For example you need to use some Mojo Max library to have fast optimized LLM kernels, but at that point you can just stay with Python that has fast optimized LLM kernels, like flashinfer.
8,476 karma · joined September 7, 2018
For example you need to use some Mojo Max library to have fast optimized LLM kernels, but at that point you can just stay with Python that has fast optimized LLM kernels, like flashinfer.
Jensen Huang says that the same thing can happen to NVIDIA. Tinygrad is stopping exporting Blackwell based cards to most countries except the 10 Tier 1 countries, as all others need a few months of paperwork. Huawei GPUs don't have these export restrictions.
It would help a lot actually for protecting people's money instead of driving it offshore.
But it doesn't look like making USA compete in this $15B market is NPR's goal with this article.
I'm not sure I would take this kind of path, I would much more focus on refactoring the project to small and easily translatable components with small boundaries, but it's cheap to try things.
Of course I believe oil lobby doesn't want competition, so it will be a rich guys' fight.
Of course that's how they will make lots of money, not by rides to Mars.
Margins are going up for the 2 frontier model providers like crazy, and I don't expect it to go down more, I think we have seen the cheapest token prices already.
I hope the Cursor guys help them catch up to be closer to frontier models because they badly need help in it.
There was a very interesting point when people who were creating Rust interfaces were asking hard questions about ownerships and lifetimes in driver interfaces from the C linux maintainers and they didn't really care to answer (just wanted to wish Rust away).
Now with AI these questions are getting practical. Fortunately big companies have big stake in keeping linux secure, so I'm not worried about it being addressed at least.
Isn't Linux planning to do the same?
On the other hand all companies know that optimizing their own infrastructure / models is the critical path for ,,winning'' against the competition, so you can bet they are serious about it.
This is interesting, doesn't greedy-only decoding slow down speculative decoding significantly?
In theory the probability of needing resampling (rejection) is (p_real-p_sample)+, which should be much smaller with non-greedy distribution
https://z-lab.ai/projects/dflash/
And DDTree?
The best thing they did was that they bought an amazing domain name.
The syncopath style is clearly categorized as more liberal (do what you feel is good).
I think OpenAI tried to diversify at least the location of the raters somewhat, but it's hard to diversify on every level.
What's much worse is that Tesla is hiding the reasons for the crashes in the crash reports (the only company with FSD).
Some people in Europe were not that happy when Biden told on public television that the Nord Stream pipeline will be blown up somehow, but luckily the media was good in not talking too much about it and later he listened to his own advisors better about how to communicate.