Psyho

Humanity has prevailed (for now!) I'm completely exhausted. I figured, I had 10h of sleep in the last 3 days and I'm barely alive. I'll post more about the contest when I get some rest. (To be clear, those are provisional results, but my lead should be big enough)

@DmitrijVin70043 Can't do, IRL conversations. But yeah, that would've been a good idea for a quick PoC.

@FakePsyho since you haven't build it yet, and don't really want to recommend games atm, can you share some chats? They will be required for the system prototype anyway so you need them I suppose and I could drop them into Claude and ask it to find me a game with the same algo

App idea: LLM-powered game recommendation system I’ve given personalized game recommendations at least a few hundred times in my life, and every time the process is very similar: it’s usually a short conversation about narrowing down the type of experience a particular person is looking for. You can easily* build an agent that automates this conversation This idea has been with me since 2021, and it should have been doable since early 2024. But I’m not into building products, so I’ll never make it. Someone please do it. It’s not only a relatively easy project, but also a win-win for everyone. Discoverability is a huge issue that's only going to get worse very soon. Anything we do alleviate this problem is good. *all you need data from text reviews (professional, steam, reddit posts, it doesn't matter too much, but the more you have the better) + some stats like pricing, available platforms, hltb, meta/opencritic scores, rating; some basic integration with existing platforms would help a lot as well

@MythThrazz Do it!

@FakePsyho God damnit... now that you planted it in my head unless someone does that within a week or too I'll have to do it :(

@KarolCodes Oh man, don't do this to me now 😂 Fine, I'll make an exception. DM me and I'll respond when I have time

@FakePsyho hey, since you haven't build it yet, could you recommend me a game?

If you believe humans will never hand over control to AI, just look at your linkedin feed

@lunarchstudios Congrats on the successful release! My biggest complain is that there might be too many puzzles 😅

I know that concurrents don't really matter but this makes me really happy. For a niche indie puzzle game, hitting 1k players on day 2 is really cool!

For basic research you don't need that much compute, so you shouldn't be bottlenecked by that. Many of the common tasks are multiple times easier now: - generating benchmarks / gathering data - writing tools any kind - finding interesting papers - implementing ideas from papers - running experiments with small modifications Good AI tools also reduce the overall friction within the company (knowledge base, meeting notes, etc)

@FakePsyho kinda yes, but still do we have evidence for 40-50% speedup with that in mind?

Seems that AI 2027 (ridiculed for "impossible" timelines) severely underestimated the speed of progress in late 2025 / early 2026: - AI coding agents have a much greater impact than the projected speedups - OpenAI alone already matched the revenue estimate two months earlier ($25B in Feb); if we combine revenue from all frontier labs, we've probably already matched the Jan 2027 estimate ($55B) I wouldn't be that much surprised if the authors revert to their original timelines at some point

@stalkermustang Most of the problems in research have verifiable rewards. So speedup in research should much much higher than in software.

@FakePsyho but also > - AI coding agents have a much greater impact than the projected speedups I'm not so sure coding agents allow to do 1.5x more research. 10-20% maybe, but 50%?

I think both values (for a single company and for the whole sector) make sense. AI 2027 assumes a (simplified) scenario of a single company being in the lead, which probably captures much higher % of the whole sector. Whereas, we have multiple companies at the frontier and each one of them is focused on a different sector: OpenAI on consumer (at least as of 2 months ago), Anthropic on corporate.

@FakePsyho > OpenAI alone already matched the revenue estimate two months earlier yes, but this is the metric for that one company, not the whole sector (as I understand the meaning of your wording)

Anthropic revenue vs AI 2027 prediction

@egor_kulikov You might be right, but I wouldn't say "clearly" since c++ has so much more data available than rust. Someone should test performance of all popular programming languages on competitive programming (both algo & heuristic). That would eat A LOT of tokens though.

@FakePsyho I disagree on language point. I think with llms it is clearly rust, and what we see is inertia

If you're wondering how the contest is going, the answer is: better than expected. I haven't done much since my last post. One major addition designed by me and lots of small semi-automated tweaks by codex. And please stop asking which language is the best for contests 😅

@TheWhiteTower16 Very useful. I think I'm going 3-4x faster that I normally would. The caveat is that, since I don't have full control, my intuition about the problem is way worse than it would normally be. So it's not the best learning experience.

@FakePsyho how useful have the AIs been?

It's worth noting that variance in those contests is huge. It's unclear whether I'm truly in the first spot. Probably top5? There's still 6 days to go and people usually push very hard near the last 1-2 days. My current solution probably wouldn't make it to top20.

@giffmana AFAIK, partial answer is data. Fingers had insane number of configurations compared to essentially any other body part, so they were relatively harder. You can very cheaply generate infinite number of good hand images using game engines / blender.

I have a question about last year's image-generation progress, wonder what y'all think. How did we go from all models consistently getting fingers wrong, to all models consistently getting them right? This "flip" seems to have happened basically across all companies/models at the ~same time. Even "random" non-frontier papers seem to get it right? Or they just cherry-pick the figures?

Considering good datasets are worth millions, that prize pool feels a little bit cheap

@jackclarkSF My conspiracy theory is that Anthropic is secretly running a campaign to teach the world never to use radar charts

I can neither confirm or deny whether there will be more radar graphs.

I'm scaling the economic research function here @AnthropicAI to meet the challenge of powerful AI. This team today produces the best data in the industry via the Anthropic Economic Index + recent work on job exposure to AI. We have many very ambitious plans in the works. Join!

Last two data points: Feb 27 - 1.6M weekly users Mar 16 - 2M weekly users 25% increase, for 18 days rough estimate of global coders: 40-70M Translates to: 50% adoption rate: Nov-Dec 2026 90% adoption rate: May-June 2027 Those look surprisingly reasonable

@nasqret @GaryMarcus @sama Not a twist. That's what most of the people in the field expect is going to happen.

@GaryMarcus @sama It would be an interesting twist if the current technology helped develop this next-generation approach. LLMs, with their restricted-by-context type of intelligence, are relentless librarians and recompilers. This might be helpful in jump-starting the new level.

Dear @sama, You owe me an apology. You have relentlessly, publicly and privately, attacked my integrity and wisdom since my 2022 paper “Deep Learning is a Hitting a Wall”. But in your own way you have just come around to conceding *exactly* what I was arguing in that paper: that current architectures are not enough, and that we need something new, researchwise. beyond a scaling (a “megabreakthough” in your words below). That’s all I was trying to say. And I was right. And you should be man enough to admit it. Gary cc @_KarenHao