"the only reason why there is even a 12-month gap between open source and frontier models is US export controls." Perplexity’s CEO Aravind Srinivas on why China’s open-source AI may become more powerful than ever. And why Anthropic has lobbied very hard for export control. "There is a chance... that, because of the export controls, they now get really good at the physical layer. One advantage they (China) have is that they can actually build data centers a lot faster. Power is not a problem. Permits are not a problem. People are not a problem. Labor is not a problem. Expertise is not a problem. And so, by forcing them to go out there and build all this, you are converting them into a far more potent competitor." --- From "20VC with Harry Stebbings" YouTube channel (Harry Stebbings ), link in commentshow more

These time metrics (12-month gap, 4 months behind, etc) are meaningless. This whole industry is absurd

DeepSeek are the absolute goats. If they had just 30% of Anthropic or OpenAI compute capacity, they would steamroll everyone else

The gap is training compute. frontier models burn billions on RLHF that open source skips to get from good to great but not godlike

No doubt, China now ships good enough weights into every cost sensitive lab on earth, while the US frontier stays rented, metered, and politically kill-switchable! The potency is different here while model advances still track effective training compute, data, and post-training search and not the number of permitted buildings! More Chinese data centres change the price and availability of inference and mid-tier training and they change the frontier only to the extent those halls are filled with chips that can actually run a single, dense, high-bandwidth job! But Srinivas is mixing three claims here that should stay separate like what export controls actually did, why a gap still exists, and whether more halls moves the frontier and to my understanding only the first is mostly right while the third is the one that matters, and it is only true under a narrow definition of compute!

We tariff ourselves.

export controls are only half the story though. open weights from labs with zero chip restrictions lag for the same reason: nobody ships frontier weights for free.

While the discussion on how Open Weights is getting closer to Closed Weights capabilities, are we missing woods for trees. Most discussion of AI adoption focuses on the models: which is smartest, cheapest or most open. The quieter bottleneck is that there is no common yardstick for deciding which model is fit for which job. Two things are missing. A shared definition of the task. Work breaks into thousands of tasks, and every evaluator names and groups them differently. One calls it "claims triage", another "document classification", a third splits it into five sub-steps. Results measured against different definitions can't be compared. A shared way of judging performance. Even for the same task, evaluators test differently and report differently. The bar also changes by industry: "summarise this document" means one thing for a marketing team and something much stricter for a lawyer or a bank's risk desk. A model that tops a general leaderboard may fail the version of the task that matters to you.

Power is not a problem? Lol

Xuất khẩu với chả nhập khẩu

Export controls are one lever, but the 12-month gap also depends on how much compute Chinese labs can actually deploy at scale. Restrictions slow hardware access more than they block algorithmic progress.

US export controls created a short-term buffer, but they inadvertently forced rapid domestic innovation at the physical compute layer. With zero permitting bottlenecks and unlimited cheap power, China's rapid infrastructure buildout could create a far more formidable long-term competitor.

the quiet version of this cost me a month once

The fact deepseek and GLM are running on huawei chips as of today and no one can tell the difference is all you need to realise

they don't need to close the gap. they need the gap to not matter.

These two will get wrecked in a market correction.

12 months behind, ha? As of Sep-2025 - Anthropic had Sonnet 4.5, OpenAI had GPT. Is he saying kimi k3 and GLM-5.3 are equal to those? These guys seem to lie intentionally and misguide people.

The export controls are doing free advertising for open source again.

it takes them a while to distill off the newst models.