Today I’m chatting with my friend Connor Blake, who graduated from UChicago this year and is now at MIT for his PhD in Prof. William Oliver’s Engineering Quantum Systems group. I love his blog Boson Cutter, and especially enjoyed his posts on deep work and the information theory of conversations. Here is our conversation!
MM: One of my friends who is doing a bioengineering PhD at MIT told me that ideas are the currency of graduate school and that grad students and professors optimize their lifestyle and workflows for ideas — using Obsidian for notes, perfecting their diet and sleep schedules, etc. Can you share some of your own such strategies for “creative insight optimization”, and those you have observed in others like your friends/colleagues?
CB: Here’s an anecdote about a friend who seems to have very good taste: Rohan Mehta (former UChicago undergrad, now at Lukin Group at Harvard) has an encyclopedic knowledge of what everyone else in the field is working on at any given moment.
I had just gotten back from APS March 2025, and he hadn’t gone. Our friend was going through her notes on the talks in his subfield, and as she listed out the talk names, he was able to guess just from the title which group and PhD student had given each talk, even though most were presenting totally new work not yet published. I’m pretty sure he built a feel for which areas are most active and worked on by whom by reading every new abstract published in his subfield on arXiv. This monitoring of what others are doing and why they’re doing it is something I’d like to cultivate, and watching him do this made me realize just how well this can be done.
Personally, I don’t know if I have very good taste yet, but I do have a pretty unique system in my opinion. I try to keep track of all the “interesting thoughts” I have (loosely defined) in one long spreadsheet with dates, categories and statuses. Here’s an example of a question I posed to myself that I was then able to answer with a subsequent paper in a row in my spreadsheet.
Start: 1/27/25, End: 1/28/25, Category: Paper, Status: Done, Note 1: Question I need to answer: Can I connect the notion of a distance in information geometry between Petz channel and some finite approximation for QEC and then figure out how to optimize it? Link: Frank Nielsen information geometry https://arxiv.org/abs/1808.08271, Note 2: No, not really - there are fidelity measurements but info. geom. is not helpful abstraction here
I keep track of mathematical tricks, papers I read, conversations I have, questions I hear, questions I write, and anything I think I’d like to remember in the future. A lot of the questions I have in this document go unanswered for months, and many of them are still unanswered. By building this web of ideas and questions, I have a queryable table of all the branches and chains of thought I’ve explored. It’s sort of like making a map of a cave to be explored so that whenever I exhaust a chain of ideas for how to solve a problem, I can trace back up the spreadsheet to find the nearest unexplored idea I had along the way, DFS style.
By mapping out the territory of the research problem I’m working on, I was able to identify more tangential applications of one of the computational techniques I’ve been developing that I otherwise would have totally forgotten about. It seems useful to occasionally revisit open problems from many months past in case something I’ve recently learned has resolved one of the older questions.
MM: In his book Deep Work, Cal Newport suggests productive meditation, i.e., thinking about a problem while one is away from the desk (e.g., while walking, waiting for the elevator, etc.) There are also numerous creative thinkers who suggest thinking about a problem before sleeping and reflecting on it after waking up. Do you have any personal experiences along such lines?
CB: I have some techniques I’ve used to “engineer obsession” (perhaps I’ll write a post about them soon), but I find this easier than trying to set aside time for explicit productive meditation. If you’re obsessed with something, you think about it all the time, so you’re always sort of meditating on it. I’ve definitely had dreams about research I’m working on, so it sort of naturally happens that I’m thinking about it as I’m falling asleep and waking up.
I think my biggest single tip for cultivating obsession is to make your environment very conducive to being reminded of what you want to be obsessed with. I used to make my phone background have a list of problems I was thinking about, and my phone home screen has an app for Quantum Physics arXiv and my Anki cards. I’ve tried very hard to replace all idle doom-scrolling with arXiv (I vibe-coded this very bright and shiny Instagram-like arXiv knockoff as a joke once; unfortunately I don’t use it much) and Anki which has flash cards for papers and equations I always want to be top of mind.
