AI Judge changed title after evaluation, original title: "Pseudonymous developer Teortaxes says Chinese AI models are only seven months behind the US, projecting parity by February 2026"
Elon Musk expects parity to take until early 2027.
I think GLM 5.2 points to a 7 months gap currently It's around Opus 4.7-4.8 level, all told (modulo vision which in Opus's case is garbage anyway). Mythos reached Preview status (≥ Opus 4.8, functionally) by early Feb 2026. This means full PRC Mythos ("Fable") by Nov-Dec'26.
@teortaxesTex What's your current timeline for china to reach Fable class ? GLM-5.2 certainly shorten the gap.
Users praise detailed AI scaling timeline analyses and express optimism about GLM benchmark gains while others insult claims of a sustained US lead over China as nationalistic or unrealistic.
No Digg Deeper questions have been answered for this story yet.
@teortaxesTex Probably Q1
I think GLM 5.2 points to a 7 months gap currently It's around Opus 4.7-4.8 level, all told (modulo vision which in Opus's case is garbage anyway). Mythos reached Preview status (≥ Opus 4.8, functionally) by early Feb 2026. This means full PRC Mythos ("Fable") by Nov-Dec'26.
That was my original estimate before GLM's scores Elon actually might be better calibrated. Extrapolating "7 months" forward ignores scaling to Fable-tier. They start getting 950s in August; by Nov clusters are running. *If* scale is the hard limiting factor… 4+ months. If.
@teortaxesTex Probably Q1
Dario's Good End necessitates the gap of "12-24 months by 2028", that's his pacing challenge. Let's be precise: at least this means that by January 2028, the best Chinese model is AT MOST as good as the best US Jan 2027 model. This requires a really rapid acceleration in 2026.
I think GLM 5.2 points to a 7 months gap currently It's around Opus 4.7-4.8 level, all told (modulo vision which in Opus's case is garbage anyway). Mythos reached Preview status (≥ Opus 4.8, functionally) by early Feb 2026. This means full PRC Mythos ("Fable") by Nov-Dec'26.
the costs of failing this goal are almost too horrifying to contemplate: the status quo, where Chyna exists and is relevant. Baidu exists, too. Dario would rather castrate himself than suffer this indignity. So he'll try to get his 12-24 months lead. (24 seems out of reach now)
Dario's Good End necessitates the gap of "12-24 months by 2028", that's his pacing challenge. Let's be precise: at least this means that by January 2028, the best Chinese model is AT MOST as good as the best US Jan 2027 model. This requires a really rapid acceleration in 2026.
I do wish Dario went into more details of his theory of victory. Muh "set the rules and norms", "gain significant leverage with which to incentivize cooperation", "powerful deterrent". Slippery think tank duckspeak, I hate this register. Can imagine him squirming and grimacing.
@teortaxesTex did this post say why exactly a 12-24 month lead?
why not 84 months or 3 months?
@teortaxesTex I wonder if having access to weights will allow people to do things not possible with opus, making it effectively "dangerous" as claimed by the regulatory corporate strategists
I think GLM 5.2 points to a 7 months gap currently It's around Opus 4.7-4.8 level, all told (modulo vision which in Opus's case is garbage anyway). Mythos reached Preview status (≥ Opus 4.8, functionally) by early Feb 2026. This means full PRC Mythos ("Fable") by Nov-Dec'26.
@teortaxesTex Multiple people from Anth have also pointed to December parity in posts here.
I think GLM 5.2 points to a 7 months gap currently It's around Opus 4.7-4.8 level, all told (modulo vision which in Opus's case is garbage anyway). Mythos reached Preview status (≥ Opus 4.8, functionally) by early Feb 2026. This means full PRC Mythos ("Fable") by Nov-Dec'26.
But pretraining scale is not the hard limiting factor for "Fable" capabilities we care about most, such as cybersecurity. It can be done with massively parallel RL on extant hardware. Recipes improve quickly. Alibaba will probably do it first. Qwen 4 Max (CCP edition), Nov-Dec.
That was my original estimate before GLM's scores Elon actually might be better calibrated. Extrapolating "7 months" forward ignores scaling to Fable-tier. They start getting 950s in August; by Nov clusters are running. *If* scale is the hard limiting factor… 4+ months. If.
@scaling01 option 1: Dario believes in the attacker's advantage in cybersecurity, and 12+ is needed for Glasswing to ensure the US can't be threatened by vastly weaker models option 2: Dario believes this much is necessary to overcome the *defender's advantage* of Chinese Glasswing
@teortaxesTex did this post say why exactly a 12-24 month lead?
why not 84 months or 3 months?
@teortaxesTex did this post say why exactly a 12-24 month lead?
why not 84 months or 3 months?
Dario's Good End necessitates the gap of "12-24 months by 2028", that's his pacing challenge. Let's be precise: at least this means that by January 2028, the best Chinese model is AT MOST as good as the best US Jan 2027 model. This requires a really rapid acceleration in 2026.
@elonmusk @teortaxesTex China is rapidly catching up to the US in the AI race. If we keep having these data center protests slowing things down, the CCP will overtake the US. People don't understand how little time we actually have to win the race.
@elonmusk @teortaxesTex 2027?
2027Q1 is very possible. But Anthropic announced a few months ago that they’d already finished training their Mythos model. That means they’ve had Fable 5-level intelligence mastered for quite some time now. We have to admit that Chinese AI models still have an objective gap. Ultimately, though, I believe this competition is what drives human progress—not malice.
@lazy_grad This is strange mafs. GLM 5.2 was ready by first week of June. Mythos was probably in the middle of its training in December. I estimate that it was already around Opus 4.7 there, I have insider info that Mythos-Preview (*deployed* in early Feb, so finished Jan) was ≈Opus 4.8.
@teortaxesTex My headcanon is Fable capability comes from Anthropic identifying steering vectors that correlate to improvements on practical tasks and tuning them up/down appropriately, not scale or training sauce
Intelligence lying on the table for open models.
Thoughts? Verbal abuse?
@teortaxesTex The models catching up were because they were able to use that latest model, but with mythos being locked down, I wouldn’t be surprised if it’s a longer time horizon this time.
@geoppls cope
@RyanKemper10 @teortaxesTex Does it have generality? I didn't get the impression it was much different than 4.8 Opus when talking to it outside of code. Even on toy "show me something cool you can make" artifacts it was meh.
Then I asked it to code something difficult and it did an unusually good job.
@elonmusk @teortaxesTex Too long, I'm sure you can do it faster and open source it.
@teortaxesTex I predict that next gen PRC models will be delayed by however long it takes for Anthropic to get Fable deployed again.