Quick Insight: Michael Bhaskar on AI Can Improve Itself
One of the promises of artificial intelligence is that it can be so smart it can identify its shortcomings and avenues for improvement – and then act on that. The benefits of “recursive self-improvement” sits at the heart of this video from the Quick Insight series from Michael Bhaskar, the author 2013’s The Content Machine: Towards a Theory of Publishing from the Printing Press to the Digital Network and 2017’s Curation: The power of selection in a world of excess, as well as the current head of Microsoft’s Futures Team. His 2023 book, The Coming Wave, Technology, Power and the 21st Century’s Greatest Dilemma, written with Mustafa, was shortlisted for the Financial Times Business Book of the Year Award.
A transcript of his Quick Insight appears below the video.
Hello, my name is Michael Bhaskar. I lead the futures team at Microsoft AI where something like an in-house think tank of interdisciplinary researchers and just interesting people. We’re trying to figure out the future of AI, the future of society, and then use that understanding to hopefully change outcomes in the present and nudge them in a better direction.
What I want to talk about today is the idea that I think is at the heart of why AI is so important.
And you know, I think that there are kind of two views on AI. On the one hand, people sort of talk about it in these cosmic terms. They’re thinking about it as this sort of earth-shattering thing, the new fire, the biggest invention ever, this great technological revolution.
And then other people sort of say, hold on a minute, actually, isn’t AI just a bit like another normal technology? It does some things well, you know. Is it really bigger than the Internet or the printing press? It hardly works.
So why are there two different opposing views? And what might sort of nudge that normal technology view into the more grandiose view? And the answer is very simple. AI becomes this kind of insane huge technology at the point at which it can improve itself or even just help improve itself. And this is an idea of recursive self-improvement, or RSI.
And really it changes everything because if an AI can speed up or improve or do things better than any of the human researchers who are working on it, and then that AI can then do that again. That is the sort of the kernel of what I.J. Good, the statistician, called the “intelligence explosion.”
So this idea is really at the heart of everything about AI. To what extent is recursive self-improvement possible? If so, when is it going to happen? What might be the rate limiters on it? So, you know, fine, AI might be able to improve software, but can it really go and throw up a new semiconductor factory?
There are a lot of questions about it, but the basic fundamental is once we get to a point at which this is happening, things get interesting, things get crazy. It becomes very hard to predict just how far the AI will go. You know, will it get beyond human level intelligence?
Almost certainly. There’s no kind of natural law of the universe that human intelligence is the limit. And again, recursive self improvement is the key. Why it’s so important now is that we’re already seeing the first signs that this is happening.
So you know, OpenAI’s latest models, substantially built with AI Anthropic’s latest products, are substantially built with AI.
We’re in the kind of the foothills of this loop whereby AI is going to build AI and then improve AI. So, you know, when people talk about AGI and they talk about super intelligence and they talk about stuff that sounds really far out and really science fiction, actually, it’s just something really, really basic. It’s the point at which AI can accelerate in any way the process of building AI.
And the crazy thing here is I’m talking in early 2026. This process has happened, and that means I think we’re leaving the era of normal technology behind. So if there’s one idea that I think everyone should get on board with, it’s that.
