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/862…

all these "ai changes everything" threadboys look like they fuck with airpods on

fragments: the consequences of absolute skepticism, cognitive debt with AI, will AI remove fun of programming, the future of source code, how to survive AI contributions to open source, explaining transformers to 12/62-year-olds martinfowler.com/fragments/2…

Software development quietly flipped from popularity by authority to authority by popularity. Most popular technical opinions now carry the same credibility as influencer-marketed VPNs and fitness supplements.

AI is giving us countless different ways to skin a cat on top of the countless previous ways we could skin a cat. The actual superpower might be to leave the cat alone.

This.

Customer email done right vs done wrong. @laravelphp: clicked, read, felt respected (great update!). @HarvardBiz: signed up, asked for essential emails only, got buried in spam, unsubscribed from the service entirely. Respect people's time.

"Cursor Killer" - this is exactly the mentality that is keeping OpenAI in catch-up mode. Others think in fundamentals, OpenAI is stuck replicating. Another wrong direction we'll all have to dig out of. Get people in a room. Understand users and come up with well-thought-out ideas.

We're definitely not going back, but we're going forward somewhat blindly. That's OK for now. Our tooling, however, is not catching up to heavily automated workflows. A year or two from now we'll be talking about human factors engineering in software development.

I was sceptical about the framework direction a couple of years back but all recent additions, from small things to bigger features, are helping it beat itself to excellence. Seems funding helped get over resource constraints and growing pains. Ripe for AI productivity gains too.

LLMs are exceptional productivity tools, but pathetic thinkers. If you realized that, Codex would actually be more useful. They have the potential to both start and derail projects. If you were more grounded and not peddling to investors 24/7, you would not have been outclassed by Gemini in Google Search so fast.

You can count on all LLM models, no matter how "state of the art" and the "best model yet" they are, to be always completely oblivious of their ignorance, and fill the gaps with nonsense assumptions. Models are functional illiterates. Some are so less than others.

Not a lot of people understand this... but you actually don’t have to have an opinion about everything. You don’t have to decide if something is good or bad. Marcus Aurelius says limiting the amount of opinions we have is one of the most powerful things we can do in life.

Today Thinking Machines Lab is launching our research blog, Connectionism. Our first blog post is “Defeating Nondeterminism in LLM Inference” We believe that science is better when shared. Connectionism will cover topics as varied as our research is: from kernel numerics to prompt engineering. Here we share what we are working on and connect with the research community frequently and openly. The name Connectionism is a throwback to an earlier era of AI; it was the name of the subfield in the 1980s that studied neural networks and their similarity to biological brains. thinkingmachines.ai/blog/def…