The Meaning of Accelerated Discovery

Preface

This article is mostly rambling. While it takes on an essay-style appearance and tries to make various points, none of the points are particularly strong or fleshed-out. I don’t offer any type of conclusive statement in this post, though it is evident how I feel about the issue.

One more comment: Some of the examples in this article are about slightly technical things. I’ll not elaborate on what these things mean in the greatest detail. The first technical concept you’ll see is something called “(dynamical) criticality.” To explain this and all its “backstory” would take too long, so I recommend reading this quite short Wikipedia article on the central notion: Critical brain hypothesis.

The Meaning of Accelerated Discovery

I believe I am now a witness to a disturbing shift of priorities in the field of physics (perhaps of science more broadly). The experience that I will elaborate on momentarily, I believe, serves as a microcosm of a broader change in the way that physics research, or research more broadly, is conducted. Let me now explain.

I recently had the fortune to participate in a “neurophysics” project with a great friend and colleague of mine. The objective of this project was to measure a “distance to (dynamical) criticality” in the human brain using both a theoretical model of neural connectivity and live EEG (electroencephalography) data. We were offered this opportunity all because of a wonderfully generous individual we met at the American Physical Society conference in Denver, Colorado . Early in the project, this individual (leading the project and initiating us into the field) presented us with MATLAB code—and I don’t know MATLAB.

I thought I ought to start learning it. After all, if I am to transition into the field of computational neuroscience and the standard is to use MATLAB for analysis, then I must be sensitive to the traditions of the field and try to acquire this skill. (As always, when it comes to learning, I parade around and bear close in mind the several other reasons for taking a particular learning approach: it is important to challenge yourself; learning something new generates new perspectives that may help you solve other extant problems; the effort expended in learning is the reward that keeps us vital and ready to charge forward. This list goes on and is informed by the book :[How We Learn] by Benedict Carey, among many others.) I spent some time getting acquainted with the basics of the software, including learning how to install the custom toolboxes that “sit on top of” MATLAB itself, which include hctsa and the Criticality library. After going through a lot of documentation—and even enlisting some help in understanding MATLAB from ChatGPT—I started writing the code that I knew would generate the results that we were looking for.

Suddenly, my collaborator presented me with the main results of the study: eight desired plots that quantified the distance to criticality. There they stood: beautiful, and probably completely correct. And I was still struggling with getting MATLAB to load the 20 huge data files that we needed to perform time-series analysis on in the first place. I was upset, defeated, confused. When I asked how he got the results that he did, he explained that he passed the work to Claude. “It would take me months to learn MATLAB,” he said. “I didn’t want to spend months on the project.”

Make no mistake here: Claude likely did the entire analysis correctly. It really did beat me, then. It really did “accelerate discovery,” the phrase that we will hear time and time and time again when we read grants and proposals about using AI technology in scientific research. Consider, for example, Dr. Matthew Schwartz’s invited talk at the APS 2026 conference: a 30-minute talk explaining how he used artificial intelligence (chatbots) in uncovering new results in physics. But more than once he emphasized the time saved by using AI technology: “Two years (me + student) or 4 months (just me) => two weeks (me + Claude)” read the slide.

Now, nobody likes to say this outright, but let us momentarily analyze the implied equivalence here: If it takes an advisor and their graduate student two years to conduct a research project that could’ve been completed in only two weeks, then evidently we ought to “replace” the graduate student with Claude. Since there are 52 weeks in a year, then Dr. Schwartz working alongside Claude would be able to complete 52 research projects in that time that he would otherwise work with a graduate student and only complete one. The equivalence relation implied here emerges when we compare the number of research projects conducted in two years: working with the graduate student for two years to conduct one piece of work is equivalent to using Claude to produce 52. Great!

We absolutely must pause and reflect: Is this the equivalence relation that we wish to ordain upon the scientific process? What are we losing in this rush, if anything?

Dr. Schwartz and his chatbots were able to triumphantly reach the finish line in record time. But have you ever heard the classic aphorism: “Life is about the journey, not the destination.” I acknowledge that this perspective may not be widely shared in this nation, but let us entertain it. (And, after all, even the richest executives still want more.) How would the words “accelerated research” sound to you if you chose to live bearing in mind the perspective that life is about the journey (the living) rather than the destination? It of course runs diametrically opposed to the notion suggested by saying that we ought to accelerate research.

What we must contend with nowadays, as AI technology finds ample seating within our workforce and research-force, is the meaning of the words we once thought we knew. The task is menial and irritating; arguing over definitions can be a bothersome, vapid process. But what do we mean by “we” whenever we claim that “we know” something? We need to ask ourselves if “we” are comfortable with AI systems contributing to “our” epistemology: Does/will it matter that we can disentangle the knowledge “we” know versus the knowledge that AI has contributed? When it comes to the pursuit of research, “we” need to determine what fundamentally consitutes that pursuit: Do we wish to exclude befuddlement, inefficiency, irritation, and turmoil?

