10/11/2026

What if we could get any answer to any question, instantly?

“There are only two tragedies in life. One is not getting what you want. The other is getting it.” This is a quote from an Oscar Wilde play called Lady Windermere’s Fan. What if we could get any answer to any question we had, instantly? I would certainly feel more capable. I would feel like I have the world at my fingertips, and that I can solve problems with ease. Most of you might feel the same way, that this would be an incredible human invention. What if it would be the most harmful step humanity could take? This quote describes the paradoxical nature of desire, that regardless of the outcome, we will not be fulfilled if we measure it based on what we desire. Fulfillment, therefore, cannot be measured by or described through desire. Many of the foremost inventors and leaders of the “AI frontier” of our time, though, are leading with the underlying implication that fulfillment is based on desire, and more specifically, achieving what we desire. Namingly, to answer any question that we can think of, and find a solution to any problem that we can come up with.

Science has always been about solving the problems that our complex universe presents to us, with the underlying goal of understanding the world around us. But what if it was never about solving the actual problem? In 1736, Leonhard Euler solved the problem of the seven bridges of Königsberg. This was the question of whether one can cross all seven bridges of Königsberg, a small town in Russia, exactly once. Euler proved that it was mathematically impossible to do so without walking on a previously taken path or walking over a bridge twice. Taking this question at face-value, it is arguably the most useless question toward science and mathematics. We could ask ChatGPT or Claude this question, and get a yes/no answer in milliseconds. So what? The point is that in the process of analyzing this problem and looking for a solution, Euler provided a foundation for graph theory and graph topology that we now use every single day in our lives (Google Maps, LLMs, Google search, and much more…). If Euler pulled out his favorite LLM and asked whether it was possible to cross over each bridge only once, and got “No”, then graph theory and topology might not have ever been invented, let alone by him. It was never about the question, though.

Recently, OpenAI solved the Navier-Stokes Millennium Prize Problem. While the Navier-Stokes equations are widely used in various fields like aerospace, the Navier-Stokes Millennium Prize Problem itself is not very useful to solve. The reason it was chosen as a Millennium Prize Problem is because, in exploring the question, people are expected to do research and discovery that is meant to further the fields of mathematics and science. Instead, we’ve gone and asked 10,000 AI agents, and gotten an answer back in 88 hours. Many leading AI researchers and scientists might say “wow, but lets do better” and try to reduce the number of AI agents and the amount of time it takes for them to solve these kinds of questions. But that is completely missing the point!

Terrence Tao, one of the world’s greatest contemporary mathematicians and a professor at UCLA, the second most prestigious public university in the world next to UC Berkeley, stated recently in a video by Big Think that if we use AI to replicate the graduate student-level papers and research that it now can solve/answer, and replace grad students with AI agents, then we won’t get the next generation of students and thinkers. That we might even stagnate as a society. He says:

“We will need to really have a much more open discussion about what basic science is and what it's useful for, and why it's still important to have curiosity-driven research, why we still need a community of humans to explore things sometimes slowly and in ways that are not as efficient as the latest models of AIs.”

We live in a kind of world where it is intrinsic to human nature to continue to want to solve problems, achieve goals, and optimize in the process. But these problems and goals will never cease to arise. No matter how many questions about the universe we answer, there will always be more that come up. This is a fact that comes with the nature of human thinking and curiosity. So, should we be trying to answer more questions, at a faster rate? (Recursive Superintelligence, Ineffable Intelligence, Periodic Labs) Or should we slow down, continue to explore what we are curious about, and enjoy it, instead of racing toward a world where we no longer have the option to do so?

Lots of inspiration in my thinking has originated from Solving Math Problems Was Never the Point and The paradox at the heart of AI and science | Terence Tao. I highly recommend watching these videos if you have the time.

08/05/2026

Thoughts on Automating Science

I spent the weekend at the Berkeley RDI Agentic AI Summit, which was an incredible experience that I am grateful for, and one session has stayed with me longer than the rest.

Richard Socher, Co-Founder and CEO of Recursive Superintelligence, talked about automating science. My first reaction was resistance. Penicillin came from a contaminated plate that Fleming happened not to throw away, and a lot of the history of discovery looks like little accidents or byproducts of experiments that flourished. If the anomalies are where the findings live then it seems to me that speeding everything up would cause them to disappear!

The more I sat with that, the less sure I was. I learned that automated sky surveys are the most serendipity rich instruments we have ever built because they scan everything in the sky rather than pointing where someone thought it would be interesting. Genome sequencing looks similar too. In both cases, the volume that is achieved by a computer/AI seems to have widened the aperture instead of narrowing it.

So, maybe the concern isn’t speed. It might be what function we are achieving. A closed loop system that is asked to maximize one number will tend to discard an instance that did something strange but off metric, because filtering is a critical byproduct of what optimization does. The anomaly does get seen sometimes, but gets scored low and dropped.

I also learned about Michael Polanyi, a polymath who coined the phrase “tacit knowledge”. He said that “we can know more than we tell.” Tacit knowledge is what a scientist accumulates that doesn’t make it into writing or the final result. In order to succeed we must fail, and through various failures the scientist accumulates an understanding that is not describable or reproducable otherwise. A similar phenomenon occurs in the creative process. If that is real, then the slow part of science is doing the actual work.

Now, the part I have not been able to resolve. A system that runs millions of experiments, then, is accumulating something in its weights that nobody wrote down and nobody can articulate. That could be considered a form of tacit knowledge, and it is not bounded by a career or a lifetime as it is with humans.

The question I left the summit with is whether we as a humanity get to a point where different models hold genuinely different tacit “knowledge”, shaped by whatever each one happened to try and fail at. Can models build tacit “knowledge”? How would we see this, given the current limitations on transformers and the attention model?

How do us as humans maintain one, whole consciousness and knowledge base, with our limited attention? Are we just a unique collection of experiences, information, and tacit knowledge? Does that make answering the above questions easier?

I would love to hear your thoughts.

Final note: I worked through most of this thinking with Claude Opus 5, which I think adds to the complexity of these questions.

Originally posted on LinkedIn