It's funny, I found exactly the same thing when I asked about P=NP. The models outright refused to attempt to solve it, claiming it was too hard. I had to really battle to get it to suggest some promising suggestions.
The point in the article that 'It's not about optimisations' really got my attention. I've previously done some work where we wrote an analysis pass under the assumption that it executed last in the transformation pipeline and this was needed for correctness. The assumption was that since no further optimisations happened it was safe. Now I'm not so sure...
I recall seeing some pretty damning reports from a security pentester that was able to escape from a container on Azure and found the management controller for the service was years old with known critical unpatched vulnerabilities. Always been a bit sceptical of them since then
Where did the 4.8% number come from? Is it based on the validator stake? How does that compare to the number required to fork Bitcoin as a function of it's supply?
Where we've had some success is with heterogeneous agents with some cheap quantised/local models performing certain tasks extremely cheaply that are then overseen or managed by a more expensive model.
Struggling to find anything interesting or non-obvious about this article. You give a bunch of LLMs various parallelizable task and some models manage to do it well but others don't. No insights as to why. As someone with a distributed systems background the supposed 'insights' from distributed computing are almost trivial.
Interesting perspective re CXL synchronous API. Wouldn't things like OOO execution and speculation help with that? And anyway the latency is supposed to be comparable to NUMA latency, is that really such a deal breaker?
On a related question, I'm in the market to buy a new laptop for development and want to get something with good support for local models. What is a good recommendation in terms of GPU support etc? I currently have a Dell XPS 13. Should I just get a MacBook? Or are there good non-Mac options?
Sure - I was just referring to the use of a Vagrantfile to configure the VM. The start of the article seemed to be pushing the Dockerfile itself as a big innovation.
You should read them as publicity to convince stupid politicians to continue to fund basic research when they are more inclined to go for tax cuts for billionaires. Annoying, but a necessary evil.