China Ships Three Major AI Releases in a Single Day
Gary Marcus Calls Out OpenAI's Math Paper for Marketing Over Method.
China Ships Three Major AI Releases in a Single Day
While Washington was getting a private preview, Chinese labs went the opposite direction and shipped in public, all at once. MiniMax released an open-weights multimodal video model (H3), ByteDance dropped Seedance 2.5 for 2K video generation, and DeepSeek pushed a public beta API for V4-Flash — a sparse mixture-of-experts model running just 13B active parameters out of 284B total, with a V4-Pro variant coming [4][5][6].
The efficiency angle matters more than the video demos. A 13B-active-of-284B MoE architecture, some of it under MIT license, is a direct shot at the cost-per-token economics that US labs have been slow to compete on. Chinese state media framed it as a "period of concentrated breakthroughs," and for once that's not just spin — three separate labs, three separate product categories, same day [4].
Open licensing plus real efficiency gains is the combination worth watching. It's the same playbook that made Llama and Mistral relevant in the West, but running at a pace that suggests coordination, not coincidence.
Gary Marcus Calls Out OpenAI's Math Paper for Marketing Over Method
OpenAI published a 249-page paper detailing new math results, including a chain-of-thought solution to an open unit-distance problem from May 2026. Gary Marcus tore into it publicly, pointing out the paper omits raw chains-of-thought (only cleaned-up summaries), doesn't detail proof verification, and is vague on how much human involvement shaped the "solution" [7][9].
Marcus called OpenAI a "sock puppet" on X, and while that's typical Marcus hyperbole, the underlying critique lines up with a broader pattern this year: Apple's reasoning-limitations research, Quanta's ongoing "is AI reasoning right for the wrong reasons" line of inquiry, and now this [8][9]. The question isn't whether the model got the right answer — it's whether we can verify how.
This matters beyond academic point-scoring. If you're deploying reasoning models for anything with real stakes — financial analysis, legal review, medical triage — "it got the right answer in our internal test" isn't an audit trail. The gap between marketing claims and reproducible rigor is exactly where enterprise trust breaks.
What This Means For Your Business
Three stories, one thread: the industry is quietly moving past "does the model work" and into "who's watching it work, and can we prove it." Astra's multi-day autonomous agents, DeepSeek's cheap-at-scale MoE architecture, and the transparency fight over OpenAI's math paper are all symptoms of the same shift — model capability is becoming a commodity, and the real differentiation is moving to orchestration, verification, and judgment layers sitting on top.
If you're a company deciding how to build with AI right now, the lesson isn't "wait for GPT-6." It's that the products winning in 12 months won't be the ones with the best raw model — they'll be the ones that can explain, audit, and control what their agents actually did during those hours or days of unsupervised work. That's an orchestration and governance problem, not a model-selection problem, and most teams aren't building for it yet.
The China releases are a warning on cost curves: if frontier-adjacent capability keeps getting cheaper and more open, the moat isn't the model — it's what you build around it. Vertical integration, workflow specificity, and trustworthy verification will matter more than which lab's logo is on the API call.
Key takeaway: The frontier is shifting from "can the model do it" to "can you prove what it did" — and that's a judgment problem, not a coding problem.
Sources
- https://www.theinformation.com/briefings/exclusive-openai-previews-astra-ai-model-dc
- https://the-decoder.com/openai-is-reportedly-building-astra-a-model-family-designed-to-work-on-problems-for-hours-or-days/
- https://aiweekly.co/alerts/openai-previews-astra-model-to-us-senators-in-washington
- https://www.globaltimes.cn/page/202607/1367274.shtml
- https://openrouter.ai/models
- https://www.chinatalk.media/p/deepseek-v4
- https://garymarcus.substack.com/p/checking-the-math-behind-openai-and
- https://www.quantamagazine.org/is-ai-reasoning-right-for-the-wrong-reasons-20260731/
- https://x.com/GaryMarcus/status/2083568386072887529
Stay ahead of AI
No spam. Unsubscribe anytime.
Want to go deeper?
Reading the news is one thing. Exploring the frontier is another. See what we're building.