Vitalik Buterin points out a limitation of AI: despite its extensive knowledge, it struggles to assess the depth of a text, especially when it discovers it for the first time. For the co-founder of Ethereum, this difference between knowledge and wisdom deserves special attention.
In brief
- On October 29, Vitalik Buterin described the AI as “extremely knowledgeable”, but “not very wise”, particularly when it evaluates a text.
- According to him, the models encounter more difficulties when faced with unpublished writing, which does not benefit from the human comments present in their training data.
- Finally, he underlines an essential difference: we can easily verify the knowledge of a model, but much more difficult to measure its wisdom.
AI knows the references, but does it understand the ideas?
Vitalik Buterin regularly expresses his reservations about AI agents, as in October with his “paranoid mode”. On October 29 at 4:11 UTC, he posted a new reflection in response to writer Ruxandra Teslo.
The discussion followed a Wall Street Journal article about Humanities 10, a freshman course at Harvard that bans students from using phones and computers for a year.
In his message onVitalik explains that AI can mobilize knowledge on almost any topic, without requiring extensive judgment. She then risks stopping at the most superficial aspects of a text, without fully understanding the scope of her argument.
This remark comes in a debate on the quality of exchanges with chatbots. Some consider them to be better conversationalists than most humans when discussing the great classics. Conversely, one writer believes that their comments on her essays often amount to empty responses.
The first reading
For Vitalik Buterin, this limit appears especially when the model discovers a text for the first time. Indeed, a famous work benefits from decades of reviews, analyzes and comments that the AI can find in its training data. A recent trial, on the other hand, does not have the same history.
The model must therefore rely more on the text itself, without being able to mobilize as much pre-existing analysis. This difference particularly matters in crypto, where teams use AI agents to review white papers, audits or governance proposals.
However, some developers are already warning about flaws that AI could exploit without fully understanding the implications. Under these conditions, the summaries produced by a model should not replace human verification, especially when the technical or financial issues are important.
Wisdom, a quality difficult to measure
Vitalik Buterin finally raises a more subtle problem: the knowledge of a model is verified quite easily, unlike its wisdom. You can check a date, a reference or a number. On the other hand, judging the relevance of reasoning requires more than simple factual verification.
An episode from July illustrates this distinction. A competition organized around the identification of the author of an anonymous Ethereum document rewarded, on July 6, the team of Franklyn Wang, from Co-Invest. After examining 27 texts, she obtained a result associated with a probability of approximately 20%.
This experience reminds us that artificial intelligence can recognize stylistic characteristics without necessarily grasping the ideas hidden behind a text. The same caution is required when analyzing code or technical documents: identifying patterns does not always mean understanding the intentions of their author.
Maximize your Tremplin.io experience with our ‘Read to Earn’ program! For every article you read, earn points and access exclusive rewards. Sign up now and start earning benefits.
