Personal Notes on DeepSeek WenFeng's Leaked Internal Discussion

I’ve read DeepSeek founder Liang Wenfeng’s full internal discussion from start to finish. Here are my personal notes:

□ DeepSeek’s Core Vision

Benefit humanity by open-sourcing their best models and making them affordable.

Stay relentlessly focused on achieving AGI. Everything else is secondary.

□ What Focusing on AGI Means

Build general-purpose agents, especially Coding Agents, because they can automate and accelerate the widest range of tasks.

Drive down training and inference costs through core algorithmic and infrastructure improvements. Lower cost means the same compute budget can be used to train larger next-generation models.

Solve Continual Learning. Today’s LLMs cannot truly learn or remember after deployment. Everything has to be placed into the context window. If AI can continuously learn from experience, it can enable recursive self-improvement and potentially accelerate progress toward AGI.

□ Interesting Take on Embodied Intelligence

Wenfeng believes embodied intelligence should come after Continual Learning and Recursive Self-Improvement. This is a much more conservative view than the current mainstream narrative.

I tend to agree with him, based on my brief experience experimenting with VLAs last year and my close tracking of recent developments:

• VLAs and world models are still far from working reliably.

• Frontier robotics labs, such as Physical Intelligence and NVIDIA, are increasingly moving toward more scalable and economical ways to collect real-world data, such as egocentric data collection.

□ Focusing on AGI Requires Restraint

Do not compete with ByteDance, Tencent, or Alibaba on consumer super apps.

Do not chase video generation, 3D generation, world models, or multimodal capabilities for their own sake.

Stay focused on AGI. Everything else is noise.

□ Competition with the US

Compute, not talent, is the primary bottleneck. DeepSeek is extremely compute-constrained.

DeepSeek currently has roughly 20,000 H-equivalent GPUs, while Wenfeng estimates that frontier-scale training requires around 50,000 GB300 GPUs.

DeepSeek V4 was trained using GPUs roughly two years behind the frontier and with only around 1/20 of the compute used by leading US labs. This was likely made possible by algorithmic and infrastructure optimization.

Roughly speaking, 1 NVIDIA GB200 ≈ 4 Huawei 950s.

Huawei 950 also trails by around two years, since GB200 was released roughly two years before Huawei 950.

However, Wenfeng believes CUDA’s moat is shrinking rapidly:

• With the help of Coding AI, efficient DSLs for GPU kernel development, such as TileLang, can be built much more quickly.

• Starting with DeepSeek V3, training ran on NVIDIA GPUs without relying on CUDA as the primary software stack, instead using TileLang.

• They are now actively porting TileLang to Huawei chips, further reducing dependence on the CUDA ecosystem.

□ Wenfeng’s Endgame Prediction

In the long run, frontier models will converge toward very similar capability levels.

The real differentiators will be:

• Cost, which he sees as the most important factor.

• Time, meaning who reaches AGI first.

If DeepSeek can maintain its current team and execution discipline, Wenfeng is optimistic that achieving AGI is ultimately a matter of when, not whether.