For my Italian followers: would you pick this name for a restaurant? It just opened down the block from my office.
English
Tal Linzen
16.2K posts
Tal Linzen
@tallinzen
Professor @nyuling and @NYUDataScience, research scientist @GoogleAI, inventor of the word "bertology"
NYC Katılım Nisan 2009
937 Takip Edilen19.1K Takipçiler
Every once in a while people will rediscover that no one actually uses the term "middle class" to mean "in the middle of the national income distribution" and then they'll get mad about it.
English
Tal Linzen retweetledi
CDS Assoc. Prof. Tal Linzen (@tallinzen) and Google Research researchers have introduced "Bayesian teaching" to help LLMs reason.
Fine-tuning LLMs to mimic a Bayesian model allows them to adapt to users much more effectively.
In Nature Comms: nature.com/articles/s4146…
1/2
English
NYC slaps when you have a $700 a month rent controlled apartment in the west village
Rhys@rhysrmann
London absolutely slaps in the sun — after almost 30 years of being here, it still doesn’t fail to amaze me
English
incomplete list
himanshu@himanshustwts
There’s been plethora of work going on to understand models through lenses of science. The Neuroscience of Transformers: arxiv.org/pdf/2603.15339… Biology of LLMs: transformer-circuits.pub/2025/attributi… Physics of LLMs: physics.allen-zhu.com Neuroscientific approach to interpretability of LLMs: arxiv.org/pdf/2502.12131 Entropy, Thermodynamics and the Geometrization of LLMs: arxiv.org/pdf/2407.21092 LLMs and Cognitive Science: arxiv.org/pdf/2409.02387 A Statistical Physics of LLM Reasoning :arxiv.org/pdf/2506.04374
English
TL;DR: meta learning
(bilevel optimization: an outer loop trains a controller that decides how the system learns from observations & actions, simulated in the inner loop)
Josh Wolfe@wolfejosh
1/ New paper from @ylecun et al on alternative approach for AI to learn more biologically... paper basically says AI is super smart but still can't learn like a toddler can... the main critique
English
amazing what proportion of my "for you" feed is now people dunking on dumb engagement bait tweets I never would have come across (or wanted to). how do I mute this?
English
@Jonathan_K_Cook @jeffrey_bowers This statistic is only correct if you go fairly far south of the border. The border towns (which the story is very clearly about) are primarily Jewish.
English
More than half the population in northern Israel belong to Israel's much-discriminated Palestinian minority. It doesn't occur to the racist Guardian to ask them what they think of Israel's latest illegal war.
English
got it! I haven't read all of the many references from Owain's group in that website, certainly sounds like there are cool empirical findings in this line of work! but this is still framed in an odd way ("I have discovered that LLMs can correctly answer multi-hop questions even when the hops are not provided in context")
English
@tallinzen @glnmario Even if one is a kind of the other (that is multi-hop of OOCR), oocr lit showed its emergence in LLMs in examples beyond a -> b and b -> c therefore a -> c, like when the b’s are only implicitly related, or when the reasoning pattern is described instead of demonstrated
English
@LauraRuis @glnmario interesting, though I'm not sure I follow, what's the difference between OOCR and multi-hop reasoning done in the forward pass?
English
@tallinzen @glnmario OOCR is the kind of thing that seems obvious because it’s so natural but the extent to which owains definition of it has demonstrated surprising generalizations (as well as its limitations wrt in context reasoning) far beyond 2-hop reasoning shows its usefulness
English
@ChenhaoTan absolutely. we need a lot more desk rejections. though this may drive out good ACs...
English
Low-quality submissions will also drive out good reviewers, another mechanism contributing to the death spiral x.com/chenhaotan/sta…
Adrià Garriga-Alonso@AdriGarriga
I never want to review an ML paper ever again. Most of the good ML researchers go work in industry instead of submitting public papers, so ML conference papers are adversely selected and are on average terrible. This mood brought to you by: having to review ICML.
