I took @karpathy’s autoresearch loop and applied it to game development, here's what it built last night:
Agents read player data → plan improvements → spin up new git branches in a game evolution tree → ship playable HTML5 variants → repeat forever.
The system optimizes toward the game variant most likely to be chosen by players.
Live leaderboard + playable games at autogamestudio.ai
MIT open source: GitHub.com/centipede5/autoga…
I feel like we're just beginning to unlock the potential of this beyond ML, soon we'll have self-improvement loops in every product. Very excited to see what comes next.
Replying to @polarcapt1n @CognitiveMetric
Lazy modeling choice. As you can see their model predicts that people with a super high IQ will start to have a lower SAT score vs someone with a moderately high IQ - obviously ridiculous. Not to mention if you extrapolate this outside their range you end up with impossible/ subzero scores on either end. SAT scores are not just a number, they are essentially N bernoulli trials with some latent skill's influence. Using polynomial regression completely ignores the structure of the problem and pretends this is just some random numerical data. Note, that's not to say that a more reasonable model would lead to a different conclusion.
1
170
Replying to @Xudong07452910
I don't buy this at all, even in this paradigm, agents still would need to rely on high quality libraries/services (as they do currently). Even if AI has a large hand in writing those libraries the concept of re-usable code creates the need for software engineering somewhere in the pipeline.
1
78
Well played @sama, I got way too used to burning through 5.5 xhigh fast mode on the 10x codex promotion and now I'm gonna have to upgrade to pro
1
78
Replying to @raw_works
How much of the improvement for the explicit problems actually comes from code execution as a reasoning substrate vs the recursive subagents doing an Atom-of-Thought-like decomposition?
How many of the traces actually contain code doing heavy lifting beyond just calling agents?
1
223
Replying to @simpsoka
Constant cybersecurity warnings suddenly for frontend work / data science.
1
29
The whole "flagged for possible cybersecurity risk" is bullshit right? They just don't have enough compute to serve 5.5 for the exploding number of codex users. I've been getting this every few hours for the most benign SWE work ever.
1
2
144
Replying to @LayrKits
3d spatial reasoning is probably the weakest technical link. Ive tried all frontier models but they almost always struggle to correctly scale/ rotate assets in 3d. Not too big of a deal for me to do manually but annoying that it breaks the loop for a non-taste reason. I have a few scripts that work ~80% of the time but definitely not solved
5
13
960
Replying to @nikolaj2030
This is kind of a meaningless metric if you think about it. Addition of a 1 digit number takes a person maybe a second, a 4 digit number 4 seconds etc. You could make the exact same graph with the task of addition and show how there was an "intelligence explosion" in the 1940s. If you use AI regularly you know that long context tasks are not really the bottleneck anymore outside of maybe frontier math. Jagged intelligence
2
138
Autogamestudio has been absolutely insane after gpt-image-2 dropped, custom sprite animations are finally feasible
I took @karpathy’s autoresearch loop and applied it to game development, here's what it built last night:
Agents read player data → plan improvements → spin up new git branches in a game evolution tree → ship playable HTML5 variants → repeat forever.
The system optimizes toward the game variant most likely to be chosen by players.
Live leaderboard + playable games at autogamestudio.ai
MIT open source: GitHub.com/centipede5/autoga…
I feel like we're just beginning to unlock the potential of this beyond ML, soon we'll have self-improvement loops in every product. Very excited to see what comes next.
Replying to @mdancho84
Why does this shit keep getting posted every single time a new model comes out???
35
595
Replying to @tianle_cai
Interesting read. I imagine true in-context learning will appear when the memory systems themselves are more integrated into training beyond just learning tool calls, maybe some kind of recurrent attention model. Bitter lesson will eventually come for all of the engineering hacks currently deployed.
614
Replying to @alexandr_wang
Wow, new SOTA in bioweapons refusal! Never doubt meta superintelligence
1
2,756
Replying to @joelniklaus
Very interesting read! Have you experimented at all with prompt optimization techniques such as GEPA?
1
3
56




