Psyho

2.4K posts

Psyho

@FakePsyho

Humanity's Last Programmer; Game Designer; Problem Solver; past: OpenAI (Dota), Pro Competitive Programmer, Poker

I don't know anymore Katılım Mayıs 2012

396 Takip Edilen29.5K Takipçiler

Sabitlenmiş Tweet

Psyho@FakePsyho·16 Tem

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)

English

553

1.1K

13.3K

2.3M

Psyho@FakePsyho·2h

@Nim_game_btc re: harness, imho it's still required for bigger/harder/more esoteric ones; but yeah observations in those puzzles should be way easier than stuff in AWTF

English

Psyho@FakePsyho·2h

@Nim_game_btc My tests were more qualitative than quantitative, i.e. I was looking at solving paths checking etc. The difference is huge, because basically none of the earlier models were able to give me satisfying results before. I'd need to finish my whole benchmark in order to give stats.

English

247

Psyho@FakePsyho·3h

gpt 5.6 sol pro is the very first model that routinely solves some of the harder logic puzzles and so far looks like it actually uses logic rather than pure guessing so far my experience is that the improvement 5.5->5.6 is bigger than 5.2->5.5

English

414

17.6K

Psyho@FakePsyho·2h

@Meronix15 My diplomatic answer is: I don't think "computer science" people should fear more than any other knowledge workers.

English

251

Meronix@Meronix15·3h

@FakePsyho Do you think in 2-3 years that there will be any jobs left in computer science fields?

English

310

Psyho@FakePsyho·2h

@jurgen_kleis I used webui for this

English

400

Kleis Jürgen@jurgen_kleis·2h

@FakePsyho what is pro? you mean ultra or xhigh?

English

426

Psyho@FakePsyho·2h

@gurselcakar Most of the puzzles that I use for testing are not freely available online. You can find similar puzzles at gp.worldpuzzle.org/content/puzzle… (requires account)

English

116

Gürsel Çakar@gurselcakar·3h

@FakePsyho is there a public repo or website where i can find some of these puzzle sets out of curiosity?

English

121

Psyho@FakePsyho·3h

classic sudoku is an easy case, because even the hardest ones are generally easily solvable via backtracking (or more precisely by applying basic patterns + backtracking) I'm mainly testing on puzzles that require advanced break-ins and are hard to "brute force". I'd say we're not there yet, but we got very close in the past few months.

English

520

Gürsel Çakar@gurselcakar·3h

@FakePsyho what is your knowledge with respect to llms ability to solve for instance logic based puzzles like sudoku? e.g. can it now solves those extremely hard sudoku (or similar) puzzles?

English

575

Psyho@FakePsyho·3h

I can no longer claim that if I ever finished that benchmark, the solve rate would be essentially 0% maybe this will motivate me to actually finish it, since it's probably few more months until it would get saturated

English

1.8K

Psyho@FakePsyho·3h

for the record, that was the experience with fable 5 on the same set of puzzles x.com/FakePsyho/stat…

Psyho@FakePsyho

I tried testing Fable 5 on a logic puzzle in webUI: > Claude reached its max length for this message. > Claude reached its max length for this message. > Claude reached its max length for this message. > You're out of usage credits > Claude reached its max length for this message. > Claude reached its max length for this message. > Claude reached its max length for this message. > You're out of usage credits > Claude reached its max length for this message. > Claude reached its max length for this message. > Claude reached its max length for this message. > You're out of usage credits > Claude reached its max length for this message. > WebUI froze I don't think we're there yet

English

2.1K

Psyho@FakePsyho·4h

@anpaure x.com/anpaure/status… we've exchanged a lot of comments and like no meaningful information 😅

anpaure@anpaure

@FakePsyho is that true? by hiring employees and directing effort towards anything but building rsi you're making an explicit bet against it unless there's an amdahl law kinda thing, you're compute constrained, and marginal capability researchers' value is 0, you're unserious by definition

English

anpaure@anpaure·4h

@FakePsyho wait which post, like the very first one?

