How close are current AI models to AGI and self-improvement to superintelligence?

Are we seeing the first steps towards AI superintelligence?

By Deni Ellis Bechard edited by Eric Sullivan

The Matrix, The terminator– much of our science fiction is built around the dangers of super-intelligent artificial intelligence: a system that surpasses the best humans in almost all cognitive domains. OpenAI CEO Sam Altman and Meta CEO Mark Zuckerberg predict that we will realize such AI in the coming years. Still, machines like the ones depicted in those movies as warring against humanity would have to be far more advanced than ChatGPT, not to mention better able to create Excel spreadsheets than Microsoft Copilot. So how can anyone think we are even close to artificial superintelligence?

One answer goes back to 1965, when statistician Irving John Good introduced the idea of ​​an “ultra-intelligent machine.” He wrote that a computer would quickly improve itself once it became sufficiently advanced. If this seems far-fetched, consider how AlphaGo Zero – an AI system developed at DeepMind in 2017 to play the ancient Chinese board game Go – was built. Without using data from human games, AlphaGo Zero played itself millions of times and in a few days achieved an improvement that would have taken a human a lifetime to achieve and made it possible to beat the previous versions of AlphaGo that had already defeated the best human players in the world. Good’s idea was that any system intelligent enough to rewrite itself would create iterations of itself, each smarter than the last and even more capable of improvement, causing an “intelligence explosion.”

So the question is how close we are to that first system capable of autonomous self-improvement. While the runaway systems Good described aren’t here yet, self-improving computers are—at least in limited domains. AI already runs code on its own. OpenAI’s Codex and Anthropic’s Claude Code can run independently for an hour or more writing new code or updating existing code. Using Codex, I recently entered a prompt into my phone while out for a walk, and it created a working website before I got home. In the hands of experienced programmers, such systems can do considerably more, from reorganizing large codebases to sketching out entirely new ways to build the software in the first place.

About supporting science journalism

If you like this article, please consider supporting our award-winning journalism by subscribe. By purchasing a subscription, you help shape the future of impactful stories about the discoveries and ideas shaping our world today.

So why hasn’t a model powering ChatGPT quietly coded itself into ultra-intelligence? The problem lies in the sentence above: “in the hands of experienced coders.” Despite AI’s impressive improvements, our current systems still rely on humans to set goals, design experiments, and decide which changes count as real progress. They are not yet able to evolve independently in a robust manner, which makes some talk of impending superintelligence seem blown out of proportion – unless, of course, current AI systems are closer than they appear to being able to improve themselves in ever-increasing proportions of their capabilities.

One area where they already look superhuman is the amount of information they can absorb and manipulate. The most advanced models are trained on far more text than any human could read in a lifetime – from poetry to history to the sciences. They can also keep track of much longer stretches of text as they work. With commercially available systems like ChatGPT and Gemini, I can already upload a stack of books and let the AI ​​synthesize and critique them in a way that would take a human week. That doesn’t mean the result is always correct or insightful, but it does mean that a system like this can essentially read its own documentation, logs, and code and propose changes at a speed and scale that no engineering team can match.

However, it is the reasoning where these systems lag behind, although in some specific areas that is no longer the case. DeepMind’s AlphaDev and related systems have already found new, more efficient algorithms for tasks like sorting, results that are now used in real code and go beyond simple statistical mimicry. Other models excel at formal math and college-level science questions, which resist simple pattern matching. We can debate the value of a particular benchmark – and researchers are doing just that – but there is no doubt that some AI systems are capable of discovering solutions that humans had not previously found.

If the systems already have these capabilities, what is the missing piece? One answer is artificial general intelligence (AGI), the kind of dynamic, flexible reasoning that allows people to learn from one field and apply it to others. As I’ve written before, we continue to change our definitions of AGI as machines master new skills. But for the superintelligence question, it’s not about the label we put on it; what matters is whether a system can use its skills to reliably redesign and upgrade itself.

And this brings us back to Good’s “intelligence explosion.” If we build systems with that kind of flexible, human-like reasoning in many areas, what will separate this from superintelligence? Advanced models have already been trained in more science and literature than any human, have much larger working memories, and exhibit extraordinary reasoning skills in limited domains. Once that missing piece of flexible reasoning is in place, and once we allow such systems to apply those skills to their own code, data and training processes, could the leap to fully superhuman performance be shorter than we imagine?

Not everyone agrees. Some researchers believe that we have not yet fundamentally understood intelligence and that it will take longer than expected to develop this missing piece. Others say that AGI will be realized within a few years, leading to further advances far beyond human capabilities. In 2024, Altman publicly suggested that superintelligence could arrive “within a few thousand days.”

If this sounds too much like science fiction, consider that AI companies regularly run security tests on their systems to ensure they don’t fall into a runaway self-improvement loop. METR, an independent AI safety group, evaluates models based on how long they can reliably perform a complex task before failing. Last November, GPT-5.1-Codex-Max tests lasted approximately two hours and 42 minutes. This is a huge leap from the few minutes where GPT-4 performs on the same metric, but it’s not the situation Good describes.

Anthropic is conducting similar tests on its AI systems. “To be clear, we are not yet in the ‘self-improvement of AI,'” company co-founder and policy head Jack Clark wrote in October, “but we are in the ‘AI that improves parts of the next AI, with increasing autonomy.'” phase

When AGI is achieved, and we add human-level judgment to an immense information base, enormous working memory, and extraordinary speed, Good’s idea of ​​rapid self-improvement begins to look less like science fiction. The real question is whether we will stop at “merely human” – or risk going too far.

It’s time to stand up for science

If you liked this article, I would like to ask for your support. Scientific American has served as an advocate for science and industry for 180 years, and today is perhaps the most critical moment in that two-century history.

I am one Scientific American subscriber since I was 12 years old, and it has helped shape the way I view the world. SciAm always educates and delights me, and instills a sense of awe at our vast, beautiful universe. I hope it does that for you too.

If you subscribe to Scientific Americanhelps ensure our reporting focuses on meaningful research and discovery; that we have the resources to report on the decisions that threaten US laboratories; and that we support both aspiring and working scientists at a time when the value of science itself is too often unrecognized.

In return you will receive essential news, engaging podcasts, brilliant infographics, newsletters not to be missed, videos not to be missed, challenging games and the best writing and reporting in the world. That’s even possible Give someone a subscription as a gift.

There has never been a more important time for us to stand up and show why science matters. I hope you will support us in that mission.