DeepSeek R2 launch stalled as CEO balks at progress

Original Article ↗ Hacker News Discussion ↗

Article framing & sources

  • Several commenters note a mismatch: the headline blames “lack of progress,” while the body focuses on GPU shortages from export controls and never mentions “progress.”
  • The article itself attributes the delay partly to the CEO’s dissatisfaction with R2 performance, citing a secondary report that relies on anonymous sources.
  • Many are skeptical that these outlets have real insiders at DeepSeek; they see a pattern of rumor, speculation, and self-referential “China news.”

Reasons for R2 delay (speculative)

  • Hypotheses include: GPU export constraints, poor speed/performance trade-offs, internal reallocation of hardware, or waiting for Chinese-sourced silicon.
  • Others think GPUs are a weak explanation (hardware needs should be predictable), and that the real risk is reputational: after R1‑0528 raised expectations, a flat R2 could damage the brand.
  • Some suggest DeepSeek may be “waiting out” Western labs, letting them burn GPU money and saturate evaluation metas before dropping R2.

DeepSeek’s openness, business model & data sources

  • Debate over why DeepSeek keeps releasing weights and technical reports in such a cut‑throat space. Proposed reasons:
    • Branding, recruitment, and influence (“if you’re not appearing, you’re disappearing”).
    • Confidence they can out-iterate themselves, making current tech expendable.
  • Others argue open weights erode moat and predict DeepSeek will eventually close models once profits from hosting/API become more compelling.
  • Some speculate (often skeptically) that R1/R1‑0528 used outputs from Western reasoning models (OpenAI, Gemini) as training data; others counter that DeepSeek’s RL approach and thinking traces predate comparable Western releases and that concrete evidence of “misuse” is lacking.

Export controls, geopolitics & military use

  • One camp: GPU export restrictions harm global AI progress and delay competitive open models that benefit everyone.
  • Another: controls are justified because of China’s expansionism, Taiwan posture, and military ambitions; the US is not restricting allies like France or Thailand.
  • This is met with counter‑accusations of US hypocrisy (Iraq/Afghanistan, Latin America coups, Guantánamo) and arguments that an all‑powerful US AI is more frightening than an open Chinese one.
  • Some note DeepSeek is reportedly used by the Chinese military, making US/EU hosting politically implausible.

Model quality, usage, and censorship

  • R1‑0528 is widely praised as a big step up from original R1 and “roughly on par” with top proprietary models for many everyday tasks, especially writing/editing.
  • Others find it weak for coding, especially at very large context lengths; several note that all current LLMs degrade as context grows, even within limits.
  • Some say they now mainly use just o3/o3‑pro and R1‑0528, dropping Claude/Gemini; others insist OpenAI still dominates real B2B use due to quality and reliability.
  • Concern raised over DeepSeek’s built-in censorship (e.g., refusing to discuss Tiananmen even when run locally), leading to distrust about unseen omissions.

Hardware ecosystem & Nvidia/AMD

  • Discussion of how a more globally distributed chip supply (EU, others) might reduce dependence on Nvidia and US export policy.
  • Comments highlight ASML export limits and US pressure not to service China’s most advanced tools; others argue China will eventually build its own full stack.
  • Some predict Nvidia’s vulnerability if China mass-produces competitive GPUs or leans on AMD designs, though others say export controls are porous anyway (underground H100 clusters in China).

Media trust & anonymous sources

  • Extended debate about anonymous sourcing, conflicts of interest (e.g., outlets funded by investors closely tied to Western AI competitors), and widespread distrust of mainstream media.
  • One side stresses that without some trust in vetted anonymous sources we “live in a world without facts”; the other argues systemic failures and incentives have made journalism unreliable and in need of serious accountability.