Quasars, AI, and the Vera Rubin Data Flood

Editor's note: This piece covers astrophysics and AI in scientific research — territory that overlaps with our tech and AI beat. The reassignment request is noted; however, the AI-in-science-infrastructure angle sits squarely within this desk's scope, and the piece proceeds on that basis.

A quasar is, at its core, a feeding event. A supermassive black hole — somewhere between a million and a billion times the mass of the sun — has been perturbed, usually by a galaxy collision, and gas is falling toward it in a screaming vortex. The gas can't all get in at once; the black hole is a choke point. So it piles up, forms an accretion disc, heats by friction to temperatures that make the sun look tepid, and radiates somewhere around ten percent of its infalling mass as pure energy. Not visible light. X-rays.

That's the picture astrophysicist Matt O'Dowd lays out in a recent episode of StarTalk, alongside Neil deGrasse Tyson and comic co-host Chuck Nice. O'Dowd, an associate professor at CUNY Lehman College and host of PBS Space Time, has spent his research career staring at these objects. The conversation ranges from the basic mechanics of quasars to gravitational lensing to the question that now shadows every corner of observational astronomy: what happens when your instruments generate more data than any human team can process?

The Physics, Translated

The name "quasar" — quasi-stellar radio source — emerged from confusion. When these objects were first catalogued, they looked like stars: pinpricks of light with no obvious structure. But their spectra showed enormous redshifts, meaning they were moving away from us at speeds that placed them at cosmological distances. At those distances, even a faint-looking object is putting out something like a thousand times the luminosity of an entire galaxy. From a point source.

The debate over what could produce that much energy from that small a region ran for years. Swarms of neutron stars were proposed. Chains of supernovae. O'Dowd frames the resolution simply: "To get a lot of energy out of a very condensed region of space, the black hole is a good way to do it."

What makes a quasar different from, say, the Milky Way's own central black hole isn't the mechanism — it's the scale. Our black hole, Sagittarius A*, masses about four million suns. Quasar-class black holes run to a billion. The Milky Way's black hole almost certainly had its own active phase early in the universe's history. A distant galaxy whose light is only now reaching us would, O'Dowd confirms, see our galaxy in that phase — though technically as an "active galactic nucleus" rather than a proper quasar, the distinction being largely one of size. M87, the galaxy whose black hole the Event Horizon Telescope imaged, is a more apt candidate: a billion solar masses, and still producing a visible jet, even if its full accretion disc has quieted.

The Event Horizon Telescope's data, incidentally, could not be transmitted over the internet. The volume was too large. The raw data was physically flown to processing centers on hard drives aboard planes. This detail, almost a footnote in the conversation, turns out to be relevant to where the field is headed.

What Gravitational Lenses Actually Do

Einstein predicted that mass bends spacetime, and therefore bends the path of light passing through it. The first observational confirmation came from Arthur Eddington's 1919 solar eclipse measurements: stars behind the sun appeared shifted from their expected positions because the sun's gravity had curved their light. Einstein himself thought cosmological-scale gravitational lensing would be too rare to observe. He was wrong, partly because in 1915 the existence of galaxies beyond the Milky Way wasn't yet established.

What lensing means in practice for quasar research is this: a distant quasar, a foreground galaxy, and Earth can align such that the quasar's light takes two or four distinct paths around the intervening galaxy. The observer sees two or four separate images of the same object. The most famous example is the Einstein Cross, a quadruply-lensed quasar.

O'Dowd's description of what makes this useful goes beyond simple magnification. Because the foreground galaxy is made of stars — and those stars are in motion — they sweep across the line of sight to the quasar like a slow radar scan. Different parts of the quasar get magnified at different moments. "You see different parts of the quasar change over different times," O'Dowd explains. "You can even see when it sweeps across the black hole in principle, and see it darken for a little bit."

The path-length difference between the multiple images also creates a time delay — hours to weeks. If the quasar produces a detectable flare, you can watch it arrive at each image in sequence. A natural replay function, written in geometry.

