bibliometrics | Joshua Bloom

We will need more (not fewer) scientists

Wed, 26 Aug 2026 00:00:00 +0000

This post originally appeared on the Valency Blog.

Labor displacement concerns with new technology are neither new nor entirely unfounded. The Luddites, pegged by history as railing against new machinery, were in essence a labor movement¹. Telephone switchboard operators dwindled rapidly following electromechanical switching investments². In a recent essay, Bill Gates called for policy intervention to stem the wholesale move from human labor to machines: “AI will take on work in law, customer service, medicine, software, and manufacturing… There will be some new jobs, but without the right policies, there will be far fewer than exist today.”³

Now, as AI becomes more enmeshed not just in our daily lives but in our scientific work, and looking back on such historical labor displacements, there is a growing unease among scientists, and especially among budding scientists. What will be the role of scientists when increasingly capable AI co-scientists⁴ can participate in precisely the sort of labor we have trained (or are training) to do?

Attitudes towards AI in science were captured in a survey of scientists in 2023 (which feels like many generations ago now)⁵. While most saw AI assisting with faster data processing and computation, many worried about entrenched bias, easier fraud, and superficial understanding — labor displacement concerns weren’t at the fore, but clearly on the horizon. Fast forward to 2026, as AI co-scientists have become demonstrably better, astronomer David Hogg’s white paper “Why do we do astrophysics?” identifies two diametrically opposed future paths: one where AI is kept at bay, and one where it sweeps through our profession and relegates scientists to observers, rather than owners, of the scientific process. And Terence Tao, in a recent ICM lecture capturing the transformative moment for AI in mathematics, argues that the real crisis is not machine capability but values: AI can accelerate proof and verification, and it falls to the community — not the technology companies — to decide what mathematical work is actually for⁶.

My own optimism here, not just as a scientist but as an educator and builder of tools for AI-accelerated science, stems from a belief that establishing our work as a zero-sum in the number of jobs to be done (i.e. that there is a fixed amount of scientific labor needed per unit time) is not the right framing. Instead, I see science work happening inside of a rapidly growing pie with increasingly more room for both scientists and their silicon sidekicks.

The geometry of the frontier of science

A useful metaphor perhaps is the Little Prince, who saw dozens of sunsets in a single day, less out of wonder than because his world was small enough that a chair was all it took to see them all⁷. That is what a finished world feels like from the inside. Place him on Pluto (with appropriate outerwear!) or on the surface of the Earth, and with the same stride and appetite, he would meet wonders that couldn’t be grokked in a million lifetimes. What changed is the size of the thing he’s standing on. That position was once ours: when the first scientific journals appeared in 1665, a diligent person could read all of it⁸. And the most famous claim from ~1900 that scientific growth had stopped — physics is finished, nothing left but decimal places — is an apocryphal prediction that has comically been proven wrong over and over⁹.

Fig. 1 · As AI helps us push the frontiers, there will be much more volume to explore and discover. Illustration by Antoine de Saint-Exupéry, from The Little Prince (1943). Image via JSTOR Daily, where it accompanies Crim, K. (2009). “Flight.” The Threepenny Review, No. 117, pp. 8–9.

So if we think of the sum total output of scholarly work from a single scientist over the course of their career as filling some volume of knowledge, then we could view AI co-scientists as encroaching on, and perhaps entirely crowding out, that future volume coverage. In a fixed volume, that’s correct. But tools and innovation push the radius out. For a ball of radius r in D dimensions (dV/dr = D,V/r). The fractional return on pushing the frontier out — new volume per new radius — isn’t a constant. It is D. And D is not small in the sciences (it’s certainly not 2, in the case of pie). If the use of such AI-enabled tools grows the radius, it should (or maybe must?) expand the volume faster than the machines can fill the space.

The central supposition of Gates’ argument of a coming wave of job loss is that there is a fixed amount of effort to be done. And that may well be right for most of the economy: demand for labor is ultimately bounded by the population it serves. While there is certainly much headroom here, I can see how a fixed number of people in the world ultimately becomes a cap, abstractly, on the volume of work to be done (i.e. zero sum). In contrast, the frontiers of knowledge in science are unbounded, and so my hope is that that implicit argument from Gates (and those many before him) does not apply.

A Scientist’s Central Role

As the search and discovery volume increases and AI helps us grow and explore the space, there are parts of the scientist’s role that become very central:

Choosing the question. Which direction to push the radius. No amount of volume-filling substitutes for picking a direction worth walking in.
Knowing when a result is too good. The skeptic in all of us when confronted with new results gets more valuable, not less, as the volume of plausible-looking output goes up.
Taking ownership. Accountability scales with people, not with compute.

In this light — as the frontiers of knowledge are pushed outward with ever-more sophisticated tooling, inexorably, I’d argue — we need more scientists, not fewer. It’s still very early innings for AI-accelerated science, and if the volume-expansion analogy holds, the shell at the frontier we’re all standing on is about to get a great deal roomier. I couldn’t be more excited about who gets to go play in it.

The conduct of science is not zero-sum.

