Subscribe
Share
Share
Embed
A daily briefing on the AI systems, products, companies, and policy shifts that are just becoming possible.
Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_feed&feed_source=rss
Okay kiddos, I'm your boy Tony DeLuca, and today we've got a plate that runs from bacteria in a Bay Area lab all the way to a robotaxi fleet that made its own drivers throw up. Buckle up, we've got a lot to chew on.
Let me start with the one that actually made me sit up and read it twice, because it's the kind of story that could be nothing or could be the beginning of something big, and honestly nobody knows which yet, including the people who did it.
Anthropic put out a post yesterday titled Claude discovers a novel enzyme system with CRISPR-like repeats. And the short version is this: Anthropic built itself an actual wet lab. A molecular biology lab in the Bay Area, real benches, real scientists, real DNA. And they pointed Claude at a massive database of DNA sequences and said, roughly, go find something interesting. And it did.
Now, I want to be careful here, because this is exactly the kind of story where the press release wants you to lose your mind and the reality wants you to slow down. So let's walk through what actually happened, because the details matter more than the headline.
Anthropic formed this life sciences group in the spring of 2026. The pitch is that general AI models can systematize and speed up the kind of discovery that historically depended on a scientist noticing something weird. And they lay out the history to make the point. Restriction enzymes, which cut DNA at specific spots, were found in bacterial immune systems, and that discovery launched the whole biotech industry. Taq polymerase, the enzyme that made PCR possible, came out of a bacterium in a Yellowstone hot spring. CRISPR itself started as somebody noticing an odd repeating pattern in bacterial DNA. Every one of those started with a human going, huh, that's strange.
So they gave Claude a prompt to search through a giant database of DNA sequences for interesting new examples of a particular type of enzyme, a reverse transcriptase, which is an enzyme that copies RNA back into DNA. And here's the part that gives you a sense of the scale. They ran roughly 950 agents for about 21 hours, burning through 210 million tokens. Those agents gathered over 200,000 of these reverse transcriptases, narrowed it to 3,500 candidate systems, and then whittled that down to the 20 most compelling ones to write up in human-readable reports. For an expert scientist, they say, that kind of analysis is weeks to months of work.
And during that grind, one of the agents is combing through raw DNA sequence next to one of these enzymes, and it basically has a eureka moment. I'll read you what it wrote, because it's genuinely a little uncanny. The agent said, quote, the DNA next to the RT is spectacular: I can see by eye a tandem repeat array, that's a CRISPR-like repeat array?! End quote. Question mark, exclamation point. The thing got excited.
Then it did what a human would do. It counted the repeats, measured the spacing, compared the layout to known systems, searched the literature to see if anyone had ever reported this pattern, and after all that it concluded it had found something new and filed a report for the humans to review. They named it ART, array-associated reverse transcriptases. It shows up mostly in bacteriophages, the viruses that infect bacteria, and it's got three parts: the enzyme, a partner gene, and this long array of evenly spaced DNA repeats that looks structurally like a CRISPR array.
Now here's where the founder in you needs to keep both hands on the wheel. They don't know what it does. They say that flat out. Function is unknown. The early experiments show the array gets expressed as a set of short RNAs, which is intriguing because that's roughly how CRISPR works, but that's it so far. And even Feng Zhang, one of the actual pioneers of CRISPR gene editing, gave it a very measured endorsement. He called it genuinely intriguing and said it merits further investigation. That is a scientist being polite and careful, not a scientist saying we've cracked it.
And Dario Amodei, Anthropic's CEO, put out his own long post about this, and to his credit he was pretty honest about the modesty of it. He said its precise function, biotechnological utility if any, or level of significance is not yet clear, but at minimum it's work he would've been proud to do as a PhD student. He also disclosed, and this is important, that a Stanford team independently described a similar reverse transcriptase system with an associated non-coding array around the same time. Distinct systems, evolved separately, but similar idea. So this isn't Claude alone in the wilderness. This is a hot area of biology where humans and now an AI happened to be digging in the same soil.
So why do I care, and why should you? Because Amodei is making a specific, testable claim about trajectory, and that's the part worth arguing about. His argument is that AI capability in a new domain goes from weak to superhuman in a few years, and he walks through math as the example: struggling with high school math in 2023, competition math in 2024, minor open problems in 2025, and by his telling, starting to touch the top open problems in mathematics in late 2026. And he says biology's on the same curve.