MM: I am curious to hear about your journey — you started college wanting to work in semiconductors, ventured into effective altruism along the way, and are currently on track to do a PhD from MIT. Can you elaborate on your journey and key influences (including books like Deep Work and Ultralearning and others)?
CB: During senior year of high school, I saw an Asianometry video on ASML and was totally blown away by how much engineering and weird physics goes into the technology, so this was how I decided on quantum engineering at UChicago since I wanted to bet that this technology and the other parts of the manufacturing process would keep getting more complicated and smaller.
Around that time, I was introduced to effective altruism and was really interested in the idea of a philosophically rigorous movement that was not just totally academic but impacted how you live your life. I was quite involved with EA my first two years — I was on the board of UChicago EA, ran an introductory fellowship (i.e., facilitating a reading group of EA ideas), was vegetarian, and went to EA retreats. I would say that the epistemic rigor of the EAs has been very influential in how I think, and people will often tell me I speak in a very EA way even if I’m not really one. I was more drawn to the rigor and rationalist thinking and less to the utilitarianism, but they get sort of muddled in the movement as I wrote about in this article.
I wouldn’t say at any point I was planning on doing a career in EA during undergrad (i.e., AI safety or animal welfare roles) since I was always more interested in being an engineer and didn’t fully buy some of the arguments. For example, I think software people are used to a world in which you basically have deterministic execution, and scaling is a matter of copying a Docker container and doing a new experiment in 5 minutes.
My AI timelines for full automation and such are much much longer than most of the EA/rationalist people because I don’t think they fully appreciate how hard it is to make physical stuff in the world and how slow diffusion will probably be. “Hardware is hard” as they say, and I think one of the most impactful problems looks like automating science. I think the group I’m joining at MIT to work on autonomous fabrication and defect reduction is a really exciting place to do this, so that’s what I plan on working on for a long time.
MM: You wrote this wonderful article about why you switched from theory to experimental research. It makes sense that experimental research labs have a great store of tacit knowledge from their proprietary experimental data. This seems to be a point more about the tacit knowledge involved in experimental setups, automation, minor refinements and optimizations etc. — but just how valuable do you think this tacit knowledge will be? One could argue that theory groups also have tacit knowledge in the form of research taste? Since the final results of both theory and experimental breakthroughs are published, it seems like the tacit knowledge claim really stems from dead-ends (ideas tried that didn’t work) — so is the claim that a database of experimental dead-ends is more valuable than a database of theory dead-ends? It would be great if you could elaborate on this, ideally with specific examples if possible.
CB: Having tried to use AI to work on theory research, they absolutely lack good problem approach taste and telling them to “solve this problem” without a ton of detail and intuition has been entirely fruitless in my experience, so it’s a very good point. It seems in the recent (claimed) Navier-Stokes case that the human mathematicians were able to create blowups of similar equations with significantly fewer resources than the swarms OpenAI used and with a worse model (Sol and Claude vs some internal OpenAI model). However, it really does appear you can just throw more compute at sufficiently well-posed open problems.
Taste definitely matters a lot in getting problems into solvable forms. Posing problems well is very non-trivial since different descriptions of a problem lend themselves to answers with different shapes. In a lot of quantum physics research, it is possible you literally just need to pick a better “perspective” by choosing a good basis in which to describe your problem, and then the solution falls out of the description.
But even finding the right abstraction with which to describe your system is very difficult — there have been two different Nobel Prizes in physics awarded for theories of superconductivity (Bardeen, Cooper, and Schrieffer for the eponymous BCS theory in 1972 and Ginzburg in 2003), one of which describes the bottom-up microscopic particle interactions, and the other is top-down about the macroscopic field theory of superconductivity. They’re both nice theories, yet it’s pretty unclear how you would pose a problem that would lead to new models in this vein without experimentalists noticing weird things and needing new levels of description.
