跪拜 Guibai
← All articles
Agent · AIGC

DeepSeek's Liang Wenfeng on AGI, Restraint, and Why the Company Finally Took Capital

By 李剑一 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

DeepSeek's decision to raise capital marks the end of its no-funding era and signals that even mission-driven AI labs must use equity to retain talent when competitors are poaching aggressively. Liang's candid admission that team stability was the 'biggest risk' reframes the financing as a defensive talent play, not a pivot to commercialization.

Summary

In a closed-door investor call, DeepSeek founder Liang Wenfeng detailed a stepwise path to AGI: chain-of-thought reasoning last year, agents this year, and a coming singularity where agents achieve self-iteration and continuous learning, unlocking embodied intelligence. The company maintains strict restraint, refusing to build image or video generation tools to stay focused on the AGI main line. Liang pegged the gap with the US at roughly two years, driven entirely by a 20-to-1 compute disadvantage, and predicted Nvidia's CUDA moat will erode as dedicated AI chips and domestic Chinese silicon mature.

The most revealing operational detail was the motive for DeepSeek's first-ever financing round. Liang identified team stability as the company's single largest risk and said the funding, which delivered substantial option grants, directly addressed that risk. The calculus was not about buying loyalty but about removing financial anxiety so researchers could stay focused in a relaxed, KPI-free environment.

On competition, Liang framed DeepSeek as a challenger willing to take a thinner slice of the value chain than OpenAI, betting that lower costs and comparable product quality will win. He projected that even a worst-case scenario of selling API access could sustain a public company, though B2B demand alone may take a long time to push the firm to net profitability.

Takeaways
DeepSeek's AGI roadmap treats 2024 as the year of chain-of-thought and 2025 as the year of agents, with a self-iterating agent loop as the prerequisite for embodied intelligence.
The company deliberately avoids image, video, and 3D generation to preserve focus on the AGI main line.
Liang estimates China trails the US by roughly two years in AI, a gap he attributes entirely to a 20-to-1 compute disadvantage.
Nvidia's CUDA ecosystem will eventually fragment as gaming and AI compute cards diverge, and US export controls will accelerate a domestic Chinese chip ecosystem.
DeepSeek's first financing round was primarily aimed at issuing employee options to stabilize the team against aggressive industry poaching.
Internal culture rejects KPIs and overtime; Liang argues that research requires a relaxed environment for genuine interest to develop.
Even a worst-case business model of selling API access could support a listed company, though B2B demand alone may delay net profitability.
DeepSeek remains committed to open source, viewing closed source as offering no benefits and open source as imposing no cost on its AGI mission.
Conclusions

Liang's framing of team stability as the company's single largest risk, and financing as the fix, reveals that even a mission-driven lab cannot ignore the market price of AI talent when competitors are offering large packages.

The explicit refusal to build multimodal generation products is a bet that AGI will subsume those capabilities later, making today's image and video tools a distraction rather than a moat.

Liang's claim that China can compete by taking a smaller share of the value chain than OpenAI is a deliberate strategy of undercutting on margin, not just on price.

The assertion that Scaling Laws have not hit a wall for DeepSeek, while Silicon Valley debates their limits, suggests the compute-constrained Chinese labs are still climbing the curve that well-resourced US labs may be plateauing on.

Treating open source as cost-free to the AGI mission assumes that commoditizing model weights does not erode any future pricing power DeepSeek might need, a bet that only works if the real moat lies elsewhere.

Concepts & terms
AGI singularity (in DeepSeek's framing)
The point at which an AI agent can self-iterate, continuously learn, and develop its own next version without human intervention, enabling a transition to embodied intelligence.
Scaling Laws
The observed relationship where increasing model parameters, training data, and compute leads to predictable improvements in AI capability and the emergence of new abilities.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