I resisted this for a long time. A poorly executed blog is worse than no blog, in my opinion.

However, a few changes in my life have made the idea more palatable recently:

  • AI has made it much easier to build and maintain a website, especially when I’m not sitting at a computer.
  • Time for free inquiry is precious now that I have a job; I find my mind wandering toward ever more ambitious questions when it can.
  • I feel like I finally have interesting things to say online, especially as the discourse surrounding AI research has largely narrowed in recent years.

What I mean by the last point is that it has become increasingly difficult to disseminate research on fundamental topics decoupled from LLM applications. By no means am I opposed to LLMs—far from it. I am repeatedly stunned by what the models can accomplish, and I get great satisfaction from building my own little agents to automate the routine aspects of my work. And ultimately, the LLM revolution is what has enabled me to start this blog.

What I do advocate for in research is radical open-mindedness (Ray Dalio’s term). Too often, researchers dismiss ideas simply because they are unpopular or old. How foolish and shortsighted this is, in the very field that once rejected deep learning. Perhaps the assumption is that more people would be working in a particular area if it were promising? Yet to accept this perspective means to concede that others’ research judgment is superior to your own, and to relinquish your agency as a scientist.

In reality, the academic terrain is far richer than that. The totality of human knowledge is like a fractal; there are infinitely many hidden frontiers lost to time and ignorance. During my PhD, I was amazed to discover fundamental ideas in reinforcement learning from the ’80s and ’90s that stood unexamined, untested, or unchallenged. Science only meaningfully advances when people are bold enough to forge ahead into the unknown, but the unknown is not necessarily a single point in time or space.

I recall the concluding paragraph of my PhD thesis:

… [I]t is important not to get stuck in old ways of thinking. Machine learning is a fast-moving field; algorithms that are convenient for today’s hardware or perform well on today’s benchmarks lead us into a tempting local optimum which ultimately may not stand the test of time. Above all else, general principles of learning and intelligence must be pursued—independently of the modern specifics of function approximation and backpropagation. Continued progress toward the development of real AI requires us to constantly question conventions and ingrained views—especially our own.

While I did intend this mainly as a criticism of the pervasive groupthink in AI research today, it cuts both ways. Many remain skeptical of LLMs, yet their incredible capabilities are undeniable at this point. The Bitter Lesson rings true in that how we develop capable AI systems may not always align with what we personally find satisfying, but the role of a truth seeker is not self-gratification.

Instead of picking ideological sides, let us stay radically open-minded. Let us combine the best elements of our known algorithms in our common pursuit of real AI.1 We still have a lot of work to do.

To conclude this post, let me properly set expectations for the new blog. Time is always scarce. If I can write one or two interesting posts per year, this project will be a success.

I also tend to be a contrarian. I am inspired to write only when I feel I have something unique or unusual to add to the conversation. Or maybe even something controversial, in order to provoke thought and discussion.

Some potential topics:

  • My personal philosophies on AI, research, and software engineering.
  • My experience in industry vs. academia.
  • Accessible breakdowns of research papers—especially my own.
  • Deep-dive tutorials on niche research topics: e.g., TD learning, multistep returns, eligibility traces, value-function operators.

Of course, the most important thing to me is writing content that people find useful. If you have requests, I would love to hear them! I can’t promise anything, but I am happy to take suggestions into consideration. Message me on LinkedIn or send me an email.

  1. Elaborating on what I mean by “real” AI could be an interesting future post. ↩