en.wikipedia.org/wiki/Liang_Wenfeng
3 corrections found
Liang, being considered as an industry expert, was asked to provide opinions and suggestions on a draft for comments of the annual 2024 government work report.
The year is wrong here: this January 2025 symposium concerned the draft 2025 government work report, not an 'annual 2024 government work report.'
Full reasoning
Official Chinese government material about the January 20, 2025 symposium does not describe it as a meeting about an 'annual 2024 government work report.' Instead, it says Premier Li Qiang convened the symposium to hear suggestions on a draft government work report, and that attendees offered suggestions for government work in 2025.
A companion page on the Chinese government website explicitly invited comments on "2025 China's Government Work Report." That makes the article's wording incorrect: while the final report reviewed work done in 2024, the report under discussion in January 2025 was the 2025 government work report, not a '2024 government work report.'
2 sources
- Chinese premier chairs symposium to hear suggestions on draft gov't work report
BEIJING, Jan. 20 -- Chinese Premier Li Qiang on Monday presided over a symposium to hear opinions and suggestions on a draft government work report... They offered suggestions on addressing the current challenges in development and ensuring government work in 2025.
- Share your views on 2025 China's Government Work Report
Whether you are investing, living, working, studying and traveling in China or doing business in China or with Chinese companies, the Chinese government welcomes your input on the 2025 China's Government Work Report to better serve your needs.
The model was built using just 2,048 Nvidia H800 GPUs at a cost of $5.6 million, showcasing a resource-efficient approach that contrasted sharply with the billion-dollar budgets of Western competitors.
This sentence conflates DeepSeek-R1 with DeepSeek-V3. The 2,048-H800 / ~$5.6 million figure refers to V3’s training, not to R1 itself, and even that figure covered only the final pretraining run rather than the full development cost.
Full reasoning
The article attaches a specific hardware-and-cost figure to DeepSeek-R1, but the primary papers do not support that.
- The DeepSeek-R1 paper describes R1's reasoning-training approach, but its abstract does not say R1 was trained on 2,048 H800 GPUs or for $5.6 million.
- The DeepSeek-V3 Technical Report says DeepSeek-V3 required 2.788 million H800 GPU-hours for its full training. That is the source of the widely repeated H800-based cost narrative.
- Independent analysis from CSIS explains that the V3 paper's 2,788,000 H800 GPU-hours works out to about $5.576 million at an estimated $2 per GPU-hour, and stresses that this was only the final successful pretraining run, excluding earlier experiments, post-training, and inference costs.
- Al Jazeera likewise notes that DeepSeek said it used about 2,000 H800 GPUs and spent $5.6 million to train R1's foundational model, V3 — not R1 itself.
So this sentence is inaccurate because it assigns the V3 training figure directly to R1, and it presents the number as though it were the full build cost of R1 rather than a narrower estimate tied to V3's training run.
4 sources
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
Abstract: ... we introduce DeepSeek-R1, which incorporates multi-stage training and cold-start data before RL. ... [The abstract does not state that R1 used 2,048 H800 GPUs or cost $5.6 million.]
- DeepSeek-V3 Technical Report
Abstract: ... DeepSeek-V3 requires only 2.788M H800 GPU hours for its full training.
- DeepSeek: A Deep Dive
DeepSeek's V3 research paper states that their models were trained on 2,788,000 GPU-hours using Nvidia H800 chips, which, at an estimated cost of 2 dollars per GPU-hour, equates to $5.576 million. This was only the final, successful pretraining run.
- China's DeepSeek faces questions over claims after shaking up global tech
... they had used 2,000 Nvidia H800 GPUs ... and spent $5.6m to train R1's foundational model, V3.
The development of DeepSeek-R1 occurred amidst U.S. sanctions where Trump limited sales of Nvidia chips to China.
This misattributes the Nvidia-to-China chip restrictions to Trump. The key A800/H800 export curbs cited in DeepSeek coverage were imposed under the Biden administration in October 2023.
Full reasoning
The sentence credits Donald Trump with limiting Nvidia chip sales to China during R1's development, but the relevant export controls on Nvidia's China-specific AI chips were imposed before Trump returned to office.
- A Reuters report published by Yahoo Finance states that the restrictions on Nvidia's advanced AI chips for China, including the A800 and H800, went into effect in October 2023 after regulators accelerated the deadline.
- CSIS testimony by Gregory Allen likewise says the Biden administration changed the export-control thresholds in October 2023 to block exports of A800s and H800s to China.
So the article's attribution is wrong: the chip-sale limits central to DeepSeek's story were a Biden-era policy, not a Trump action.
2 sources
- Nvidia says U.S. speeded up new export curbs on AI chips
Oct 24 (Reuters) - Chip designer Nvidia said new U.S. export restrictions blocking the sales of its high-end artificial intelligence chips to China had gone into effect on Monday... The restrictions impact shipments of the company's modified advanced AI chips A800 and H800.
- DeepSeek: A Deep Dive
The Biden administration ultimately realized that continued sales of the A800 and H800 chips meant that their policy would not have the intended impact on China's AI ecosystem... The U.S. government modified the export control performance thresholds in October 2023 to block exports of A800s, H800s...