How to Achieve Artificial Superintelligence
Be nice when you talk to the AI.
I took only half-seriously the reports that an AI model (LaMDA) become sentient. But the release of the o1 model has moved us much closer to superintelligence. Only a few steps remain to achieve it.
Early language models were happy generating new tokens and didn’t bother much about thinking and reasoning. One could argue that this is what most of us expect of the generative AI. The drive to make AI good at reasoning was initially coming from researchers who were not satisfied with the dumb hallucinating language models that were spitting mindlessly tons of words on you. However, there are practical applications for reasoning AI. The ability to generate computer code is well-known, but programmers often must correct mistakes and hallucinations even in the code. Despite the general acceptance of AI-aided programming, the language models still could not work on difficult problems that required a lot of reasoning.
The model o1 addressed it by forcing the language model to slow down and think. This is done by adding special reasoning tokens. Solving a complex problem, the model switches into the reasoning mode, when it generates new tokens not to send them to the user, but to go through the logical steps. These reasoning tokens are used then to generate a final output.
