I think a lot of LLM progress in the last 2 years or so is really from densifying the training data to niches like this and not so much from general problem solving capability. I guess the hope is that the deluge of super specific training data will pay off through better overall competence. cf https://www.dwarkesh.com/p/the-sample-efficiency-black-hole
This is a pretty stunning result. The time series looks really convincing. Is the way the detector itself is trained orthogonal to this or could there be some "leakage" in that the pre-chatgpt text is in the (positive) training data?
For Switzerland, that's only true for the special tourist train panoramic carriages on the route, there are normal trains running on the same routes, too, that don't require reservations.
I think this is a complete misunderstanding of what people mean by taste in software engineering. Taste is more like the System 1 response one builds to code over time, which (ideally) captures the quality of the software beyond surface level, so things like maintainability, composability, readability, likelihood of hidden bugs. This is completely different from the question if the code fulfills the immediate task at hand, but also not the same as pure aesthetics.
Was just going to ask if anyone remembered the name of this one. It was amazing! 15 year old me felt very important/futuristic syncing the news before school in the morning and reading during school break!
Edit: after looking it up with plucker as a starting point I think the one I actually used was AvantGo, same idea though I think.
I ran Claude Code on my ca 2015 ThinkPad which was having wifi issues and asked it to fix them. It diagnosed the problem and applied some obscure kernel flag which fixed the issue.
There are plenty of expenses in this order of magnitude that are not tied to direct increases in productivity. I think it may become a serious hiring impediment for companies to be really skimpy on these budgets for example.
Is it really that different with the current iteration of AI compared to what was possible 10 years ago? There may be some new awareness at the executive level of what is possible, but I feel like a "slacker detector" or whatever would have been possible with xgboost or lstms.
The most amazing thing to me about Cider-V was that Cider (without the V) actually went away after a relatively short amount of time, when virtually every other internal service that is officially EOL-ed lives on essentially forever.
That seems possible for generating completely new proteins.
Do you think it's also the case for lead optimization where you typically have some degree of measurements around your starting point, and you are expecting to stay in that local neighborhood for the generated candidates, too?
Would you be open to sharing a version of your pitch deck? The main question in my mind is what kind of exit the VCs have in mind when they give you this money.
That's definitely true for some of them, but for others it's not so clear, like the Apollo or Manhattan projects? Those of course also have lasting impact but it's more in terms of knowledge, which at least arguably we are also accruing with these data centers.
i intermittently blog at phonons.wordpress.com
Working on making better proteins with machine learning at cradle.bio
[ my public key: https://keybase.io/tmco; my proof: https://keybase.io/tmco/sigs/P9gtvzmXYJ6NlWm18exm-PIA-JC6RWYr0qUDqxgitSs ]