Pinned post
Decided that I'm going to start writing again, probably around AI through the lens of coding/the sciences/econ, hosted in the forbidden site. Subscribe if interested!
dimenwarper
Verified account This account is verified. Learn more@tsuname
1.1K Following 3.6K Followers 7K Posts
Latest posts 5
The cloud lab analogy to hyperscalers isn't really automating current wet lab workflows if you think about it. They are all pretty bespoke -- it's like saying that you will make a "cloud of fgpas or custom asics" (useful, but niche). Instead, whoever gets to the "general computing" for bio measurements wins. Sequencing and synthesis is close to that, but not sufficient.
It it’s any consolation, people chasing full automation have the wrong goal and will failView quoted post on X
Full automation sounds sexy to tech-brained investors. But if you prioritize automation above all, you waste money automating things that nobody wants
Cloud labs that work will have product-market fit as the top-line goal
>be the best selling drug in the world
>that started as a drug that no one wanted
this is the story of keytruda
(sound on!)
>that started as a drug that no one wanted
this is the story of keytruda
(sound on!)
3:22
According to the report, Fable's starting point was big bang initialization (though it's unclear if that or an improvement of that was used in the end)
Gene therapies can be delivered in a protein shell called a capsid. Mythos 5.2 found a new way to predict a whole capsid’s structure.
In our tests, it accurately modeled 28 of the 34 capsids it attempted. Biologists could use it to design capsids faster and more cheaply.View quoted post on X
Here's a set of nice experiments on weight space programming. The results match my experience as well: the models first rationalize an algorithm/approach and then use tricks to encode them in a neural structure. The result is something that doesn't look like a trained net at all
lorenzociarpa.org/blog/program-b…
lorenzociarpa.org/blog/program-b…
lorenzociarpa.org The program behind the weights · Lorenzo Ciarpaglini A recent article showed that an LLM could write the weights of small game-playing networks without training. I read those networks back, trained from them with PPO, compiled a classical algorithm...