AI commentator Teortaxes claims GLM 5.2 narrows the gap between Chinese and Western frontier models to seven months

@teortaxesTex i think its hard to overstate how much the harness / provider makes a difference when evaluating these models. ~7 months seems reasonable tho (i'd say its closer for coding and further away for things like medical / legal / etc)

@_xjdr there is decay the further you move from the main dimension of competition, ofc. Ultimately the big 3 invest a whole lot more into long tail data. Still, GLM kinda holds up even on legal, it's surprising.

@teortaxesTex Are you ready to admit you were wrong about Mythos? You said it was more than an overtrained code monkey, yet it hasn't made a single scientific discovery in any field in over a month.

@teortaxesTex Also a better metric, imo, is the cost-adjusted performance. If Chinese LLMs are delivering 80-90% of US LLM potency at 5-10% of the token cost, then are they really that far behind?

Is'nt it possible that there is some (unvolontary) backchannel from private evals to the big western labs.

While *models* may not be universally superhuman at pattern extraction from low bitrates, deep learning *systems*, if you include the backpropagation phase absolutly are.

Actually excluding all bits of the test set in the training set is more or less impossible without extremely strict date cutoffs, and assuming big western labs have access to private western corporate datasets that aren'nt available to chinese labs, I would assume more bits are leaked (and the opposite being true for chinese specific data of course).

@teortaxesTex 能力分水岭其实比算力更深

@teortaxesTex At least gpt 5.5 solved some Erdos problems

@teortaxesTex @_xjdr I think when, the harness becomes more important for legal and other wordcel tasks because legal questions get much easier once you have the right info in context.

Are you expecting legal knowledge (as a prerequisite of expertise) to be baked in the weights alone?

@teortaxesTex In my personal use, I do codex + DS for equity research, and ds is miles ahead in terms of hard data digging, cleansing and structuring versus GPT 5.5 high simply because the context mgmt is so much better

@teortaxesTex I've always found the 6-7 months gap a bit of apple-to-orange comp. Constrained by computes, Chinese LLMs by default are not as multi-modal as GPT/Opus, thus they tend to max out the highest value-added verticals, where in many cases they are head to head with US ones.

The compute point is the important one here. If the gap closed even on hard private evals, then "it's just compute" was always too clean a story.

Compute is necessary, not sufficient. Method, data curation, and post-training do work that more TPUs alone don't buy — which is exactly why a lab with less raw compute can compress a 7-month gap.

Worth flagging though: "7 months" is a snapshot, not a trajectory. Gaps this elastic can widen again the moment the next frontier release lands. The closing is real — just not necessarily monotonic.