Every few weeks, another AI model does something that would have sounded impossible just a year or two ago.
One model beats some of the world’s brightest young mathematicians. Another helps conduct scientific research. And some are writing software that experienced engineers say would have taken them days or even weeks to finish.
There are even AI systems today that can work on their own for hours before asking a human for help.
Now, I’m not pointing to any one of these breakthroughs on their own and declaring that we’ve reached artificial general intelligence (AGI).
But I think it’s time to stop looking at these breakthroughs in isolation.
Because by doing so, it paints a clearer picture of just how close we are to AGI.
Beyond Chatbots
Back in February, I argued that the core ingredients of AGI were finally starting to come together.
First, AI learned how to answer questions. Then it learned how to reason through complicated problems. More recently, it’s started working on its own for much longer without needing constant supervision.
At the time, I suggested these things were all building blocks of general intelligence.
Looking back over the past six months, I believe we’ve added several more.
Take mathematics.
Last summer, an advanced version of Google’s Gemini Deep Think earned a gold medal score at the International Mathematical Olympiad, widely considered the world’s most difficult math competition for high school students.
Only about 8% of contestants earn a gold medal in a typical year.
Google’s AI solved five of the competition’s six problems, earning 35 out of a possible 42 points after its work was graded by the same judges who score human competitors.
It was an extraordinary achievement.
But Google says newer versions of Deep Think are now helping mathematicians and scientists tackle real research problems in mathematics, physics and computer science.
In one internal benchmark involving advanced mathematical proofs, the system reached roughly 90% accuracy.
To me, that’s a big piece of the AGI puzzle.
Solving problems with known answers is one thing. But helping solve problems that nobody has answered yet is something much closer to general intelligence.
Anthropic recently revealed a similar milestone.
Researchers gave teams of Claude AI agents an open-ended machine-learning research problem with no answer key.
The agents had to come up with their own ideas, write code, run experiments, analyze the results and keep improving their approach until they found a solution that worked.
According to Anthropic, the best AI team outperformed experienced human researchers working on the same problem.
Again, general intelligence is about tackling problems you’ve never encountered before, figuring out what works and adapting as you go.
That’s exactly what these AI agents did.
And we’re seeing the same thing happen in software engineering.
Anthropic recently asked Claude to build a brand-new C compiler capable of compiling the Linux kernel, one of the most complex open-source software projects ever created.
Over the course of two weeks, Claude worked through nearly 2,000 coding sessions. It processed roughly 2 billion input tokens, generated 140 million output tokens and ultimately wrote about 100,000 lines of code.
And it worked.
Claude successfully built software that could run one of the world’s most complex operating systems on several different types of computer processors.
Even more remarkable, the entire experiment cost less than $20,000 in API usage.
That’s a shockingly small price tag for a project of this scale. It could easily require hundreds of thousands of dollars in engineering salaries to accomplish a project of similar size and scope.
To me, this represents another step toward AGI.
To think like a human, you need to be able to tackle large, unfamiliar projects, stay focused over long periods of time and solve problems along the way.
Speaking of which, AI has been getting much better at staying on task.
As I wrote about earlier this year, the research group METR measures how long an AI can work on a real problem before getting stuck or needing human help.
Image: metr.org
And as you can see, the trend is remarkable.
According to METER, the amount of work frontier AI models can complete on their own has been roughly doubling every seven months.
Anthropic says it’s seeing the same thing.
After studying millions of Claude Code sessions, the company found that some of the longest autonomous coding sessions nearly doubled in just three months, growing from less than 25 minutes to more than 45 minutes before the AI asked for help.
If that trend continues, those forty-five minutes will eventually become four hours. Then a full workday.
That’s why I think it’s a mistake to think of artificial general intelligence as a finish line.
There won’t be a day when somebody suddenly announces that we’ve officially reached AGI.
Instead, we’ll just keep seeing more and more pieces fall into place.
Here’s My Take
In my view, the breakthroughs we’ve witnessed over the past six months suggest we’re getting closer to AGI.
Intelligence isn’t just about knowing the answer. It’s about using that knowledge to accomplish something useful.
And that’s exactly what today’s AI is becoming surprisingly good at.
But that raises an important question.
If AI has become this capable, why aren’t businesses seeing bigger results?
We’ll explore that question further in tomorrow’s Daily Disruptor.
Regards,
Ian King
Chief Strategist, Banyan Hill Publishing
Editor’s Note: We’d love to hear from you!
If you want to share your thoughts or suggestions about the Daily Disruptor, or if there are any specific topics you’d like us to cover, just send an email to dailydisruptor@banyanhill.com.
Don’t worry, we won’t reveal your full name in the event we publish a response. So feel free to comment away!
