Andrej Arpas

I packaged up the "autoresearch" project into a new self-contained minimal repo if people would like to play over the weekend. It's basically nanochat LLM training core stripped down to a single-GPU, one file version of ~630 lines of code, then: - the human iterates on the prompt (.md) - the AI agent iterates on the training code (.py) The goal is to engineer your agents to make the fastest research progress indefinitely and without any of your own involvement. In the image, every dot is a complete LLM training run that lasts exactly 5 minutes. The agent works in an autonomous loop on a git feature branch and accumulates git commits to the training script as it finds better settings (of lower validation loss by the end) of the neural network architecture, the optimizer, all the hyperparameters, etc. You can imagine comparing the research progress of different prompts, different agents, etc. github.com/karpathy/autor… Part code, part sci-fi, and a pinch of psychosis :)

Me realising I can run these models locally on my M1 MacBook Air for free

🚨 META’s head of AI safety and alignment gets her emails nuked by OpenClaw​​​​​​​​​​​​​​​​ >be director of AI Safety and Alignment at Meta >install OpenClaw >give it unrestricted access to personal emails >it starts nuking emails >“Do not do that” >*keeps going* >“Stop don’t do anything” >*gets all remaining old stuff and nukes it aswell* >“STOP OPENCLAW” >“I asked you to not do that” >“do you remember that?” >“Yes I remember. And I violated it.” >“You’re right to be upset” LMAOOOOOOOO

Two developers coding in the same branch.

I think it must be a very interesting time to be in programming languages and formal methods because LLMs change the whole constraints landscape of software completely. Hints of this can already be seen, e.g. in the rising momentum behind porting C to Rust or the growing interest in upgrading legacy code bases in COBOL or etc. In particular, LLMs are *especially* good at translation compared to de-novo generation because 1) the original code base acts as a kind of highly detailed prompt, and 2) as a reference to write concrete tests with respect to. That said, even Rust is nowhere near optimal for LLMs as a target language. What kind of language is optimal? What concessions (if any) are still carved out for humans? Incredibly interesting new questions and opportunities. It feels likely that we'll end up re-writing large fractions of all software ever written many times over.

Super cool! It was a bit chatty, but it focused on getting across the main idea, its motivation, and the thinking behind the results, failed attempts, and fixes. I'd love all ML/AI results presented in such an intuitive way.

New art project. Train and inference GPT in 243 lines of pure, dependency-free Python. This is the *full* algorithmic content of what is needed. Everything else is just for efficiency. I cannot simplify this any further. gist.github.com/karpathy/8627f…

From the xAI all-hands (just released publicly, uncut): Guodong Zhang: "I can already feel the AGI, at least for coding." He told kernel and compiler engineers to seriously ask whether it's still worth it, or whether it's time to join the effort and automate yourself. "What a year to be alive." Elon Musk: by end of year, AI won't even write code. It'll generate optimized binaries directly. Bypassing compilers. Bypassing coding entirely. "Just say 'create optimized binary for this particular outcome.' That intermediate step will not be needed." Grok Code expected to be state-of-the-art in 2 to 3 months. This is an internal all-hands they published raw.

If you're a fan of optimization puzzles or coding, be sure to check out the MLSys 2026 Graph Scheduling Competition! 🧩 The challenge: You have a complex AI workload and a tiny high-speed memory scratchpad. Your goal is to design a schedule (just a JSON file) that moves data in and out efficiently to minimize latency. It’s like a high-stakes game of inventory management. Open to students, researchers, hobbyists and puzzle-solvers worldwide. Register now and start coding → goo.gle/4ktgyyb #MLSys2026 #CodingChallenge

The timeframe you seem to refer to is from the distinct, though related, Belgian-French occupation of the Ruhr Valley. The Rhineland occupation did not fully end until 1930. This picture is widely dated to 1929. I stand corrected on the troops’ origin, however; most colonial troops were allegedly withdrawn by 1925, so these should indeed be metropolitan French. (Not that one can tell for sure from the angle.)

