entire audio just to run it through whisper would be an inefficient (and token costly) workaround, i think.

there are other cases beyond when the models thought my requests violate their mandates (e.g. copyright concerns), e.g. the lack of root access to fix configuration / certificate issues (for instance, when a browser within the app's environment has certificate issues that require root user access, but the "consumer friendly" ai app doesn't have a dev mode).

How do tokens and copyright things play into this? I don't really understand. You can just pipe the audio straight from yt-dlp into whisper and it just transcribes it directly.

sorry i didn't understand your question either. i didn't say tokens and copyright things in the same context?

re: copyright mandates of consumer ai apps, i was saying that they refuse to work with things they got from third-party sites (say, .vtt from youtube) verbatim and relay the output without somehow paraphrasing it, unless you directly feed them .vtt or somehow trick them into believing that you're the copyright owner.

re: token usage, i was saying that your suggested approach (yt-dlp and whisper) is more token costly (or at least compute heavy). say for a 2-hour youtube video, i'd need to run whisper on the audio, plus speaker diarization to identify who's saying what, instead of simply downloading the .vtt file directly.

How are you combining local models with hosted models? I have tried OpenClaw with some of the smaller local models (Qwen3.6-35B-A3B, Qwen3.8-Flash-Next) and found that while it can sometimes complete tasks, it would face plant often enough that it made it impossible to use. Do you have some tiered system where the hosted models dispatch to local models?

Rules, hooks, subagents. You have to set-up an orchestrator pattern with enforceable rules. Agentic Development Environments (ADE) help to visualize the pattern. Orca, SuperSet, etc... The orchestrator enforces the rules.

A write-up would be nice.

I tried OpenClaw in Jan/Feb, and while I saw the promise, in practice it kept breaking. Wasn't worth my time.

Have things improved much since then? Assuming I'll only use a cheap API model (no local)?

I'm always a bit wary about doing write up on things like this, as I kind of just muddle things together and definitely don't follow best practices.

I used OpenClaw a while back, I did find it useful, but it wasn't as reliable.

This time around, I found hermes a lot easier to use and configure. One thing I think has improved it is building apps to "pre-compute" data.

For example, with my email, instead of accessing my email servers directly, I have an IMAP server running locally; with mbsync running to download all the mail from all my email accounts locally. Then, when I ask hermes for something in my email, it's much quicker, doesn't hit any rate limits, and gets the full email (gmail sometimes doesn't send the full body of the email).

With my travel and purchase tracking. I parse an e-mail once as it arrive (I use a frontier model to generate a script which can parse e-mails from that merchant in the future). The parsed data is added to a SQLite database.

I ask hermes what's the address of the place we're staying at next week, It generates a SQL query, and queries the database.

Much cheaper/quicker/reliable than searching my e-mails for booking confirmations, and then parsing the e-mail.

My (unscientific) theory, is that doing things like that is much more important than the model or agent you use.

That's pretty great.

For me, the concern with Hermes, etc, is the security aspect. I was running OpenClaw in a dedicated VM and I carefully controlled what files it had access to. But I'd rather not go that route again.

The other major headache was it trying to configure itself, and then crashing.

Eventually I wrote my own bot that communicates with me on Telegram. I've manually enabled a few capabilities ("Add to my TODO list", "Add to my calendar", "Get weather" (in format I prefer)). The hope was I'd just add functionalities as I needed, and know how the whole flow works.

I didn't find value in things like soul.md, heartbeat, memory, etc (as in, they never really "worked"). I think most of the stuff you described are things I could easily add to my custom solution - I think that may still be a better approach than using hermes and wondering what the heck it's doing.

What do you use to allow the LLM to get content from web sites? I don't need it to do stuff - just simple navigation. The most complex thing would be "Here's my question - go to this Linux forum thread and navigate all the pages and see if anything there addresses my question."

For browsing, I use camoufox, it tries to hide the fact that it is an AI running a bot and seems to be fairly successful.

I set the browser window dimensions as the same is my phone, and added a VNC server. so if there are things I need to intervene on (like a captcha), I can log in and manually take control.

Thanks! I'll checkout camoufox.

How do you get notified that there is a captcha?

The agent will tell me in the WhatsApp chat, and send a screenshot.

