causalmodels

The article's Karim Khan example pretty deeply undercuts the thesis. Losing access to your bank account is the actual coercive power. Losing a Microsoft email is an inconvenience in comparison.

The real 'military bases' are banks.

Brand name pharmaceuticals are sort of a different thing. Brand names must comply with the naming guidelines of the FDA, European Medicines Agency, and HealthCanda simultaneously. In practice, this makes it tricky to use actual words. So my companies adopt an 'empty vessel' naming approach. The empty vessels are nonsense words that (1) invoke an emotion (wegovy is a good example), (2) can be trademarked, and (3) it can survive brand pressure.

No, it did not make it up. I was curious so I asked it asked it to imitate a posh British accent imitating a South Brooklyn accent while having a head cold and it explained that it didn't have have fine grained control over the audio output because it was using a TTS. I asked it how it knew that and it pointed me towards [1] and highlighted the following.

> As of May 29th, 2025, we have added ElevenLabs, which supports text to speech functionality in Claude for Work mobile apps.

Tracked down the original source [2] and looked for additional updates but couldn't find anything.

[1] https://simonwillison.net/2025/May/31/using-voice-mode-on-cl...

[2] https://trust.anthropic.com/updates

I just asked it and it said that it uses the on device TTS capabilities.

The dashed lines on top of the data points and labels is making me wince

The ads you're going to need to worry about are not going to be shown on webpages.

These numbers aren't that crazy when contextualized with the capex spend. One hundred million is nothing compared to a six hundred billion dollar data center buildout.

Besides, people are actively being trained up. Some labs are just extending offers to people who score very highly on their conscription IQ tests.

Given how things are going, I expect to I'll need to age verify accounts that are almost twenty years.

Why does an ecommerce website need a profit sharing agreement?

Sorry but I'm not going to pay 56 dollars to read a 17 year old paper with 18 citations.

Or like as therapists as we continue to 'do the work' via buying new and increasingly strange forms of self exploration.

But also get real if you don't believe that kill bots won't take all the cannon fodder jobs. Computer programs don't have the nasty habit of disobeying orders or revolt.

Lump of labor fallacy strikes again.

This is really nice. Any chance you have conversation branching on the roadmap?

Personally I think people should generally be polite and respectful towards the models. Not because they have feelings, but because cruelty degrades those who practice it.

It's just a different set of in group // out group signals, not some sort of moral failing. You're well within your rights to not like the signals though.

Who is this Publius guy anyway?

You can absolutely sue Wikipedia [1]

[1] https://en.wikipedia.org/wiki/Asian_News_International_vs._W...

edit: My bad, I get the joke now.

Safety and capabilities research are pretty much two sides of the same coin.

Doesn't the autobahn have unrestricted sections with no speed limit?

A good sales person on a golf course with a VP can do a lot of damage

repo link: https://github.com/XuezheMax/megalodon

Full disclaimer up top: I have been working on agents for about a year now building what would eventually become HDR [1][2].

The first issue is that agents have extremely high failure rates. Agents really don't have the capacity to learn from either success or failure since their internal state is fixed after training. If you ask an agent to repeatedly do some task it has a chance of failing every single time. We have been able to largely mitigate this by modeling agentic software as a state machine. At every step we have the model choose the inputs to the state machine and then we record them. We then 'compile' the resulting state-transition table down into a program that we can executed deterministically. This isn't totally fool proof since the world state can change between program runs, so we have methods that allow the LLM to make slight modifications to the program as needed. The idea here is that agents should never have to solve the same problem twice. The cool thing about this approach is that smarter models make the entire system work better. If you have a particularly complex task, you can call out to gp4-turbo or claude3-opus to map out the correct action sequence and then fall back to less complex models like mistral 7b.

The second issue is that almost all software is designed for people, not LLMs. What is intuitive for human users may not be intuitive for non-human users. We're focused on making agents reliably interact with the internet so I'll use web pages as an example. Web pages contain tons of visually encoded information in things like the layout hierarchy, images, etc. But most LLMs rely on purely text inputs. You can try exposing the underling HTML or the DOM to the model, but this doesn't work so well in practice. We get around this by treating LLMs as if they were visually impaired users. We give them a purely text interface by using ARIA trees. This interface is much more compact than either the DOM or HTML so responses come back faster and cost way less.

The third issue I see with people building agents is they go after the wrong class of problem. I meet a lot of people who want to use agents for big ticket items such as planning an entire trip + doing all the booking. The cost of a trip can run into the thousands of dollars and be a nightmare to undo if something goes wrong. You really don't want to throw agents at this kind of problem, at least not yet, because the downside to failure is so high. Users generally want expensive things to be done well and agents can't do that yet.

However there are a ton of things I would like someone to do for me that would cost less than five dollars of someones time and the stakes for things going wrong are low. My go to example is making reservations. I really don't want to spend the time sorting through the hundreds of nearby restaurants. I just want to give something the general parameters of what I'm looking for and have reservations show up in my inbox. These are the kinds of tasks that agents are going to accelerate.

[1] https://github.com/hdresearch/hdr-browser [2] https://hdr.is

Subscription services like Netflix, not emails.

I’m talking about subscription services a user signed up for at one time

Bots acting on the behalf of users should not be blocked but we have spent several decades treating bots (except for the googlebot) as bad.

Like if I want to programmatically unsubscribe from a subscription, why should I have to do it myself?

An LLM is a Markov chain in the same sense that a cat is a tiger, technically true but it misses the qualia.

> As a researcher, I've always wanted some guy with an econ degree and a year of ML eng to recommend me papers after glancing at them for maybe 30 seconds.

This kind of elitism is baffling given the quality of work from independent researchers recently.

And no, I have no I have no “vested interest” (whatever the hell thats supposed to mean) in someone’s Twitter.

If you know your buckets you can pretty easily get most LLMs to generate structured outputs using something like zod-gpt [1].

[1] https://github.com/dzhng/zod-gpt

This got me curious, so I looked into it. The only post ChatGPT dataset was MOSS-3-SFT, which was used for dialogue fine tuning. MOSS-3 was generated using ChatGPT-3.5-turbo [1]. They explicitly call out their fine tuning data in footnote 13 so I wouldn't say it is faked per say, but it isn't a great look.

[1] https://huggingface.co/fnlp/moss-moon-003-base#data