English
@glnmario to be clear it's important to point out the limitations of chain-of-thought monitoring but it's weird to frame this as a "discovery"
English
@glnmario I think in this community there's a lot of alpha for coming up with a new term for an obvious thing, or an unusually scary term for a not-actually-scary thing
English
Tal Linzen retweetledi
I'm a prof but not in a research dept . We mostly teach undergrads and don't offer a PhD. AI agents are letting me tackle projects I've had in mind for years but haven't had the resources to put any sustained effort into. Transformational (maybe, we'll see how it goes)
Tal Linzen@tallinzen
I think we're going to need CS PhD students to do far more than provide accountability, by which I think Sayash means do code review for AI agents and make sure the agent isn't making silly mistakes. The main value of a strong PhD student for a PI is that they're immersed in a problem, a method, an application, a collaboration with another field; they are obsessed with finding the next question to ask, not just executing the experiments their advisor asks them to do. I simply wouldn't be able to work on the range of things I'm able to work on if I were going it on my own, even if all of my code was generated instantaneously by an agent.
English
This sounds great of course but in practice in engineering and many sciences the point of a PhD has for a long time been mixed: you get training in research on the one hand, sure, but you also help your group build systems, datasets and other "deliverables" for sponsors (companies, government agencies). Many applied labs in CS operate in this way and I think that's what the original post was referring to. It's not a great model and maybe we shouldn't be too sad if it's disrupted.
Caglar Gulcehre@caglarml
Coding is only a small fraction of what a CS PhD student actually does; perhaps 10–30% of their time. The real goal of a PhD, and of being a professor, is not to outsource research work but to educate and train the next generation of scientists: people who deeply understand their field, can think critically about it, and ultimately become experts capable of pushing the frontier of knowledge forward. Coding is probably one of the least interesting part of being a CS PhD. We need human experts even more than before at the age of AI.
English
@tallinzen Sayash has not been closer than 500m to the significant other of a CS PhD student in his life
English
I think we're going to need CS PhD students to do far more than provide accountability, by which I think Sayash means do code review for AI agents and make sure the agent isn't making silly mistakes.
The main value of a strong PhD student for a PI is that they're immersed in a problem, a method, an application, a collaboration with another field; they are obsessed with finding the next question to ask, not just executing the experiments their advisor asks them to do. I simply wouldn't be able to work on the range of things I'm able to work on if I were going it on my own, even if all of my code was generated instantaneously by an agent.
Sayash Kapoor@sayashk
In the last few months, I've spoken to many CS professors who asked me if we even need CS PhD students anymore. Now that we have coding agents, can't professors work directly with agents? My view is that equipping PhD students with coding agents will allow them to do work that is orders of magnitude more impressive than they otherwise could. And they can be *accountable* for their outcomes in a way agents can't (yet). For example, who checks the agent's outputs are correct? Who is responsible for mistakes or errors?
English
@yoavgo I don't know but I bet if you search for "Bayesian model of confirmation bias" you'll find ten different papers from Tom Griffiths explaining why confirmation bias is rational under certain assumptions 🙃
English
is there a variant of bayesian inference where evidence that is aligned with the prior is weighted non linearly more than evidence that contradicts it?
English
I'm not sure why these CS profs are even at universities? Why not just join the frontier labs where you can just do your research yourself with unlimited compute and as many agents as you like? Training is a fundamental responsibility of universities. Or am I just stupid & naive?
Sayash Kapoor@sayashk
In the last few months, I've spoken to many CS professors who asked me if we even need CS PhD students anymore. Now that we have coding agents, can't professors work directly with agents? My view is that equipping PhD students with coding agents will allow them to do work that is orders of magnitude more impressive than they otherwise could. And they can be *accountable* for their outcomes in a way agents can't (yet). For example, who checks the agent's outputs are correct? Who is responsible for mistakes or errors?
English
While this is true but I'm not completely convinced CS will stay a relevant category (or much tinier, almost back to pre-1990s). With effective agents, computation is just too ubiquituous.
Sayash Kapoor@sayashk
In the last few months, I've spoken to many CS professors who asked me if we even need CS PhD students anymore. Now that we have coding agents, can't professors work directly with agents? My view is that equipping PhD students with coding agents will allow them to do work that is orders of magnitude more impressive than they otherwise could. And they can be *accountable* for their outcomes in a way agents can't (yet). For example, who checks the agent's outputs are correct? Who is responsible for mistakes or errors?
English