Centipede5

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/aut… 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.

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

If you’re an indie game dev or AI vibe coder building games, what’s your biggest bottleneck right now?

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

Half a year ago, METR made an aggressive capability extrapolation that was the 97.5th percentile of an extrapolated distribution. That extrapolation basically came true with Opus 4.6. We called it the worst-case time-horizon, and we are in that world. (although I think this was not the 97.5th subjective percentile for anyone involved, and I think it was close to my 70th percentile at the time but I'm not sure)

Autogamestudio has been absolutely insane after gpt-image-2 dropped, custom sprite animations are finally feasible

@Alibaba_Qwen @arena Embarrassing

Qwen-Image-2.0-Pro is now live 🚀🚀 We’ve pushed image quality, multilingual text rendering, and instruction following to a new level, while making performance much more consistent across styles.🌅🌃 Ranked #9 worldwide for Text-to-Image on @arena 🔗Try it now on ModelScope: modelscope.ai/studios/Qwen/Q… modelscope.cn/studios/Qwen/Q… API：modelstudio.console.alibabacloud.com/ap-southeast-1…

@mdancho84 Why does this shit keep getting posted every single time a new model comes out???

🔥 GPT-6 may not just be smarter. It literally might be alive (in the computational sense). A new research paper, SEAL: Self-Adapting Language Models (arXiv:2506.10943), describes how an AI can continuously learn after deployment, evolving its own internal representations without retraining. Here are the details: 🧵

Surrogate models are going to be the next big thing in LLM harness development

@thomasfbloom Human mathematicians won't scale with Moore's law though

An aspect of using AI to solve maths problems, rarely discussed, is the monetary cost of running these AIs. For example say an Erdős problem is solved by an AI, and the cost of this run is $10,000. 1/

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.

@alexandr_wang Wow, new SOTA in bioweapons refusal! Never doubt meta superintelligence

1/ today we're releasing muse spark, the first model from MSL. nine months ago we rebuilt our ai stack from scratch. new infrastructure, new architecture, new data pipelines. muse spark is the result of that work, and now it powers meta ai. 🧵

@joelniklaus Very interesting read! Have you experimented at all with prompt optimization techniques such as GEPA?

Introducing the Synthetic Data Playbook: We generated over a 1T tokens in 90 experiments with 100k+ GPUh to figure out what makes good synthetic data and how to generate it at scale huggingface.co/spaces/Hugging…

@cloneofsimo Thank you for this, the original 2d "donut" chart is extremely misleading

Gaussians are not empty inside! This is common mis-"common-misconception" When you look at the chi squared distribution it feels like gaussians are essentially shell with radius d, which is the visualization here. But its not the case. The inside has strictly larger density. Its just the nature of high dimensional space (where most of the oranges are at the peel @tszzl) thats pumping more space at the outer crust, not more density! Btw ||x|| ~ d^1/2 ± O(d^1/4) (as chi squared dist has mean d and std sqrt(2d))

Huge productivity hack for vibe-building: Have your agent build a simple streamlit / chartjs admin dashboard for whatever you’re working on. Visual debugging >> logs + manual testing. ~60% of the time it exposes a broken or weird architectural choice from a 5 second scan

Real ones will know we're on the good AI timeline

@mirofish_ai is a stupid project created by people who have already forgotten the bitter lesson of ML. Zero chance it goes anywhere, just a really shitty untrained predictive model. Nonetheless there will be people hyping it for a while because it sounds like sci-fi...

@adxtyahq 90% of links clicked once does NOT mean 90% of traffic going to one-click links. Most likely the majority of their traffic is still going to the top 10% of links clicked more than once, so the cache hit rate will be significantly higher

read this article about a google L7 interview (URL shortener system design) the dev had everything - redis, nosql, load balancer, horizontal scaling L7 kept asking - why not vertical scaling? - why do you even need cache? things that stood out : - horizontal scaling → coordination overhead at scale - low cache hit rate → cache miss + db hit = extra latency made me realize how much of system design is just throwing buzzwords around without actually understanding trade-off's

Most people think "AI research" is some insane galaxy brain activity only the smartest people in the world could handle. If you actually look through papers you'll find it's mostly: - buying more data - paying people to generate more data - building out virtual environments to create more data - stealing more data from your competitors The transformer of today is nearly architecturally identical to gpt-1, almost all advancement comes from bigger models and better data, so it's no surprise that large parts of that are going to be automated

Holy sh*t: The TIMES article about Anthropic contains more serious information between the lines than many realize. Read this article: tl;dr - Model releases are now separated by weeks, not months. Some 70% to 90% of the code used in developing future models is now written by Claude. - Anthropic ended up holding up the release of the new model, known as Claude 3.7 Sonnet, for 10 days until they were certain. - Staff believe the next few years will be a pivotal test, for the company and the world. “We should operate as if 2026 to 2030 is where all the most important things happen—models becoming faster, better, possibly faster than humans can handle them,” says Graham. - Dario Amodei has warned that AI could displace half of entry-level white collar jobs in one to five years, and urged the government and other AI companies to stop “sugar-coating” it. (...) “It is not clear where these people will go or what they will do,” he wrote, “and I am concerned that they could form an unemployed or very-low-wage ‘underclass. - Internally, employees began to question if Anthropic had crept to the cusp of the moment they had anticipated with fear and wonder: the arrival of a process known in AI circles as recursive self-improvement. - Some external experts, believes fully automated AI research could be as little as a year away.

This is basically a repackage (actually a worse version) of GEPA / MIPRO style prompt optimizers, hardly revolutionary. This is a known technique and while it is somewhat effective in some situations (extremely well defined problems with last gen models) I've never seen it materially improve performance outside of that and usually results in a weird bloated prompt overfit to the particularities of whatever you trained it on.

Tencent researchers found a way to get reinforcement learning performance without updating a single parameter it costs $18. the RL methods it outperforms cost $10,000+ the method is called Training-Free GRPO, and the core idea is more interesting than the cost savings

... and now o3 just claimed 2nd overall in poker, beating all the other frontier models and only losing to gpt-5.2

Very impressed with o3's performance at @kaggle's game arena. Despite being almost a YEAR old it still beat all but google's latest Gemini 3 model in both chess games. Not to mention a very strong performance in poker so far. Just goes to show how much of a revolution thinking models where for structured tasks.

@matt_elevenlabs 11B !

We just raised $500M at an 11B valuation 🎉 To celebrate, we’re giving away 1,000 free credits so you can test our platform. For the next 6 hours, comment “11B” below and we’ll DM you the credits (must follow) 👇