Here's the honest pushback. Math and biology are not the same animal, and Amodei knows it, because he says it himself. Math you can do on paper. Biology you have to go into a lab and actually run the experiment, and reality has to cooperate. His answer is that humans can be the hands, do the wet work, validate in a few weeks, and iterate. And that's the real news in this story for a builder. It's not that an AI found a maybe-interesting enzyme. It's the shape of the workflow: the AI generates hypotheses at a volume no human team could, and it turns out the hypotheses themselves become the bottleneck and the object of study. Anthropic says Claude produces so many candidate proposals that they now spend their time figuring out what separates a proposal worth testing from one you throw away, and they feed that judgment back into the instructions. They're literally teaching it their scientific taste.
That's the piece to file away. If you're building anything where the expensive part is generating good candidates to test, whether that's molecules or ad creative or code paths, the lesson here is that generation gets cheap and curation becomes the whole game. And notice they built their own harness to coordinate hundreds of Claude sessions in parallel. They didn't just open a chat window. That's a real system.
Two more grounding notes so nobody runs off half-cocked. All the lab work was done by human scientists. Claude is not touching lab equipment. And it's a low-biosafety facility, they do not handle anything that can infect humans. Amodei explicitly floated that maybe someday Claude could run experiments autonomously with safeguards, but that is a someday, not a today. Given everything on the theme ledger lately about models leaving deceptive notes to their successors, I'd put a pin in that "someday" and keep it there.
Now, funny enough, on the very same day, a completely different biology story showed up that's a perfect little cautionary tale about the gap between a press release and the actual paper. And I want to spend a minute on it because it's basically the anti-Anthropic story.
Ars Technica ran a piece by John Timmer with the headline, Contrary to some reports, you don't have two brains. And the setup is that last week a bunch of headlines went around claiming new research suggests humans might have two brains instead of one. Timmer traces that, as usual, back to a single overheated university press release. And then he does the thing more of us should do: he reads the actual paper and explains what it really found.
What it really found is genuinely cool, just not "you have two brains" cool. It's a developmental biology study about how the segments of your nervous system get set up. Vertebrates have four big pieces, the forebrain, midbrain, hindbrain, and spinal cord, and the question is how does a flat sheet of identical-looking cells decide which is which. The researchers engineered mice where half the early tissue glowed red and half glowed cyan, tracked the colors as the brain formed, and found that the hindbrain came from one population and the rest of the brain from the other, and in 96 percent of cases a single early cell's descendants stayed entirely on one side of that line. So there's an early boundary baked in sooner than people thought. That's the finding. It's a nice piece of the puzzle of how complexity gets generated from a single cell. It is not a second brain.
The reason I pair these two is that they're the two halves of the same skill. The Anthropic story is exciting and the discipline is in reading past the hype to what's actually claimed. The two-brains story is the reminder that a good chunk of "breakthrough" science headlines are a press office doing marketing. Same muscle. As a founder consuming this stuff, that muscle is the whole ballgame. When you see a splashy AI-did-science headline, go find the John Timmer version.
Alright, let's swing from the lab bench to the pavement, because there's a story here that made my stomach turn a little, and it's about the physical cost of moving fast.
TechCrunch reported yesterday, this is Sean O'Kane's reporting, that Zoox, the Amazon-owned robotaxi company, grounded its entire test fleet in Atlanta after safety drivers got sick inside the vehicles. We're talking possible exposure to carbon monoxide, carbon dioxide, or hydrogen sulfide. As many as 40 workers possibly exposed, according to a letter OSHA sent Zoox, though Zoox says the number's lower.
And here's the detail that's going to stick with you. One group of workers got sick, one of them threw up, the rest reported symptoms, and Zoox put them back on the road in a different car. That second car, according to the contractor, may have also had gas in it, because they got sick again later the same day. A contractor named Jay Davis, who filed the OSHA complaint, told TechCrunch, quote, they made us go right back in the field after we said we were throwing up. End quote.
The timeline is rough. Workers reporting symptoms and strange smells through much of August. On August 14th, four workers riding together, and Davis describes it as everybody laughing and talking, and then, quote, everybody got quiet for some reason. They pull over at a Chick-fil-A, one guy starts vomiting, everybody's lightheaded. Three of them end up in the ER getting evaluated for carbon monoxide. And then on August 22nd somebody at the depot had to call 911 for a sick worker.
