Not to mention that they are imperfect and often have very serious complications which we can’t predict very well (that’s what my research is on) and we still do them
We’ve gone decades doing thousands of cellular therapies/transplants/etc. without knowing the mechanisms which drive how they work. Still do them every day!
This is a slide from Terence Tao's 'Math 2.0' public lecture yesterday evening at Caltech. I am highly confident that, if the cure actually worked, almost any Stage 4 cancer patient would tell you that human comprehension of its mechanism is completely irrelevant to them.
As a former Rocket League tryhard and robotic surgery student, this article inspired me to combine the two.
Playing by the rules of @gen_intuition: trained on hours of labeled robotic surgery inputs and no coded physics.
Created a world where a tissue can be manipulated.
Thank you @mims & @WSJ for the thoughtful writeup about the new frontier we're trying to push in AI.
Acting realtime in space is so simple for humans, yet so hard for machines, and we believe our approach has a chance to bridge that gap.
wsj.com/tech/ai/ai-wor…
@khookhar (Update) Was inspired to continue this project after reaching out and hearting from Amigo's CMO, so I tried combining Amigo's patient-simulation idea with this project's clinical consensus idea.
Found that mixing them brings the most improvement (would love to share more)
Getting physicians to respond to surveys is one of the most annoying parts of clinical research and innovation.
So I built a way to synthetically get evaluations from physicians, and compared it to my real evaluations. open.substack.com/pub/rishpasuma…
Still new and learning more about AI evals, so would appreciate any feedback
(Shoutout to @ayushnoori's research and HealthBench for ideas/graph inspo)
It's a new take on OpenAI's existing HealthBench and MedAgentBench that can genuinely accelerate clinical language models.
There are so many more interesting results, would love to share them!
[A THREAD]
AGI, especially in medicine, is held back by the inability of models to reliably evaluate themselves.
So, drawing from my validated physician-simulation platform, I benchmarked 4 leading models clinical language quality and ability to re-evaluate themselves.
There are genuinely so many scientists who refuse to adopt Claude Science/Codex because the language is not built for them.
After speaking with a few amazing people in the AI/language field, I don't think this problem isn't limited to science professionals either
Open-sourcing has always been the standard of “democracy” in tech.
But could equipping people with the knowledge that one report/mute = worth the support of 100-400x individuals actually hurt dialogue?
Muting if u disagree jk
12K Followers 4K FollowingLife Sciences @OpenAI | prev. ML @Meta, AI for Drug Discovery @Deloitte | Writing at https://t.co/uv52PYIrey | Building https://t.co/qCRvEsjHUb
4K Followers 216 FollowingLead Technical Recruiter for @microsoftAI Team EMEA - building the teams that are building the next phase of humanist superintelligence
9K Followers 2K FollowingHead of Drug Discovery & Development Research @AnthropicAI | previously CTO & Co-Founder @CoefficientBio (acq. by @AnthropicAI) | ex-@Genentech | @MIT @Penn
14K Followers 2K Followinginvesting in the future of biomedicine
partner @BreyerCap + medical geneticist @harvard
editorial @NEJM_AI @decodingbio | bayesian
12K Followers 4K FollowingLife Sciences @OpenAI | prev. ML @Meta, AI for Drug Discovery @Deloitte | Writing at https://t.co/uv52PYIrey | Building https://t.co/qCRvEsjHUb