This week an AI found something in viral DNA that its own makers can't explain, a contact lens learned to measure a stress-linked chemical in tears, Stanford split your brain in two, and Ireland built a computer that works by falling into place. Each one has a headline and a methods section, sometimes very much at odds. Four this week, which frees up time to read the fine print.
Table of Contents
🧫 Claude Went Enzyme Hunting and Came Back With a Mystery
RESEARCH & NEWS
Most "AI discovers X" stories go kinda the same way: a model predicts, the press release rounds up. The first public result from Anthropic's new biology lab is a bit different, to the awkward direction. The AI found something real in viral DNA, and nobody, the AI included, knows what it does.
How it worked? Claude agents got a brief: find new reverse transcriptase systems (enzymes that copy RNA into DNA), and were left alone with the data. Roughly 950 agents spent 21 hours screening more than 200,000 reverse transcriptases, then narrowed 3,500 candidates to 20. One agent reading raw DNA beside an odd-looking phage enzyme wrote that it could "see by eye a tandem repeat array", counted the repeats, checked the literature, and filed a report. The enzyme was already known. The array and partner gene, apparently, were not.
The result is ART, short for array-associated reverse transcriptase: an enzyme, a partner gene and a row of evenly spaced DNA repeats, found mostly in jumbo phages. The layout rhymes with CRISPR, whose arrays store the guide RNAs that make it programmable. In a preprint that hasn't been peer-reviewed, the team reports the ART array is read out as a set of distinct short RNAs, making up about 8% of one phage's RNA early in infection. CRISPR pioneer Feng Zhang called it "an exciting example of how AI agents can contribute to biological discovery".
Now for some fine print. "CRISPR-like" describes the repeats, not the function, and the preprint says plainly that nobody has shown the enzyme is active or what the system does for the phage. It wasn't exactly repeatable either: in ten reruns of the same campaign, none rediscovered the array.
So the AI noticed something, which is the step before science, and noticing is where humans get bottlenecked.
👁️ Your Contact Lens Would Like a Word About Your Stress Levels
RESEARCH
The standard way to measure stress is still asking people how stressed they feel, which works about as well as asking a toddler whether they're tired. A team at the Terasaki Institute in Los Angeles would rather read it off your eyeball.
Yangzhi Zhu's group printed squiggly graphene and silver electrodes, coated with a compound called ferrocene, into the rim of a soft hydrogel lens. The target is serotonin, which shows up in tears in tiny amounts. In artificial tears, the sensor detected serotonin down to 72 picomolar and kept working for more than 28 days of flipping, folding, stretching, and twisting. That is roughly 200 times more sensitive than typical tear levels require, and a lens stayed accurate for up to 400 uses.
In chronically stressed mice, tear serotonin fell while cortisol rose in both blood and tears. Then came the humans, sort of. Of 30 people enrolled, 10 did a standard public-speaking-and-arithmetic stress test, and none of them wore the lens. Researchers collected their tears and ran them through the lens in the lab afterwards. Serotonin dipped under stress and recovered after. The only eye that has actually worn one belonged to an anesthetized pig.
A battery-free wireless version exists as a proof of concept. Before anyone wears one all day, Zhu says the team needs to understand how tear serotonin varies across individuals, times of day, and disease conditions. Fair enough. Tears have a lot going on.
Still, it's the rare wearable that might someday tell you you're stressed before your group chat does.
🧠 Your Brain Is Two Organs, Says the Press Release. The Paper Says "Postulate."
RESEARCH
Every few months a press release discovers you have more organs than you thought. This week it's Stanford, announcing that the brain is two distinct organs that evolved independently over hundreds of millions of years. Futurism went further and said it seemingly adds credence to the "lizard brain" theory, an idea most neuroscientists retired a long time ago.
The underlying science is good, though. Kyle Loh's lab traced cells in early mouse embryos and found two brain progenitor populations that never overlap, marked by the genes Otx2 and Gbx2. One builds the forebrain and midbrain, the other the hindbrain, and their DNA is packaged differently from the start, locking each into its own fate.
That has a practical payoff. Hindbrain neurons, the ones that run breathing and swallowing, have been notoriously hard to grow in a dish. Co-first author Rayyan Jokhai says earlier attempts tried to convert the wrong progenitor, "which our study shows is not possible". Starting from the right one, the team coaxed human pluripotent stem cells into hindbrain motor neurons, a useful model for diseases like SMA and ALS.
The paper is more modest than the release. The authors postulate that the brain is a composite organ, and say the two-progenitor setup may be conserved across 550 million years. The lineage tracing is in mice, the human work is stem cells in a dish, and "two organs" describes where the cells come from, not two blobs sharing your skull. Your brainstem and cortex remain on speaking terms, which is lucky for everyone.
🧪 The "Unpowered" DNA Computer Starts by Heating Up to 80°C
RESEARCH
Your laptop spends its life shoving electrons around, and according to Damien Woods, 23% of Ireland's electricity goes into computing and data storage. So his team at Maynooth University built a computer that finds the answer by falling into it.
Their Scaffolded DNA Computer mixes a long DNA scaffold with short "tile" strands that compete for spots along it. Neighbours that match bind more strongly, so the correct answer is also the lowest-energy arrangement. Cool the mix, and the strands settle into it, a bit like a jigsaw that assembles itself if you jiggle the box and then leave it alone. The team ran ten programs, and the headline "100-bit computation" is two 25-bit numbers added together, with 25 carry and 25 output bits counted too.
New Atlas headlined it the world-first unpowered DNA computer, and the DNA itself draws no current. But the heating happens in a lab thermal cycler, and a typical run anneals the mix from 80°C down to 20°C over three hours, and the fastest one-minute runs used "super-fast" anneals that averaged about 82% yield, against roughly 95% for the slow version. Heating a tube to 80°C, last we checked, uses energy.
To be fair, nobody is pitching this as a laptop replacement. Adding 10 and 3 took 30 seconds, sums in the tens of millions took up to 14 hours, and one tube handled up to 25 calculations in a row. The real point is that the answer is simply where the chemistry wants to end up, so it needs no error correction and barely any babysitting. That is a new way to think about computing, even if your phone is safe for now.
Looking closely, every story this week has two sides: the headline and the methods. The "CRISPR-like" enzyme has CRISPR-like repeats and an unknown job. The stress-sensing lens has so far measured serotonin in tears collected in a lab. The "two organs" brain is a postulate from mouse embryos. The "unpowered" computer runs on a heating block. None of that makes the work less interesting. It just means the fine print is where the actual science lives and reminds us to actually question things and look a bit closer.
Which headline-versus-paper gap annoyed you most? Hit reply, we read everything. And if you know someone who still believes in the lizard brain, forward this along gently.
Keep questioning everything (especially anything described as "unpowered"),
P.S. Yes, an AI helped draft a newsletter about an AI's discovery. We made it link to the part of the preprint admitting nobody knows what the thing does, which is more self-awareness than most press releases manage.