John Tian

Actually a great question; I was working on my own chess benchmark a couple weeks ago and I measured the number of moves top models could make against Stockfish in a random position that's equal (+-0.5 centipawns) given the legal move options every turn. If I recall correctly, top models like 3.1 Pro were making it 10-20 moves before losing via checkmate. I couldn't continue and publish my results due to it becoming prohibitively expensive, unfortunately.

@johnrtian Are current LLMs good at Chess and Go? I wonder if we benchmark those? Answering my own question: reddit.com/r/ChatGPTPro/c…

Also, is playing games a valid benchmark for LLMs? Seems like a problem for Alpha Go or whatever

@mikeysee If it is AGI, I expect it to be quite good at Chess and Go!

@johnrtian Alright fair enough, I guess to be "AGI" then you need to be good at everything humans do and that includes playing games. We should expect LLMs to be good at Chess and Go too then right?

@max_spero_ When you focus solely on SEO, you don't have much bandwidth left for making the product actually good! If numbers keep going up through SEO, it doesn't really make sense to do anything else. Until Pangram ups its SEO game, these crappy detectors have no incentive to improve.

Please please please PLEASE stop using crappy AI detectors that sell humanizers. Is there some psyop that makes people use the worst possible tools instead of the accurate ones?

@gum1h0x here's my implementation with 1.26 bpb! main difference is i have a decoder to go from latent to bytes, and i patchify before the encoder instead of doing byte-level attention first and then mean-pooling into chunks. github.com/openai/paramet… would love your thoughts :)

implemented a naive baseline at 2.3 bpb which maybe has some small issues. seems quite hard ngl. would be very interested in a solution. i unfortunately don’t have the capacity rn to spend a deserving amount of time on this chall’s version. gl to all who try :)

here's the pr for anyone curious: github.com/openai/paramet…

got a bpb of 1.26 with JEPA! for context: every single entry on the parameter golf leaderboard is a GPT. this is a JEPA encoder-decoder that predicts latent representations, not next tokens. it uses a pure byte-level tokenizer (vocab 260 vs 1024 BPE). there's no tokenizer: the model has to learn everything from raw bytes. even with untuned hyperparams, it's within 0.04 bpb of the GPT baseline. the gap to close is small and i have a ton of ideas about how to beat it! (would love some credits)

@theo @crierlon Talk about it next stream

@crierlon Long story

T3 Code now supports Claude. If you have the Claude Code CLI installed and signed in, you can use it with T3 Code. Hopefully the lawyers won't make us remove this 🙃

@jskoiz @cheatyyyy love*

@jskoiz @cheatyyyy amazing project, look the hover effects!! fyi tho hovering over this icon causes it to re-render, resetting the animation

can someone update hascodexratelimitreset.today

@jonathanzliu @heyruchir ohh... what's ur churn %?

@heyruchir LTV per subscriber is $21 so my CAC needs to be below that

just burned $1.4k on paid ads and only made ~$600 back to organic it is

@heyruchir @jonathanzliu yeppp! a CAC of $46 is not bad at all if the app is sticky

@jonathanzliu What’s your LTV for every subscriber? If it’s high enough than this is worth it.

@jonathanzliu $600 LTV?

@mikeysee @theo @OpenRouter @convex @thsottiaux

So just following up yesterdays discussions with @theo here are the results testing various reasoning efforts through @OpenRouter on the @convex evals. The GPT 5.4 xhigh result was the most surprising to me, so I re-ran it again to check and it got the same result which is inline with what Theo was saying that xhigh is worse than high.

@jackfriks @subtlebytes you mean 5x?

@subtlebytes 10x

claude code VS codex codex is quite good, 100x better than anything i used a year ago. but coding with claude makes everything feel like a video game, and i get things done in seemingly less time while having more fun?

@angehyc the colors are pristine

testing posters

@0xajka @theo @t3dotchat pretty sure @r_marked added it!

@theo “Wait Sonnet 4.6 dropped?” as if you didn’t already have it available on @t3dotchat

Wait Sonnet 4.6 dropped? Is it worth a video?

@theo exa instant for the win!

T3 CHAT LAUNCH WEEK DAY 2 Search is now 10x faster, and models can do multiple searches per request

@theo It's genuinely so evil that I have to believe it's ragebait..

I miss likes being public. Dude literally said god killed my best friend to make a point, and 20 people agreed enough to hit the like button. If you're one of those 20, speak up below. I want to know who you are.

@wilsonhou @lowercaseclub @angehyc perhaps.. do you guys do pseo?

@johnrtian @lowercaseclub @angehyc would love to work together sometime !!

in a month @lowercaseclub will be 6 people, 5 fulltime. incredible, talented people who for some reason or other joined the lil studio @angehyc and i started after quitting our jobs ~2 years ago. i feel like an imposter but also more excited than ever for whats to come life is surreal