China's Evs, Llms And The Stopping Power Of The Real World

Latest stories Houthis have further tightened their grip on the Red Sea Resistance airbase hit heralds a wider drone war in Myanmar Fed, battling persistent inflation, ups rates 1st time in 3 years

Bessent's lament that BYDs are“heavily subsidized” doesn't pass the“best $70,000 car $35,000 can buy” test when per vehicle purchase subsidies are on average just over $2,000 in 2026. Even layering in a 30% appreciation of the Rmb doesn't get us anywhere near BYD's (and domestic competitors) price advantage.

On January 20^th, 2025, an eccentric startup attached to a Chinese quant fund upended the AI world by releasing DeepSeek R1, an open-source near-frontier reasoning model trained on a minuscule budget. Since then, dozens of open source labs have sprouted in China pumping out near-frontier models priced at a fraction of US offerings. China's open source challenge to the economic viability of American frontier labs has implications well beyond IPO valuations.

Secretary Bessent was selling BYD short. A 2026 ~$35k BYD Seal 08 would sell for well over $70,000 – in 2022. The Seal 08's specifications are superior to a 2022 NIO ET7 which sold for $78,000 (longer range, faster charging, faster acceleration, four wheel steering) and far superior to a 2026 US market BMW i5 M60 whose $84,000 price tag and mediocre specs are scandalous in comparison.

Since 2022, China's domestic car companies have been hedonically improving at +20% per annum rates. Hedonic (or quality) improvements give consumers better products over time. In the past four years, new cars in China have gotten more than twice as good – as in a 2026 spec EV would command more than twice the price in 2022. Or, conversely, a 2022 spec EV would sell for less than half the price in 2026.

Hedonic improvements are not nearly as noticeable in the US car market. Since 2022, offerings from legacy car makers, including Tesla, appear frozen in amber. The 2026 Tesla Model 3 is essentially the same car as the 2022 version. Tesla Autopilot may have improved substantially but so have Chinese self-driving systems. Self-driving is also usually included in the price of Chinese EVs – unlike Tesla's subscription service. To Tesla's credit, they have reduced the top of the line Model 3's price by 7% versus the 2022 version.

Honda's 2026 Pilot is one generation ahead of its 2022 version. It looks different. On specifications and features, however, one would be hard pressed to find any substantive improvements. Honda's SUV is a little bigger with a little more power and perhaps a slightly more capable lane keep and adaptive cruise control system. And they increased the price by 9%.

Tracking hedonic improvements in cars is easy. Historical model specifications and prices are readily searchable.

But this isn't just cars. There have been noticeable hedonic step changes across almost all consumer and industrial segments in China. While difficult to quantify, domestic travelers now expect luxury accommodations at budget prices. Leaving food quality for experts to judge, this non-foodie can confidently state that the interior décor arms race has flooded the country with trendy looking restaurants.

OpenAI publicly launched Astra on September 5.

On September 8, the company published a solution to the Navier-Stokes equation, completed by an as yet unreleased internal model.

The AI world has since spun itself into a frenzy about impending RSI and the end of the world while the math world sunk into a depression in all likelihood deeper than that experienced by the chess and go world after their grandmasters were felled by AI.

Not speaking for all of us but many mechanical engineers surely raised an eyebrow and wondered whether we are too dumb to understand the brouhaha. Many mechanical engineers likely had no idea that our lowly Navier-Stokes equation was held in such esteem by mathematicians – at least yours truly was genuinely dumbfounded. Our humble little equation to simplify fluid dynamics was never meant to be“solved.” It was always a scratchpad exercise, a rule of thumb, a mental model for how fluids“are supposed to” behave. Think of it as the Black-Scholes model for mechanical engineers.

Whatever“solutions” were useful in a mechanical engineering context could have been simulated by computer programs for decades (e.g. Ansys, Simcenter, OpenFOAM). The ultimate arbiter of whether the solution is correct is a working physical prototype. That's how engineering in the physical world works.

The OpenAI“solution” of Navier-Stokes that satisfies mathematicians has zero, repeat, zero engineering implications. Real world fluids are affected by countless parameters – not the overly simplified four unknowns and four constants of the Navier-Stokes equation. The mathematician's solution of Navier-Stokes would fall apart instantly in real world tests. Not one plane will fly more efficiently and not one submarine will travel more silently because of OpenAI's“solution” of Navier-Stokes.

Ten thousand agents running a state-of-the-art LLM have produced a solution as useful to a mechanical engineer as a chess engine. Hey, America has the best chess engines but the country's lead in LLMs over the past four years hasn't helped Ford and General Motors close the gap with BYD and Geely – a gap that has, in fact, widened considerably.

The stopping power of the physical world has, so far, humbled America's frontier AI labs. The frontier labs have been busy bolting physics, chemistry, engineering and biological science capabilities onto their LLMs. This has, so far, not had a measurable effect on real world science and engineering as China's universities widen their lead in the Nature Index (see here ) and the country's industries out-compete global competitors using inferior open source LLMs.

In the paper“The Last AI Built by Humans: Towards Genuine Recursive Self-Improvement,” written by a team of AI researchers across China's AI landscape (e.g. Tsinghua, Bytedance, Xiaohongshu), recursive self-improvement (RSI) is categorized into five stages summarized below:

L1, Execution stage – Human researchers fully design the improvement pipeline. AI only executes the pre-defined improvement steps; it cannot decide what to improve.

L2, Improvement-selection stage – AI autonomously selects which components to improve (prompts, weights, tools, memory). Humans still define the overall objective and evaluation system.