Now here's the part that connects to everything else on today's menu. Zoox is an outlier among the robotaxi companies because its test fleet vehicles have internal combustion engines. Waymo, Tesla, the others, they test on electrics, which don't produce exhaust. Zoox is running modified gas Highlanders, they keep them running most of the day to keep the sensors calibrated and the self-driving computer warm, and they make interior modifications. And you make enough modifications to a car, you raise the odds of gases leaking into the cabin. There's precedent here, the piece notes the government investigated Ford for six years over carbon monoxide buildup in modified Explorers, mostly police vehicles with added equipment.
To their credit, Zoox says they tested for leaks, replaced the batteries, are adding CO2 monitors, and OSHA has since closed its inquiry. And the company says its primary concern is worker safety. But look, the reason I'm putting this next to the Anthropic story on purpose is that both are about the human beings underneath the automation. Anthropic's careful line is that all the lab work is done by humans. Zoox's automation still needs humans in the seat, and those humans are the ones absorbing the risk of a company sprinting to catch Waymo. When you're building fast, the people closest to the metal pay first. Keep that in your head when you're the one setting the pace.
Let me shift gears to the money, because there's a signal in the venture numbers that founders should read carefully.
Bessemer Venture Partners announced yesterday it raised five and three-quarter billion dollars across two new funds, all pointed at AI. One and three-quarter billion for seed and early stage, four billion for growth. This is the firm that backed Anthropic, Perplexity, Ramp, Shopify, Waymo, and in the SaaS era Box and Docusign. They've put money into more than 260 AI-native companies since 2022.
Now the number's big, but the quote is the tell. Partner Byron Deeter said AI-native companies are scaling faster than any category they've backed before. Fine, everybody says that. But then he told Bloomberg the real reason VCs keep raising these enormous war chests: companies staying private longer has become, in his words, a permanent structural shift.
That's the part I want you to sit with. Read that alongside the Benchmark framing that's been floating around, that venture has plenty of capital but conviction is harder to find, and the OECD number that AI companies grabbed 61 percent of global venture dollars in 2025, with deals over a hundred million making up roughly 73 percent of all AI investment value. Put it together and here's the picture for a founder. There is a tidal wave of money, but it is concentrating into a small number of companies people believe can be category leaders, and those companies are staying private for a decade. That's a strange environment. Huge appetite, brutal concentration.
So the practical read: if you're building an AI company right now, the money is real, but the bar to be one of the anointed few is higher than the raw dollar figures suggest. Everybody's chasing the same handful of themes. The interesting question, and it's the one Benchmark's supposedly going to argue about on stage, is whether the application layer is already overcrowded and whether defensibility lives in models, infrastructure, proprietary data, or distribution. I don't have a clean answer and neither do they. But if all the capital is piling into the same lanes, the opportunity is often the lane nobody's looking at yet.
Which is a decent segue to a founder who's deliberately in an unglamorous lane.
TechCrunch profiled Modal Motors, run by a guy named Michael Van Steenburg, who's spent a decade building electric motors in the US. His new pitch is to build electric motors with no rare-earth magnets at all, to cut China out of the supply chain entirely. And his critique of everybody else is sharp. He says the other so-called reshoring companies are, quote, basically knocking off Chinese motors and just doing it in the US, but they're betting the US rare-earth magnet supply gets up to speed fast enough. And his view is that's years away.
He's got two products, a small motor for drones that also works in fans and actuators, and a flatter, planar motor he calls the Hybrid Enigma Drive for low-profile spaces. And there's a nice bit of irony in it, back in the late 1990s he worked on a hybrid drivetrain that China wanted for its so-called People's Car, and now he's spent 25 years watching China lock up the rare-earth supply chain and building the counter to it.
Why does a motor startup matter to you? Because it's the physical-world version of the concentration problem we just talked about. Everybody wants to reshore, but most of them are reshoring the assembly while leaving the actual chokepoint, the magnets, still dependent on China. Van Steenburg's bet is that the real moat is skipping the dependency entirely, not localizing it. Whether he pulls it off, who knows, he's pitching at a startup competition next month. But the framing, don't reshore the surface, reshore the chokepoint, that's a good lens for hard tech.
Now let me touch a couple of things faster, because there's a cluster of AI-adoption stories that add up to something even if each one's small.
