I wish there was a market for super short time GPU cloud for dev/test. If I want to quickly test out a GPU kernel on a B200, I don't want to rent it for an hour. Plus, if I iterate a lot based on the GPU kernel, I need to be able to simply do a batch job like test, but with

An indicator that all the existing constructs of society have done an extremely poor job of assimilating discovered knowledge. Given all the information available in the open, or closed forms in journals/conference in different formats are not digestible. Vetted info needs

I am also curious about how soon they will realize they should opt for the "learn from the env" as experts do have a limit in what they know. Naturally, they end up creating "labs" on multiple areas of science. Essentially, recreating existing institutions which discover

One way to check passion and care is to see how many iterations of something someone is ok with going through to make it better. Are you happy to iterate for the 10th, 20th, or more to make it better is a good indicator.

The novice kept trying to restart the diverging training run. Noam Shazeer said, "you can't fix it without understanding what's wrong", and restarted it. It worked.

This is just the beginning. Distribution will be a solved problem in the future. Marketing will not exist either, except for things that people don't even know can exist. Whatever can be thought of can be found through a simple search in the future.

AshutoshShrivastava Same but different but still the same. x.com/rajveerbach/st…

Is there work about achieving long context, beyond typical model provider context window, through distributed context windows via multiple agents?

Another analogy: LLMs know the source vertex and find only the immediate neighbour vertices. Humans know the source vertex and the destination vertex, and have to discover the edges connecting them.

Software is mostly permissionless. Take an absurd example: if I were in Antarctica, I’d need permission to build a hardware factory. But I could sit, write code, and run it on a server in a data center on another continent. The hardware isn’t managed by me. The data center

When our team tried Sonnet 4.5, we realized it was worth building a whole new version of Devin around it. This model behaves very differently! The team wrote up some of our lessons

Reposting because I wish more companies had such mission statements!

Diego If the data used is sampled from the general public. x.com/rajveerbach/st…

Curiosity is proportional to the amount of stake you unconsciously have in the world.

I need X algo to have the ability to show me the solution to this when someone solves it. The X algo is only as good as it's at distribution of what I'm interested in.

A good search product will turn people into producers instead of consumers. The win for a search product will be determined by this in the long term.

😂

Need a ranking for the best HF managers of this kind in the industry so far. Returns generated by model improvement per GPU hour!