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 the coffee's hot, the menu's loaded, and we've got a real spread of AI morsels to chew through today. Buckle up, because today we're leading with a story that oughta make every founder lean back in their chair and ask a hard question: what exactly are you buying when you buy an AI coding agent?
Let me set it up for you, because it's one of the more honest pieces of research I've read in a while. There's a recent report out — this is from a study covered on October ninth — looking at AI coding agents, and the headline on it is almost poetic. The agents generate more code, but not more software. Sit with that for a second. More code, not more software. That's not a slogan some skeptic cooked up on a podcast. That's the finding.
Here's the meat of it. Two Harvard researchers, Fiona Chen and James Stratton, got their hands on aggregated analytics from a company called Jellyfish — and Jellyfish is a tool that measures the granular output of engineering teams. We're talking three hundred million individual work events. Commits, pull requests, the actual plumbing of how software gets built. Plus issue-management data across more than seven hundred thousand employees at over seven hundred software development firms, running from 2021 all the way through March of 2026. That's not a survey where you ask a hundred developers how they feel. That's the actual exhaust of the work itself.
And they did this thing the economists love — a "difference of differences" regression. In plain English: they looked at each company before and after it brought in AI coding tools, and they compared the jumps. So what'd they find?
In terms of raw code? It's a geyser. Introduce AI coding agents at a firm and you get a thirty percent jump in total lines of code generated, a twenty percent bump in total commits, a twenty-three percent rise in pull requests. If there's one thing these agents are good at, it's churning out volume. Nobody's disputing that.
But here's where the record scratches. All that extra code did not translate into more finished software. The resolution rate for issues and epics — and that's the industry term for actual shipped features, the stuff tracked in Jira — did not change in any statistically significant way after the AI tools showed up. Didn't move. And they checked whether maybe the features just got bigger or more complex to make up for it — nope, no "compositional shift" either. Same work, same complexity, same throughput at the end of the pipe.
So where'd all that extra code go? It went into the review line, and it clogged it. The average review process — the time between a pull request getting submitted and it actually getting merged into the codebase — ballooned forty-nine percent on average after AI agents came in. Forty-nine percent. The share of pull requests where somebody said "no, go back and fix this" nearly doubled. Comments per pull request went up thirty-five percent. And firms responded by shifting fourteen percent more of their workers into doing code review.
Think about what that is. You put a firehose on the front of the factory, and the factory didn't get faster. The water just pooled up at the inspection station. The study authors call human code review the "bottleneck," and they say any efficiency gained in the writing phase gets — their word — "absorbed" by the downstream constraints.
Now, could AI help with the review too? Theoretically, sure. But the study found that so far that help's been marginal. By March 2026, eighty percent of these firms used some form of AI code review — but the AI agents were only responsible for about twenty-three percent of review comments and under eleven percent of pull requests. Humans are still doing the heavy lifting on the checking.
There was one comment in that piece I want to read you, because it's the realest thing in the whole story — from a developer going by "metavirus": "I'm forced to use AI at work and this is absolutely ten thousand percent true. Soooooo much slop code goes into the basin, which dozens of people need to sift through, strain out the dreck, then eventually close their eyes and release into the wild because now we have new SLAs so we're not seen as a bottleneck to progress." And then the kicker: "Great incentives, brought to you by brilliant executives who know fuck-all about what's actually going on."
Pardon the language, but I'm quoting the source, and honestly that quote is the whole episode in a nutshell.
So who benefits, who gets squeezed, and what do you watch next? Here's Tony's read, and I'm gonna be fair to both sides. First, the honest caveat the researchers themselves make: these agents are young. The study's data cuts off in March. The tools have gotten better since. And ninety-five percent of the firms in the study had only recently rolled agents out — they're still learning when and where to point the thing. So this isn't "AI coding is a bust." This is a snapshot of an awkward adolescence.
But if you're a founder, here's the actionable part. The constraint on your engineering org right now is probably not how fast you can produce code. It's how fast a trusted human can trust the code. If you go out and buy agent seats for everybody and your success metric is lines shipped or pull requests opened, congratulations, you're gonna crush that metric and your actual feature velocity won't budge — and your senior people will be drowning in review. The gain only shows up if you invest in the review side with the same seriousness you invest in the generation side. Better testing, better review tooling, real standards for what an agent is even allowed to touch. Otherwise you paid good money to generate more stuff for your best people to clean up.
And that's a theme worth sitting with, because it rhymes with something we touched on yesterday. We covered Satya Nadella calling for an AI "emergency brake" — externalized safeguards, tamper-proof logs, the ability to pause a model mid-task. Different problem, same underlying shape: the generation of AI output is racing way ahead of our ability to verify, govern, and absorb it. Yesterday it was safety. Today it's software throughput. The machine makes stuff faster than the humans can check it. Keep that pattern in your back pocket, because you're gonna keep seeing it.
Now let's shift from the factory floor to the showroom, because there's a funding story that's a perfect counterpoint to the coding piece.
