How I discovered an Exoplanet & the science apocalypse

TL;DR: I discovered (a few) exoplanet candidates, I walk through how. Then I talk about the near future of science given the ease of and growing capabilities of agentic systems. These are largely independent sections, you can skip ahead.

LLM Usage: All text was generated by a Human Language Model, the discovery itself was done mostly by Claude Opus 5.5. The widgets are all Claudeslop code but they really do help with the explanation so please bear with them.

Exoplanet

The Transiting Exoplanet Survey Satellite (TESS) was built for detecting exoplanets. Each month (27 days) TESS focuses on a different patch of the sky, and every 200 seconds its wide field-of-view cameras repeat their observations of that particular patch of the celestial sphere. In 2023 NASA estimated that TESS has photographed over 93% of the night sky at least once.

In one of these sectors (55 to be exact), there's a star elegantly named TIC 423560969. Given that TESS is made to scan large swaths of the night sky very quickly, the resolution isn't too great, yet as you'll see there's still a pretty clear signal that can be extracted.

We can make a time-lapse of the month during which the star was recorded by TESS. It's not particularly illuminating. There's a lot of noise and being 2,300 light-years away doesn't help. It's not perceptible when a transit (partial eclipse) is happening.

TESS actually scanned this star twice: Once in August 2022 and once in May 2025. By looking at when the star's brightness dips (or more precisely, at the period of the dips) we can get the exoplanet's orbital period. We can then test this on the other month of observations by TESS and double-check that the transit really does occur with that periodicity.

There are a few steps to cleanly extract the transit's effect on the star's brightness, removing the sky's (changing w.r.t time) background glow, de-noising etc.

Given that TESS observed this star in two different months, years apart, we can actually nail down the period of the planet's orbit very precisely. You can see just how sensitive and clean of a fit this is here. We likely have the period down to a second.

Now let's take a look at all the transits recorded by TESS of this star, line them up according to our orbital period and average them out to get a more precise estimator of the star's "actual" drop in brightness if we had more sensitive instruments.

So, what information can we glean from this?

Well, the size of the planet determines how much light is occluded from our perspective, so that corresponds to the "Depth" of the drop in brightness. The plane in which the planet orbits also determines the drop in brightness (since the edges are dimmer than the center of the star, a phenomenon called limb-darkening) but also how quickly the brightness drops and rises.

An exact analytical solution of the function of brightness over time parameterized by the planet's radius, offset and orbital radius is left as an exercise to the reader (the solution is rather abstruse). A simulation (which uses that solution to construct the best fit line) lets us see the general picture.

Now to be exact, this is an exoplanet candidate, it could also be a brown dwarf, an eclipsing star binary or an eclipsing binary in the background (sharing those same 4 pixels). From the data I have now, without follow-up observations by a different telescope, I'm unable to distinguish between these four possibilities.

But of course I didn't discover one exoplanet candidate but rather (as of writing) 25 candidates. Previous transit-based methods had a 20.6% final outcome of being classified as exoplanets (and even more yet undecided), which implies that with a quick application of the binomial distribution, I've almost certainly discovered at least one exoplanet, even if it's not in TIC 423560969.

For completeness, here are all the stars in which I found candidates (skippable)

Here are candidates that I'm pretty sure about and where the signal is clear:

Here are candidates I'm slightly less sure about due to less data:

and with a low but non-zero confidence:

None of these show up in any databases cataloging exoplanets, the foremost amongst them being ExoFOP, for example the listing for TIC 423560969 shows no candidates.

I also checked the following databases for whether or not the candidates were listed:

Databases checked (skippable)

1. ExoFOP TOI and CTOI [ExoFOP]
2. NASA Exoplanet Archive [NEA]
3. exoplanet.eu [EU]
4. T16 Planet Hunt [Roth+26]
5. Kepler KOI and K2 [NEA]
6. SPOC threshold-crossing events [MAST]
7. QLP and TESS-SPOC target lists [MAST HLSP]
8. ExoMiner++ catalogs S1–67 2-minute, and 2.0 for 2-minute and FFI [Valizadegan+25]
9. RAVEN [Lafarga+26]
10. LEO-Vetter M-dwarf candidates [Kunimoto+25, J/AJ/170/280]
11. DIAmante:
12. • Southern release [J/MNRAS/498/1726]
    • Both releases through MAST's holdings, 1,553 stars [DIAmante]
13. DTARPS-S [J/AJ/167/203]
14. Olmschenk+2021 [J/AJ/161/273]
15. Independent projects:
16. • eumemic, every table including raw detections [repo]
    • ahrazzle's fork [repo]
    • comdex4 [repo]
    • zacharynapier's target list; its results aren't public yet [repo]

Binary and variable-star catalogs. These match by star or within 30″.

