Who would have thought that the game of go would play a role in the development of AI?
The CEO of DeepMind, an artificial intelligence laboratory belonging to Google, announced that it had brought out a technology from the boxes to improve an AI with characteristics comparable to those of ChatGPT. And he is sure, once developed, the project could well exceed the capabilities of its rival from OpenAI.
AlphaGo’s story
In 2016, when such a performance was still impressive, an artificial intelligence program called AlphaGo caused a sensation by beating a go game champion on its field. If AI was still far from what we know today, this feat was made possible by two systems invented by DeepMind: reinforcement learning and tree search. The first allows you to solve hypothetical problems by testing each of the solutions, while the second, for its part, allows you to remember past decisions and define those that are possible.
It was these systems that, at least partially, motivated Google to acquire DeepMind in 2014, as the company struggled to get computers to master video games. And, according to Demis Hassabis, the boss and founder of DeepMind, they are also the ones who could allow Google to take the lead in the field of AI.
What are Gemini’s promises?
Hassabis, after teasing the development of such a tool last month at a Google developer conference, said a bit more about what to expect from Gemini, the name of its large language model using AlphaGo advances. It should not be fundamentally different in nature from GPT-4, since it is trained with texts and can be used in comparable situations. But the real difference, namely the lessons that have been learned from the AlphaGo experience, should allow it to be even more versatile.
Thus, for Hassabis, these technologies could give Gemini the ability to plan actions and solve complex problems, beyond the simple generation of textual answers. Where GPT-4 becomes more relevant by taking into account feedback from researchers, reinforcement learning could automate this process. At this point, it might be important to clarify that Hassabis was one of the signatories of the communiqué calling for the development of AI to be considered a risk for humanity, to be classified on the same level as a pandemic or the nuclear war.
Gemini is still in development and could take months to be available to the general public. But, proof that Google really believes in the project, DeepMind works with another laboratory dedicated to AI at Google, Brain. And while it’s possible the technology will catch up with ChatGPT, it’s also not certain that OpenAI will sit idly by waiting for that to happen.
Source : Wired
4