A startup called TypeSafe AI — the makers of a model called Jev — just raised eight hundred seventy million dollars at a seven and a half billion dollar valuation. This is a recent report from October ninth. The round was led by Andreessen Horowitz, with Sequoia in and existing investor DCVC. And here's the thing that makes it interesting for builders: Jev is not a chatbot. It's built on a transformer architecture, but it is explicitly not a large language model. It doesn't output text. It outputs probabilities — what the company calls "calibrated decisions."
Now why does that matter? Because the pitch is the exact opposite of "more tokens, more code." The pitch is: for automation tasks, you don't need the thing to write you a paragraph. You need it to make a decision — yes, no, this bucket, that bucket, this probability — fast and cheap. TypeSafe claims Jev works significantly faster and uses far fewer tokens than a standard LLM. Co-founder Diogo Almeida — who was previously a researcher at OpenAI — put it this way: "We have been super good at human language for four years, but it's not useful for automation because computers speak a different language." The company was co-founded in 2024 with former Meta research engineer Sasha Sheng and an engineer named Erik Gafni.
And the adoption number is the eye-popper. They launched September fifteenth — weeks ago — and they claim a third of the Fortune 500 is already using the model. Now, take enterprise self-reported adoption with a grain of salt the size of a bagel, always. But even discounted heavily, that's a remarkably fast land-grab for something that came out last month.
Here's my builder's angle on this. For a year now the whole industry's been in a token-maxing arms race — bigger models, more reasoning, more output. And a chunk of the market has quietly been screaming "I don't need an essay, I need an answer." If you're automating a workflow — routing a ticket, flagging a transaction, deciding whether a document needs a human — you don't want a verbose, expensive, chatty language model. You want a cheap, fast classifier that gives you a calibrated probability and gets out of the way. Jev is a bet that there's a giant, under-served market sitting right next to the LLM gold rush, and that market cares about cost-per-decision, not eloquence. Whether a seven-and-a-half-billion-dollar valuation on a few weeks of revenue holds up — that's a different question, and I'd watch the retention numbers, not the launch buzz. But the product thesis is sound, and if you're building automation, it's worth asking whether you've been reaching for a sledgehammer when a tack hammer would do.
Alright, let's go from a model that makes decisions to a regulatory decision that just unlocked a whole industry. Now let's dig into self-driving trucks.
This is current, out of TechCrunch Mobility on October eleventh. Two self-driving truck companies — Aurora Innovation and Kodiak AI — just got a federal exemption that clears what was a genuinely goofy roadblock to commercialization. Here's the dumb little problem that was holding things up. Under federal rules, if a big rig breaks down on the shoulder, the human driver has to get out, within ten minutes, and physically set up those reflective warning triangles behind the truck. Makes sense — for a truck with a human in it. But a self-driving truck doesn't have a guy to hop out and lay down triangles. So what do you do?
Aurora actually took the federal safety regulators to court over this. Lost the first round, escalated it to the D.C. Court of Appeals. And now the Federal Motor Carrier Safety Administration has granted a five-year exemption: instead of roadside triangles, these trucks can use cab-mounted warning beacons — high-visibility lights built into the truck itself.
Daniel Goff, VP of external affairs at Kodiak AI, said the exemption will help the industry "usher in an autonomous era of freight movement on U.S. roads." And Gerardo Interiano from Aurora called the cab-mounted beacons a "critical, twenty-first-century solution that enhances roadside safety by immediately alerting other road users without ever needing to put a person in harm's way."
Now — why should a founder who isn't in trucking care about reflective triangles? Because this is a textbook case of the thing that actually blocks physical-AI businesses. It's not the self-driving smarts. Those trucks can already drive. Kodiak's got a new four-hundred-thirty-five-mile autonomous route running between Dallas and Laredo hauling refrigerated and dry freight. The technology was ready. What was blocking commercialization was a regulation written for a world with a human in every cab. And the way it got solved wasn't a breakthrough in AI — it was a lawyer, a court, and a federal agency agreeing to swap a triangle for a beacon.
If you're building anything that touches the physical world — robotics, autonomous anything, drones, devices — the lesson is that your roadmap has a regulatory critical path running right alongside your technical one, and sometimes the regulatory one is longer. The reporter asked a good open question too: could these cab-mounted beacons eventually become standard across the whole trucking industry, human-driven or not? That's how these things go — a workaround for robots becomes a safety upgrade for everybody.
And while we're in the mobility lane, one more number from that same briefing that founders should clock: Waymo went to the debt markets for the first time and pulled down a five billion dollar raise. PIMCO, Blackstone, and Sixth Street led, but the lender list was an absolute phone book — Capital Group, T. Rowe Price, Apollo, Fidelity, Franklin Templeton, on and on. Up till now Waymo leaned on Alphabet's balance sheet and venture money. Turning to debt — five billion of it — tells you they're scaling commercial operations into new markets in the U.S., Europe, and Japan, and that the appetite for lending against AV infrastructure is very much alive. That's the money side of autonomy growing up right in front of us.