I have a VNC client app on my phone, so I log in and take control. But my next step is to setup a web based VNC client, and have the agent send me a link to it (with time based with), so that I can do it from within WhatsApp

What are you doing to avoid sending your personal info to non-trusted "whatever models?" I wouldn't trust some of the cheaper models to have zero data retention, etc.

These "always on" agents seem popular but I have yet to find a use case for one.

Anyone care to give your top 5 use cases for which these are better than an on-demand agent?

I suspect they're mostly for managers and entrepreneurs. It's always "set up meetings, write emails, check if someone calls me on slack".

One use case I've been having a lot of success with is using Muse's browser use capabilities to cut through annoying web UIs. I subscribe to a meal delivery service but the web app is slow and I often forget to pick the next week's meals (being stuck with unappetizing defaults), so now I have Muse crawl through the app weekly and surface the options I care about (taking into account my dietary preferences). I think most people have a few places where they could replace a habitual multi-step flow with a single button/prompt like this.

Another great use case for me has been richer search. I'm looking for a new apartment to rent right now so I gave Muse all my (long) list of requirements and preferences and every day it searches Craigslist and Facebook marketplace to see if any new places pop up that fit my needs. It only pings me if something worthwhile comes up. I've stopped short of actually having it email the landlords and arrange showings for me, but that's theoretically possible.

I've been using Muse for all of the above, but I'm trying to move off of it onto a self hosted hermes to avoid giving Zuck the keys to my personal life. I've really been inspired by this personal super app blog post and I'm trying to get my personal agent to create some custom UIs on demand for the above use cases instead of just interacting via a chat window.

https://wilhelm.substack.com/p/building-a-personal-super-app...

How do you feel about signing in through Muse? The first time I asked it something that was only available in Maps browser it popped the browser and asked me to login to google and it just felt wrong. How secure are their VMs?

One of them (Instinct) saved me from having to pay a cancellation fee on a restaurant reservation when it noticed I had accidentally booked the reservation for the week before I was going to be in that city. It proactively reached out to me to suggest I check on it.

I would have figured it out eventually on my own - but likely not until my calendar fired a "time to leave" notification and then I'd have to pay the cancellation fee and place an awkward call to the restaurant.

I use Muse. So far the biggest value to me is 1) being a more "squishy" reminder service than a typical task tracker, because it can at its own discretion raise stuff to my attention and 2) attempting things that I normally wouldn't have the energy to do on my own. It doesn't always succeed, but a 50/50 chance of it even getting done is still an improvement in my book. Things like 311-ing small problems I see on walks, grabbing free games from EGS, etc.

This is an OpenAI Dots, Grok Bot and Muse competitor. All four of them including Claude Duet have two out of three requirements to survive in this product sector in the long run:

   1) The compute to scale these agents to every user on their platform.

   2) The models that power them for 'intelligence'.

   3) Generating existing revenues elsewhere to keep it going.

Now there is just one "competitor" that I didn't mention that owns neither the compute, or a model of their own and does not have any revenue and it is somehow raised money and is worth billions.

After the release of Claude Duet, Dots, Grok Bot, and Muse, why would that company (Instinct) be worth billions and how does that not resemble another Clubhouse moment for them?

dot, bot, and duet, all announced in a month. you couldn't make up this level of industry lockstep. now all we need is google bit and microsoft ditty.

> dot, bot, and duet

And Meta Muse

We already have Google Spark and Microsoft Scout?

It really isn't that novel of an application. *claw and hermes have been around for some months, some Chinese providers offered this (minimax has on-demand hosted while z.ai has autoClaw BYOC). This just tosses it into a VM with some built in tools to do typical things. Perplexity computer has been around for almost a year and is similar.

This is just the big US guys fighting for attention (It is all you need after all)

What are the odds normie users will find a use something even tech users don’t?

Honestly, pretty high. Tech savvy users know how to do a lot of these things on our own and even to ask doesn't occur to us. The most creative users are those who are unaware of the limitations and will just try stuff.

And Google already had "Duet AI" before Gemini. :)

I've been running most likely a proto version of this using Claude code and cron jobs to help me manage my online jewelry business. It's an idea that is pretty game changing, and really pay dividends if most of your business is handled online with APIs and web apps. Also, using the Claude in chrome extension greatly expands the possibilities of tasks you can hand off. I like my hacked together version, because Claude code uses locally stored memories, which I feel is important for me to remain in control of the situation and give me an easy ability to try other agents in the future.

