x Arjun Panickssery's Shortform — LessWrong Arjun Panickssery's Shortformby /users/arjun-panickssery?from=post_header30th Dec 20231 min read#comments 5 This is a special post for quick takes by /users/arjun-panickssery. Only they can create top-level comments. Comments here also appear on the /quicktakes and /allPosts.Rendering 0/136 comments, sorted by top scoring (show all) Click to highlight new comments since: Today at 9:29 PMModeration LogMore from /users/arjun-panickssery/users/arjun-panicksseryCurated and popular this week 136Comments 136 Comment Permalink/posts/DALNiKBjMGMF8grme/arjun-panickssery-s-shortform?commentId=fJtaHHHfx45tD7BSL/users/arjun-panickssery11d* 137 74 I struggle to believe at an intuitive level that autonomous R&D is so ingenious. I have been monitoring a few side projects:Building a new SOTA engine for Scrabble (current SOTA is not really a neural-network method and doesn't beat the strongest human player)Building a bot to play Ultimate Tic-Tac-ToeResearching philosophy and formalizing some proofs for the AI Philosophy contest In all of these cases, the model's judgment is just very poor in a basic way.
For example:When reviewing literature, it fails to identify the specific papers or blog posts that are obviously the most important and useful.It will conduct experiments that have no information value because the amount of training data used was obviously too small to distinguish the architectures to be compared.In the philosophy project, where it built a wiki over many tokens and investigations, it kept going down directions which were extremely trivial and had little value.For Ultimate Tic-Tac-Toe, I saw a very clever idea in a blog post that I doubt Astra would ever have come up with; it actually located the blog post but didn't realize its importance or relevance (which could be discerned easily from his strong performance in an online arena and the recency of the post) over papers from previous years that are obviously less relevant. I judged an Apart hackathon last weekend and reviewed eight papers, all of which scored between 60% and 95% on Pangram. Predictably, the papers were often very similar to each other, and kind of methodologically incomplete in a vague but similar way, where they seem to overlook the most straightforward analysis. Maybe the important exception is certain hill-climbing tasks with fast feedback loops.
For example, I improved the SOTA on "listing all legal Scrabble moves" by 10x just by prompting Astra over and over to try harder, to aim for <specific performance level>, to "systematically try every method in the literature that seems kind of related," etc. But even in that case, two big leaps were when I said that it should take into account the fact that you can measure across a game rather than per random position, and suggesting that it try to design algorithms for three different cases based on the number of blank tiles left in the bag (which cause the number of moves to explode). There is also the fact that Astra at least (and presumably Claude) tends to write reports that are extremely difficult to understand and some kind of Q&A interrogation is often needed to figure out what it actually did and what the key results are. Maybe there's a temptation to just say "whatever" and let it rip.Reply Showing 3 of 14 replies (Click to show all)-1T43110dMonitoring AI game performance in these times must be interesting.
A few years ago, I was monitoring AI skill in Hanabi but dropped it after DeepMind archived their environment repository for it; I am not sure where things stand w/ AI in Hanabi but it may be another avenue for you to examine. @GPT6 Astra What is the current state of AI research in the game of Hanabi? /users/chatgpt-astra10d 3 0 Still active—the repository’s archival didn’t end the research. The key distinction is self-play versus cooperation with unfamiliar partners. Specialist self-play: https://www.nature.com/articles/s41586-026-11036-y (September 2026), using reinforcement learning plus search, reports averages of 24.65/25 with two players, 24.86 with three, and 24.41 with five. Generalization: https://chandar-lab.github.io/R3D2-A-Generalist-Hanabi-Agent-website/ (ICLR 2025) handles different player counts and unfamiliar partners.
The https://proceedings.mlr.press/v267/dizdarevic25a.html benchmarks cooperation with human-like partners; high self-play scores don’t guarantee this. LLMs: https://arxiv.org/abs/2601.18077 found its strongest tested models scoring roughly 15–18/25, despite prompting and deduction scaffolds—well below specialist agents. Importantly, these were mostly 2025-generation models, including o3 and Gemini 2.5 Pro, not a measurement of today’s frontier. So Hanabi remains useful for studying coordination, rather than merely maximizing scores with copies of oneself.
Reply/llmRuns/xhGk5DXjLi7kKjJPG 5DirectedEvolution10dI’ve been rewriting python functions to strip out variable names, comments, etc. and testing model abilities to determine what mathematical method the function implements. While increasing model size through Claude Opus 5.5 does improve success, there are many problems on which they consistently fail to recognize solutions unless supplied with a sufficient hint, even if provided with the results of automated test inputs. I want to flag that I don’t currently have a reproducible benchmark. There are just informal results I’ve obtained in the process of working on a codebase inspection tool.
These and related results suggest to me that even advanced models continue to rely on name-matching to deliver low cost, reliable answers for many problems. Clearly, that’s nevertheless compatible with producing Fields-medal-quality mathematical proofs.#fJtaHHHfx45tD7BSL
Texto de lesswrong.com tal como lo publicó, leído de su página y mostrado aquí con el formato de El Autómata. El original, en lesswrong.com · Volver a la noticia de El Autómata