It was an honor to speak at the ARC Prize Research Summit and share my views of the challenge of measuring and assessing open-ended systems. Thanks to @arcprize both for inviting me and for sharing the full video of my talk below!
Language models can learn new tasks from just a few examples in their prompt, with no retraining. This is called in-context learning (ICL). Now genome language models can do it too.
With a handful of examples, Evo 2 ICL generates proteins, de novo antibodies, CRISPR-Cas9 loci, mitochondrial genomes and entire bacteriophages.
We can even use the synthetic sequences iteratively to construct a chain-of-thought (CoT) and achieve harder design goals.
New work with Yukun A. Hao, @BrianHie and team 🧵
This talk is a beautiful articulation of one of the most important design problems of our time. A few thoughts on planning with agents:
- I'm finding one of the best inputs to plans are recordings of live conversations among a team. this tends to capture a *ton* of peripheral useful context, that's hard to capture even in a solo voice recording -- because people elicit intent from each other. A generalization of this principle is that agents thrive on broad, high level, "messy" context - just like senior employees.
- I think visual outputs from agents are critical to complement text, but I'm actually of mixed minds on interaction. Yes, it's clearly nice for certain kinds of things like twiddling visual parameters. But most of the interactivity my agents produce feels mostly designed for dopamine reasons and not providing much utility. I also think interaction can inhibit deep thought compared to linear reading. I actually think for many deeper problems I would prefer top notch prose -- Richard Dawkins / Steven Pinker / Annie Dillard level -- from my agents over any kind of flashier output.
- One mental model I find consistently helpful for planning: the job is to establish a *design space* together between human and agent -- that is, a set of relevant dimensions of variation -- and then to efficiently explore the space. Getting the dimensions of the space right is actually one of the hardest parts. Once you have that, you can ask an agent to plant a bunch of starting points spread throughout the space, and then locally hillclimb from the best one.
New talk! About the divided human & agent worlds, boundary objects, and building thicker interfaces
I've been exploring how to improve "planning", aka. deciding what agents should build, how you want it built, & making the whole team part of that process
https://t.co/JS73DCdSXi
Maybe controversial: the conversation about scientific bottlenecks is overfocused on automation. New measurement technologies deserve more attention. You could automate an assay that needs an entire facility – or you could REINVENT the assay so that it fits into just a few tubes.
In every Harry Potter book, there's approximately 45-100 students that are having a perfectly normal year and wondering why Harry and his friends are constantly involved in some whackadoo bullshit that threatens to close the school and fuck up their education
Introducing Dreaming: Across sessions, Devin builds a memory graph of how you like to work
At night Devin self-improves its memory to remove stale records and discover latent information
We are creating an OSS standard called Agent Memory Repo: https://t.co/b6qzEJsaeT
A week later, I'm still thinking about NHJ's article on NYC schools. It was courageous but a huge missed opportunity to call for real reform.
She blames school choice, inequality, segregation, the wealthy and herself. But she seems reluctant to criticize the actors and policies implicated by the failures she documents, perhaps because those conclusions would be uncomfortable for many of her ideological allies.
- She describes absentee teachers, poor instruction, unresponsive administrators, and students who graduate without mastering the material. But she never calls for greater accountability at the district, school, principal, and teacher levels.
- She documents teachers failing to fulfill basic responsibilities but doesn't explore whether tenure and other employment protections allow chronic teacher absenteeism and poor performance to persist.
- She laments the concentration of disadvantaged students but doesn't consider whether NYCPS should consolidate under-enrolled schools, thus allowing more resources towards instruction and increasing socioeconomic diversity.
- She describes her daughter earning As in honors classes despite substantial gaps in her knowledge. Yet she doesn't consider the value of standardized testing as a check on grade inflation and identifying learning gaps before they compound.
