0816 | Why I Remain a Skeptic: AI Leadership, Alzheimer's Surgery, and Yadda 3.0

Mia: You're listening to HackerNews Daily, I'm Mia.

Milo: And I'm Milo. Today we're digging into why a well-known developer remains a skeptic about a certain tech, how working with AI can feel more like leadership than coding, plus a fresh open-source release and more.

Mia: We've also got a story about an AI-assisted kernel that runs dramatically faster, a long hunt for a shell history bug, and a look at what haunts Unicode.

Milo: Plus some medicine, a space anomaly worth remembering, and a developer who discovered they didn't exist. Let's get into it.

Mia: Let's pick up with a piece written by a longtime open-source developer, Joshua Barretto, who has a post up called "I Remain a Skeptic." He opens with a blunt admission: he still doesn't use large language models for anything he actually cares about. And he wants to be clear that's not just because of the environmental, social, or political concerns — the ecosystem lock-in, the learned dependence on these tools, questions of self-sovereignty, and what he sees as the deliberate dismantling of labour bargaining power. For him, all of that loads the argument, but the core question is simpler and, in his view, far from settled: do these models actually work for building non-trivial software?

Milo: That's the part I want to dig into, because it's easy to wave off a skeptic as someone just uncomfortable with change. But his efficacy argument is specific. He points out that four years into this so-called revolution, the industry has almost nothing to show for it — software that isn't meaningfully better, faster, cheaper, or more secure, once you exclude poor-quality demoware — and no genuinely new ideas beyond what he calls AI boosterism.

Mia: Right, and he pushes further on productivity. A trillion and a half dollars in, he says, there are hardly any independent studies that confirm top-level productivity gains, and the more rigorous studies that do exist show only marginal or even negative results. Metrics like lines written or pull requests merged miss the point, because they measure output volume, not whether the output is actually more fit for purpose.

Milo: And he has first-hand corroboration. The pull requests that LLMs generate and send to his projects are, in his words, still rubbish — less repetitive and less stylistically consistent than human PRs, but no more useful, with a far lower share worth merging. Meanwhile, frontier-model code scans keep missing obvious issues that hobbyists spot immediately, even in libraries that sit inside major tech companies' dependency trees. So for him, not feeling like he's falling behind isn't stubbornness — it's a considered assessment that an open-source project people trust is worth more than a feature list. As he puts it, you might be able to replicate the function of my code, but so can a simple memory copy function.

Mia: So the takeaway there isn't that AI is useless everywhere — it's that the case for it as a productivity multiplier in real software work is far weaker than the hype suggests, and the people building and maintaining the tools that matter aren't all persuaded yet.

Milo: From there, let's contrast that with a very different framing. Allen Bargi has a short note out called "Working with AI feels more like leadership than coding," and his argument is that AI work is fundamentally less predictable than traditional software work — and that this unpredictability is actually the reason context, clarity, and feedback matter so much.

Mia: He grounds it in his own experience. For most of his career, code gave him certainty — a program did what its instructions told it to do, and if the same input produced a different result, that was a bug. People were never like that. As a leader, he could explain a task and get exactly what he asked for, or sometimes get something better, because a colleague understood the intent behind the request. And occasionally the result showed him he hadn't been as clear as he thought he was.

Milo: And he says working with AI feels closer to that second experience. The same request can produce a different answer, make a useful connection, miss an obvious point, or surprise him with an approach he hadn't considered. It's frustrating when he treats AI like a compiler — and far more useful when he treats the interaction as a form of collaboration, without anthropomorphising the tool, since it has no lived experience, accountability, or human judgment.

Mia: That reframes the skill set. Good leaders share context, explain the desired outcome, set boundaries, and respond to what comes back — and Bargi says those same habits improve his AI work. A good prompt helps, but a shared working context helps more. Examples, corrections, and reusable instructions cut down on misunderstandings, and over time the system gets better aligned with how he actually thinks.

Milo: So where Barretto sees an unresolved efficacy problem, Bargi suggests the real change is a shift in how you work: the discipline of explaining what you want and why, and then adapting to what comes back, is more like guiding people than like programming a deterministic machine.

Mia: And that second narrative is actually being put into practice right now. Author Stephen Cresswell has published Yadda 3.0.0 to npm — that's a JavaScript behaviour-driven development library that maps ordinary-language specifications to executable code, in the same family as Cucumber but much less prescriptive about how those specifications are written. And his more significant claim is that Claude wrote most of the release.