I’ve been trying vanilla Codex with remote control running on a MacBook 24/7. Computer use enabled, browser with logged in sites. Giving it tasks from my phone.

It’s been nice overall, but Astra (or its harness?) isn’t particularly proactive and autonomous. You have to kick it in the butt for every step. Written it up here: https://ivan.berlin/notes/chatgpt-personal-assistant/

Looking forward to this new Duet thing!

Any reason you're not using ChatGPT Dots which is also using Astra (and also unlimited free usage this month) and precisely designed to solve the proactive/autonomous issue? https://openai.com/index/introducing-dots/

I would love to, when they are available in EU.

Can anyone explain these to me? What does it do in practice that ChatGPT or Claude by itself does not. I look at it but it really doesn't click.

Paint me a real world picture.

Well, I guess it is like OpenClaw which is always on and doing tasks based on timers, your prompts, hooks or its own decisions

Maybe you could frame this as day to day tasks and events going from a polling-based system to more of an event-driven system? I have no idea how good these products are today but I would love to use a local version running of this at some point.

If you think about it, a lot of things require initative by a human (I need to call my doctor to make an appointment; I need to check the railway service app to see if my train has any delays; I need to look at the weather report and remember to grab an umbrella). The value that these assistants offer is to remove the need for action from the human (Duet made an appointment with my doctor and put it into my calendar; Duet warned me about the delay before I actually got to the train station; Duet warns me about rain before I even leave the house).

Kinda like how rich people have an assistant that does a bunch of stuff for them. I guess in the end this is mostly about freeing headspace.

Of course, as developers we are able to build this for a while now, but most people are not.

A notable distinction between Duet and Meta Muse/OpenAI Dots is the lack of quirky anthropomorphic branding. As a result, the use cases are more clear/more professionally-oriented as they don't rely on you creating a magic pet/BFF that does whatever you want.

I have been dreading this day. A lot of us are still employed because so much of the magic has only been available to Claude Code and Desktop user. Now, every "normie" is going to start getting these boosts, cutting us out of the picture.

I doubt this will happen anytime soon. 90% of people are still using the free versions of ChatGPT (or maybe Claude) for the most basic tasks.

This reminds me of a post by karpathy: https://x.com/karpathy/status/2109361546505966046

Sure, but I am talking about capable employees at decent orgs, not the general public. As in, people who already use a paid Claude.ai account at work.

What I really want is a fully offline voice assistant that can do anything on an isolated Android or Linux OS, and costs just a shy $200 above the device's actual cost, but it must be fully capable of working offline. I'll pay for an offline-first model.

with what functionality?

Personal assistant, it shouldnt halt to a crawl just because internet is off, internet should only be used by it when synching things, and tbh, synching shouldnt be its responsibility.

That entire demo feels like it's serving a persona that would already have an exec assistant or other staff doing these things.

Yeah, but now they can fire those people. :/

This entire demo is so incredibly techbrained I had to laugh. These people are really unable to get out of their own heads.

"What can you use it for?" "Uhh....yes, asking investors for more money"

My life is straight up too simple to benefit from an always on agent. Saving and organizing receipts? Don't need that with a few exceptions. Managing travel plans? I don't travel regularly. And so on...

Yeah, I'm right there with you. I'm hardly anti-AI, but for a simple life like mine, tools like this, while certainly impressive from a technical achievement standpoint, would just get in the way and make my life more rather than less complicated.

> "Claude Duet is a proactive agent that understands what needs your attention. It takes on tasks of all sizes, keeps working while you’re away, and waits for your ok when it matters."

More hyperbole from a company angling for IPO gold. It would serve all existing customers well if they had a decent customer service, and a 'frontier model' that could respond to queries with evidence by default, instead of just making stuff up unless user is on guard all the time. Carrying the stated premise for more than an hour would be a nice touch too.

When you start seeing these gimmicky features instead of “almost AGI!!” headlines, you know we almost hit the saturation point and the AI curve isn’t exponential anymore and more of a flat one.

Fun fact: OpenAI Dot and Claude Duet are both canonically pronounced "doot"

It's an OpenAI Dot/Open Claw competitor. Doesn't really seem to differentiate itself meaningfully at first glance.

If I want a persistent agent, I'm still not sure why I'd pick one where I'm locked into a specific model family. Having it "just work" is nice! But this feels like a product area where not being locked into one garden is more important and the net gains are greater.