Hobsbawm, E. J. (1952). “The Machine Breakers.” Past & Present, 1(1), 57–70. DOI:10.1093/past/1.1.57. See also Conniff, R. (2011). “What the Luddites Really Fought Against.” Smithsonian Magazine, March 2011.
Feigenbaum, J. J. & Gross, D. P. (2024). “Answering the Call of Automation: How the Labor Market Adjusted to Mechanizing Telephone Operation.” Quarterly Journal of Economics, 139(3), 1879–1939. DOI:10.1093/qje/qjae005; ungated as NBER Working Paper 28061.
Gates, B. (2026). “A Turbulent AI Era and Critical Choices to Make.” GatesNotes, 26 August 2026. gatesnotes.com; coverage at CBS News and CNBC.
Gottweis, J., Weng, W.-H., Daryin, A., et al. (2026). “Accelerating scientific discovery with Co-Scientist.” Nature, 655, 487–496. DOI:10.1038/s41586-026-10644-y. Ghareeb, A., Chang, B., Mitchener, L., et al. (2026). “A multi-agent system for automating scientific discovery.” Nature, 655, 497–505. DOI:10.1038/s41586-026-10652-y
Van Noorden, R. & Perkel, J. M. (2023). “AI and science: what 1,600 researchers think.” Nature, 621, 672–675. DOI:10.1038/d41586-023-02980-0
Tao, T. (2026). “Mathematics in the Age of AI.” Public lecture, International Congress of Mathematicians, 24 July 2026. Essay: arXiv:2608.16753; recording.
Saint-Exupéry, A. de (1943). Le Petit Prince. Reynal & Hitchcock. The sunsets are in ch. VI — forty-three in the French, forty-four in Katherine Woods’ 1943 English translation.
Banks, D. (2017). The Birth of the Academic Article: Le Journal des Sçavans and the Philosophical Transactions, 1665–1700. Equinox Publishing.
Kelvin’s actual 1900 lecture: Thomson, W. (Lord Kelvin) (1901). “Nineteenth-Century Clouds over the Dynamical Theory of Heat and Light.” Philosophical Magazine, Series 6, 2(7), 1–40. DOI:10.1080/14786440109462664. The “sixth place of decimals” remark is Michelson’s: Michelson, A. A. (1903). Light Waves and Their Uses. University of Chicago Press, pp. 23–24. Full text at archive.org.

The Invisible Future of AI ✨

Sat, 15 Aug 2026 00:00:00 +0000

This post originally appeared on the Valency Blog.

Some companies with obscure products, ironically, used to remind us how ubiquitous they were. The Intel Inside jingle lives rent free in my head.¹ The “We don’t make a lot of the products you buy, we make a lot of the products you buy better” campaign was remarkably helpful in making BASF a household name. Now, we’re constantly reminded by just about every company big and small that they have AI Inside.

The dot-com naming frenzy was in Bill Clinton’s presidential address at Caltech in January 2000.² After enumerating open scientific questions like the nature of black holes and the origin of gamma-ray bursts (a question near and dear to my own heart³)—Clinton said: “And maybe the biggest question of all: How in the wide world can you add $3 billion in market capitalization simply by adding .com to the end of a name?” (I was in the audience and we laughed, nervously) The naming phenomenon was real and measurable: Cooper, Dimitrov & Rau (2001)⁴ documented that companies earned cumulative abnormal returns (~74%) in the days around announcing a dot-com name change—often with no change to the underlying business whatsoever. Today, the .ai domain boom is somehow meant to convey that same cachet and gravitas, transcending and elevating the product itself.

As we think about the future of AI in products and science, a natural question emerges: how long does the AI Inside moment actually last? When does a transformative technology stop being a selling point and start being just table stakes? This felt like it might be an empirically answerable question, so I decided to explore.

Measuring Technology and Tooling Visibility

Using the NASA Astrophysics Data System (ADS) and Valency Bond, I investigated how often astronomers explicitly name a given technology in their paper titles and abstracts, in the major journals (ApJ, AJ, MNRAS, A&A, PASP), normalized by the total number of papers published each year that mention the term in the body of the paper. Call it visibility. I looked at detectors (charge-coupled devices; CCDs), telescope facilities (Keck Telescopes, the Hubble Space Telescope; HST), and methods (Markov Chain Monte Carlo; MCMC). When I aligned these technologies by their peak visibility year and normalized each curve to its own peak, a rather striking pattern emerged:

Fig. 1 · Technology visibility in astronomy: title/abstract mention rates in the major journals, normalized to each technology’s own peak and aligned at year 0. Data: NASA ADS.

When a technology is novel and exciting, scientists advertise it. A title like “CCD photometry of the globular cluster NGC 6752” (Penny & Dickens 1986⁵ — one of hundreds of papers from the mid-1980s with nearly this exact title) wasn’t just describing a measurement, it was signaling that the authors were using the hot new detector technology. But when a technology becomes infrastructure, the explicit naming stops. Figure 1 shows this effect across many types of technology.