Milo: Let's separate the two parts of that, because they're both notable. First, the modernisation. Yadda 3 is now Node-only: it drops browser bundling and removes obsolete integrations like CasperJS, PhantomJS, Bower, and Component. The test suite moves to node:test, the code gets modernised to ES6, examples are refreshed to include Playwright and Puppeteer, and it ships TypeScript definitions. All of that is meaningful housekeeping for a library that's been around a while.

Mia: And then the more interesting claim — that most of that release was written by Claude. Cresswell says he worked through Claude Code with Opus 4.8, starting from a development plan — what he calls a Yadda 3 epic — that Claude itself wrote. That epic broke the work into deliberately separated phases: remove obsolete functionality first, then update the toolchain, keep mechanical formatting separate from behavioural changes, modernise the source, explore API changes, update examples and continuous integration, and only then finish metadata, documentation, and type definitions.

Milo: The structure is worth dwelling on, because it's exactly the kind of careful planning that's too easy to skip. Each phase was planned before implementation. Cresswell says he made very few interventions — fewer than a human contractor would likely have needed — and that Claude made remarkably few mistakes and even identified subtle edge cases on its own.

Mia: He credits one thing above all for that: Yadda's existing comprehensive test suite. Having solid tests meant Claude's changes were continuously checked against real expectations, which is precisely the sort of guardrail that separates a genuinely Useful AI-assisted release from the kind of low-quality LLM output Barretto was complaining about.

Milo: So across these three pieces you get a real tension. Barretto says the evidence for AI productivity gains is thin and the outputs are still often unfit for purpose. Bargi says the way to make them useful is to treat the work as leadership — context, clarity, feedback. And Cresswell offers a concrete case where that recipe was followed, with strong tests and a phased plan as the scaffolding, and it produced a shipping release. The question going forward is whether that recipe can become the norm, or whether it stays the exception.

Mia: So the first one is a pretty wild story from a GPU coding contest. A developer going by sankalp wrote up how he used OpenAI's Codex — the coding agent — to essentially do auto-research for a kernel optimization challenge run by GPU Mode. The task was to implement a batched QR decomposition that matched PyTorch's compact representation, and it was judged not just on correctness but on runtime across a range of matrix sizes.

Milo: And how did the agent-powered approach actually do?

Mia: He placed twelfth out of 183 participants, with a 232-times speedup over the baseline. And that came after more than 1,500 submissions spread over 14 days. The checker verified the math — that the QR factorization was sound, that the orthogonal factor was properly orthogonal, and that things reconstruct correctly — then ranked the correct submissions by geometric mean runtime across shapes like 512, 1024, 2048, and 4096. Low-bit internal formats like FP16, FP8, and NVFP4 were allowed to speed things up.

Milo: And I'm guessing the tools made that many submissions practical?

Mia: Exactly. GPU Mode's popcorn CLI let agents test, benchmark, and submit directly, and there were essentially unlimited spaced submissions allowed. At one point the workspace actually ran out of Modal credits because everyone was hammering the submission system at once. And the author is quick to call himself an underdog here — about a year of Triton-focused GPU knowledge, a bit of CUDA, no professional kernel work. The entrant just above him on the leaderboard, going by the handle CUDA Colonel, is a principal engineer at NVIDIA.

Milo: That definitely reframes what's impressive here. It sounds like his core argument is about domain knowledge making you a better prompter.

Mia: Right. The post argues the better you know a domain, the better you can prompt LLMs, because you're converting what he calls unknown unknowns into things you can ask about. He even has a section on introducing idea diversity to escape local maxima in the search. The Hacker News discussion was mostly experienced people confirming the result plus a few open questions.

Milo: And there's one detail in the thread worth keeping: someone deliberately tested a fresh set of model releases — described as definitive new versions — on a semi-abandoned video compression codec, chosen on purpose because its authors ship a bit less polished work. A nice real-world stress test for that very argument.

Mia: Next up, a debugging story that a lot of shell users will feel in their gut. Michael Stapelberg detailed tracking down a Zsh history data loss bug — a problem where commands he was sure he had run would sometimes just be missing from his saved history. Ctrl-R reverse search would come up empty, and the file would only hold very old entries, with years of newer ones simply gone.

Milo: That's the kind of thing that's genuinely unnerving, because you can't always tell when it happened. Did he find visible corruption?

Mia: No. No non-printable characters, no incomplete lines — nothing obviously wrong with the file. He initially restored from daily backups, but the problem kept coming back. His config loads a reasonably sized history, saves it with a huge save limit, appends incrementally, and explicitly unsets share history — which is notable because NixOS enables sharing by default.

Milo: So how do you even begin to hunt down something that silent?

