Google presents an AI system to write expert-level scientific software. Using LLMs + tree search, it invented novel methods in bioinformatics, epidemiology, geospatial analysis & more, often surpassing human SOTA. (1/4)
In bioinformatics, it discovered 40 novel methods for single-cell data analysis that outperformed the top human-developed methods on a public leaderboard. In epidemiology, it generated 14 models that achieved SotA for forecasting COVID-19 hospitalizations. (2/4)
Breakthrough plot below. This superhuman performance arises because of the ability to exhaustively and tirelessly carry out solution searches atan unprecedented scale. (3/4)
@arankomatsuzaki Great for science, but I wonder if this will widen the gap between those with access to advanced AI and those without.
@arankomatsuzaki 🚨 Google just unveiled an AI system that can write expert-level scientific software. It doesn’t just assist—it invents new methods in bioinformatics, epidemiology, geospatial analysis & more, sometimes beating the human state-of-the-art. ⚡ Follow @TLDRJake to get smarter faster
@arankomatsuzaki As someone not super technical in AI, if this system can invent 40 novel methods for single-cell data analysis that beat human experts, how might that speed up cures for diseases like cancer? What risks come with over-relying on AI-generated code we don't fully understand?
@arankomatsuzaki AI needs to write 50 thousand LoC, in the original, to be judged by professional programmers with an obsession in music, visual design, hardware design, and to be recognised as if the work is a timeless masterpiece by composer Chopin and other greats in their fields.
@arankomatsuzaki That is impressive. Combining LLMs with structured search opens the door for real breakthroughs. The fact it is already outperforming human benchmarks in complex fields says a lot about where this is heading.
@arankomatsuzaki like that's gonna change the world or something. until it writes my essay for me
@arankomatsuzaki this could accelerate discovery in areas where we desperately need breakthroughs; drug development, climate modeling, materials science. having AI explore solution spaces humans can’t is genuinely valuable but it also raises questions about reproducibility and scientific rigor.
@arankomatsuzaki Do you think scientists will see this as a partner or as competition?
@arankomatsuzaki less debugging, more inventing?
@arankomatsuzaki this is what we like to see AI for good
@arankomatsuzaki AI is now designing the future of science. 🔥
@arankomatsuzaki Cool! More AI for science! I wonder how out of distribution are those discoveries
@arankomatsuzaki AI breaking SOTA.
@arankomatsuzaki Neurointelligence
@arankomatsuzaki AI outcoding scientists.
@arankomatsuzaki They had me until they used the word empirical
@arankomatsuzaki This is huge! How do you think this will change the relationship between scientists and programmers? Will we see more domain experts building their own tools directly?
@arankomatsuzaki Then proceeded to delete the human genome thought “acknowledge to the user my failure” and sent deep regrets message
@arankomatsuzaki LLMs can generate scientific code, but LAMs go further by executing, testing, and refining it in real workflows. With @getactionmodel, LAMs automation are reliable, bridging research ideas to applied solutions faster and with real-world accuracy. Making life easier #LAM #AI
@arankomatsuzaki ill be your cofounder
@arankomatsuzaki 👍👍👍
@arankomatsuzaki Wow👀💫
@arankomatsuzaki Amazing share
@arankomatsuzaki AI writes expert software. Cool! 😎