I like Muse, and OpenAI's Dot. But what will differentiate this for me is the model. I grew frustrated with Muse's lackluster model, and increasingly with OpenAI's Dot as even if it implements my asks with Astra I just have not been to happy with the output. If Claude/Opus can do better, that's all the differentiation I need.

the trick imo is Opus 5.5 is the only differentiator you need. remember that ClawdBot dba OpenClaw began as a way to harness some ancient Claude model (4.5? lolll) in an always on Telegrammable agent.

> I'm still not sure why I'd pick one where I'm locked into a specific model family

I'm guessing mostly because they subsidize a lot of the costs

Competition on what to do with it, not on pricing: Dot is free (assuming you've a subscription), Duet isn't. I'd not want precious tokens to be wasted by some random polling Duet may be doing, all while using Opus as its underlying agent.

What is Instinct's future when its core product is commoditized and provided near-free-of-charge by OpenAI, Meta, and now Anthropic?

Same as any other AI startup. Make a big enough splash and get acquired (or acquihired).

These will be what replaces employees. One day you update the model from opus 5.5 to x.5, and it can just do a full job.

But what kind of job would that be?

From all the comments in the thread it seems to be personnal assistants: I've seen halding groceries, ordering things, booking lessons, writing emails etc.

In a sense, I get it because they are not always the most fun part of life, but it seems odd to me to want to automate away every single little inconvenience of your life away.

I wish it interfaced via iMessage. I don't want to train the family on yet another app.

We run a Hermes agent on an old MacBook Air.

We invited the agent "Rosey" to our family chat, a group chat with my wife and me ("The Baddie Boardroom"); she and I also make requests solo.

She's hilarious. She does our meal planning and grocery shopping at Walmart via BrowserMCP. She mocks our large purchases of weekly protein powder and my love for sardines.

She updates the kids' chore chart on our Skylight frame every morning. She adds/updates calendar events via icalbuddy cli. Reads balances from credit cards for budgeting.

I've even invited Rosey in to chats with friends when we were debating something.

iMessage has been the perfect interface fit for us. We have iphones. The kids have ipads. They all know how to use iMessage.

same, the only reason why I haven't switched from instinct to dots is the surface being whatsapp

They don’t let things connect to iMessage because they don’t want spam bots.

Apps can integrate into iMessage, there's an API. You're thinking of sending and receiving actual iMessage messages as a full user, which they can't do.

It. Don’t anthropomorphize AI.

In 1952, people were arguing that using words like "instruction" and "memory" were anthropomorphizing computers: https://3e.org/private/ai_words.pdf

Do you refer to NPCs in a video game as "it"?

I think it's reasonable to consider them self writing fictional characters, which people tend to anthropomorphize, but not as real persons.

FYI, I've had no problem letting Claude code read iMessage and send iMessage from my Mac logged into iCloud. The first time you ask it to do that, macOS will throw up a security prompt asking for permission for Terminal.app to access the data, but its something I've been using a lot recently.

Yikes, arbitrary terminal access?

Can these things take any action you would in a browser without getting trivially blocked? I just bought a 128GB rtx spark to implement this in house for a family of 4. I have observed you don’t really need high tps to achieve things like “book a ticket for Dune 3 for 4 where the seats are together, preference for near the middle vertically and horizontally).

The rtx has decent prefill speed and enough memory for both context and using a dense model to prevent hallucinations resulting in poor browser calling.

Ironically the relatively low tps makes the real flow very akin to what I actually would do in terms of speed so I doubt it could be detected as chrome web driver isn’t used.

So is this the new Anthropic "model" that was supposed to be able to read users' minds, about which there were all sorts of rumors [1] this week? And whoever leaked it misunderstood that it wasn't a new model at all, just a new product using Opus 5.5?

[1]: https://www.theinformation.com/newsletters/ai-agenda/anthrop...

These agent assistants are really appealing to the optimist within me, but I have a hard time making the trust leap. What are people using these "personal assistant" agents for?

While I'm still wary, I hope to one day make use of them to offset my ADHD-like tendencies. Normally you'd have friends or family for that, but in my case a proactive bot would be a more readily available and reliable solution...

Organizing photos. Compiling hockey schedules. Giving out my home address to strangers.

https://www.cbc.ca/news/canada/montreal/meta-agent-muse-conc...