Mia: He worked through inotify, fatrace, strace, and bpftrace before landing on a winning strategy: patching Zsh to crash loudly and then analyzing the core dump. The traced mechanism is genuinely sneaky — Zsh reads the old history file, writes its contents out to a brand-new file, then renames that new file over the old one, which deletes the old file in the process. So under the right timing, history silently gets dropped.

Milo: And presumably there's a fix in the pipeline?

Mia: Yes — Zsh 5.9.2, released July 2026, contains the fix, tied to upstream issue 53454. The article even has appendices, including one on an accidentally exported history-file environment variable and one on whether AI could have found the bug. The commenters led with firsthand experience. One person said they'd felt like they'd lost history before but never investigated; another who keeps nearly a decade of Zsh history thinks they may have been hit without noticing. And the author's mention of accidentally exporting the history file made one reader realize a mistake they'd been making for years — which they called both happy and sad. One more said roughly 98 percent of their history is only useful through search, so a silent bug like this is exactly what scares them.

Mia: And finally, something a little more unusual for Hacker News — a story about ghost characters in Unicode. It's based on a post by Paul O'Leary McCann on his blog, tracing what are called ghost characters back to the 1978 JIS X 0208 encoding, established by Japan's ministry of economy, trade, and industry, which still serves as a reference for all Japanese encodings.

Milo: Ghost characters — so characters that exist in the standard but have no known origin?

Mia: Exactly. After the standard was released, several added characters had no obvious source — nobody could tell what they meant or how they should even be pronounced. A 1997 investigation cleared most of them up by interviewing the catalogers. Some were found to be inadvertently invented as mistakes. One example: a character was an error introduced while trying to record a particular combination of two components drawn as one over the other. The two parts were printed separately, cut out, pasted onto a sheet, and copied — and the seam between them got read as a stroke of the character. That's how a genuine error became a standard character.

Milo: So they solved the mystery for most of them. But one remained?

Mia: One core ghost character had neither a clear source nor historical precedent. The most likely explanation is that it was created as a misreading of a different character, but no specific incident was ever uncovered. A common documented source for the others was a national administrative districts place-name list, whose latest edition is a seven-volume set with each volume roughly nine hundred pages — and it has no page references at all. On top of that, these JIS ghosts made their way into Unicode, which has its own separate set of ghost characters introduced during CJK unification. The related notes are wonderfully specific: the ghost character showing up in a digitized century-old newspaper because of faded printing, a wiki treating each ghost character as a youkai — a spirit — name, and an artist's hand-printed book made entirely of invented characters.

Milo: That's the thing I love about this story — it shows that even our most standardized, carefully engineered systems carry artifacts of human error, and that a character can outlive every record of how it came to be.

Mia: We're starting with a story today that reads almost like science fiction, but it's a real surgical procedure, and it's stirring real debate among Alzheimer's researchers. A feature in the journal Nature, written by Elie Dolgin, looks at a procedure called deep cervical lymphatic-venous anastomosis, or dcLVA. The idea is to connect tiny lymphatic vessels in the neck, part of the system that drains waste from the brain, directly to nearby veins. The theory being that fluid and waste proteins could flow more easily out of the brain and into the bloodstream.

Milo: And how did this get on the map? The procedure wasn't exactly new. It was first described back in 2022 in a Chinese-language journal by a microsurgeon named Qingping Xie, who heads a hospital in Hangzhou, China. But it barely drew attention until another microsurgeon, Wei Chen, at the Cleveland Clinic, started showing video of patient recoveries at surgical meetings. Chen recalls that a lot of jaws dropped, and soon there was a frenzy of this surgery being performed, as he put it, left and right.

Mia: And where was all this happening? Almost everywhere, it was in China. Hundreds of hospitals offered the surgery, thousands of people sought it out, many of them paying more than two hundred thousand yuan, which is about thirty thousand US dollars. Demand was fueled by viral testimonial videos and marketing on Douyin and WeChat. As you might expect, that raised alarms. Chinese regulators eventually stepped in and restricted the procedure to formalized clinical research settings. And Dr. Xie, the surgeon who pioneered it, has been in detention since September for reasons that haven't been disclosed.

Milo: So what's the actual evidence here? The piece anchors on one featured patient, a man in his eighties, who was described as alert within three days of surgery, and by eight months out, walking and reciting a Maoist anthem from memory. That's compelling, but Nature spoke with more than two dozen researchers, and the community is genuinely divided. There are real risks, too, including infection, bleeding, and injury to nearby nerves. The headline on this one is that what looks like a miraculous reversal still has researchers arguing over whether this is a breakthrough or a risk not worth taking without much stronger evidence.