Chuck Nice's characterization — "a telescope booster" — is accurate enough that O'Dowd adopts it, though he adds that the lens is, in his words, "crappy." It's not a precision optical instrument; it's a galaxy full of stars with uncertain positions. Modeling it requires statistical inference about the distribution of mass in the lensing galaxy. But if you can do that modeling, you can map the inner structure of the quasar at resolutions otherwise unachievable. O'Dowd claims resolutions comparable to the Event Horizon Telescope, across thousands of objects rather than one.

That last clause is where everything gets complicated.

The Data Problem, Which Is Also an AI Problem

When O'Dowd started his career, the catalog of known gravitationally lensed quasars numbered in the hundreds. The Vera Rubin Observatory, which achieved first light earlier this year and is currently in engineering mode ahead of its full survey, will find thousands. It will image the entire southern sky every three nights for ten years. One image from its camera requires the equivalent of 400 HD televisions to display. The southern sky totals three to four thousand such images. By orders of magnitude, the data rate exceeds every previous telescope in operation.

"We could barely do one," O'Dowd says of the complex modeling required for a single lensed quasar system. The prospect of doing thousands demanded a different approach.

The AI tooling O'Dowd's group uses centers on variational autoencoders — neural networks that compress the time-varying brightness curves of quasar light into a compact "latent space," then attempt to reconstruct the original data from that compressed representation. Within that latent space, further networks try to extract physical parameters: black hole mass, spin rate, the geometry of the lens. The premise is that if the network has learned the physics of what generates those light curves, it can invert the problem and read the physics back out.

The conversation on StarTalk gets genuinely useful when it moves from what the tools do to what their failure modes look like. Tyson raises the analogy of a neural network trained only on cats and dogs that confidently classifies a chipmunk as one or the other. The concern applies: a model trained on expected quasar physics will interpret novel phenomena through the lens of what it already knows. It will find what it's looking for.

O'Dowd's response is that his team deliberately expands the training distribution beyond what they believe to be physically realistic — "let's go much bigger than that to make sure we encompass the true space" — and tests brittleness by deliberately breaking inputs and checking whether outputs remain reasonable. There are also unsupervised approaches where the network is asked only to find structure, without being told what structure to expect.

But the deeper question, which neither O'Dowd nor Tyson fully resolves, is whether AI can find patterns for which it has no template at all. "Can you find a pattern for which there is no template?" Tyson asks. The answer, O'Dowd suggests, is theoretically yes — unsupervised learning can surface unexpected regularities — but interpreting what those regularities mean still requires a human who understands the underlying physics.

What Gets Lost

The most interesting friction in the conversation isn't about quasars. It's about the graduate students who used to do the work AI now does, and whether something is lost when they don't do it anymore.

O'Dowd's initial instinct is that the answer is no. He cites spherical trigonometry — once a required graduate course for calculating telescope slew paths and angular positions on the celestial sphere — as an example of knowledge that has simply become unnecessary. Computers do it. Nobody is harmed by not knowing it.

Tyson pushes back, not by defending spherical trigonometry specifically but by describing what it felt like to spend hours classifying galaxies in a course he considered pointless. "When I look at a galaxy that you just took a picture of, I have a whole other relationship with it that you don't. I'm feeling it." The muscle memory of having processed thousands of objects by hand produced an intuition that no amount of algorithmic output replicates — not because it was better information, but because it was embodied differently.

Neither position is fully satisfying. The spherical trig example is probably O'Dowd's stronger case: procedural computation that was only ever instrumental. The galaxy classification example is probably Tyson's: pattern recognition that built judgment, not just knowledge.

The practical question for the next generation is whether future researchers, trained to query models rather than build them, will know enough about what's inside the black box to recognize when it's wrong. O'Dowd frames this gently. Tyson frames it as "vibecoding into a black box" and appears genuinely unsettled.

The Vera Rubin Observatory will start producing data at scale soon. The AI pipelines will be ready. Whether the people interpreting the output have the right kind of intuitions to catch what the models miss — that part is still being designed.

Samira Barnes covers technology policy and AI infrastructure for Buzzrag.