Do Science in Bed

4 min read Original article ↗

On April 8, 2026, I had a conversation with Claude (Anthropic) that went from a loose intuition about dark matter to a testable astrophysics prediction in peer-reviewable form, in one sitting. From my phone. I am not an astrophysicist. I am a Dunning-Kruger case, and the tools now exist to work with that.

I am writing this to encourage anyone interested to go explore. I have no doubt that this back-and-forth will eventually be available in the equivalent of a Claude skill. There are massive public datasets full of patterns that have yet to meet us. A language model will do the math you cannot, check the literature you have not read, and stress test your ideas in the same sitting. Here is what that looked like.

The question

I had a feeling that the cosmic web (the large scale structure connecting galaxies) looks like electrical circuitry. If it does, can signals travel through it?

The math

Cosmic web filaments are magnetized plasma. Magnetized plasma carries Alfvén waves. This is textbook MHD (Alfvén, Nobel Prize 1970). We computed the propagation speed using observed filament conditions.

Quantity Value
Magnetosonic speed 91 km/s
Maximum reach in 13.8 Gyr 1.3 Mpc
Damping length ~4 Mpc

Signals travel slowly but do not fade. The reach is about 1.3 Mpc, which is galaxy group scale. The Milky Way and Andromeda (0.78 Mpc apart) are within range. The Virgo Cluster at 16.5 Mpc is not.

An AGN outburst at 1% coupling dominates the filament electromagnetic background by 10,000× and delivers energy comparable to a galaxy's total magnetic field energy over 5 Gyr.

The test

We proposed searching NANOGrav pulsar timing data for directional noise correlations aligned with known cosmic web filaments. Nobody has done this. The data is public. We ran a pilot on 67 pulsars.

Direction Correlation Significance
Virgo Cluster r = −0.184 1.24σ
Perseus Pisces r = +0.156 1.39σ
Coma Cluster r = −0.178 0.73σ

Not significant. Consistent with noise. But the galactic latitude control was clean (r = −0.03), confirming it is not an ISM artifact.

Where it broke

This is the most important part.

At this point we believed we had something. A novel physical mechanism, a pilot result hinting in the right direction, and a test nobody had run. It felt like a finding.

So we stress tested it. We computed the expected signal timescale. A magnetosonic wave crosses 100 kpc of circumgalactic medium in about 1 Gyr. The DM variation rate over a 15 year observing window is 1000× below NANOGrav precision.

Our own math killed our own proposal. The test we designed would not work as framed. The signal is real but too slow to detect as a time variation in any current dataset.

We also found that AGN driven wavelengths (~1 kpc) fall below the WHIM turbulence outer scale (~500 kpc), meaning wave coherence over megaparsec distances is not guaranteed.

We revised the proposal: look for static directional excess in noise properties rather than time varying DM. A weaker claim, but one that survives the math.

The review

I sent the writeup to ChatGPT for adversarial review. First version was rejected (ChatGPT reviewed the title, not the content). Second version received “major revision, borderline publishable.” I fed each review back to Claude, which revised accordingly.

Critiques that landed: the timescale problem (which we had already caught), missing damping physics, and the waveguide analogy being too strong. All incorporated.

Critiques that missed: the claim that filaments provide no wave confinement (density gradients do), and the claim that shocks do not generate Alfvén waves (they do, observed in the solar wind).

What came out

The output is a set of testable hypotheses, not a discovery. None have been confirmed by observation or human peer review. What the session produced was a peer-reviewable proposal grounded in public data, with limitations honestly identified, including limitations we discovered ourselves during the session.

How it worked

I contributed the physical intuition and the research direction. Claude did the computation, literature search, and drafting. ChatGPT did the adversarial review. I routed between them.

The tools were a language model, Python, public data, and web search. From a phone.

Why I am publishing this

Most attempts like this will be wrong. And mostly scifi, which at the very least is a lot of fun. If enough curious people run this loop on enough public data, someone may find something genuine. Worst case you may be taking your first steps into your future field. This is vibe coding applied to science. Go dabble.

Reproduce it