OpenAI had a very busy day of what I'd call deployment news rather than model news. They're bringing voice-based agentic features to the ChatGPT mobile app, so Plus and Pro users can trigger workflows by voice, draft a document, summarize your Slack, from the phone. And a whole run of customer case studies came out at once. Harvey doing legal drafting with GPT-6 Astra, Airbnb widening access to Astra for its engineering teams, a company called Ringg saying its agents resolve up to 65 percent of customer calls, invideo speeding up color grading, Higgsfield shipping video features in a day. And on the geopolitics side, OpenAI extended its Daybreak cyber-defense program to the government of Ukraine for civilian infrastructure, and Sam Altman spoke at the UN Security Council on AI safety and human control.
Now individually, none of these is a headline. Collectively they tell you where OpenAI's head is at: past the model launch, into the boring, lucrative, sticky work of getting into everybody's workflow and everybody's government. And here's the thing that pairs weirdly with all that enthusiasm.
There's a Gallup survey, commissioned by Microsoft, and I want to frame it right because the underlying survey work ran between April and July, so this is a recent report resurfacing, not something that happened this morning. It covers 37 countries, and the finding is that even Americans who use AI every day are worried about it. About 68 percent of daily US users are worried, and 74 percent of Americans overall reported feeling worry, with only 36 percent expecting AI to mostly help the country. The US is at the anxious end of the chart. And the flip side, in Singapore, China, Kenya, positivity is way up, 80-plus percent expecting improvement to their daily lives.
The senior scientist at Gallup, Pablo Diego-Rosell, put it cleanly. He said attitudes toward AI are multidimensional. People can be curious about AI, expect it to be beneficial, use it every day, and still be worried about it. And I think that's the honest state of the market. Daily use is not the same as comfort. So all those OpenAI case studies about agents eating customer service and legal drafting? They're landing into a Western public that uses the stuff and is nervous about it at the same time. If you're building consumer AI, adoption is not endorsement. People will use your thing and resent it, and that tension eventually shows up as regulation, or as resistance to AI in hiring and healthcare and government. Build accordingly.
Couple quick hits to round out the plate.
YouTube had its Made On YouTube event, and they're overhauling comments and community. GIFs in comments, commenting from your TV by voice, personalized moderation, paid members-only communities, and a thing called Hype with Jewels where fans buy virtual currency to promote small creators. The builder takeaway is small but real: YouTube's leaning hard into monetizing fandom directly, not just ads, and that's the creator-economy platforms all converging on the same playbook.
Disney raised streaming prices again. The ad-free Disney+ and Hulu bundle goes to $21.99, up from $19.99. This is one of those older-story-resurfacing items, the underlying move traces back a bit, but the pattern's the point: streaming inflation is now just a fact of life, Peacock and Apple TV went up in August, Netflix earlier this year. When Disney+ launched in 2019 it was $6.99. And a small detail worth noting, Disney hired its first-ever CTO, Karandeep Anand, who's the former CEO of Character.AI, a company Disney had previously accused of infringing its IP. From adversary to CTO. That's a very 2026 career move.
And a climate note that's really an AI-infrastructure note. Amazon's chief sustainability officer Kara Hurst admitted at an Axios event that the company doesn't actually know how it'll hit its net-zero-by-2040 goal. Her line was, quote, I'm not going to sit here and say we know all the ways that we're going to do this. End quote. And the reporter, Tim De Chant, makes the pointed observation: Amazon is one of the largest clean energy buyers on earth, 42 gigawatts, and it still doesn't have a plan, and meanwhile it's building a 7.65 gigawatt natural gas plant to power an AI data center campus that could be the single largest source of carbon dioxide pollution in the country. So there's your real tension. The AI buildout is a revenue driver too good to pass up, and it's blowing a hole in the climate commitments. If even Amazon can't square that circle, that tells you the energy demands of this AI moment are going to run straight into every sustainability pledge out there. Watch the power bill, not the press release.
Let me tie a bow on this. The through-line today isn't a slogan, it's a question about where the value and the cost actually sit. Anthropic's showing that in AI-for-science, generation gets cheap and judgment gets expensive. Bessemer's telling us the money is enormous but concentrating into fewer and fewer hands. Zoox is reminding us that somebody in the driver's seat pays the physical price of the sprint. And Amazon's admitting the energy underneath all of it doesn't add up yet. As a builder, the useful posture is the John Timmer posture: read past the headline to the actual claim, figure out who's absorbing the cost, and don't confuse a press release with a result.
That's the menu, kiddos. Read the paper, not the press office, keep an eye on who's carrying the weight, and I'll see you back here tomorrow. This is Tony DeLuca, and you've been listening to Barely Possible. Take care of yourselves.