If every lab is already using AI to improve AI, what makes Discovery Loop different?
The first reaction is fair. OpenAI, Anthropic, Google, xAI, take your pick. Everybody talks about recursive self-improvement. Everybody wants a model that writes a better model. So why would four people leave the richest lab on earth to do the thing the industry already claims it is doing?
Because they are not building the next general assistant. They are building the loop that runs science, and they are starting with a customer of one: their own research stack.
The frontier labs train one giant foundation and hope AGI falls out of it. Chat first, science later, maybe. Discovery Loop flips the order. Automate the experimental cycle itself. Hypothesis, implementation, evaluation, next bet. Thousands of those cycles at once. Use that machinery on machine learning until you get architectures that are not just a fatter transformer. Then carry the same loop into drugs, chips, materials, biology. The model is supposed to be the researcher, not the intern who writes the email after the researcher is done.
Closer to AlphaFold than to ChatGPT
That is the part people keep missing. AlphaFold was not a chatbot that happened to know proteins. It was a system pointed at one ugly, high-value problem until it beat the field. DeepMind later dissolved a lot of that specialised shape and pushed people toward Gemini. The argument inside the industry is now whether the next leap looks like another general model or like a factory of narrow, deep loops.
Vinyals has been blunt about the real bottleneck. Models already code. They already run experiments reasonably well. What they lack is research taste: the instinct for which idea is even worth the compute. Nobody has that solved. So the early Discovery Loop plan is unromantic. Humans and machines invent hypotheses together. Full automation is the later chapter, not the press release.
Wait. Is this not just AutoML with better branding and a $50 billion rumour attached?
AutoML-Zero is literally on Quoc Le's resume, so the family resemblance is real. The difference they are selling is scale plus target. Not a search over network layouts for one benchmark. A closed loop that proposes, implements, scores, and repeats across a whole research agenda, first on ML itself, then on physical problems where an experiment is not just another training run. Chip design. Molecules. Materials. If that sounds like the same sentence every lab now uses in a keynote, look at the org chart. A keynote can say "AI for science" while the actual GPUs stay on the chatbot.
There is also the structure. Public benefit corporation. Dean has said they might make calls that are not in the company's purest financial interest. That is either sincere or very good fundraising copy. Either way, it is a different constraint set from a lab that has to ship Gemini 4, keep Search ahead, and explain capex on an earnings call.
So the difference is not "they discovered self-improvement and nobody else did." The difference is what they refuse to be. Not a general model with a science sidebar. A science machine that treats better AI as the first experiment, not the product.