A kernel engineer at DeepSeek has published a strikingly frank account of why he keeps building the technology he expects to replace him. Shengyu Liu, who writes as intlsy, argues that competitive pressure makes stopping impossible, and that open AI is the only guard against a future run by a handful of companies.
DeepSeek Watch analysis
What happened
Liu's post lands in the middle of the sharpest safety argument of the year. Dario Amodei called at the weekend for frontier development to slow. Jensen Huang pushed back, Elon Musk and Sam Altman voiced support for pacing, and Mark Zuckerberg wrote that Meta has quietly slowed its own model work under safety constraints.
Into that debate walks a working engineer. Liu describes AI's move from assistant to platform, one that can now read complex codebases such as NVIDIA's CUDA and work with them through operational tools. He estimates that within six months to a year, model output could match the quality of his own. The arithmetic is his: a system thinking at 300 tokens a second, typing a command in one, writing a function in twenty.
The heart of the post is why he carries on. Pride and the thrill of the work play their part. The load-bearing reason is competitive: if he slows, rivals will not, and the replacement gets built either way. He would rather it were built in the open.
There is an ideological edge, too. He frames the endgame as a choice between two societies: a Cyberpunk 2077 world in which a few players control the best AI while everyone else gets the budget tier, or something closer to a communist model. He asks whether a world where Anthropic holds the most powerful AI becomes one or the other, and writes that Anthropic mastering AGI first would be like handing a weapon of mass destruction to Hitler before the Allies had one. Since he believes neither OpenAI nor Anthropic will supply cutting-edge intelligence openly and cheaply, working at DeepSeek is, in his view, the only way to keep the best technology open source.
Wccftech, which carried the post, notes the author's pedigree: Liu captained Peking University's supercomputing team to first and second places at top competitions, a record that lends this particular blog post rather more standing than most.
Why it matters
This is the clearest first-person statement yet of the trap Amodei's slowdown call is trying to escape. Essays can describe race dynamics from altitude. Liu describes them from the engine room: a specialist who believes his replacement is six to twelve months out and keeps building anyway, because unilateral restraint only hands the future to someone else.
It also sharpens the open-versus-closed axis of the argument. The pacing debate has mostly run between American labs. Liu reframes it geopolitically: concentration of frontier AI among a few companies is itself the danger, and open weights the counterweight. Whatever one makes of that, it helps explain why a slowdown pact that excludes China's labs may be structurally stillborn.
Our read
Take the argument seriously and the predictions lightly. The race-dynamic logic is coherent and honestly stated by someone with standing to describe the work. The capability estimates are one engineer's back-of-envelope, not benchmarks. The wartime analogy is inflammatory, and we would file it under rhetoric rather than analysis, but the question beneath it, who controls frontier intelligence, is the right one.
Read the post, then ask what would falsify it. If frontier labs demonstrably slow without a binding agreement, Liu is wrong. If they cannot, his post may be the most honest job description in the industry.
What to watch
- DeepSeek's response, or silence
Whether the company endorses, ignores or distances itself from an employee framing its mission this starkly. - Echoes from other labs
More frontier engineers writing versions of this post would harden the race-dynamic argument from opinion into testimony. - The six-to-twelve-month clock
Liu's own timeline makes the claim testable: kernel-engineering quality against frontier models is now a dated bet. - A mechanism, not a mood
Amodei's proposal, Altman's endorsement and Liu's fatalism all await the same thing: something outsiders can inspect.
Discussion spark: Liu argues open frontier AI is the safeguard against a concentrated one. If he is right that nobody can afford to stop, are open weights the safety valve or just a faster road to the same place?
Sources and evidence
not affiliated with or endorsed by DeepSeek