James MacGlashan

Florence ICE detainee dead after untreated tooth infection, official says An ICE detainee, who had been at a Florence detention center for four months, died Monday following an untreated tooth infection.

An ICE detainee in Arizona has died of a TOOTH INFECTION after it went untreated for weeks, a local official says. He was a Haitian asylum seeker imprisoned in Florence, Arizona. @emilybregel.bsky.social reports.
tucson.com/news/local/b...

[TW: graphic fracture, sound of breaking bone]

Sen Tim Sheehy (R-Montana) badly breaking the arm of a Marine veteran protesting the war Iran.

Absolutely grotesque.

Jayapal to Noem: "I want to introduce you to just four of the US citizens unlawfully detained by ICE ... they're in this room with us."

Man, "rewrite it in rust" is being made pretty easy with llms.

(Ideally there'd be more significant design changes for a difference in language, but still...)

What a fucking psychopath.

Indeed! I have a preview of it!

TIL Yann and I have similar ideas about how to change publishing.

The additions I'd make are
- Make conferences smaller w/ presentations by invite (similar to @togelius.bsky.social suggestion)
- Build social media tooling for sharing/discussion and paper meta data to support it.

Once again, a long post with strong opinions. It's probably twice as long as it should be, it's also repetitive and written in affect. And you probably disagree with my argument. So maybe you shouldn't read it. On the other hand, most things worth reading are written in affect.

The loss of agent is personally painful. I used to use that term to differentiate the kind of AI research I do from genAI. "I work on intelligent agents."

Now people assume that means "I wire up APIs to LLMs."

A clip from the moment Hillary Clinton found out Lauren Boebert took an unauthorized photo of her deposition: 'I'm done!'

I cannot sufficiently convey how horrifying this graph is.

I'd be happy to have it with you should the opportunity ever arise! :)

I opined because decision making agents is literally my area of expertise. It is of deep importance to me to build machines with human-like cognitive abilities & I am painfully aware of how our current methods are lacking.

I don't need to be a philisopher to have a relevant view on it.

If you really want to insist on embracing vague and useless deifnitions, have fun I guess.

It doesn't change the fact that there are meaningful distinctions you can draw for properties people posses. I guess you'll have to come up with a new name for these important differences.

I'm not concerned with philosophers bc they often insist on vague definitions as a matter of course.

Computer scientists and mathematicians have been able to separate agency/decision making straightforwardly for decades and made progress on them as a result.

The definition is clean, and useful.

A running average is not sufficient for the class of problem I described. It's lacking a goal for the environment and an optimization process for that goal with the model. Similarly to how I pointed out elsewhere that deploying a frozen policy from RL loses the agency. Same with a classifier.

Rule 110 - Wikipedia

Likewise, that the conditions for some form of agency is simple does not mean everything posses agency.

(And if you're not familiar with my analogy to rule 110, see here. It's a super cool result.)
en.wikipedia.org/wiki/Rule_110

Yes, it is easy to define programs with limited agency. RL research started with toy problems that were exactly that! Nevertheless, it remains distinct.

Analogously, rule 110 is crazy simple and Turing Complete. Despite that modest requirement, most CA or otherwise are still not Turing Complete.

I would start w/ the Sutton, Barto RL book. The second edition is free online here:
incompleteideas.net/book/the-boo...

Video interviews with Rich Sutton might be a helpful, gentler, intro too.

This post on my RL website here might also be helpful:
www.decisionsanddragons.com/posts/should...

That kind of system is where people might argue LLMs capture some form of agency implicitly. But it's weak and the lack of clear objective during inference really hinders it and is why its susceptible to project injection attacks.

Human-Timescale Adaptation in an Open-Ended Task Space Foundation models have shown impressive adaptation and scalability in supervised and self-supervised learning problems, but so far these successes have not fully translated to reinforcement learning (...

MetaRL where you deploy a model that has internalized an RL optimization process with environment interaction for a well defined and regularly measured objective can be said to retain agency. For example, see the AdA work (arxiv.org/abs/2301.07608)

Contrast with classic classification and a deployed model which does not possess that. There is no internal model being computed and optimized in the deployed classifier.

Similarly, if you deploy a frozen policy learned from RL, you've lost the agent part. Still useful, but not an agent anymore.

Agency is best described as behavior governed by optimizations of models of an environment in which the system is embedded. This is principally what classic RL focuses on.

People and animals certainly. Probably insects in limited ways. I'd be surprised to find much if any in plants/fungus but the biological world is often more complex than appears and I'm not a biologist so I won't comment on that.

Really? Do you think RNA compute models of their environment and solutions to maximize objectives of that environments that govern their behavior?

So even if we granted some kinds of limited forms of agency to LLMs through in-context processing, its limitations make it easy to contrast as a significant difference relative to human and even animal agency.

The best argument for LLMs having agency is a hope that it emerges as a function of the context (and trained as a process w/ e.g., RL). But

1. engineering by hope is terrible; and
2. we can poke enough holes to show this isn't strong if it exists at all. E.g., consider prompt injection attacks.

From that, we can show that LLMs don't have agency in the way they appear to. You can gesture at RL post training, but people interact with a frozen model, not the RL system.

This is too reductive. Agency isn't hard to define as a mathematical problem statement and its trivial to observe that people and animals have it whereas other systems do not.

One of the early distinguishing factors of RL research vs other AI work was tackling the agency problem.