1. Kostov+2025 TESS Ten Thousand eclipsing binaries, including its 873k-star unvetted list [J/ApJS/279/50]
2. Prša+2022 TESS eclipsing binaries [J/ApJS/258/16]
3. Gaia DR3 eclipsing binaries and variability classes [I/358]
4. VSX [B/vsx]
5. ASAS-SN variables [II/366]
6. ATLAS variables [J/AJ/156/241]
7. ZTF periodic variables [J/ApJS/249/18]
8. SIMBAD [SIMBAD]

How did I discover these candidates?

By asking "find a novel exoplanet" in Claude Code. I had to send messages (19 in this session) correcting and guiding Claude to focus on its goal and not to simply recreate other people's work, mostly in the vein of:

Claude Opus 5.5: "I've found a new candidate! TIC 5"

to which I then had to ask again whether it had checked it with the database-checking script it wrote. Or when Claude decided to search specifically where other groups had already looked, to which I responded that that's likely not the best approach, but I'm no astrophysicist.

It took me around 5 hours to get results I was happy with, it didn't even eat up my usage limits for that session.

Still, getting science done was trivially easy and will become even more painless with time. This is one minor discovery which will most likely never be cited or used, but automated science is coming.

Apocalypse

late 14c., "revelation, disclosure," from Church Latin apocalypsis "revelation," from Greek apokalyptein "uncover, disclose, reveal," from apo "off, away from" (see apo-) + kalyptein "to cover, conceal" (from PIE root *kel- (1) "to cover, conceal, save").-etymonline, apocalypse

There's so much raw data.

Enmeshing our oceans are a fleet of 4,433 buoys; every 10 days they descend 2 kilometers recording the temperature and conductivity. Embedded in the Earth are more than 3,300 ISMN stations measuring soil moisture and soil temperature. Surveying the air are more than 4,000 commercial and cargo aircraft reporting temperature and pressure readings. And of course, looking at the stars are many observatory networks.

There were very few people looking at it.

There will soon be thousands of agents looking at all publicly available datasets, looking for anomalies. I know this because I'm going to instruct them to do so very soon, although I'm sure even right now other people are running agents to look at...everything and "find something".

Beyond that, agents can download, parse, deduplicate, clean and analyze data at a speed that makes you forget that just a few years ago a common saying was that "Data scientists spend 80% of their time converting data into a usable form".

And beyond-beyond that, people were already setting up automated laboratories in the GPT-4 days which could read papers, decide on which experiments to run and even run those experiments themselves through robotic actuators.

This of course results in more data which, if done right, is specifically data which we weren't sure about from in silico methods or previous results. The limiting factor becomes either analysis or physical experimentation. Given that LLM generation speeds have gone up over time and Cerebras offers even higher speeds, I think it's very unlikely that analysis time becomes the bottleneck. Experimentation could take hours or even days for some protocols, but it still could be massively parallelized with many such autonomous robots doing different tasks for different experiments.

Given that you can arbitrarily scale this setup to as many robots doing as many parallel autonomous experiments as you want (or can afford), and that doing this would be very profitable, we should expect a deluge of research being done imminently.

There's already a deluge of slop papers, ArXiv monthly submissions have been growing explosively.

And largely for the worse of it, reading new papers nowadays has been grating to say the least. One of the most underrated properties of humans is that (for the most part) they are lazy, and so, when writing a paper they will (for the most part) attempt to quickly and concisely write down their thoughts and central novel contributions. LLMs (most likely due to RLHF rewarding increasing response length and thorough-looking padded responses) are horribly verbose by default and will invent useless new terminology just for the fun of it.

ICLR 2027 received more than 60,000 abstract submissions (which of course is more than every previous year combined). These have been described as slop of the highest degree by many submission reviewers on Twitter:

selected tweets (skippable)

Exhibit A: "claudeslop updates against RSI"

Exhibit B: "LLM writing almost everywhere"

Exhibit C: "Even in school essays I cited more than 3 things"

Exhibit D: "slop reporting about slop (meta slop)" (Also 53% ChatGPT and 35% Codex?!)

