Tokens measure "how much work the model did" in the same way that step counts measure "how far the person went"
GPUs "generate tokens" in the same sense that human feet "generate steps"
You can't compare token counts across different providers to get an absolute measure of "total work done" for the same reason that you can't compare step counts across different people to get an absolute measure of "total distance traveled"
Supposedly it was named "caml" because the author was a smoker who enjoyed camel cigarettes, hence the joke that [later implementations](https://en.wikipedia.org/wiki/Caml#History) were named "Caml Light" and "Caml Special Light"
This is more or less my belief too. But I think there are probably domains/companies where the "promo video" is actually good enough to ship (or at least, where that's the quality that devs were shipping before)
People don't write blog posts about how they wake up at 3AM to assign new tasks to their intern, nor do they build "orchestration frameworks" that involve N layers of interns passing tasks down between eachother
Also the process sounds like a nightmare: "it broke and I asked 4 different LLMs to fix it; my `AGENTS.md` file contained hundreds of special cases; etc." I thought this article was intended to be a horror story, not an advertisement
Money allows for economic specialization, leading to greater productivity. You don't build an economy that produces cell phones when you're stuck bartering goats for wheat.
I thought that Fortran was traditionally faster than C++ for numerical code due to stricter aliasing rules in the language, which I wouldn't expect to carry over to an IR?
Techies (at least at meta) are owners in the sense that a large % of their comp is in the form of equity, so it is in their interest for the company to maximize stock price.
There are overheads in server workloads that scale with the number of machines (network traffic, serializing/deserializing requests). There are also fixed costs per server that don't scale with core count, or at least scale sublinearly (storage, physical data center space, motherboard, ease of maintenance). So running 10 machines with 100 cores can be cheaper and more performant than running 1,000 machines with 1 core even if $/core is higher. And of course individual cores can be beefier: wider SIMD units, application-specific extensions like bfloat support for ML workloads, etc.
Of course Moore's law is slowing down, but cores/$ is an extremely silly metric to use
We're not talking here about being overly critical in a code review, or even criticizing someone who pushed bugs to prod. This is someone who intentionally violated company policy in a way that harms everyone who works there. It is not business as usual. Getting mad about this sort of thing is like getting mad about somebody keying your car or spreading false rumors about you: entirely reasonable and justified and not an interaction where professionalism is expected.
It's unclear to me how making these divisions dramatically increase their infrastructure costs (by either pulling them in-house or paying a premium to the original parent company) will "help them compete"
GPUs "generate tokens" in the same sense that human feet "generate steps"
You can't compare token counts across different providers to get an absolute measure of "total work done" for the same reason that you can't compare step counts across different people to get an absolute measure of "total distance traveled"