“I put a strawberry on the table…”
For example : “I put a strawberry on the table, then an upside-down cup on top. I put the cup in the microwave; what happens to the strawberry?” Asking this question to Microsoft Copilot, Google Gemini, or even ChatGPT in its recent version 40, plunges them into abysses of perplexity or absurdity, such as: “The strawberry is enclosed under the cup and will undergo a mini steam cooking”. We understand that these limitations are severe: how can we entrust a medical diagnosis, a legal recommendation to an “intelligence” whose reasoning is so distorted?
The new version “o1” of ChatGPT (also called Strawberry) largely addresses these types of challenges and inaugurates a new type of architecture, which takes advantage of already known elements. We will use a multi-agent system: where a first draft of a response by an AI will be challenged by a second AI, which will order it to resume reasoning step by step.
The “chip war” has only one winner, and it’s American
Furthermore, we will use the process of Chain-of-thought (chain of thought) that had been described by Google researchers in early 2022. We will in a way resume a path of reasoning in very small steps. This allows us to correct the absurdities and discover that the strawberry did not arrive in the microwave…
These new LLMs are much slower but show a form of “reasoning”, going up to a PhD researcher level. There is no doubt that these architectures will expand and change the game again.