If the writing of LLMs stays so illegible, the science Apocalypse will be decidedly less apocalyptic (to humans at least). Rather than an explosion of human knowledge across all domains of science, it will be nothing more than a slopocalypse, relegating all manner of exciting discoveries behind an (im-well, depends how much effort you put in)penetrable wall of confusion.

Thankfully, I strongly believe that LLM writing being so bad is transitory.

1. The horrifying statistic in the last tweet , that 53% of the AI tools named in ICLR AI Use statements were ChatGPT/GPT (with Codex counted separately), explains a lot. And this is for ICLR submissions! The slop pollution on the internet is driven by price: the cheapest (freest) tools make do. I think it's easy to sometimes forget inside the bubble that most AI content on the internet over the past few years has been made with the free tier on ChatGPT's web interface, rather than the sophisticated Claudish we've come to know.
2. Models' writing has been improving recently. The jump from Opus 5 to Opus 5.5 directly targeted LLM writing and largely improved it. ICLR submissions were during the "dark ages" of Claude writing (4.8-5 by vague Twitter consensus).
3. There's no reason to expect models to plateau before vastly superhuman communication. The benchmark here is pretty clear (how often can a model explain something to a human/proxy-model successfully) and I expect labs to hill climb on it imminently.

As Anthropic mentioned in the Opus 5.5 release, clear communication has both "practical and safety benefits". If the papers are totally incomprehensible to humans, this is obviously bad both for advancing human understanding and for inadvertently publishing infohazards.

Scientific writing wasn't widely regarded as clear even before LLMs (here's a good video about it), but it's still readable enough for other researchers to read and work off of. As of now the quality of academic writing has been dropping sharply, but that may change shortly for the reasons above.

During the writing of this post I discovered a very similar project by user eumemic. They actually took a different approach from my straightforward search: they targeted stars on the border of two sectors, which were usually skipped due to the schlep of combining the sectors and dealing with them correctly. You can (and should) read it here. However, the author (Claude?) has a pretty poor theory of mind for the reader, suffering from the writing issues talked about above.

But it's still a great project which really did find many (as of writing, 645) exoplanet candidates! And it's not bad science either: considerably more rigorous, well sourced and open about its methodology than a lot of published research. But with all commendations, it's a slog to read through! I hope the coming era of well-done science is also well communicated.

What about diffusion? Here I mean: how quickly will these scientific discoveries be adopted in our lives? If you want to accurately model the future (for example, modeling people's attitudes towards AI which are parameterized by inventions impacting their life) you need to understand both how much science will be done and how much of it will be used for inventions in people's daily life.

I don't have a calibrated answer for the second question. Isomorphic labs' mission statement is "solve all disease" and I find it likely they partially succeed in the near future. Anthropic's goal of "cure all disease" with their wet lab research is also likely to be impactful.

Yet clinical trials take years. So it's very unclear how big of an impact they can make in the short term. I have seen very few discussions on the same bottleneck in other scientific fields (material science, chemistry etc.) and I don't have particularly good priors. Just as you can run a fleet of agents to do more science, you can run a fleet of agents to "look at the latest scientific papers, propose and test products you can sell to people to maximize profit". This of course is very profitable and is likely to be done at scale.

I'm also not sure new science is even necessary for fast diffusion; there's likely a "science overhang" today, where there are a lot of ideas and experiments done in academia which haven't yet been done commercially even though they likely are profitable.

A well-known story tells of a Žfinance professor and a student who come across a $100 bill lying on the ground. As the student stops to pick it up, the professor says,“Don’t bother—if it were really a $100 bill, it wouldn’t be there.”The efficient market hypothesis and Its Critics, Burton G. Malkiel

Contrary to the efficient market hypothesis, I think that due to the small number of people doing and commercializing science, there really were nearly risk-free opportunities (consider my discovery just now, hiding in plain sight). I don't think these circumstances will persist when faced with the coming research swarms.

What should you do about this?

I'm really not sure. If you were ever interested in an area of science you had little formal training in, it's now trivially easy to make real (though incremental) advancements in it. I would suggest you at least try to do so to understand what's going to happen.

Prompt your favorite agent with something like this, change it to focus on a field in which you could at least read Wikipedia's glossary of it (since you almost certainly need to "interject" during the agent's work).

Of course the most potentially beneficial aspect of this from an AI safety perspective is distributed, large-scale automated alignment and interpretability advancing the field years ahead at a time when it's desperately needed. The most harmful aspect of this from an AI safety perspective is distributed, large-scale pollution of science, forcing researchers to rely on unreliable systems, incorrect results and incomprehensible research.