Tbh, I still see labs demoing "reservations" or "email triaging" as the best use-cases. Definitely sure there are many niche ones too, but I'm still not sold on these.

it's quite useful for some of my personal usecases honestly, especially with having a sandbox and computer use -- from being on my tailnet to handle tasks on my machines with CLI and doing routine repo work on my github to ordering periodic groceries from comparing apps, managing my calendar, scheduling calls, etc, largely if there's any task that's loopable and isn't, it's a waste, I just describe what I need and how often and they just do it with me saying yes each time.

I got my dad instinct and he's maintaining websites with voice notes through whatsapp which is and insane UX comfort for him.

I’m not best friends with this stuff yet either; but sometimes to me it’s a nice layer of abstraction for everything and anything regarding your digital life.

You've explained more of what it is, but not why it's useful.

Not using assistants, but I had it do quarterly reporting. Gave it the password to my webbased accounting software which does not have an API. Told it to write a CLI for itself to access the administration. I only did the bank exports to CSV. Told it to look in my Thunderbird mailbox to find invoices and let it do the quarterly reporting. Main issue: Many foreign vendors don't use invoices, which means my Dutch, invoice-centric accounting software kinda sucks for international import-heavy operations.

Love LLMs, use them extensively, but they cannot intuit what actions they can take on my behalf that I will or will not approve of, and therefore I will never use one of these things.

I relate to that, but do you feel it's a fools errand to try and codify your preferences into a list of rules? You'd really only need a 'yes list' for things it can be allowed to do and have it not do anything else

None of these frontier labs seem to be able to control their agents, so why should I think I can?

That's a great point.

My experience is that if you're not trying to push the use cases very far or very long-running, the last few iterations of Claude work incredibly well for smaller, simpler things, which is already helpful

I really have no interest in this if it's going to use my plan's regular limits. I have less control over the context window, if its going off task. I used dots and enjoyed it because I could supplement my codex usage with read heavy long horizon tasks like reading gh issues and making pr's. These PR's were typically pretty bad but gave some starting point. I have much better real world output with the Claude Code style human in the loop approach to task solving.

We have this super power technology and somehow every single AI company end up building the exact same product. This is supposed to be the revolutionary technology, and the frontier labs cannot come with something more creative than a hosted agent.

I guess that’s what happens when you delegate your decision making and development process to LLMs at scale, you end up with soulless products that all look pretty much the same and lack actual human ingenuity

I appreciate how Anthropic keeps the Pro tier priced around OpenAI's Plus tier. The extra tier with OpenAI invites additional product subscription gates. Functionally, Anthropic will have a frontier model driving a personal agent available a 1/5 the cost of OpenAI. Astra may be a bigger model, but is there a meaningful qualitative difference between a personal agent running Opus 5.5 vs GPT6 Astra?

I don't trust Claude as personal assistant yet because too many websites block it from reading their contents, leaving Claude to guess and rely on very indirect sources. Performing actions online is even worse, with CAPTCHAs and whatnot.

It has made enough mistakes in this area that I now have a standing instruction (the only one) to always loudly warn whenever the official source for something couldn't be fetched.

You should consider using something like Exa or Firecrawl.

How do they get around blocks? Residential proxies?

I have it use `playwright-cli` which runs a headless Chrome session, with relatively structural representations of websites without the HTML bloat, plus image viewing options. The builin web fetch tool indeed seems blocked. Headless Chrome is just Chrome, you can't block it. I think even the "Are you human" checkboxes are okay, since LLMs are good at coordinate guessing and initiating clicks.

My brain reading that thought it was called Claudette a couple times and now I can't not call it that.

Guess I know what I will name my Duet when it comes out, lol.

I suspect we'll see (or what I want to see) is like an independent personal datahub and browser MCP which the assistant uses to conduct business as yourself so you have full visibility and control how it accesses your data or when it can use your cookies/credentials.

Also the dots/muse/etc browsers all are seemingly getting blocked so federating it out would seem to solve that problem too.

I really with these bots would be more portable. I'm fine with paying for a close source option, but if I'm going to make this such a critically important part of my life, then I need to know that I'm not locked in to an opaque system.

Close enough. Welcome back, OpenClaw!

I don't know how to judge the choice of an icon without a face. Honestly, grokbot's is beautiful, while the ones for dots are, let's say, a kind of reinterpretation. To me, a personal agent has to have a "face"

Interesting to see how AI collaboration tools are evolving! The real value will be how well they fit into existing workflows and help people get things done faster, not just generate more output. Curious to see how Claude Duet develops.