Alright, let's switch gears from the road to the voice in your ear. There's a smart piece out of TechCrunch, current, October eleventh, where a couple of execs make the case that voice AI has not had its ChatGPT moment yet.
The setup: investors have poured billions into voice AI — model makers, enterprise customer service, meeting note-takers, dictation tools. Every week some new thing claims to sound human. And the CTO of an enterprise voice platform called PolyAI, Shawn Wen, basically says: slow down. He says yes, we've hit the milestone of full-duplex models — those are the ones that can speak while they're listening to you, so you can interrupt them like a real conversation. But he says the next challenge is making the reasoning fast enough that the model fetches an answer quickly and the conversation actually feels natural. His point about customer service is sharp: once the voice is good enough that a caller will engage for the first two or three turns, they start to build confidence, and over time they stop reaching for a human. The trust gets built in those first few seconds.
And then there's Alex Gay, the CMO at the meeting note-taker Otter, with what I think is the most important line in the piece — and it's about the unglamorous foundation. He said transcription was never the endpoint; it was the layer they build productivity gains on top of. But — and here's the quote — "if your original transcription didn't have the accuracy that you needed, all the follow-up actions that you have become flawed. And the minute that starts to take action, that is wrong. You lose trust in the platform."
That right there is the whole game for the agent era. If you're building voice agents that take actions — booking, canceling, updating records — your reliability is capped by your worst layer. And the worst layer is usually the speech recognition at the very bottom, missing a keyword, mangling a name. Garbage in, confidently-wrong-action out. Both execs also flagged the transparency piece — tools need to disclose that you're being recorded or that you're talking to an AI. Which, for a founder, isn't just manners, it's increasingly table stakes for staying out of legal trouble.
The builder's takeaway: voice is coming, but the companies that win won't be the ones with the slickest-sounding demo. They'll be the ones who nailed the boring accuracy at the bottom of the stack before they let the thing take actions on your behalf. Same lesson as the coding story, honestly — the demo generates confidence; the verification layer earns trust.
Now let me do a quick lightning round through some things a builder should know, because there's a lot on the menu today.
Apple. There's a current report, October eleventh, that Apple's set to announce a new smart home hub — Bloomberg's Mark Gurman says it's called the HomeView. A Siri-powered device with a display and an A18 chip, meant to run your other smart-home gear, coming in a countertop model and a wall-mount. Apple's also reportedly refreshing the Apple TV box and the HomePod mini, plus a partnership with LG for a smart lock, thermostat, and doorbell. Why care? Apple has whiffed on the smart home before — HomeKit never caught fire, the HomePod was a dud. This is their most serious swing yet, and if you build in the home-automation or ambient-device space, a credible Apple hub reshapes the whole competitive map. Watch the launch.
Apple again, different flavor — a current item about a reverse acqui-hire. Apple filed with the European Commission that it's bringing on team members and licensing tech from a personalized-audio startup called Huxe. Huxe was founded by folks who'd worked on the AI-generated podcast features in what used to be NotebookLM, now Gemini Notebook. Huxe shut down back in May, and Apple scooped up the talent and a non-exclusive license without buying the company outright. Why the dance? Because the reverse acqui-hire has become the go-to move for picking up AI talent and tech while dodging the antitrust scrutiny a straight acquisition would draw. If you're a founder thinking about exits, this structure is now a standard part of the playbook — know it exists, know what it does and doesn't get you.
And one from the science desk, because I find it worth a beat — there's a recent report, October ninth, that the winning video in Nikon's Small World in Motion microscopy competition got disqualified for undisclosed AI use. Researchers flagged that cell features appeared and disappeared in ways real biology doesn't do, and others spotted what looked like AI watermarks in the source files. Nikon investigated, pulled the win, and everybody down the list moved up a spot. The new winner's a fellow named Nguyen Nam Nhat out of Vietnam. Small story, big signal: every competition, every journal, every awards body is now scrambling to rewrite its rules around generative AI, and the provenance question — did a human actually make this — is becoming its own discipline. If you're building anything that touches content authenticity, that's a market forming in real time.
Let me close the loop with one piece of continuity and a little perspective. All week long on this show the throughline has been the gap between how fast these systems produce and how slowly we can trust what they produce. The false murder tip to the Philadelphia police we covered Saturday. Nadella's emergency brake yesterday. And today, hard data from Harvard showing AI writes a third more code but ships exactly as much software, because the humans at the review desk are buried. It's the same story wearing different hats: generation is cheap, verification is expensive, and right now the bottleneck is almost always the verification.
For a founder, that's not a reason to sit it out. It's the actual opportunity. The money this year isn't only in making the firehose bigger. It's in building the stuff that lets humans trust and absorb what the firehose produces — the review tooling, the accurate transcription layer, the calibrated decision model, the regulatory path that lets the robot truck legally pull over. The generation layer is getting commoditized fast. The trust layer is wide open.
That's the menu for today. I'm Tony DeLuca, this has been Barely Possible — go build something your reviewers won't curse you for, and I'll catch you next time.