The example of artificial intelligence accelerating discovery on everyone’s mind now is the recent report of an explicit counter-example to the Jacobian Conjecture in mathematics. According to a researcher at Anthropic, it was the LLM Fable (presumably Fable 5) that provided the counter-example. This contribution to our knowledge of mathematics was provided by a human aided by AI technology. Human mathematicians checked the result and it worked. Now that humans have performed the validation, “we” can say that “we” know something new. But was it “we” who discovered it? Nevertheless, it is now encompassed within our knowledge.

I am not neutral on the issue of integrating AI technologies into our research. However, I hope to have conveyed that I believe their application can indeed not only speed up research but also contribute significantly to achieving its objectives. This proposition is one I concur with, actually. I am merely commenting on the disquiet that has arisen within me recently. I think that we humans have still not figured out how to live with ourselves and understand how knowledge is passed around. The introduction of a technology that dabbles in the same epistemic domains as the ones we do is woefully untimely.

What really did Claude “beat me” at, exactly? Did it learn some MATLAB? Did it go through the process of running and re-running code with a host of errors and rectifying them each time? Did it enjoy the learning? Did it dislike the debugging? Did it learn something about the human brain through its analysis of EEG data? It arrived at the destination before I did, but how was its journey?

Notes/References

1. “Neurophysics” is a neologism used to refer to the nascent research discipline that entails understanding mammalian nervous systems using a variety of frameworks, formalisms, and mathematics that has seen extensive application in physics.

2. MATLAB is the name of a piece of software that is used all across quantitative research. It is a software, and it also has its own scripting language. And I didn’t know any of that before starting this project.

3. hctsa stands for “Highly-comparative time-series analysis.” It is a collection of (custom) MATLAB scripts. The main idea behind the package is that it helps somebody analyze time-series data by computing a whole host of various mathematical quantities that may be relevant in describing the data. If you’re actually curious about what this package looks like, see its GitHub page. Criticality is another MATLAB software package that we are using for the analysis—you can see its GitHub page if you want, too.

4. “APS” is an acronym that stands for the American Physical Society. I went to its 2026 meeting and documented my experience extensively in the posts Traveling to Denver, Colorado (APS Part 1), I Got Into a Fight at APS (APS Part 2), The Friday Hangover (APS Part 5), with the other parts awaiting “peer review.”

5. An “equivalence relation” is a relation (predicate of two variables, $P(x,y)$ ) that is (i) reflexive, (ii) symmetric, and (iii) transitive. These words probably mean little if you haven’t studied (first-order) logic. What you must understand, though, is that an equivalence relation is a logical (and mathematical) notion, and I borrowed that notion to accentuate an argumentative point. A good example of an equivalence relation is: “ $p$ is of the same opinion as $q$ ”; a good anti-example is “ $p$ is in love with $q$ .” Based on the three required properties defining an equivalence relation, can you figure out why the former qualifies but the latter does not? (I got those examples from Classification of sets - Lec 03 - Frederic Schuller.)

6. I wish I could say a lot more details about this, but to do so may jeopardize the identities of the individuals involved. I will still mention this story in passing because it is relevant here. Recently, somebody came to my university to discuss agentic workflow in gravitational wave-signal modeling. During his presentation, he suffered a damning Freudian slip by saying outright the word “replace” when describing how agentic workflows can automate graduate student-level work. It was along the lines of, “…and it can replace—well, not ‘replace,’ but you know what I mean.”

7. “According to a researcher at Anthropic…” Levent Alpöge, mathematician, now technical staff at Anthropic, reported, in an irritatingly-worded post on Twitter at 2 o’clock in the morning, an explicit counter-example to the so-called Jacobian Conjecture. We’re not super concerned with what the Jacobian Conjecture actually is; the main point here is that mathematicians thought the conjecture to be true for a long time, but along comes Mr. Alpöge and Fable and gives a bite-sized counter-example. If you are interested in the details, first start with the Wikipedia article on the subject, and then proceed to digest Terence Tao’s digestion of it on his blogpost A digestion of the Jacobian conjecture counterexample.

8. “…presumably Fable 5…” We don’t actually know because, to many mathematicians’ distate, Levent Alpöge has refused to elaborate more significantly on the details of the interactions with the chatbot that led to the counter-example. Mathematician Abhishek Saha of Queen Mary University of London is reported to have said in AI’s solution to 87-year-old riddle takes mathematicians by surprise:

“I don’t know how he did it, what exactly was the prompt to give Fable, because if one were to search everything, it wouldn’t quite work, so obviously there was some insight also which is not currently published.”

Updates:

1. 2026-09-08: Added hyperlinked references.