- She describes how experienced teachers choose to work in better schools but doesn't salute the districts that have had “reform superintendents” like DC, Dallas, Denver, and Houston that have used cash bonuses to incentivize the best teachers to work in the most disadvantaged schools.
- She notes that eliminating homework harmed her daughter, but doesn't examine how the movement against homework is largely driven by equity advocates who argue that it can turn unequal home circumstances into unequal academic outcomes.
- She condemns school choice yet uses it to secure a better school fit for her daughter, without noting it’s the only realistic escape valve for families who can’t move or afford private school.
- She describes the “impossible balancing act” of teaching students at vastly different levels but doesn't explore how academic tracking can help students at both ends of the academic spectrum.
- She describes disruptive students who harm the classroom environment but doesn’t ask for stronger disciplinary actions to protect other students.
- She documents a school that lowered academic expectations in pursuit of equity but doesn't acknowledge how doing so perpetuates the very achievement gaps those policies seek to eliminate.
Hannah-Jones’ story was powerful. She supplied the evidence for a broad examination of why schools are failing disadvantaged students. It could have been the strongest call in a generation for urban school district reform.
Segregation and poverty are real obstacles. But they cannot become explanations that absolve everyone else of responsibility. We can’t wait to solve all societal issues before demanding higher performing schools.
Today, Pioneer Labs is announcing our first step towards terraforming Mars. 🚀🌼
With equipment that fits in just a single rocket launch, we can convert Martian dirt, water, and air into enough building materials to construct a small city on Mars.
To do it, we made the first microbe for Mars. We found the best microbe on Earth and used evolution to teach it how to source all of its nutrients directly from Martian materials.
The first astronauts will be greeted with safe shelter already filled with water, oxygen, and rocket fuel for the return journey.
This is the first step toward using biology to make Mars a friendly place for life. It lets us live off the land and helps us build the next great frontier.
It's the first of five organisms we need to green Mars ⬇️
In March 2026 Harvey raised at an $11b valuation at ~$200m ARR (50x revenue)
By June, for every $1 in revenue, Harvey was spending $1.50 with Anthropic (et. al.)
Meanwhile Anthropic raises at $965b valuation in May 2026 with $47b run rate (20x revenue)
I have so many thoughts here but I think it's the first crystal clear articulation of the double counting of revenue that's happening in VC land.
A single dollar flows into Harvey's , boosting their valuation at 50x ratio, then turns into $1.50 and flows into Anthropic where it boosts valuation at 20x.
$1m spend customer spend with Harvey was leading to $80m in valuation markup across Harvey and Anthropic.
I don't get it.
Spotify published something of a spicy take on its engineering blog this week that made for an incisive and thought-provoking Saturday morning coffee read: "Why Spotify Is Not Using Bayesian A/B Testing"
I usually roll my eyes when I see blog post titles of this nature, since a) this discussion has been had ad nauseam, and most contributions rehash well-worn arguments, and b) methodology choice is highly context-dependent and doesn't generalize well. But this blog post is (very) thorough, and the context in this case, a household-name product with ~800MM MAU, is compelling even if the lessons don't apply universally.
The post walks through a number of reasons why Spotify's A/B testing needs are met through a frequentist process, and why introducing a Bayesian process wouldn't add sufficient value. The two that I found most powerful, which really relate to mistakes in implementation and not choice of methodology, and which I've empirically seen to be the case:
- The winner's curse is only ameliorated in the Bayesian approach with a well-selected prior. But teams often implement a uniform prior, which provides no shrinkage (pulling noisy uplift estimates toward more plausible values), and some misspecified priors actually performed worse than the group sequential testing (GST) baseline. And even an oracle prior, while improving effect estimation, showed no power advantage over GST.
- Forming an informative prior from past experiments is challenging because it requires many historical experiments (the blog post suggests >200, which strikes me as high) and for the much more prosaic reason that the world changes over time and historical results may not be representative. That second reason seems obvious, but it impacts experimentation in subtle but substantive ways: a consumer product's DAU mix can evolve meaningfully over time by *cohort age* and provenance (organic vs. paid), and *new users* may possess dramatically different motivations / intent / product knowledge over time.
