ChatGPT n'est pas devenu mauvais (ma réponse à Micode/Underscore_)

Item: ChatGPT n'est pas devenu mauvais (ma réponse à Micode/Underscore_)
Rating: 3
Author: PostSingularity Institute

by Ludo Salenne

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence

🎙 Ludo Salenne 👥 267K 📅 January 26, 2024 ⏱ 24 min 👁 17K 📄 expert opinion 🧭 2026-08-21

Available in: English (current) Français

Keywords

ChatGPTGPT-4performancequantizationprompt engineering

Summary

In this video, Ludo Salenne responds to claims that ChatGPT has become less capable, particularly referencing a video by Micode on the Underscore_ channel. He argues that ChatGPT is not inherently dumber but has become more demanding due to resource constraints at OpenAI. He explains that the rapid growth in users and the addition of new features have strained resources, leading to a perceived decline in performance. He discusses the concept of quantization, which reduces model size but may affect precision, and emphasizes the importance of crafting optimized prompts to get better results. He introduces the ‘RCT’ method for prompt engineering and advises breaking tasks into smaller subtasks. He also addresses the theory that user feedback may be misleading AI training, referencing OpenAI’s research on weak-to-strong generalization. He concludes by announcing updates to GPT-4 Turbo that should improve performance and encourages viewers to adapt their usage to maintain quality.

149 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides a clear and structured argument, offering plausible explanations for the perceived decline in ChatGPT’s performance, such as resource allocation and quantization. The creator uses analogies (e.g., the box of crayons) to make technical concepts accessible. However, the argumentation is largely based on personal experience and inference rather than empirical data. The video does not present original research or systematic analysis, but it does offer practical advice for users, which adds value for a non-expert audience.

Scientific Rigor, Source Quality, Title Accuracy

The video cites several sources, including an article from Inc.com about ChatGPT’s ’laziness’, a blog post on quantization, OpenAI’s research on weak-to-strong generalization, and the LMSYS Chatbot Arena leaderboard. These sources are relevant and lend some credibility to the claims. However, the creator does not critically evaluate these sources, and some are anecdotal or from non-academic outlets. The title accurately reflects the content, and the video is well-structured with clear chapters. The creator also acknowledges the collaborative nature of the response, avoiding a confrontational tone.

178 words

Title / Content Match

The title accurately reflects the content: a direct response to the claim that ChatGPT has degraded, arguing that it is not inherently worse but requires better prompting.

Quality & Reliability

6/10

The video presents a plausible hypothesis (resource constraints, quantization, prompt quality) but relies on anecdotal evidence and lacks rigorous empirical validation. The creator's expertise is practical rather than academic, and the argument is persuasive but not scientifically robust.

Chapters

Cited Sources

ChatGPT Is Showing Signs of Laziness. OpenAI Says AI Might Need a Fix. — Cited as evidence of public complaints about ChatGPT's performance.
ChatGPT Users Statistics — Used to illustrate the rapid growth in ChatGPT user numbers.
Due to high demand, we've temporarily paused upgrades... — Referenced to show OpenAI's resource constraints.
What are Quantized LLMs? — Explains the concept of quantization in large language models.
Weak-to-strong generalization — Cited to discuss the idea that AI models may need to train themselves.
New embedding models and API updates — Mentioned as an example of OpenAI's updates and improvements.
LMSYS Chatbot Arena Leaderboard — Used to show rankings of AI models and the paradox of newer versions performing worse.
Micode/Underscore_ video: ChatGPT est devenu mauvais — The video to which this is a response.

Concurring Sources

LMSYS Chatbot Arena Leaderboard — Shows that some newer models perform worse than older ones, supporting the claim of performance variability.
ChatGPT Is Showing Signs of Laziness. OpenAI Says AI Might Need a Fix. — Reports on user complaints and OpenAI's acknowledgment of issues.

Dissenting Sources

No specific discordant sources cited — The video does not present any sources that directly contradict its claims.

External References

Contribution & Novelties

The video offers a practical perspective on the perceived decline in ChatGPT’s performance, attributing it to resource constraints and the need for better prompting. It provides actionable advice for users, such as using the RCT method and breaking tasks into subtasks. The discussion of quantization and weak-to-strong generalization adds depth, though these concepts are not new to the AI community.

Pour aller plus loin :

Quantization in deep learning — Provides background on quantization techniques.
Weak-to-strong generalization — OpenAI’s research paper on this topic.
Prompt engineering — Overview of the field and best practices.

94 words

Radar Profile

The radar profile shows moderate scores across all dimensions, with slightly higher scores in information quantity and quality, reflecting the video's informative but not deeply technical nature. The low technical level score indicates that the content is accessible to a general audience, while the fiabilite_globale score suggests a moderate level of trustworthiness.

Reliability 5/10

💬 Équilibré : Sur les 30 commentaires analysés, les avis sont partagés entre ceux qui apprécient l'analyse et ceux qui restent sceptiques, certains soulignant que l'expérience utilisateur ne devrait pas se dégrader pour les abonnés payants.