Austin Tripp's personal website
Preamble This post was the result of me putting Claude Fable to the test in early July when it was made available to the public again and was included in my personal Claude Pro subscription. I gave it a question and some hypotheses about a paper that had been on my radar for a while, and it ran experiments and produced the report below. I reviewed it with roughly the same thoroughness that I would…
I recently read the paper 1 “Grand Challenges for Predictive Modeling in Small Molecule Drug Discovery”, which aims to survey all the challenges in the field I work in. About half of the paper was what I expected, the other half was either new or something I was exposed to but didn’t really think about. In this post I’ll briefly outline what I learned. Things I learned New aspects of binding While…
I recently found myself wanting to use the Vecchia GP approximation for a problem at work where low-rank approximation methods were a poor fit. I had vaguely heard about this approximation from a conversation with Marcus Noack , and I understood that it was a sparse approximation to the GP covariance. I always assumed this meant rounding near-zero entries in a GP covariance matrix to be exactly…
I just wrote an article How AI Experts Can Drive Better Collaboration in Drug Discovery on Valence’s substack, reflecting on some experiences I’ve had at Valence/Recursion. Please read it on substack! Essentially I argue that, on the margin, AI experts should be more assertive when formulating the problem, since domain experts don’t always get it right. Even though they know the problem well,…
A colleague recently introduced me to Quarto and I can’t believe I didn’t know about it sooner! Quarto is a document compiler- you write plain text files in .qmd and compile them to many output formats- pdf, html, slides, etc. For the kinds of writing I do, it feels like a strict upgrade from a set of tools I’ve already been using: Markdown : quarto’s syntax is a superset of markdown. 1 Your…
After switching my website’s build from Jekyll to Nikola in 2024 (more details here ) I’ve decided to switch again to Quarto . Unlike Nikola or Hugo, Quarto is designed for scientific publishing, so equations, cross-references, and code cells are all first-class citizens. See my parallel blog post about Quarto for more about why I like it. Technically speaking, this was much more challenging than…
I’ve been a somewhat late adopter to coding agents: until ~1 month ago I was mainly using cursor and asking it to suggest local edits, or occasionally bigger refactors. In the past 1 month I’ve been trying to change that. In my work this has mostly taken the form of giving coding agents bigger open-ended tasks (e.g. “plan out this big refactor”). However, outside of work I’ve also tried just doing…
Don’t write your paper for somebody to read it, write for somebody to skim it. This is my version of a rule I’ve heard many times before in different forms. The justification for this rule is that most papers are very niche, and few people that will care enough to read every little detail. However, a greater number of people will want to read the paper at a higher level to understand the key idea,…
I recently ran an LLM deep research query about large matrices of Tanimoto coefficients, 1 and one of its references was the paper “Bandwidth adjustable Tanimoto kernel: a smooth alternative to the Gaussian kernel” ( https://doi.org/10.1007/s41237-026-00296-7 ). Skimming this paper, I found a citation whose title and publication venue matched my own paper, but with a completely different author…
This is a re-post from my book summaries repo , which is my main home for book summaries. I’m posting here because this book is more work-relevant than usual. Authors: Donald Kirsch and Ogi Ogas Link: https://www.amazon.com/Drug-Hunters-Improbable-Discover-Medicines-ebook/dp/B01HDVCRY0/ This book is the authors’ attempt to explain what the job of “drug hunter” is, specifically to explain how hard…
I think the Bayesian optimization (BO) research community needs to change its evaluation practices. In a previous post I explained how I think users should generally bring their own model to BO and try to calibrate it to their beliefs/expectations. In this post, I explain why existing evaluation practices are not well-suited to this stance on BO, and outline a different approach to BO evaluation…
In a recent blog post ( link ) I described my “model-centric” view of Bayesian optimization (BO), essentially arguing that the model is the most important component of BO and BO users (and researchers) should do more to get it right. Under the assumption that the reader broadly agrees with the content of that post (or at least thinks this view of BO is one of several valid views), here I want to…
The most common kind of Gaussian process (GP) model is: Essentially, the function of interest is sampled from a Gaussian process, and observations are noisy versions of the function values with Gaussian noise of standard deviation . Model fitting is done either by maximum likelihood or MAP (maximum a posteriori), with the parameters fitted being: Parameters of the mean function and kernel function…
Cross-posting my summary of the book “Talent: How to Identify Energizers, Creatives, and Winners Around the World” from my GitHub: link . This is a good book, I appreciated the heterodox advice about how to run interviews and spot “undervalued” candidates, even if only some of it is backed by data (this is a field with generally poor data). Citation BibTeX citation: @online{tripp2026, author =…
I think the Bayesian optimization (BO) community is vastly underselling one of BO’s most practically appealing features: modularity . A “modular” algorithm is one composed of individual parts which are largely exchangeable. I think BO is extremely modular, and this is something users want in practice, but BO papers rarely describe it in this way. This post outlines these points in more detail and…
A lot of Bayesian optimization (BO) papers include experiments on “toy” functions. Examples of such toy functions are Ackley, Rosenbrock, Hartmann, and Branin. I basically ignore these experiments when I read papers and skip to the next section. When I review papers (and therefore cannot simply skip sections), I find my eyes glazing over and feel bored: I don’t care about these functions, don’t…
Bayesian decision theory is one of the best justifications for BO- particularly for myopic acquisition functions like expected improvement. However, these acquisition functions are only “optimal” if one’s utility function is (the identity function). Have BO researchers (and BO users) basically forgotten to swap this out for “real” utility functions in practice? In this post I argue that we have…
A lot of conversations at NeurIPS this year made me think that I view the role of the surrogate model in Bayesian optimization a bit differently than many other researchers in the field, and this profoundly impacts my view of many other aspects of BO. Therefore, the purpose of this post is to explain my view and contrast it with what I believe is the more mainstream view. Disclaimer: this post is…
Last year I made a bunch of predictions about ML , and since 2025 is over it’s time to grade myself and repeat this exercise. This will be a lower effort version than last year’s: partly because AI is growing and I don’t think I’m doing a good job keeping track of it, partly because a lot of progress is happening behind closed doors of frontier labs (so less info is available to me as an…
Happy 2026! Here is my self-assessment on last year’s goals and my goals for this year (2026). Review of 2025 goals Good news: all goals were met 🙌. In detail: Do amazing work at Valence. Assessment : pass Comment : overall I’m happy with the work I did. More writing. Want to average 1 blog post every 2 weeks (= total of 26 blog Assessment : exceeded Comment : 35 total posts 🥳 Use LLMs more for…