I’m also not sure “negative result” is exactly the right frame for theory research since you often don’t have hypotheses and evidence in the sense of experimental research. For example, if you can prove that an assumption widely held in the literature is false or breaks down badly, this is a very informative result even if it’s “negative”. On the other hand, if you propose some complicated setup that somewhere along the line violates a no-go theorem and doesn’t work, it’s probably not that useful to anyone since your idea was provably wrong the whole time. You can also very easily end up at expressions that are correct but never simplify to something useful — it’s not exactly a negative result, but it’s not clear it’s useful or worth communicating without more work.
So I suspect a much higher proportion of “informative results” in theory end up in the literature than experimental research. Properly taken experimental data is rarely totally uninformative on the other hand. It definitely has noise, but negative results for some hypothesis often have evidence of multiple phenomena at once, and disentangling all the competing effects can be valuable even if the experiment “failed” in a normal sense of a device failing etc.
MM: There is much talk about quantum computing (QC) helping advance AI: for example, access to a QC could be incorporated as a resource in the AI’s training and tools (e.g., an AI with access to quantum simulations could innovate in materials). I would love to hear any thoughts you have on this.
CB: It looks like quantum computers will function like API calls within larger algorithms, all of which AI will definitely reshape, but I’m not sure. Right now it’s looking like quantum computing is best-suited for solving very hard chemistry and physics problems which are themselves quantum.
These are nice for current quantum computing since they are configurationally intensive as contrasted with data intensive — you have a relatively small number of particles in your system with an exponentially complex configuration space you care about and yet arranging them in the right way is both, a) very nontrivial and b) valuable (for drug design or something).
Machine learning is famously data-inefficient and requires a lot of it, at least compared to humans. Quantum ML requires getting that data into quantum states, doing some computations, and getting it back out. This is an enormous cost, so it’s unclear to me which problems could justify this cost relative to comparable classical systems. As for scaling, Google’s 2024 Willow chip has 105 physical qubits which is roughly double what the 2019 Sycamore chip had. There have been enormous advances in algorithms, materials science, and architecture since then, so the comparison isn’t “they doubled it in 5 years, wow, this is so slow”, but scaling quantum systems is very difficult (and what I’m working on in my PhD!)
MM: From being interested in semiconductors and exploring AI, how did you come to be involved in research with Prof. Shuolong Yang’s lab, and then how did you get interested in quantum error correction?
CB: I really wanted to join a hands-on project when I first got to UChicago, and Prof. Yang was open to taking first years — for which I am very grateful. The project was run by three undergrads without much graduate supervision, so it was really a crash course in learning how to do lots of things.
We built a fully autonomous physical vapor deposition machine with some Bayesian ML inside and published a nice paper about a year ago. The materials side was interesting to me, but I had read a lot about control theory and found the interplay between stochastic processes (such as in quantum noise models) and the ability to control them really interesting.
At the same time, I was taking a bunch of classes on quantum information theory, one of which was with Prof. Liang Jiang. Quantum control matters a lot to implementing fast and low-noise gates, and some systems like bosonic qubits also require really non-trivial control theory to send errored quantum states back to the codespace. This is what I’ve been working with Prof. Jiang on since then, and it actually has a lot of overlap with some work in my current group at MIT.
MM: To wrap up, I would love to hear recommendations for books, blogs, papers, or anything else that you like.
CB: I really like Charles Yang’s blog The Republic of Science. He’s now at the Anthropic Institute though, so I’m not sure how much he’ll be writing publicly in the future. He writes about metascience — how science progresses, how it could progress faster, what are the weird political/economic problems that science in and out of industry faces. In particular, I like this piece, which was even cited in a recent White House report “Science: A New Golden Age” when touching on autonomous labs and the Genesis Mission.
As for more mainstream bloggers, Jasmine Sun is an incredible writer. I discovered her blog almost two years ago. She writes very lucid pieces on the people and culture surrounding the development of AI, but always from an interesting angle. She has a recent piece about datacenters in the Midwest where she visited these small towns with new datacenters popping up and talked to the people in charge of the projects. She has such a good pulse on what is happening that the pieces always seem to arrive at the perfect time relative to the discourse.