English

Psyho@FakePsyho·5h

I really dislike posts like this Literally all of the labs are trying to build "God Himself". The difference is the marketing and that most of the labs are not good at it.

Air Katakana@airkatakana

i don’t understand how anyone could still be surprised anthropic has made it clear that they are trying to build God Himself. your usage limits are of no concern to them all claude subs are just donations to this end if you want to consume a product you should be on chatgpt

English

109

9.1K

Psyho@FakePsyho·4h

@anpaure sure, but your original post sounded like you have major doubts, so I was hoping for something more tangible

English

anpaure@anpaure·4h

@FakePsyho i'm not saying i don't believe but there's an is-ought gap, they're probably trying, but that doesn't imply they couldn't obviously be much better, and they aren't because [some random internal reason] the basis is talking to friends at labs and reading all the news on twitter

English

Psyho@FakePsyho·4h

@anpaure yes, labs (OpenAI and Anthropic) do things I'd expect them to do given that they're trying to build machine god I still don't get why you believe it's not the case

English

129

anpaure@anpaure·4h

@FakePsyho you've said it, now do you think the labs are doing a good job with these? i feel like they're not even trying properly on "people to like you" and they're struggling despite putting out *some* effort on the gov side

English

122

Psyho@FakePsyho·4h

@anpaure For RSI -> In general; I've switched my line in the middle of the sentence

English

Psyho@FakePsyho·4h

For RSI you need good researchers, compute, data, governments and people to like you, income, etc. If you're in the lead, I'd imagine you want to minimize variance as well. Any specific actions that you think go against that? One of the worst I can think of was sora app, but I don't see anything substantial lately (for OpenAI at least).

English

385

Psyho@FakePsyho·1d

@argus96_adam I hope AI doesn't surpass humans at entertaining people anytime soon

English

Adam Adamowski@argus96_adam·1d

@FakePsyho The interesting question is whether LLMs can create interesting/creative problems for programming competitions. Has anybody tested that?

English

143

Psyho@FakePsyho·1d

Back to rainy Warsaw! There are still a few things I want to write wrt AWTF: future of competitive programming, what does this result mean and maybe some analysis of the heuristic problem I have to trim my backlog though, so I'm probably back to my regular tweeting schedule

English

135

6.4K

Psyho@FakePsyho·1d

> Could you also share your thoughts on what career paths young ppl should consider? I've been asked this question easily 100+ times in the past year and honestly I don't have a satisfying answer here. My philosophy was always to avoid "career" and strive for being generalist: focus on core skills, get better at knowing yourself, find what gives you feeling of accomplishment, find your weaknesses and work on them, work on your soft skills as well. Do a lot of many things. I do feel it's more important than ever, because it's really hard to predict how the world will look in 2-3 years from now. The only thing that's guaranteed is that things will change quickly. If you're young you shouldn't have a 5-year plan that's rigid. > Which areas of math/cs will remain difficult for llm based models over the next 10 years? none

English

113

Rvess@flying_knee3·1d

@FakePsyho Hi Psyho! Could you also share your thoughts on what career paths young ppl should consider? Which areas of math/cs will remain difficult for llm based models over the next 10 years?

English

188

Psyho@FakePsyho·1d

@sheggle_ I haven't see any summaries, but I also haven't looked too much. You'll probably find a lot of scattered around info on Japanese twitter.

English

246

Sheggle@sheggle_·1d

@FakePsyho Any good write-ups that analyze the solutions?

English

262

Psyho@FakePsyho·1d

Forgive me for the selfie. I missed all of my opportunities to take a Warsaw-related photo while coming from the airport and I had to improvise while running some errands one hour later.

English

Keşfet

@Nim_game_btc @Meronix15 @jurgen_kleis @gurselcakar @anpaure @elonmusk @BarackObama @taylorswift13