Not necessarily, there could be diminishing returns on mere parameters count .

There is nothing to say for example a 1 Quadrillion parameter model will be vastly more intelligent than current SOTA especially since new training data is largely synthetic today

That's precisely what he is saying, there is diminishing returns (or optimization left on the table).

I read it as it is impressive because smaller models 2.5T are squeezing similar returns as 10T models despite being 1/4th size not that there beyond 2T today the number or parameters do not have much meaning

or the latest qwen3.8 27B doing so well at ~1/100 the size of K3

What about general knowledge you can get out of it before hallucinations start?

I do not rely on any LLM of any size for general knowledge baked into the weights, they all hallucinate and that is the wrong way to hold them imo

I think there is some merit in that smaller models cannot memorize so much of the training data, i.e. that they are less likely to do copyright infringement, and by analogy not having memorized SDK / API surfaces that have since changed from the training data

> I do not rely on any LLM of any size for general knowledge baked into the weights

You have to rely on it to a certain level for agentic/coding work, presuming that's the general subject we're talking about here... For instance I recently encountered a project where it would have been a lot worse if the LLM didn't already know "what is" xterm.js and a bunch of its associated npm-related/node related software. If it was still smart but had to google and find results for everything it would have been a lot more time consuming and risked sending it down a wrong path.

for sure, there is a minimum size and knowledge base that is required to be useful

at the same time, search may find newer or better alternatives, and you can always specify specific technologies you want to use, I typically do this when starting a new project

Storing general knowledge in VRAM has always been a dumb idea in the first place.

It did OK on schlongbench v1.0 (test of a specific niche word that doesn't make it into smaller LLMs) but it sure does love to count words

https://pastes.io/r8F1AY8h

Qwen 3.8 27B beats Opus, Fable and GPT 5.6 by a comfortable margin on the AA-Omniscience Hallucination Rate benchmark.

GLM 5.3 is "only" 753B parameters. Much much smaller.

You want to take a look at the "Scaling Laws" paper, so you can extrapolate from these numbers.

This paper, as well as the Chinchilla one, aged like milk though.