It’s pretty amazing how far AI has come. Just a few years ago, LLMs became widespread. Then came RAG, tools, and agents, and people quickly began to incorporate AI into their work. It’s a wonder how much sophisticated infrastructure pops up once there is funding for it.

Today, OpenAI published this article on advances to math and TCS achieved with their model Astra. This follows a partial resolution of the Jacobian conjecture with the model Fable 5.

Solving mathematical problems often leads to scientific achievement, so these posts went viral. They’ve attracted a lot of attention, from mockery, to confusion, to amazement, and everything in between.

People are afraid that AI will harm their stability, while others are excited by the opportunity to change the world.

Trust and Control

For me, one of the biggest fears that came with AI was admittedly the lack of control and trust.

In the past, well-trained computer science and mathematics professionals designed code and mathenatics, and there was a feeling that they could be trusted. This was because they wrote every line of the code and every line of mathematics meticulously, careful to not make mistakes.

It’s much more difficult to trust LLMs. When an LLM writes code, is it “meticulous”? Language models are prone hallucinations: errors or mistakes. Of course, humans are too, but at least we can take responsibility for them. And the best trained humans with years of experience rarely make mistakes.

So trust and a lack of true oversight and control over LLMs became the first major concern about these language models.

But to be honest, as the days go by, LLMs are getting better and better, less and less opaque, and can often even answer questions more accurately than humans. After all, they’re designed to be indistinguishable from humans when it comes to language.

Boundaries

The truth is, boundaries with LLMs and AI become weaker and weaker every day, which is what makes people afraid. AI is getting so good that people want to use it for everything: even important and life-changing things.

Today, AI is being used for science, mathematics, engineering, and anything else one could possibly imagine (even art?!). The human ego that we are the unique species capable of understanding, intelligence, and advancement is crumbling.

And to be honest, maybe it should.

For our own stability, should we prevent AI from developing sufficiently to advance the standard of living of the world? To preserve one notion of control, should we abandon any ideas which require adaptation?

Using AI to Enact Positive Change

Once we are ready to accept that artificial intelligence can surpass human intelligence simply due to our own biological energy throughput limitations, I think the conclusion is that we must focus more on ensuring that AI will truly align with human needs.

I think that AI probably has the potential to make great change possible, as long as adequate infrastructure is build to ensure it’s done safely. And that’s my goal.

My concern with existing AI infrastructure is that it focuses on speed. Silicon valley startups flaunt their token usages, and sacrifice security measures for fast product development and testing.

It might work for some things, but for science and research, I think we’ll need to build better tools that support human oversight while making sure humans don’t become the bottleneck. In the next few weeks, I kind of want to work on tools that will help me use LLMs effectively: developing materials for students I’m tutoring, writing code with complete oversight rather than in the autonomous fashion that people currently use, automatically formulating and checking mathematical theories I develop for science, and creating strict contracts that solidify the expectations for LLM use in my own life.