DeepSeek V3 : ce ChatGPT chinois m'a bluffé (mais j'ai peur)

Item: DeepSeek V3 : ce ChatGPT chinois m'a bluffé (mais j'ai peur)
Rating: 4
Author: PostSingularity Institute

by Ludo Salenne

Applied Sciences & Engineering AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence

🎙 Ludo Salenne 👥 267K 📅 January 6, 2025 ⏱ 35 min 👁 66K 📄 expert opinion 🧭 2026-08-21

Available in: English (current) Français

Keywords

DeepSeek V3AI testChatGPT comparisonMixture of Expertsdata privacy

Summary

Ludo Salenne presents a comprehensive test of DeepSeek V3, a Chinese AI model that has been generating buzz. He starts by showing the interface, which is minimalist and similar to ChatGPT, and highlights features like file upload (up to 50 documents) and web search. He then conducts several tests: summarizing a PDF, checking weather forecasts (where he finds inaccuracies), and searching for YouTube tutorials. He discusses the model’s architecture, based on Mixture of Experts (MoE), and compares its performance with ChatGPT’s o1 and o1 Pro modes on a complex protein identification task, where DeepSeek fails. He also addresses concerns about censorship, data privacy, and the model’s system prompt. The video concludes with his overall opinion, praising the speed and open-source nature but warning about data security and potential biases.

129 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video offers valuable hands-on insights into DeepSeek V3’s capabilities and limitations, based on real tests. The argumentation is structured and balanced, acknowledging both strengths (speed, open-source, context length) and weaknesses (hallucinations, data privacy, censorship). However, the analysis is largely anecdotal and lacks rigorous scientific methodology. The creator’s personal experience and subjective impressions are central, which may not be representative of all users. The comparison with ChatGPT is useful but not exhaustive, and the protein test is a single data point.

Scientific Rigor, Source Quality, Title Accuracy

The video cites several sources, including the DeepSeek technical report, the LMArena leaderboard, and an article about Hugging Face’s CEO concerns. These are relevant and add credibility. However, the creator does not always verify the accuracy of the information he presents, and some claims are based on his own tests rather than peer-reviewed research. The title accurately reflects the content, and the video is well-structured with clear chapters. The creator also mentions his own resources and other tutorials, which are promotional but not misleading.

180 words

Title / Content Match

The title accurately reflects the content: a personal test of DeepSeek V3 with both positive and negative aspects, including concerns about data security.

Quality & Reliability

7/10

The video provides a hands-on test of DeepSeek V3, with practical demonstrations and comparisons. However, the analysis is largely subjective and lacks rigorous scientific methodology. The creator acknowledges limitations and cites some sources, but the overall reliability is moderate.

Chapters

Cited Sources

DeepSeek chat — Official DeepSeek interface used for testing.
LMArena leaderboard — Used to compare model rankings.
Hugging Face CEO concerns about open-source AI — Article discussing potential risks of open-source Chinese AI.

Concurring Sources

DeepSeek-V3 Technical Report — The official paper confirms the model's architecture and performance claims.

Dissenting Sources

LMArena leaderboard — The leaderboard shows DeepSeek V3 ranked 9th, behind several proprietary models, contradicting the hype of it being the best.

External References

Contribution & Novelties

The video provides a practical, user-oriented evaluation of DeepSeek V3, highlighting both its impressive capabilities and significant concerns. It contributes to the public discourse by offering a balanced perspective amidst hype. The detailed comparison with ChatGPT o1 Pro on a complex task is particularly insightful, showing real-world performance differences.

Pour aller plus loin :

DeepSeek-V3 Technical Report — The official paper detailing the architecture and benchmarks.
Mixture of Experts (MoE) — Overview of the MoE architecture used in DeepSeek.
LMArena (formerly Chatbot Arena) — Platform for comparing AI models via user votes.

92 words

Radar Profile

The radar profile shows high scores in quantity and quality of information, reflecting the video's comprehensive coverage and practical tests. The technical level is moderate, suitable for a general audience. The global reliability is decent but not exceptional, due to the subjective nature of the evaluation and lack of rigorous methodology.

Reliability 7/10

💬 Positif. Sur les 30 commentaires analysés, la majorité exprime une appréciation positive de la vidéo et de DeepSeek, avec des remerciements et des retours d'expérience. Certains soulèvent des points de vigilance sur la confidentialité des données, mais le ton général est enthousiaste et constructif.