@ArpasAndrej @TheNameofWar Tu racontes n’importe quoi? C’est en 1923 pendant l’occupation de la Rhur. Quand on ne sait pas on se tait ?

French soldiers at Koblenz, WW1

Love this. Weighted sum (`Final Score = Σ (weight_i × P(action_i))`) ignores variance, though. Could swap weighted scores for GP over action probs like $u(c) \sim \mathcal{GP}(\mu(\mathbf{p}), k(\mathbf{p}, \mathbf{p}'))$, sample posterior for exploration, using JAX: e.g. gp_scorer(action_probs, weights) → Upper Confidence Bound ranks. Reduces echo chambers. (Where \(\mathbf{p}\) is vector of action probs for candidate \(c\), \(\mu\) linear mean (weighted sum baseline), \(k\) RBF kernel for covariance. Posterior sampling: Draw u \sim \mathcal{N}(\bar{\mu}, \bar{\Sigma}) post-observations, rank by samples (e.g., Thompson) or UCB.)

Checkout the new recommendation algorithm built by @xAI 🚀 No Scala, only Rust and Python 🔥 github.com/xai-org/x-algo…

engineers watching everybody else vibe code apps

@Object_Zero_ This is a bunch of hooey. East Greenland is sparsely populated for a reason. Yes, the waters surrounding it are full of ice floes, nearly impenetrable, dense pack ice, but it also features katabatic winds, way less fauna, highly inhabitable.

Scoresby Sund, Greenland This is the San Francisco Bay of the Arctic Ocean. Scoresby is 500-600 meter water depth, it is East facing and provides near perfect shelter from storms. It is perfect geography for a large naval base from which to control not only the Greenland, Iceland, UK (GIUK) Gap, but also the entire Arctic Ocean and also the Northern Atlantic. The only comparable geography in the world is Stavanger Fjord in Norway. The only limitation for Scoresby and why it remains undeveloped is the ice and the need for icebreakers. But submarines do not care. Subs can transit under the ice and many submarine bases are constructed as artificial sea caves. Scoresby Sund represents an incredible piece of undeveloped infrastructure as a submarine base, that would be a crown jewel in a world where the Northeast Passage and Northwest Passage open up new shipping lanes than dramatically shorten the transport distance between all the major economies. Doing this at the necessary scale would require something like 30-40,000 people, maybe more. If you map out the development is it large and expensive and could dwarf the existing GDP and population of Greenland. If someone committed to build a naval base here, they would need to ensure they could never lose it. It would provide a major security boost for Europe and the Eastern seaboard of US against Russian and Chinese subs, who would have to cross the Indian Ocean and Cape of Good Hope, or the Pacific Ocean and Cape Horn before entering the Atlantic. Scoresby Sund is a game changer for naval dominance in both the North Atlantic and the Arctic.

𝘣𝘢𝘤𝘬 𝘧𝘳𝘰𝘮 𝘵𝘩𝘦 𝘵𝘳𝘪𝘱. #pixelart #ドット絵

DeepSeek was a side project at High-Flyer Quant. Qwen was a side project at Alibaba. Twitter was a side project at Odeo. Mac was a side project at Apple. Meanwhile: Windows Phone was a core project at Microsoft. Metaverse was a core project at Facebook. Google Glass was a core project at Google. Apple Intelligence is a core project at Apple. taste, passion, agency > roadmaps.

Same Odyssey trailer.. but with Historically accurate costumes.. wasn’t so hard, was it? And it’s still badass! Super cool work by DF Fox! Kudos to the dude for the splendid AI adaptation!

One of the most common engineering mistakes is that people often believe picking the easiest path would get to the end goal faster. In reality, you would face the same hard problem at 80% of way there. Worse, you did not spend any valuable time to study. The problem would still be alien to the teams. Instead, it is often better to face the hardest problem head on so that you do not need to see it as a stranger again.

@yunta_tsai Space servers - Landauer’s principle. Lower energy bound of ~3K on how much energy it takes to store a bit. Most optimal.

Solving AGI is solving the compression of photons.