2000 word summary lightly edited from Claude https://mp.weixin.qq.com/s/bYZrKp48Y7EpsU8_vd6TcQ
DeepSeek at a Crossroads
DeepSeek, the Chinese AI lab founded by Liang Wenfeng in 2023, is navigating a period of quiet but significant internal change. Several key researchers have departed — including core contributors to DeepSeek-R1 and its multimodal work — lured away by competitors offering salaries two to three times higher and lucrative equity packages tied to upcoming IPOs. Yet there has been no mass exodus; most researchers remain, drawn by the lab’s unusual culture and Liang Wenfeng’s longer-horizon vision of AGI.
That culture is genuinely distinctive. DeepSeek operates with no mandatory overtime, no time-clock, and no formal performance reviews. Most staff leave by 6–7 PM. Liang Wenfeng believes high-quality cognitive work maxes out at six to eight hours a day. The roughly 100-person research team is flat — just two levels — and weekly meetings are open across teams, encouraging organic collaboration. New research directions often begin simply because three to five people think an idea is worth pursuing.
Liang Wenfeng himself is an unusual founder: he stopped meeting investors after his capped-return proposal was rejected, rarely socializes with his team, and focuses almost entirely on technical depth. Colleagues describe him not as a CEO but as a researcher who goes deep on architecture, infrastructure, and data alongside his team.
Technically, DeepSeek has spent the 15 months since R1’s global breakout methodically advancing efficiency (new attention mechanisms, a switch to the domestically developed TileLang operator library), model architecture improvements like mHC for training stability, and more unconventional bets like OCR-based document understanding and biologically inspired continuous learning. The upcoming DeepSeek V4 — expected in April 2026 — will likely be the strongest open-source model available, though probably not a dramatic leap beyond peers.
The tensions are real. Competitors like Anthropic, OpenAI, and Chinese peers including Kimi and MiniMax are pushing hard on agentic applications and coding — the current industry battleground — while DeepSeek’s iteration pace has been slower. Benchmark scores mean less in an agentic world where product reach and real-world usage data matter more. DeepSeek has a product team of several dozen but hasn’t yet entered hot areas like AI coding agents or general-purpose assistants in a serious way. That appears to be changing: a recent job posting explicitly named Claude Code and other agent products as reference points for a new product manager role.
Liang Wenfeng is also beginning to address the question of company valuation — a pressing issue for employees holding unpriced stock options as peers go public at soaring valuations.
DeepSeek remains one of the few elite AI labs in the world not running on 80-hour weeks and venture capital pressure. Whether that deliberate pace and contrarian focus can sustain its edge — or whether it needs to shift — is the central question the lab is now working through.
what would you add if you had 5 more minutes
2:19 PM
Weighed editorial gaps and underdeveloped content areas
Weighed editorial gaps and underdeveloped content areas
A few things the summary glossed over that are worth surfacing:
The domestic chip angle is underplayed. DeepSeek’s deliberate investment in adapting to Chinese-made GPUs — the TileLang switch, the UE8M0 FP8 format designed for next-gen domestic chips — isn’t just a technical footnote. It’s a strategic hedge against US export controls and a signal that Liang Wenfeng is thinking about infrastructure sovereignty, not just model performance. That’s a different kind of ambition than most AI labs have.
Liang Wenfeng as a character deserves more. The detail about him sitting silently at a dinner for half an hour before Yan Junjie realized who he was is one of the most telling things in the piece. So is the anonymous 138 million yuan charity donation. The summary treats him functionally — his management philosophy, his hours — but doesn’t quite capture how genuinely strange and rare this person is by tech founder standards.
The “original innovation” mandate is more radical than it sounds. Liang Wenfeng is explicitly asking his team to pursue directions that larger, better-resourced labs won’t bother with. That’s a counter-positioning strategy as much as a research philosophy — and it’s either visionary or a liability depending on how the next few years play out.
The downstream impact on the broader Chinese AI ecosystem — inspiring Kimi K2, giving rise to new teams like MiroMind — got one sentence. That’s arguably the most consequential part of DeepSeek’s 2025.