About Me
- Currently working on an AI startup.
- Before that I led a research team at a quant fund.
- Undergrad @ MIT.
- Long ago I did competitive math/physics (MOP, US IPhO team).
Feel free to contact me, I'm friendly :).
AI
I am interested in how AI systems can reason
more reliably (one of the main issues with LLMs).
I have some notes on the more classical side of AI/automated reasoning.
Some of my biases
- Compared to most AI researchers, I'm not very bullish on LLMs,
even with enhancements like CoT. I think (as of mid 2024) we
have already seen most of the use cases we are going to get out
of them—generally (1) things like search and summarization
where a high error rate is tolerable, (2) domains that are
highly rote (e.g. standardized tests, boilerplate code, front
line customer support), or (3) things that were already doable
with ML but are now easier because they can be achieved just by
prompting. But I'm only ~80% confident of the above, and even
if I'm right, there could still be other architectures inspired
by or built on LLMs that might fare better.
- I think GOFAI—things like logic and heuristic
search—is underestimated. That body of scholarship always
had its problems, especially around knowledge acquisition and
representation of uncertainty, but there was also something
right about it: it is interpretable and it directly captures
desirable invariants (e.g. in formal logic you always have modus
ponens available, as you should; by contrast an LLM might
neglect to apply, or misapply, modus ponens based on any number
of inscrutable contextual factors).
- I think psychology and linguistics are also underestimated
sources of inspiration for AI.
- I think often in AI we try to run before we can walk—that
is, to attack complex problems when we can't even solve simple
ones. It's no mystery why: a 50% solution to a real business
problem can be a lot more economically valuable than a 100%
solution to a toy problem. But the latter may be what is needed
to advance scientifically. For this reason I think projects
like the ARC Prize are valuable.
Random info
From Boston, now in SoCal. In my free time I dabble in various things
such as ultimate, pickleball, piano, and Japanese. I've written some
notes on math contests (from long
ago when I taught that material), accounting, and motivation/procrastination.