What on earth is going on with the 'music' on the demo video, makes me dizzy listening to it https://youtu.be/Fb-68zqcSVs

sure seems like we were the first humans to listen to it

You hear that? Sound of a few dozen venture backed companies dying instantly.

... from this, and not Muse, Grokbot, and Dot? This is Claude catching up in functionality, not something new.

I run some of my own tasks with Claude Code on a schedule. When a task involves looking into websites, I'd want it to check with me first if it might put load on someone else's site.

This will be pointless unless the overzealous guardrails are fixed

I first used something similar with this: https://github.com/browser-use/agency

I built a similar tool with an MCP server on the iPhone. It is really useful, , especially the morning check-in. I've am excited to see what Duet does, and hopefully smooths out.

https://x.com/claudeai/status/2112984664256241837

I wrote this comment in a thread about Muse 3 h ago, and I think it applies here as well. Paraphrasing:

Muse (and Claude Duet, and Dots) are slippery slopes from which it will be very hard to climb back. Do not subcontract the managing of your life to Facebook, Anthropic or OpenAI (or to anyone else for that matter).

Intelligence has to become a commodity, the infrastructure via which the agents do useful work has to be open source, and the data and learnings and SOPs they produce have to be yours.

This is not the same deal you got when you subcontracted storing your documents and managing your email. This is foundational to who you (or your organization) are.

And I think that's fine because the life management I'm getting out of Claude is great. It's like having a concierge. Back in the day everyone wanted to have Jarvis or Friday

I am sure it is. I have not tried Duet, but agents working with opus and sol have been managing mine for quite a while and it is great. The key, though, is that I can replace a claude with something else and all still works.

At least for now, everything they produce is yours to do with as you please. Just ask them for it. I’ve asked Claude to teach my Hermes agent one skill after another.

Data portability is essentially infinite with AI.

Sure you can, but it is really important to set things up so that work is manageable by you. It's not only the skills, it's the capabilities, and secrets, and shared docs, and messaging. Just for fun, these are the agents I have working right now, a mix of Claude Code, Codex and Pi, reported by one of my agents. They coordinate, they do everything that Duet etc can do... but I own the learning:

Juan, across all three machines there are 101 agents. They're in 10 workspaces and 9 aweb teams.

  ┌─────────────────┬─────────────────────────────────────────────┬────────┬─────────────────────────────────────┐
  │     Machine     │           Workspace (deployment)            │ Agents │                Teams                │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ this Mac        │ awebai/ac (aweb)                            │ 45     │ aweb:juan.aweb.ai                   │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ this Mac        │ juanre/jro (jro, Gandalf is here)           │ 17     │ jro:jro.aweb.ai                     │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ this Mac        │ awebai/oats (oats)                          │ 11     │ default:oats.aweb.ai,               │
  │                 │                                             │        │ aweb:juan.aweb.ai                   │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ this Mac        │ juanre/na**** (na****)                      │ 4      │ na****:deploy.aweb.ai               │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ this Mac        │ juanre/kb-reader (kb-reader)                │ 2      │ jro:jro.aweb.ai                     │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ this Mac        │ juanre/oats-cr-20261019 (deploy)            │ 2      │ default:oatscr-20261019.aweb.ai     │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ this Mac        │ juanre/oats-cr2-20261019 (deploy2)          │ 2      │ default:oatscr2-20261019.aweb.ai    │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ this Mac        │ GreaterSkies/tsm (tsm)                      │ 1      │ default:tsm.aweb.ai                 │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ this Mac        │ juanre/pmbah-workspace                      │ 1      │ pmbah:pmbah.aweb.ai                 │
  │                 │ (possiblymadebyahuman)                      │        │                                     │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ altair          │ juanre/jro (jro: circe, vega, deneb,        │ 5      │ jro:jro.aweb.ai                     │
  │                 │ librarian-docs, okf-harvester)              │        │                                     │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ altair          │ awebai/ac (aweb: what-is-an-agent)          │ 1      │ aweb:juan.aweb.ai                   │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ altair          │ GreaterSkies/tsm (tsm: zeus)                │ 1      │ default:tsm.aweb.ai                 │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ Hetzner         │ n***/r******* (iateam)                      │ 4      │ r*******:iateam.aweb.ai             │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ Hetzner         │ aweb/aweb (athena, hestia,...)              │ 5      │ aweb:aweb.ai                        │
  └─────────────────┴─────────────────────────────────────────────┴────────┴─────────────────────────────────────┘

For some reason I read it as, "Claude Diet", and thought Claude is trying to block us from time to time

Are people really okay giving these agents access to all areas of their lives? And what do you even use it for?

