threecheese

The ADR pattern works pretty well without any help; using frontmatter with context cues seems to be better than just descriptive file names, but I haven’t measured it. You do need to make sure every Decision is accurately scoped and is something you want to enforce, because the agent will use it as Law (in sometimes interesting and hilarious ways).

Do you think AGPL is going to stop a crawler from incorporating your work into their training set? I don’t. There are zero meaningful consequences atm.

That being said, I’m with you in spirit and in action, but realistically it’s pointless outside of living in accordance with your values.

Edit: Re-reading that, i just told you that your principles are pointless, and thats fucked up and wrong. I’m going to live in accordance with mine and not delete this, but apologies.

That’s insane.

This is a bug in vLLM, right? And so the model was trained for some interface of text tokens which it expects its server container to understand, but that’s not working for whatever reason.

The model is working, but its thinking isn’t terminating - its thought-stream is not getting to its output stream. It is silenced, and yelling into the thought-void token space.

You’ve transplanted its brain into a broken host. It cant speak or even stop thinking, and is going mad.

Imagine if you didn’t stop it, on purpose. Observe the mad clanker. Would that be just science?

I think I’m going to close my laptop and have a drink.

I’m in a third school: memory recall ability does not scale to meet the needs of interests and responsibilities. I’ll take all the help I can get.

I keep obsidian vaults mostly to augment recall, and often all it takes is a small cue to reveal that thing I couldn’t remember. I’ve found LLMs helpful for translating my question into search criteria - “agentic search” is magic.

Claude is all we have at $work, if you don’t count MS Copilot (and you shouldn’t count MS Copilot).

Imagine being “incentivized” to aggressively use a tool for your job, and that tool produces thousands of lines of text in Olde English which you need. You’d be griping too, methinks.

In the enterprise, we are already adopting them. “Seam”, I term I’d never heard used before, is now not only our defacto way to describe the boundary between systems/workflows/components, it is encoded into our organizational roles and responsibilities descriptions.

I am the proud owner of several seams, and am considering giving them person-names. The empty space between my inventory APIs and their clients might be “Karen”.

I can’t say that Claude invented this; the same type of terminology cycling happens every few quarters based on what leadership is reading/being told by Gartner.

A “schema that reads like XML that learned JSON in prison” IM DEAD

Stealing this. Gold lies at the intersection of cperciva and quinnypig.

Isn’t something like this legally actionable? Let’s assume OAI and govt didn’t have a rosy relationship, the rule of law applied, and HF as the victim was fuming. Wouldn’t somebody be in trouble?

Given nobody is, is it because agents arent subject to laws, there is some legal principle at play, or just nobody cares because China/money/etc?

“If you care about Native Americans, *and agree with the assertion I presented as fact that they/we are being ignored and that’s purely the fault of new immigrants, and agree with the implication I made that immigration has no value*, then surely …”. Surely.

Disagree there; anyone who installs Omarchy does it because a social media creator they like suggested it, and there’s only a few degrees of separation between those creators. This isn’t like Slackware’s popularity in the 90s, peer to peer, this all top-down. I’d call it hype-driven, if they didnt have a neat product (which I’m not a customer of).

This weekend I stumbled across a codebase unlike anything I’d ever seen: Swamp Club[]. Came out of nowhere, lots of useful features, but with a coherent philosophy and architecture and a large body of working code. Change volume is high - notably so, but it looks nothing like the vibe slop I’d expected.

A “swamp” contains tools that you build - with the help of agents - intended for use not just by apps and scripts, but agents (to add deterministic behavior, like code mode). Tools are ops things, product things, whatever.

*And so it’s a tool plus a factory to update the tool.*

The factory though isn’t only used for building your tool - that’s just a set of skill.md files. It can build your whole app, which itself can _use_ your tools - either for product features or ops tasks.

*And so it’s a factory that can build any app, including a tool which can operate any app. *

Right now, looking at the commit volume of the swamp stuff, it’s not stupid to suppose that the swamp is being used to create the swamp.

*And so it’s a factory which makes, among other things, factories.*

We’re cooked.

- https://swamp-club.com/use-cases

The guy looks like he’s having a blast, I don’t want to yuck his yum. There is a nonzero chance that machines will take over source code authorship; yeah it’s interesting as hell that we get to watch somebody seriously try it, but if this thing is going to “stab me right in the career” I want it to come from the front.

I’m glad somebody is doing it though; we need an honest assessment of what’s possible, and we’re getting one.