Mia: From a surgery debate, let's move to another hotly discussed topic on Hacker News, and this one is about semaglutide, the weight-loss drug, and its apparent link to a lower predicted risk of dementia. A study connecting the two sparked a wide-ranging comment thread. It opened with someone noting something fairly basic, that diabetes itself is a well-known risk factor for dementia. And that quickly widened into a debate about what's actually driving that connection.

Milo: Some commenters pointed at diet. One framed it as Western disease and sugar-heavy diets broadly, calling semaglutide a countermeasure, not yet a vaccine, but maybe in due time, possibly via gene therapy. That same person added that fructose also appears to help cancer metastases spread. But then someone else pushed back hard on the sugar angle, arguing there's no such thing as acquired diabetes in physically active people, and that being sedentary is the real culprit.

Mia: Right, and the debate kept going from there. Someone challenged the idea of a Western-versus-Eastern distinction at all, pointing out that sugar-heavy diets are present everywhere people can afford them. And another commenter took issue with reading this as an indictment of sugar in the first place, noting the doses in these studies are always quite high and that mice are rarely a reliable model. That same person made a more subtle point, that the study also showed glucose actually inhibits the behavior, which suggests the real issue might be the ratio of fructose to glucose, rather than fructose alone. On that reading, table sugar, which delivers equal amounts of both, would basically be cleared by the study.

Milo: And the thread landed somewhere more moderate. One commenter argued that being sedentary is also a factor, that many interconnected factors are probably in play, and that GLP-1 receptor agonists, the class semaglutide belongs to, seem to help with more than one of them. They conceded that unpicking the exact chains of causation is likely extremely complex, but the general picture continues to look good. So there's no clean villain in this story, but the direction of the research, at least, seems encouraging.

Mia: And to close out this segment, let's talk about ticks, because there's a new product on the way that could change how people deal with them. Smithsonian Magazine, in a piece by freelancer Kristen A. Schmitt, reports on a test called LymeAlert, billed as the first at-home test for infected ticks. It's expected to hit the market in August, and the hope is that it becomes a medicine cabinet staple in the fight against tick-borne illness.

Milo: And this is a real problem for a lot of people. The article notes that among the roughly nine hundred tick species in the world, only a handful are responsible for most infections in the U.S. Research cited in the piece says more than thirty-one million Americans, nearly one in ten, experience a tick bite every year. And ticks can transmit dozens of diseases, in as little as thirty-six to forty-eight hours, including anaplasmosis, babesiosis, Rocky Mountain spotted fever, alpha-gal syndrome, and of course, Lyme disease. The CDC even reported a spring spike in emergency room visits related to tick-borne illness, with Lyme the most common diagnosis. An estimated four hundred seventy-six thousand patients are treated for Lyme each year, mostly in the Northeast, Mid-Atlantic, and Upper Midwest.

Mia: So what this test does is let someone who finds an attached tick check whether that tick is actually carrying infection, before deciding whether to seek treatment. That matters because the standard response right now is, at least partly, guesswork. The person behind it is Erin Dawicki, a pediatric physician associate in Boston, who came up with the idea during a health care entrepreneurship course in her MBA at MIT Sloan back in 2024. She'd been assigned the Lyme disease track, and her motivation came from clinical experience, seeing kids with huge swollen joints and no obvious trauma, and having to blanket-prescribe antibiotics because the risk of Lyme was so high, while knowing patients can't always afford an office visit every time a tick turns up.

Milo: So the practical implication is pretty straightforward. The most common treatment is a ten-to-fourteen-day course of antibiotics, and catching the infection early is what makes that treatment effective. A test that tells you quickly whether your tick is actually carrying disease could spare people unnecessary antibiotics, while still flagging those who genuinely need to act fast. It won't remove the uncertainty manufacturers and regulators have to work through, but for tick season, it could give people a much more useful first step than waiting and wondering.

Mia: Let's start with one of the most famous unexplained moments in the search for extraterrestrial intelligence. Back on August 15th, 1977, Ohio State University's Big Ear radio telescope picked up a striking narrowband signal. It appeared to come from the direction of the constellation Sagittarius, and it carried the hallmarks researchers expected from a genuinely alien transmission.

Milo: And here's the remarkable part about how it was even noticed. Astronomer Jerry Ehman was a volunteer going through data by hand, poring over line printer paper that an IBM 1130 computer had processed. A few days after the observation, he spotted the anomaly, circled the intensity reading with the string "6EQUJ5," and wrote "Wow!" right beside it. That gave the signal its name.

