I think his point is hand wavy at best. It presupposes infinite scaling and ignores all the algorithmic efficiency wins that are being discovered. Ironically, many of which are being discovered with autoresearch style workflows, using the very LLMs that his company builds.

The #1 post on HN right now[1] is full of people jubilating about how they can run Qwen 3.8 27B on their > 5 year old GPUs. If that isn't democratization of AI, I don't know what is.

I'm sure he's smart enough to instantaneously realize this too, but as the famous Upton Sinclair quote goes, he won't mention it even if he does.

[1]: https://news.ycombinator.com/item?id=49324985

▲

pu_pe 2 hours ago | parent [-]

Which percentage of people have GPUs capable of running Qwen3.8 27B? I am one of those, and for my job I am still resorting to hyperscalers because tasks are completed faster and more accurately that way. Even if we assume that models will no longer improve and we reach a point where everyone can run Fable in their laptop, surely running 1000x Fable agents would give you an advantage.

I think access to compute will matter just as much, if not more, as access to models.

▲

woadwarrior01 2 hours ago | parent | next [-]

NVIDIA's 3090 was released in September 2020. Apple's M1 was released ~2 months later. Anyone with a 5 year old M1 Mac with 32GB or more RAM can run a 4-bit quantized version of Qwen3.8 27B on their machine. AFAICT, there are ~110m Apple Silicon Macs in the world. I'd wager that at least ~20% of those have enough memory to run this model. And if you account for gamers with NVIDIA and AMD cards, I'd wager that the segment is at least an order of magnitude bigger.

> Even if we assume that models will no longer improve and we reach a point where everyone can run Fable in their laptop, surely running 1000x Fable agents would give you an advantage.

IMO, asymptotic advantages are marginal. At least for coding, we got a glimpse into how much of an advantage it gives (or doesn't) when the Claude Code codebase leaked[1] ~4 months ago. :)

[1]: https://news.ycombinator.com/item?id=47586778

	▲	pixelesque an hour ago \| parent [-]
		As someone typing this on an M1 Mac with 32 GB of RAM who tried using 3.8 27B (Q4_K_M) yesterday in both LM Studio and llama.cpp, I wouldn't call it particularly usable in terms of token speed. (and that was with `--spec-type draft-mtp` for llama.cpp). If you want to leave it running with the fans going crazy for 40 mins or overnight or something, fair enough, but otherwise it doesn't seem worth it to me. It's certainly not "interactive", even taking into account the over-thinking it does by default. The 3.6 (maybe they'll release a 3.8?) MoE model is much more usable (but obviously not as good) on this machine spec.

▲

InsideOutSanta 2 hours ago | parent | prev [-]

That's actually an interesting question, and I don't think we have the data to answer it. But we do have the Steam data, and about 7% of Steam users have a GPU that can run it well at a 4-Bit quant (≥24 GB VRAM). About 30% can run a 3-bit quant - I'd say that's just barely usable (≥16 GB VRAM).

I'm not sure whether that's low or high, or how it compares to a general audience.