Excited for @Caltech @SAIRfoundation’s mathathon participants to benefit from the search, formalization, prover, and informalization tools that we use for our lean library building and applications.
Jake Taylor
Verified account This account is verified. Learn more@quantum_jake
Exploring the science and policy of critical and emerging technologies. Views are my own.
868 Following 4.2K Followers 3.6K Posts
Latest posts 5
Scientists are empowered by AI. But our role is changing.
We are becoming astronomers, but we observe these machine intelligence systems -- and in observing, we discover new scientific results.
To do this reliably, our thesis is simple: you must have the trust infrastructure that ensures you can trust each step of the discovery process.
youtube.com/watch?v=Brbq2h…
We are becoming astronomers, but we observe these machine intelligence systems -- and in observing, we discover new scientific results.
To do this reliably, our thesis is simple: you must have the trust infrastructure that ensures you can trust each step of the discovery process.
youtube.com/watch?v=Brbq2h…
From science.org/content/articl… , @Leonard41111588 on Lean soundness bugs: "This is going to keep happening."
Seems like the semantic problem (mapping between natural language intent and formal language) remains a big problem: reviewers found two places where Lean proved a weaker statement than the English one in a 180-page AI-generated proof, and the reviewer suspects up to 100 more.
We're continuing to iterate on the 'informalization' stage of this, which seems to be presently the weakest link.
Seems like the semantic problem (mapping between natural language intent and formal language) remains a big problem: reviewers found two places where Lean proved a weaker statement than the English one in a 180-page AI-generated proof, and the reviewer suspects up to 100 more.
We're continuing to iterate on the 'informalization' stage of this, which seems to be presently the weakest link.
www.science.org AI is shaking up math. But can its proofs be trusted? As AI firms churn out math breakthroughs, researchers question use of Lean software as proof checker At @GoogleStartups Forum, hearing all sorts of interesting things about, e.g., RSI driving Carbon development. Also local models! Lovely start by @DynamicWebPaige on great progress with Gemini Argon and Carbon.
Not yet seeing frontier edging in on the hardware verificaiton and validation problem. But definitely seeing trust infrastructure is bigger than just physical engineering.
Not yet seeing frontier edging in on the hardware verificaiton and validation problem. But definitely seeing trust infrastructure is bigger than just physical engineering.