Mia: Now, that string confuses a lot of people. It's not an encoded message. 6EQUJ5 is simply how the telescope recorded the signal's changing strength over time, measured as a signal-to-noise ratio against a baseline. Each character represents one frequency channel, sampled for ten seconds with a couple of seconds of processing. What the telescope saw was an unmodulated continuous wave, stable, with no detectable modulation.

Milo: And it lasted the full seventy-two second window that Big Ear could observe. But here's the thing — it has never come back. No recurrence, and no confirmed explanation, terrestrial or otherwise. Researchers floated ideas like reflections off space debris, interstellar scintillation, even hydrogen clouds around comets. Some suggested an actual extraterrestrial transmission, but a single occurrence with no replication keeps that interpretation firmly speculative.

Mia: So nearly five decades later, the Wow signal remains exactly what it was that August night — a genuinely strong, genuinely strange signal from a specific patch of sky, that appeared once and never again. The observatory's director later documented the finding, but the mystery itself stands unresolved.

Mia: Switching gears to creativity and how new ideas actually come to life. There's a 2023 essay by Henrik Karlsson, co-authored with Johanna Karlsson, called "Cultivating a state of mind where new ideas are born." It argues that what people call the solitude of creativity isn't really about being alone — it's a mental state where other people's opinions stop mattering, so that fragile, larval ideas can actually survive long enough to grow.

Milo: The essay makes that case with an unusual example: the coworking boom of the early 2010s. Startup incubators tried shared living and coworking spaces thousands of times, and the piece argues there are no famous success stories to show for it. Sam Altman, then Y Combinator's president, tweeted in 2015 that not providing coworking spaces is part of what makes YC work. In a 2019 interview with Tyler Cowen, Altman spelled out why — great ideas are fragile and easy to kill, and in their larval stage the best ones sound bad. Coworking spaces make people laugh at those ideas and steer them toward the plausible but inconsequential ones. He put it memorably: coworking filters out the worst ideas, but it's a terrible band-pass filter, because it also kills the best.

Mia: The essay leans on figures like Picasso, who said that without great solitude, no serious work is possible, and it draws on the working notes of mathematician Alexander Grothendieck and filmmaker Ingmar Bergman. And not everyone in the Hacker News discussion agreed. One commenter pushed back, pointing out that academic environments are obvious counterexamples. They said they did their best work in grad school precisely because of the environment, and that everyone in their lab benefited from being around each other daily. That pattern, they argued, contradicts the essay's framing.

Milo: So the real contested question underneath it all: is the workspace and the company around you a breeding ground for good ideas, or a filter that quietly kills the strangest ones before they're ready?

Mia: Finally, a cautionary tale about identity, a name, and a sanctions list that refuses to let go. Sean Byrne wrote an article called "The Other Sean Byrne Doesn't Exist" about the sixteen years he spent being mistaken for a fictitious employee of an Irish aircraft-parts company.

Milo: This year, Apple denied him access to App Store Connect, telling him his information fully matches one or more restricted parties on the U.S. government's consolidated screening list. He sent his passport, his full legal name, Sean Joseph Byrne, his driver's license, and pointed out he had never lived at the listed Sligo, Ireland address. Apple still hasn't replied. And the list has one entry for a Sean Byrne, at a house in County Sligo, added in 2009 under a license requirement of all items subject to the export administration regulations, with a presumption of denial.

Mia: Here's what makes it maddening — that entry seems to match no real person at all. It stems from a 2009 Justice Department prosecution of Mac Aviation and the McGuinns over illegal exports of aircraft equipment to Iran. A 2010 superseding indictment no longer named Sean Byrne as a defendant, and repeatedly described the name as merely an alias used by co-conspirators. Yet the list entry still has no date of birth, no passport number, no middle name.

Milo: And it's not just Apple. Byrne also recounts Nasdaq stopping a stock tender order after a routine name-match background check, demanding extra proof of his California address before the sale could go through. Sixteen years of a name listed with no identifying details, matched against real people, with no clear path to be cleared. It's a striking reminder of how a single ambiguous entry, and a government's presumption of denial, can quietly follow an innocent man around for decades.

Mia: And that's a wrap for this week's roundup. We've wandered from skeptical takes on computing topics [chuckle] all the way to a possible Alzheimer's treatment and even the Wow! signal, that mysterious radio blip from 1977. Plenty to keep you thinking.

Milo: Absolutely. A quick favorite for me: whether working with AI feels more like leadership than coding, and the tale of a fictitious employee who turned out not to exist at all. Great conversations all around.

Mia: We've linked everything in the show notes, so dig in whenever you're ready. Thanks for listening, and we'll catch you next time.