Ari Karchmer

OpenAI is really cooking lately. The difference in writing quality of 5.4 and the competitors is massive

@ylecun congrats

Unveiling our new startup Advanced Machine Intelligence (AMI Labs). We just completed our seed round: $1.03B / 890M€, one the largest seeds ever, probably the largest for a European company. We're hiring! [the background image is the Veil Nebula - a picture I took from my backyard, most appropriate for an unveiling] More details here: techcrunch.com/2026/03/09/yan…

@ziv_ravid @JudahGoldfeder @ylecun Ravid, did human intelligence arise from the constraints of earth or are humans on earth-like planet because we are the ideal intelligence and earth-like planets are the only ones able to support it? Why should we expect machine intelligence to be useful on earth?

New paper out: AI Must Embrace Specialization via Superhuman Adaptable Intelligence With @JudahGoldfeder, Philippe Wyder, and @ylecun . There is quite a lot of buzz on our paper, so here is my take. Everyone's talking about AGI, but nobody agrees on what it means, and that confusion is actively hurting the field. We surveyed the most prominent definitions and mapped them along two axes: the kind of capability they refer to (learning vs. doing) and the scope (anything, anything important, anything humans can do). The result is a landscape of definitions that don't just disagree, but they're often internally inconsistent. Our starting point is simple: human intelligence is not general. We are specialized creatures, shaped by evolution to excel at a narrow set of tasks critical for survival. We feel general because we can't see our own blind spots. Magnus Carlsen is the greatest human chess player ever, but compared to what's computationally achievable, he's not actually good at chess. That's not a knock on Magnus. It's a statement about the limits of human adaptation, and why anchoring AI's North Star to human-level performance is the wrong move. We propose the term Superhuman Adaptable Intelligence (SAI), or, in other words, intelligence that can learn to exceed humans at anything important we can do and can also tackle tasks entirely outside the human domain. The metric isn't a growing checklist of benchmarks. It's adaptation speed: how fast can a system acquire a new skill? This has concrete implications for how we build. SAI points toward self-supervised learning for acquiring generic knowledge from unlabeled data, and world models for planning and zero-shot transfer. It also pushes back against the current monoculture of autoregressive architectures, because specialization demands architectural diversity, not one paradigm to rule them all. Or as we put it: the AI that folds our proteins should not be the AI that folds our laundry. This paper grew out of a conversation with Yann on our The Information Bottleneck podcast, which led to a public exchange with @elonmusk and @demishassabis on X (not every paper can cite a Twitter feud as source material).

Ive realized things about the state of markets and the corresponding human condition that would kill an average man

test time compute indeed

My baby spends much more effort on producing output than he does taking input Lessons in that

ratio of cost of A100/hour to cost of 1 pack of premium baby wipes 📉 it's so over

@deanwball "These guys are just that good"

The U.S. government just essentially announced its intention to impose Iran-level sanctions, or China-level entity listing, on an American company. This is by a profoundly wide margin the most damaging policy move I have ever seen USG try to take (it probably will not succeed).

Has anyone tried asking the babies how they implemented their RL? @ylecun

I'm suspicious of resnets. can't say why

@roydanroy @hayou_soufiane I'm not sure this is necessarily a circular definition, though it may be refined. It reminds me of cryptography style definitions which model "knowledge" inside a "machine" as that which is extractable in polynomial time. See for example "proofs of knowledge"

@hayou_soufiane We should try to sharpen this logic. It's fabulously loose since the support has not changed. What's the right way to think about capabilities that can be extracted without it being a circular definition ("those things that RL can elicit").

It seems that most gains from RL comes from the pretrained model itself. The format reward stuff (GRPO etc) just extracts those capabilities. It helps to have good reward signal, but it's not the main ingredient.

@gavinrbrown1 Hi Gavin! Hope you're doing well. Actually, I just emailed this same man, and he also told me that he now spends more time responding to email from alumni than reviewing dissertations. Message received :)

I apologize to everyone who preordered. To thank you for your patience, you'll all receive co-first authorship on my next paper.

Boston University de facto embargoes all dissertations: a solitary hero reviews everything. He's still working through the Fall 23 grads.

@Turn_Trout This looks really cool. Is there any obvious relation to Component Modeling (arxiv.org/abs/2404.11534)?

9) Gradient Routing: Masking Gradients to Localize Computation in Neural Networks ➡️ Blogpost: turntrout.com/gradient-routi… 📖 Full paper: arxiv.org/pdf/2410.04332

6) Application 3: In a challenging toy model of “scalable oversight”, we use gradient routing with reinforcement learning to obtain a performant, steerable policy. Surprisingly, this works when merely 1% of the data is labeled, while baselines completely fail at this setting.

@liamgallagher @ManCity

@PhilFoden i just finished my phd in computer science at boston university, shoutout to you (and pep) for the inspiration

@Dan_Jeffries1 @HermanNarula Isn't their argument centered around the fact that when it comes to x-risk, evidence is no longer a prerequisite to funding and taking preventative measures? In other words, non-zero chance of extinction has infinitely negative expected payoff, so little evidence is needed.

This is exactly right. Ten years from now they will still be telling us it's all coming to an end when billions are using the tech and we've had the kind of problems we always have with tech, like privacy, overreach, surveillance, centralization, scams and the like, but not super intelligent machines deciding we're bugs. Still they'll be telling us it's surely just around the corner, you just have to keep the faith. It's basically a religion. No proof needed whatsoever. Just faith. The more proof they get that they're wrong the more it strengthens their faith that they were right all along.

Some people’s identity and careers are built on a Sky-net scenario *rooted in speculation on the behaviour of a thing that doesn’t exist* they have lost objectivity.

Want to reach out? Collaborate? My DMs are open.

What other computational separations exist? Can they be more “natural” if the computational advantage is smaller (we study super-polynomial advantages)? How else can we use formal relationships between Crypto and ML to heuristically argue about “real-world” ML phenomena?

Accepted to #ICML2024: “On Stronger Computational Separations Between #Multimodal and Unimodal #MachineLearning” (arxiv.org/abs/2404.02254)