Mustafa Suleyman, Microsoft's CEO of AI, believes AI models will act mostly autonomously within two years. However, he says it will take two more model generations and up to 100 times more computing power for AI to work consistently.
In a podcast with investor Seth Rosenberg, Suleyman discusses when to expect autonomous AI models. He says AI agents could soon act independently in narrow use cases without constant user oversight.
But truly reliable systems are still a long way off. Models need to be in a place where each output must exactly match a function call. Current models still have some variance in the correct solutions for most answers.
Today's 80 percent accuracy isn't enough for reliable action in new applications, Suleyman says. End users need 99 percent accuracy to maintain trust. Suleyman emphasizes that this is not about emerging capabilities, but about a more precise mapping between prompt and output.
He estimates models will need 100 times more computing power - at GPT-6 level, likely OpenAI's model after GPT-4 and GPT-5. This may take another two years, Suleyman says.
Until then, he sees the best prospects in areas where some inaccuracy is acceptable, such as legal research with multiple reasonable answers, while areas such as medicine, which require life-critical precision, are still too demanding.
Complete autonomy, where an AI system makes its own plans and procures its own resources, is dangerous, Suleyman says, and needs to be regulated.
Quality data is the key to success
Suleyman believes that model size is becoming less important to AI success. He points to Microsoft's open-source Phi 3 model, which is 100 times smaller than top models in terms of inference effort.
It may not be as good, but it matches or beats GPT-3.5, Suleyman says, noting that data quality, not the number of parameters, is becoming critical.
OpenAI CEO Sam Altman recently said that the future of AI training is about learning more from less data, which is the same message.
This creates opportunities for startups that may be able to succeed with smaller, well-trained models. Suleyman advises young companies to focus on collecting high-quality training data for their applications.
This includes training human experts to provide proper feedback for model training. Suleyman says his AI company Inflection, somewhat acquired by Microsoft, used strict selection for "AI teachers" when developing the Pi chatbot.
Suleyman predicts personalized AI assistants will accompany users throughout their digital lives, remembering everything they do and say while proactively making suggestions.
