> I really hope Framework or someone else changes that in the coming years.
Building a custom CPU like Apple Silicon is plain impossible for a company like Framework. The best they can do is have support for ARM motherboards and sell those with the best CPUs on the open market.
Very likely to be an actual reflection. That's probably their real achievement here and the key reason why they are actually publishing it as GPT-5. More or less the best or near to it on everything while being one model, substantially cheaper than the competition.
Don't count your chickens before they hatch. I believe that the odds of an architecture substantially better than autoregressive causal GPTs coming out of the woodwork within the next year is quite high.
Because of the niceties of Rust, combined with the widespread compatibility and architecture support of gcc / C compilers in general?
Rust is a modern language, with package management, streamlined integrated build/testing tools, much less cruft, and lots of high-level features and syntax that people actually like. C is neat but complex codebases benefit from modern languages that help in building robust abstractions while still maintaining the speed of C. Not to mention, of course, the borrow checker and memory safety.
I genuinely don't understand why some people are so critical of LLMs. This is new tech, we don't really understand the emergent effects of attention and transformers within these LLMs at all. It is very possible that, with some further theoretical development, LLMs which are currently just 'regurgitating and hallucinating' can be made to be significantly more performant indeed. In fact, reasoning models - when combined with whatever Google is doing with the 1M+ ctxt windows - are much closer to that than people who were using LLMs expected.
The tech isn't there yet, clearly. And stock valuations are over the board way too much. But, LLMs as a tech != the stock valuations of the companies. And, LLMs as a tech are here to stay and improve and integrate into everyday life more and more - with massive impacts on education (particularly K-12) as models get better at thinking and explaining concepts for example.
Unsloth also works very diligently to find and fix tokenizer issues and many other problems as soon as they can. I have comparatively little trust on ollama following up and updating everything in a timely manner. Last I checked, there is little information on when the GGUFs and etc. on ollama were updated or what llama.cpp version / git commit did they use for it. As such, quality can vary and be significantly lower with the ollama versions for new models I believe.
My friend, what exactly do you think is so energy intensive with LK99 synthesis? I've briefly taken a look and the process proposed is really not that onerous in terms of energy consumed. It is a matter of perfecting the process that is the hurdle, we already spend tons of energy happily in similar industrial processes.
Building a custom CPU like Apple Silicon is plain impossible for a company like Framework. The best they can do is have support for ARM motherboards and sell those with the best CPUs on the open market.