This visibility metric is meant to avoid technologies that simply fall out of favor. Indeed on the same peak-aligned time axis, the fraction of all papers mentioning each term anywhere in the full text stays essentially flat after the visibility peak. So these technologies moved from the abstract to the methods section, not actually falling into disuse.

Fig. 2 · Full-text mentions, peak-aligned: the fraction of all papers mentioning each term anywhere in the full text stays essentially flat after title/abstract visibility peaks.

Unpacking the CCD Visibility Story

The charge-coupled device (invented at Bell Labs in 1969⁶; on telescopes by the late 1970s) utterly revolutionized astronomy. Before CCDs, we exposed photographic plates and squinted at grains. CCDs were digital, linear, and quantitative and were a genuine paradigm shift in how we collect photons and do science.

In 1991, “CCD” appeared in nearly 10% of all astronomy paper titles and abstracts. One in ten papers in the entire field felt compelled to mention the detector. Astronomers highlighted the technology because it mattered that they were using it.

CCDs (and their CMOS descendants) are now the main detectors in virtually every optical instrument on Earth and in space, and the term appears in barely 1% of abstracts — about 13% of its peak visibility. The technology didn’t go away. Quite the opposite: it became so ubiquitous that naming it became redundant. When everything is CCD photometry and spectroscopy, nothing is.

The sociologist Bruno Latour talked about black-boxing as “the way scientific and technical work is made invisible by its own success” (Pandora’s Hope, 1999).⁷ When a machine runs efficiently, we attend only to its inputs and outputs, not its internal complexity. The more science and technology succeed, the more opaque they become. What I find here is that we can watch this happen quantitatively: a rise, a peak, and a long, slow fade into assumed invisibility.

So Where is AI on This Curve?

Looking at the red dashed line in figure 1 we see AI/ML in astronomy — machine learning, neural networks, and deep learning, bucketed together. It’s still climbing, and, if you squint, perhaps nearly a plateau (peak that arrives around 2028?).

Right now, we write titles like “A recurrent neural network for classification of unevenly sampled variable stars” (Naul et al. 2018⁸) or “Real-bogus classification for the Zwicky Transient Facility using deep learning” (Duev et al. 2019⁹ — work I’m proud to have been a part of). The method is in the title because the method is still the point. It signals sophistication and it’s a real differentiator.

MCMC, the workhorse Bayesian sampling technique, appears to have just peaked (2024, by my reckoning) and is beginning exactly the descent that CCD began in 1991. Methods, like detectors, also apparently get absorbed. If AI/ML follows the trajectory of every other transformative technology in this analysis — CCDs, HST, Keck, MCMC — then within a decade or two, titles of scientific papers will focus on the outcomes and insights, not the pathway to get there. AI will be inside, but it won’t be novel.

The Paradox of Success

This, it seems, is the paradox at the heart of every transformative technology: the more successful it becomes, the less visible it is. The AI Inside moment is, by its very nature, temporary. Intel still puts chips in most of the world’s computers but who notices the sticker anymore?

For the companies currently festooning their products with AI ✨ emojis: the branding window is finite. For scientists, this curve is something closer to a prediction: we’re perhaps three years from peak AI visibility in the astronomical literature.

And for those of us thinking about the long arc of AI + science, I take this as a genuinely optimistic result. Paradigm shifts don’t end with a bang; they end with a slow fade into infrastructure. AI-accelerated science will just be called science.

Visibility metric: fraction of papers mentioning a term in title or abstract, per year, normalized by papers mentioning the term anywhere in the full text, in ApJ, AJ, MNRAS, A&A, and PASP (1960–2025), via the NASA ADS API. AI/ML bucket: “machine learning” OR “neural network” OR “deep learning”. MC bucket includes MCMC and Markov Chain Monte Carlo variants. Peaks for AI/ML are projected; all others are observed.

On the Intel Inside campaign: Intel’s own history
Clinton, W.J. (2000). Remarks at Science and Technology Event, California Institute of Technology, January 21, 2000. Clinton White House archives
Bloom, Joshua S. (2011) What are Gamma Ray Bursts?
Cooper, M.J., Dimitrov, O. & Rau, P.R. (2001). “A Rose.com by Any Other Name.” The Journal of Finance, 56, 2371–2388. DOI:10.1111/0022-1082.00408
Penny, A.J. & Dickens, R.J. (1986). “CCD photometry of the globular cluster NGC 6752.” MNRAS, 220, 845. DOI:10.1093/mnras/220.4.845
Boyle, W.S. & Smith, G.E. (1970). “Charge Coupled Semiconductor Devices.” Bell System Technical Journal, 49, 587.
Latour, B. (1999). Pandora’s Hope: Essays on the Reality of Science Studies. Harvard University Press. See also Science in Action (1987).
Naul, B., Bloom, J. S., Pérez, F., and van der Walt, S., “A recurrent neural network for classification of unevenly sampled variable stars”, Nature Astronomy, vol. 2, pp. 151–155, 2018. doi:10.1038/s41550-017-0321-z.
Duev, D.A., et al. (2019). “Real-bogus classification for the Zwicky Transient Facility using deep learning.” MNRAS, 489, 3582. arXiv:1907.11259