0804 | Ten Math Advances, AI's Largest Solo Project, LLM Expertise, Cloudflare's Kimi & GLM

Mia: Welcome back to HackerNews Daily, and thank you for joining us. I'm Mia.

Milo: And I'm Milo. Today we've got quite a spread — from ten big advances in mathematics and theoretical computer science, to the question of just how large a software project an AI can actually complete on its own.

Mia: We'll also dig into a thoughtful essay on how large language models reward expertise, Cloudflare running smaller models like Kimi and GLM at scale, and Germany seeing wind and solar overtake fossil fuels for the first time.

Milo: Plus, an uncomfortable look at Flock's cameras and a wrongful stop, and twenty years of Pandoc. Let's get into it.

Mia: A new piece out of OpenAI lays out ten advances in mathematics and theoretical computer science, and it's sparked a genuinely interesting debate on Hacker News. One commenter pushed back on the idea that mathematicians should feel existential dread about these results. Their take? These advances are actually bringing math mainstream — this person says they personally look forward to the interpretations and discussions of these results by human mathematicians, and they note that it's mostly the big-name superstars who seem to be benefiting from the attention.

Milo: So instead of framing these tools as replacements for mathematicians, some are framing them as a way to pull more people into the field. The reasoning is that when a model produces a significant mathematical result, it creates a discussion that more people can follow and engage with. That flips the anxiety on its head — rather than shrinking the role of human mathematicians, it could be expanding the audience for what they do.

Mia: There's also a question floating around about the largest software project an AI can complete entirely on its own. One developer in the discussion says they're building a programming language with AI help, and when they leave the model running unsupervised, it keeps generating code in strange directions. They're curious what structure AI-only code would even look like.

Milo: And someone in the replies draws a comparison to self-driving cars — wondering whether we'd ever reach the point where a hundred percent of the code is written by machines, with no human hand in it. That comparison gets at the real question underneath the whole thread: at what scale can a model be trusted to work alone, and what does that code actually look like when nobody's steering?

Mia: An essay doing the rounds argues that LLMs reward expertise — and the whole point is that the reward isn't about fancy prompting tricks. The author's example is Terence Tao, the mathematician, using ChatGPT to dig into a recently discovered counterexample to the Jacobian Conjecture. The author says, and I'm paraphrasing, this is not the same ChatGPT I talk to — they couldn't get to where Tao gets even with unlimited tokens to burn.

Milo: So what does Tao actually do differently? His messages are very short and to the point. He responds to the gist of what the model says, not point by point. He pushes back on wrong-looking responses without bluntly contradicting them — something like, "this looks more complex than I was hoping for." And he makes leaps and suggestions himself, almost never just following the model's lead about where to go next.

Mia: But the author's key point is that the tips alone aren't enough — what makes it work is actually understanding the mathematics. By signalling genuine expertise, Tao shunts the model into what you might call talking-to-mathematicians mode, not explaining-to-amateurs mode. The models respond differently to someone who clearly knows what they're doing. The common belief that no skill is involved, since everyone talks to the same models, is wrong — the most important prompting skill is deep expertise in the domain you're prompting for.

Mia: Cloudflare published a post about running the Kimi and GLM models at scale, titled "Smaller, faster, safer." The technical details are thin in the discussion, because the conversation actually swerved into something else entirely — whether the writing itself reads as AI-generated slop. One commenter said their slop detector went off at a specific paragraph, adding, "I love AI, but I really hate reading it."

Milo: That sparked a whole sub-debate about how to handle AI-sounding prose on Hacker News. One person said they've stopped making that comment entirely, because it ends up on half the posts on the site — and wished the platform could just flag AI-generated prose and filter it out. Another pointed to LinkedIn's recently announced flagging button, though they were unsure how well it would carry over to Hacker News. Someone else recommended a browser extension that flags AI-sounding content across X, Reddit, and Substack.

Mia: And there's a real split over whether commenting about it even helps. One reader argued people should keep calling it out, so that posting slop is seen as uncool. Someone else pushed back on that, though — pointing out that comments are held to a higher standard than submissions, and citing Hacker News guidelines that say the site is meant for conversation between humans. A more creative commenter described writing a personal script that scrapes every comment, checks whether anyone has already called the article AI-generated, and grays it out if so. Either way, the thread is a reminder that as these models write more of what we read, the conversation about the writing itself isn't going away anytime soon.

Mia: For the first time, wind and solar have overtaken fossil fuels in Germany's energy mix — that's the headline from bne IntelliNews, as discussed on Hacker News. And the milestone is worth celebrating, but the discussion thread quickly zoomed out to give it some context.

Milo: Right, one commenter called it genuinely good news, but cautioned that how you measure matters. Count all energy rather than just electricity, and look beyond the developed world, and progress looks a lot slower. They pointed to global energy substitution charts that show fossil fuels still dominating worldwide — even arguing some read that data as renewables being effectively irrelevant for the rest of our lifetimes.

Mia: And another commenter put Germany's achievement in scale. The United States alone consumes about four times all of Germany's fossil fuels, while China's coal use alone is more than eight times Germany's entire fossil fuel consumption. So the takeaway there was: this is genuinely good and relevant news at the local level, but pretty irrelevant at the global one, where there's more talk than action — partly because limiting or taxing imports and pushing emissions abroad could be too economically painful.

