Claim: Alexi Gladstone Says Energy Based Reasoners Cut Inference Compute by 345 Times
Bye bye looped reasoning… and hello Energy-Based Reasoning⚡! Introducing Energy-Based Reasoners TLDR: - EBRs use 345x less inference compute and 2.6x fewer training FLOPs than the best baselines - EBRs are 7x-43x more sample-efficient than the best looped baselines at inference - EBRs achieve 99.9% on Sudoku-Extreme and near 100% accuracy on multi-solution benchmarks 🧵Thread
Context
A large claimed cut in reasoning compute would matter if it holds, but the figures in this post are the author's own results and are labeled here as a claim.
Alexi Gladstone posted what he called Energy Based Reasoners, and said they use 345 times less inference compute and 2.6 times fewer training FLOPs than the best baselines, are 7 to 43 times more sample efficient than the best looped baselines, and reach 99.9% on Sudoku Extreme plus near 100% accuracy on multi solution benchmarks. The post is a thread starter from the author and not an official lab announcement. Those performance figures are his claims.
Deep dive
Alexi Gladstone introduced what he calls Energy Based Reasoners, presenting them as an alternative to looped reasoning. In the post he claims that these systems use 345 times less inference compute and 2.6 times fewer training FLOPs than the best baselines, that they are 7 times to 43 times more sample efficient than the best looped baselines at inference, and that they reach 99.9% on Sudoku Extreme and near 100% accuracy on multi solution benchmarks. A video is attached, and the post says a thread follows.
All of those multipliers and scores are the author's claims. This file does not have a peer reviewed paper, a named official lab, or an independent reproduction in the lead post. The baselines are not named in the opening text, and 'Sudoku Extreme' and the multi solution benchmarks are not linked there. Until a paper or a third party test is attached, the results should not be treated as an established advance.
The reason the item is in the deck is that a real cut of that size in reasoning compute would change what people can run, and the post cleared the view floor. The reason it is ranked low and labeled as a claim is that a single author thread is not the same thing as a confirmed research result.
Readers should wait for the rest of the thread, a paper, or a code release before using the 345 times figure as a fact.
What to expect next
A paper, code, or a clearer writeup of the baselines would be the next thing to watch. Independent replications on the same Sudoku and multi solution tests would support or shrink the claimed gains. If a known lab later coauthors or disputes the results, that should be added here and the claim label revisited. Until then, the multipliers stay attributed to Gladstone and marked as a claim.