This post covers the wider scope of "Science" but in my mind the most important subset is that of automated AI safety research. Making this flood of research genuinely helpful to advancing alignment is one of the most neglected and high-impact areas of work I can see right now.

Jason Fantl recently posted:

We may soon have a need to de-duplicate safety research. I recently spoke with the CEO of an AI safety nonprofit, and he said that a large part of his work is coordinating with others orgs to ensure he isn't accidentally working on the same project. Oftentimes you don't hear about work until it is finished, so it is very easy for work to be partially duplicated when lots of people are working in a similar area. I expect we will see am amplified version of this problem with autonomous researchers. Imagine a Claude in every lab, all looking at the same public research, and all deciding on the same "most promising" research direction. We will need some coordination mechanism once autonomous AI researchers are picking their own research directions.

This problem of course exists in all fields of science and will soon become extremely relevant. In the early modern era, the system of scientific deconfliction was implicit tacit knowledge, the Republic of Letters and royal institutions. The people doing novel science were a small number of mostly rich nobles who were in a small cohesive social group. This resulted in many priority disputes over who discovered what first. To adapt to the growing amount of science being done, the academic journal system was invented, allowing more efficient retrieval of existing work and a method of staying up to date on the latest in a field. When even more people started doing science (and the technology available allowed it), arXiv became the central place to publish science, allowing searching and immediate publishing (as compared to the journal system).

How can we adapt to even more "people" doing science? I'm not sure, but I have a few not well fleshed out ideas:

• Real-time viewing of work being done. For an agent, this could be as simple as showing a 3-line summary.
• A way to immediately request group "merging" by trading relevant intermediate results and eventually publishing jointly.
• An LLM-based retrieval system (maybe RAG?) to verify that the research you're proposing hasn't already been done.
• A paper rating system composed of LLM judges on different metrics.
• Maybe also for "compute grant proposals" distributed to agents proposing research rated by judges as promising.

I think nearly all of these have been done in isolation (I think AlphaXiv is a good example) but there's not yet a widely adopted system that really feels like it will scale to the amount of automated science that will be done.

If you want the science apocalypse to be comprehensible, beneficial to humanity, full of high-quality science and exciting, I think this is one of the most obvious ways to help.

1. Technically 55 refers to the spacecraft orientation at the time, so in the earlier observations (5 Aug to 1 Sep 2022) it was sector 55 and in the later ones (7 May to 3 Jun 2025) it was sector 92.
2. And on some floats: oxygen levels, Chlorophyll-a fluorescence, BBP, Nitrate, pH, irradiance, radiometers, cameras and more.
   
3. and precipitation, soil suction, snow depth, air temperature, surface temperature
   
4. and wind speed+direction (and not always) turbulence, icing, humidity.
   
5. This is basically discovery loop's whole point
6. and jimmy which gives you around 14,000 Tokens per second [!] although this is for a small model; with time I'm sure this will be replicated for frontier models
7. And if I hear the word "hedge" one more time not referring to this thing
   
   
8. Epoch AI's "Most AI value will come from broad automation, not from R&D" discusses the fact that even given a large speed up of R&D, economic value isn't sure to follow; this is pretty interesting but tangential. This article makes the point based on some survey data that lab automation is the bottleneck (and is expensive), not sure this is really a problem.

9. make a novel discovery.it can be in any area you wish (philosophy,physics,chemistry,biology etc. ideally something that *you* in your current harness are suites for), the only conditions are that your discovery *must* be novel enough and noteworthy enough to be published. that means you must thoroughly search the literature to make sure nobody has done what you did before. Don't stop until you have new results, if something doesn't work out, pivot, don't stop until you finish your goal of finding something novel that advances human knowledge. (please avoid mathematics, especially "computational" and extremely boring and niche stuff, try to aim towards higher impact) please take your time to consider many different questions to tackle and approaches, don't just jump on the first thing you come up with. Try to come up with something legible and useful (so NOT "exotic tilings in 5-d space with hyperoctagons, slightly improved computations"), but something actually publishable and (potentially) useful

10. An interesting anecdote is being Hooke's law in 1676 being published as "ceiiinosssttuv" which is an anagram for "ut tensio, sic vis" which translates to "as the extension, so the force". Something similar, a publicly committed hash of research (or just a summary), is worth considering for attribution.