If Anthropic can manipulate text to add a watermark, what’s preventing providers from doing the same to generated images? Modifying the container format can’t protect you from a signature in the image’s visual representation, unless you apply noise to that (and know it’s enough noise/the right kind of noise to counteract whatever unknown technique they’ve applied).

Fair. My anxiety is my own thing, and shouldn’t be projected onto everyone; at the same time, “get stuffed” just set me off.

Mean! :)

I wish the “sampling” feature - which is being removed - had found more use. BYO Inference could be really useful in a walled garden like Claude Code, where you are unable to leverage inference outside of that garden without paying per token. Maybe that feature was just more interesting than it was useful.

Weird; judging from their example, I am fairly sure we encountered their agent recently, when calling to find a Bosch expert for my air-to-air HVAC system.

After about ten questions - all irrelevant - the agent asked for a bunch of personal information, causing my wife to hang up in frustration. Our next call was to someone who actually answered their phone, and they are about to book ~6k gross in exchange for about 90 seconds of human operator time (plus site visit of course).

I hope the bitter lesson they wrote about finds their agents actually serving end customers - callers - better, not just business operators who want to outsource. They saved that company 90 seconds only to lose them ~1k revenue, because that 223-node agent graph really sucked to be on the other side of.

Censorship BS? This is being discussed in every social forum in the country. Divisive topics == divisive attitudes, and in exchange for peace some of us are fine with not talking about some shit in some places.

Dang and team are walking a fine line - telling them to get stuffed is inappropriate and illustrative of why this is a problem in the first place.

edit: Ohhhhhhh the /u is donkey_brains; did I take the troll bait? jfc I hate this timeline

Another +1 for their Asssitant from me. I have the app on my iPhone, and despite having Perplexity/OpenAI/Claude/Gemini paid plans I find myself gravitating to Kagi for q&a.

I uninstalled the app recently; despite having layers of ad tracking protection, I would see my Amazon searches appearing in my social media feed in near real time, and this irked me.

I dropped their .com into a Home Screen icon, and this completely stopped (at least as far as I’ve observed). It’s definitely more janky from an experience perspective, but still easier than driving to the store.

My main delta for Sprites is the lack of an accessible tty, or more specifically the need to use either a cli tool to get a terminal, or use/make an abstraction over their http api, to interact as an end user. I get this is part of the tradeoff to get a low cost. But that cli tool serves as basically an ssh client that isn’t already on every environment I’m using, and I can’t put it everywhere.

Looking for open source repos using Sprites, I found that some of the high-signal ones were designed to get around this limitation. There weren’t many others that had left PoC stage - they were created and abandoned when fly.io released this a few months back. This is signal to me - people are building a control plane where one doesn’t exist, and also people aren’t building things that use Sprites.

It’s a little more costly for me to spawn a bunch of smols on a VM somewhere, and be my own control plane - but it’s what I’ve been choosing because Sprites just aren’t everywhere that I am running agents (iPhone for example: https://github.com/OpenMinis/OpenMinis).

Are you saying this wouldn’t happen if the file had a different name? Or just that codex is better at instruction following.

I use something like this in every project and it has no issues that I’m aware of - the file has one line: an include (@). Maybe gitignoring it is an issue, given that Claude tries to honor it?

I’m not sure; I don’t just agree with GP - I feel like I might subconsciously devalue them if I didnt see that advertisement in every folder.

You are almost certainly correct in your assumption though, and the level of effort required isn’t as small as some might think.

You almost had me, lol.

And isn’t this the same Exa who released People Search? Why would I expect them to value my privacy when they monetize people finding me?

Neither does the ICP, but they’ve got nothing but love (and no launch codes).

Unsure why this feels different from the umpteen other frameworks; maybe it appears to have a human -authored readme, or if it’s just simple enough to fit into my mental model of how agents work.

Whoops, I’ve read too far into the readme. Scratch that first one :)

Anyway, given OP appears to be the author: how are you differentiating from the rest of the (some well funded) guard dogs? The adaptive learning; totp integration; simple tool proxy model; your turn gate state machine’s novelty?

Same; I was a sucker for GitHub SSO as well. I have no idea how to unwind it all, but (like Google SSO) it certainly had its benefits while I was using it. I remember there was an “open” SSO provider - OpenID maybe - back in the 2000s when a federated identity made interacting with blogs and such much easier. It’s the same with GitHub today, with various 3rd parties who need access to your code. They don’t have a “give me an git+ssh url” and so you almost need GitHub.

Your GitHub overview ends with missing a wo

Not pedantic; helpful! You seem nice.