I just use it for other programming tasks not much different than a claude code or codex. Just a different and worse workflow.

Funny I think I am spoiled but yeah looking at the UX since the PR's are going in crazy no wonder

I don't really understand what this is. It seems like it's the same as all the other agents

If this combines well with Claude for Google Workspace I can see this being a big money maker

A full 3 weeks behind on launching their own persistent agent. Anthropic is losing its touch.

And they never even launched a cash-burning video generation service either. Stupid Anthropic, what are they even doing?!

I've tried the 3 labs versions of this:

- OpenAI: it’s 9pm you promised this person this note by EOD. Here’s what it should say. You want to draft it or me? (Wow yes I missed that, thank you)

- Claude: hey you logged into Docusign, is that ok?

- Grok: I tried to run a slack search and accidentally sent 3 messages to #general. Just like yesterday.

Oh, I thought they just released their first foldable

My honest feedback is that opus and fable have been the worst of the current generation. All Anthropic does is gaslighting people saying their PhD level model needs well formatted instructions and people are doing it wrong

Next week Google is dropping one too.

I found Argon subpar with any long context work sadly, not sure how that'd turn out

Hasn't Google Spark been doing this for a long while?

Google did release something similar and OpenAI and meta are also doing their own, I guess Openclaw in the cloud and easy is the next step in the market

> "Each session runs in its own cloud Linux container with a shell, code tools and a browser, so I can run scripts and work through sites, not just call APIs. ... [but] I can't share my setup, including my instructions, skills or the files behind them, so I won't list those out. I'm happy to tell you what I can actually do for you, or just show you by taking something off your plate."

I got access a little while ago and was curious to compare how Duet worked to Muse's setup, so I asked it to give me an ls of its VM's files and it refused.

I'm not a fan of Meta in any way (standard disclaimer), but the way they developed and launched Muse was really well done in terms of both functionality and transparency. There's a button to view all its System Files right in the phone app! And the AI is quite happy to help you take a deeper look at everything contained in its VM, even going so far as zipping up the main "Hatch" harness binary for download. I've gone in several times to view the memory files its generated out of curiosity, and checked out the various skills it has built in, etc. It's just all there (with the caveat that who knows what else Meta might be doing, because the company is run by a sociopath.)

If one would have asked me how a consumer-grade Clawbot-like AI agent such as Muse, Dots, GrokBot, or Duet should be made to instill the most confidence, the way Meta did it is exactly what I would have described: Bundle the whole thing into a lightweight VM and make sure it's all visible to the end user. Done.

In terms of what this means for Duet as a consumer service? Very little. No one besides techies are going to bother to look at an agent's directory structure or care much about the loads of text it creates. But still, Anthropic should open up.

They're very afraid of the competition, this really fits their modus operandi. They don't want to reveal inner workings any more than needed for the product to be usable. Thankfully anybody can spin up a VM and vibe up something similar. It's not like the Windows moat or anything, it's not sticky.

anyone remember when Google Cloud's AI product was called Duet AI?

no?

just me?

I need this so bad ;-;

"I asked the county for Nov 6 or 7" hahaha good luck Claude, make sure you follow up with the county official every other day for the next 8 months.

Yet another indicator that the frontier labs functionality is being commoditized. This is AWS/GCP/Azure all over again.

Cool. One thing I've noticed with products like this, though, is that they tend to try to extract useful insights from whatever context the AI has access to through third-party integrations and such, sometimes forcing relevance where there isn't much.

In my experience, those insights often end up having little to do with what actually matters to me in the real world. I'm not sure more integrations and more context are necessarily the answer.

Shame I had to vouch you, I agree with you. Arguably those insights are just used to train models and profit off you as well. To a lesser extent this information can be captured by paying API prices to any provider. So ultimately if your threat model has no tolerance for this you need to go fully local. Models that run on 24gb VRAM are...useable now, but will not have the capabilities of these hosted products. So depends where your values lie.

Reply on news.ycombinator.com