A hopeful view of large language models (LLM)

File this under hopeful naïveté, naïve idiocy, or pointlessly hopeless romanticism.

I was playing around with Gemini’s new image model, NanoBanana, and I was shocked at how good it was. That got me thinking about the present state of AI tools—by which I mean large language models (LLMs). People often use “AI” to mean magic, and no two people mean the same thing when they say it. But for this fleeting thought, AI = LLMs like ChatGPT, Claude, and Gemini that most of us have used.

AI is like a Rorschach test: no two people interpret the inkblot the same way. Similarly, views on AI diverge wildly. Some think LLMs are just statistical bullshit machines. Others think they’re essentially human brains in silico and revolutionary. I fall somewhere on the “these things are ridiculously good” side. I don’t think they’re nothing-burgers, and I don’t buy the argument that they’re just vomiting the collective mediocrity of humanity. That’s an extremist, nuance-free take.

We’ve reached the stage where a fair number of people actually use LLMs. Numbers are hard to come by, but maybe ~15% in the U.S., less in India—within just 2–3 years of availability. And because these are software tools, not hardware, adoption is easier. That’s not a trivial figure. Contrary to naysayers predicting that AI companies will burn billions and flame out spectacularly, I think LLM adoption will continue to diffuse across individuals and companies, steadily ticking upward.

What also surprised me is the explosion in open-source models, especially Chinese ones like DeepSeek and Alibaba. Benchmarks suggest they’re on par with the best proprietary models. There’s also a proliferation of smaller, specialist language models. And now, models run directly on phones. The fact that Google’s Edge Gallery can run a reasonably large model like Gemma locally is ridiculous.

All of this makes me think people are underreacting to LLMs. The collective knowledge of humanity—flawed though it may be—is now literally at your fingertips. Even if incremental improvement has slowed, today’s models already have phenomenal breadth.

One area that excites me most: education. Forget rich vs. developing countries. What about poor regions—rural India, parts of Africa—where education is dismal? Imagine a kid with little formal schooling but access to even a halfway-decent model on a smartphone. The potential is staggering. We could see geniuses emerge from places where systemic neglect has failed entire generations. Thanks to open-source, even poor countries could patch together workable models and distribute them widely.

Can LLMs produce more geniuses like Hannah?

Yet a paper posted on February 10 (opens a new tab) left the math world by turns stunned, delighted and ready to welcome a bold new talent into its midst. Its author was Hannah Cairo (opens a new tab), just 17 at the time. She had solved a 40-year-old mystery about how functions behave, called the Mizohata-Takeuchi conjecture.

“We were all shocked, absolutely. I don’t remember ever seeing anything like that,” said Itamar Oliveira (opens a new tab) of the University of Birmingham, who has spent the past two years trying to prove that the conjecture was true. In her paper, Cairo showed that it’s false. The result defies mathematicians’ usual intuitions about what functions can and cannot do.

So does Cairo herself, who found her way to a proof after years of homeschooling in isolation and an unorthodox path through the math world.

It feels like a golden age of learning. A “smart person” who knows a lot is now in your pocket. And yes, I lean toward the techno-optimist camp: we’re underreacting. Critics cite hallucinations, errors, unreliability. But compare that to human teachers: they’re not 100% accurate either, and in many places, education quality is horrendous. Honestly, DeepSeek today is better than any teacher I had in school or professor in my MBA. That’s not hyperbole—the breadth and depth are insane.

The ripple effects are underappreciated. I tried Googling for work on this and found only a few small studies in Africa and India. They highlight challenges: teachers’ lack of technical literacy, poor access to internet or devices, barriers to adoption. Those are real. But even at the bare minimum—through smartphones, basic internet, cheap computers—these models are already maximal.

And I can’t stop thinking about what the future holds.