Human wisdom began compounding when we invented language - not because human became smarter. Modern science began compounding when we invented the paper 350 yrs ago! Now as AI scientist becomes the major contributor of research, we need a new research language primitive to compound knowledge more efficiently. Thank you @AnneliesGamble for telling this story so beautifully. As AI makes ideas and experiments abundant, the scarce resources shift. Research taste decides what is worth pursuing. Evaluation tells us whether meaningful progress was made. Rigor determines whether the result deserves to enter the scientific record—and compound. These are the foundations of an AI-native research stack.

It’d be fun to host the world's healthiest hackathon - not 24 hrs no sleep but who can build the most efficient agent loop - get your agent running tasks all night and doing the best work with minimum input tokens and minimal human interaction lol 😆

AI4S research today: self-evolving loops, multi-agent systems, agent skills, science benchmarks etc. --> all to make literature review, hypothesis generation, and autonomous experimentation better. Much of today's AI4S research tries to compensate for weaknesses that frontier models are rapidly eliminating anyway. Now it's become boring... While the first half of the AI4S races on how to build individual, brilliant AI scientists The true speed limit of science isn't the brilliance of the individual scientist — it is the ecosystem they work within: how knowledge is represented, shared, verified, accumulated, and reused across generations of researchers. Making every AI scientist 10x smarter while leaving the ecosystem unchanged is like putting a Formula 1 engine on a dirt road The second half of AI4S must shift from scaling agent intelligence to build an AI-Native scientific ecosystem from first principles. Read more ⬇️