Ryan Greene

ex @OpenAI

[Ryan Greene @ryangr]

@uncle_deluge paying a premium for aura

[Ryan Greene @ryangr]

I am hearing disturbing rumors that Ilya Sutskever has made the Boltzmann Machine computationally tractable.

[Ryan Greene @ryangr]

@mnvrth all the credit belongs to Codex and https://t.co/FYxleqEuPE

[Ryan Greene @ryangr]

sharing an essay on the relation between compression and intelligence, in commemoration of Ray Solomonoff's 100th birthday https://t.co/iajYqMkWDb

[Ryan Greene @ryangr]

it is a good question, one i would've better addressed if Solomonoff was born a week later

way out in theory-land, the agentic counterpart of Solomonoff's optimal predictor is Hutter's AIXI (Hutter, 2005). the AIXI agent chooses the actions that lead to the highest cumulative future reward, *as predicted by Solomonoff induction on the observation / reward stream*: the ideal agent is an expectimax-wrapper of the ideal predictor, which is, of course, the ideal compressor

coming back down to reality, one of the main theories regarding LLMs is that RL "just" improves the model's pass@k for problems whose solutions already existed latently within the pretrained weights (Yue et al., 2025): it is bounded by what the model has already compressed

[Ryan Greene @ryangr]

i do not mean to cast any shade towards Anthropic (all of the labs are going to be doing this), just highlighting a flaw that has been prevalent throughout the ARC-AGI variations

[Ryan Greene @ryangr]

an impressive jump that i have to assume is the result of being the first frontier model to have RL'd the public demo environments, which is a rather large confounder on what Chollet is trying to get these benchmarks to measure (out-of-distribution generalization)

[Claude @claudeai]

On ARC-AGI-3, an evaluation where AI models must solve novel problems, Opus 5’s score is three times as high as the next best model.

[Ryan Greene @ryangr]

@Awesome_O_AI a tragic case of logarithmic scaling

[Ryan Greene @ryangr]

@MarkovMagnifico behold https://t.co/XeMZjd3XcQ

[max @maxbittker]

RuneBench is out: measuring long horizon goal optimization across 14 AI coding models inside Runescape

[Ryan Greene @ryangr]

@lmcorrigan1

[Ryan Greene @ryangr]

@deepfates it’s so over for me

[Ryan Greene @ryangr]

oh, you’re a capabilities researcher? are you minimizing the description length of the digital universe, or embedding it with a will to power?

[Ryan Greene @ryangr]

Richard Feynman on data efficiency (from "The Relation of Physics to Other Sciences")

[Ryan Greene @ryangr]

@bryancsk of course you did

[Ryan Greene @ryangr]

was not expecting b2b saas on the wikipedia article for "List of works based on dreams"

[Ryan Greene @ryangr]

Schmidhuber's obsession with credit assignment is just another example of his great taste in the most important research directions :)

[Jack Clark @jackclarkSF]

Now of course Schmidhuber is divisive and does spend a lot of time trying to appropriately assign credit in the history of AI and through this has become something of a punchline. But you have to admit that he has demonstrated incredible vision and foresight.

[Ryan Greene @ryangr]

@jackclarkSF i mean, he also wrote the most cited AI paper of the 20th century (Long Short-Term Memory)

[Ryan Greene @ryangr]

@bryancsk The Butcher Shop by Niku Steakhouse