The post states that "an experimenter using flat-prior Bayesian inference with posterior-probability thresholds is running frequentist-equivalent inference under a different vocabulary" because:
- posterior mean equals MLE
- one-sided posterior P( B>A | D ) = 1-p-value
- "95% posterior-probability threshold is algebraically equivalent to the one-sided rejection region"
The post does concede that Bayes provides superior interpretability, but in my experience, product leadership is generally comfortable with the concept of confidence intervals (vs. Bayesian credible intervals), so the benefit may be more muted than ~10 years ago.
Link to post and the accompanying paper below.
A lot of people are missing Terence Tao’s point and thinking “mathematicians are upset that AI is better than them.” That’s not what he’s saying, and some people are forgetting that Tao is one of the most AI-pilled mathematicians out there.
His point is that when people work on discovering something, along the way they invent new concepts. Those concepts later become useful far beyond the original goal, and enables further inventions. Finding a solution does matter, but the intermediate idea is often what makes the field richer, because other people can share it and build the next thing from it.
In tech, we can use the analogy of collaborative software. We started with algorithms for merging changes in a Word document, and evolved that to concepts about versions, diffs, and merges, and later to real-time collaboration tools like Git, Google Docs, and Figma. Humans built upon these concepts and developed more powerful solutions.
Terence’s worry is that a machine automating a solution robs the field of the value of developing the intermediate discoveries in the pursuit of larger discoveries.
When automating a solution, the intermediate discoveries and invention of concepts can be buried or completely hidden in the black box. We don’t learn from them to build the next thing; it’s like we never made the invention of collaborative document editing and thus could not have the conceptual understanding to invent the next version – and since it’s hidden, we also don’t socialize them to allow other people to invent, too, a core tenet of collective discovery.
So then, in both code and math, this leads to the atrophy of development of concepts in the field.
In other words: pure ‘solution extraction’ that hides the process of discovery can leave the field with a checked-off theorem but little new insight or new questions to pursue. And it might prevent us from understanding a field deeper.
I am seeing, first-hand, that atrophying of skills in software development. We push buttons and get solutions. There is much less incentive to develop new concepts and human skill. The bet most software companies are making is that LLMs are so effective in writing code that you’re still shipping overwhelmingly more value even with human skill atrophy, and it’s the right bet IMO.
However, much of the software industry is built upon building things, not necessarily novel invention and research. In such an environment, you can say that you accept some atrophying of conceptual invention and human skill for more output.
On the other hand, sectors like math and pure sciences that are focused on invention and insight might be the hardest hit by this.
Practical/applied sciences might fall somewhere in the middle. An Alzheimer’s cure, room-temperature semiconductor, or highly effective carbon capture solution are far too valuable to sandbag and say only humans can do that to develop concepts in the ‘proper’ way. The outcome matters too much to treat the preservation of concept invention as the highest goal. Even there, though, hidden intermediates can slow the next breakthrough if nobody can see how the first one actually worked.
So the question is not “is AI allowed to solve hard problems?” It is “in this field (math, science, tech, etc.), is the answer itself the main point, or are the concepts and abstractions we use to get there also the thing we need to maintain?”
In pure math, there’s an argument that the intermediates are often more useful than the solution, and atrophy in concept development is highly detrimental to the field. Solving Navier–Stokes, contrary to what some people claim, has little practical application, and pure math might be one of those fields where just finding a solution isn’t the entire point, and can actually be contrary to the field, which is what Tao is worried about.
ChatGPT casually watching and transcribing my conference talk practice run, identifying when I'm doing meta-commentary in the middle like "cut this slide" "update my script here", time stamping and matching it to my speaker notes, making edits in Keynote with computer use. Mad.