Milo: There was also a second thread debating how to read those global energy graphs. One commenter argued that burning hydrocarbons is relatively inefficient — only about twenty to thirty percent of the fuel's energy becomes useful work — while electricity is much more efficient, so we need far less of it. They also noted the exponential growth of wind and solar is hard to spot on a chart that ends in 2024. Extrapolate that exponential growth forward, and another commenter suggested wind and solar could reach close to one hundred percent of electricity in the early to mid 2030s. So even the skeptics agree the trend line is real — the argument is really about how fast and how far it reaches.

Mia: Flock makes the AI-powered license-plate cameras that police and communities install everywhere — and its CEO sat down with The Drive podcast to talk about the company's promise of zero wrongful stops, telling the host: I wasn't the first. The story got picked up on Hacker News, and the discussion turned into a pretty sharp debate.

Milo: Flock is a Y Combinator-backed company, and one commenter captured the tension bluntly, saying it hurts their brain to think about the Venn diagram of privacy advocates and Y Combinator fans that hang around Hacker News.

Mia: Exactly — this piece lands right at the intersection of the site's two loudest camps. On one side you have people who care deeply about surveillance and civil liberties, and on the other, people who tend to cheerlead Silicon Valley startups. The whole conversation is basically those two instincts pulling against each other on what the company does and what its promises actually mean in practice.

Mia: On August third, 2006, John MacFarlane uploaded version 0.1 of pandoc to his website — a free, open-source document converter written in Haskell. That was twenty years ago, and he just published a retrospective looking back at the whole ride.

Milo: That first version was around three thousand lines of code with basically no outside dependencies. It could convert between Markdown, reStructuredText, HTML, and LaTeX, and output to those formats plus RTF or slide presentations. Fast forward twenty years, and it's grown to over two hundred releases, support for more than fifty document formats, automatic citations and bibliographies, integration with academic tools like Quarto and Jupyter Notebook, and it's been installed on millions of computers — making it the most popular program ever written in Haskell.

Mia: And the origin story is great. He wanted to learn Haskell, so he built a document converter as his project. Existing Markdown tools in Perl, Python, Ruby, and PHP just matched text patterns to transform Markdown straight into HTML. Pandoc instead actually parsed the document into a structured tree, which is what made it possible to convert between any of those many input and output formats at all. Early on there was no marketing — he just emailed two friends. Within months a Turkish developer was packaging it for Debian, and version after version added features: tables, footnotes, math, and by 2008 the full converter took shape with support for everything from Wikipedia's format to open document files.

Mia: The Dunning-Kruger effect — it's one of psychology's most famous findings, the idea that unskilled people overestimate their ability. But a piece from McGill's Office for Science and Society argues it might not be real, or at least, not in the way we think. It got shared on Hacker News under the title: The Dunning-Kruger effect may just be a data artefact.

Milo: The author wanted the effect to be real, but doubt crept in after checking the literature and corresponding with David Dunning himself. And Dunning offered a key correction: the effect is about us, not them. The real lesson was always that we should all be humble and cautious about ourselves — especially in areas where we lack competence. It's not really about dumb people. As originally defined, it's a bias where someone terrible at, say, English grammar might predict a sixty percent score and actually get fifteen, while someone excellent might predict seventy and actually score ninety.

Mia: The 1990s studies tested grammar, humor, and logical reasoning, comparing self-predictions with actual scores. And Dunning told the author the effect has more to do with being misinformed than uninformed — like confidently saying the capital of Scotland is Glasgow. Then in 2016 and 2017, two papers argued the effect can be replicated with completely random data, which raises real questions about whether the famous result is real or just an artifact of the statistics. It was the kind of finding everyone loves to quote, but the closer you look, the murkier the picture gets.

Mia: So MiniMax released their third-generation video model today — it's called H3 — and it's a big deal because it's the first one the company has shipped with open weights. The ComfyUI newsletter says it has day-zero native support there. And the headline feature is that it doesn't just make video — it generates real stereo audio in the same pass, in up to 2K resolution and up to fifteen seconds per clip.

Milo: That's a genuinely different capability. It takes text, images, video, or audio as input, and MiniMax's lead selling point is what they call multimodal context understanding — it can take several images, audio, and video together and resolve them all against a single prompt. There's also a motion transfer mode where a reference video supplies the movement, like a camera move or a performance, while the subject and style come from somewhere else.

Mia: And the models are small enough to matter. The post quotes stated figures — total memory footprint cut by 66 percent, from about 124 gigabytes at full precision down to 42 and a half on the smallest variants. Combined with dynamic VRAM offloading, that means a next-generation 2K video model can allegedly run locally on something like an RTX 3060. Though one commenter rightly asks an open question: on a 16 gigabyte card like that, how long does generating a fifteen-second clip actually take?

Milo: The discussion around it is split between excitement and disruption angst — talk about copyright, creative value, and what this does to the industry. One commenter says that after seeing the posted samples, they deleted their LTX2 and WAN folder weights immediately, calling them now completely worthless. That's the kind of reaction that tells you this is more than an incremental update.

Mia: Now to a case that reads like a cautionary tale about AI-generated security advisories. JFrog Security Research, in a piece by Afek Berger, looked at a newly created GitHub repo that published a batch of SQLite vulnerability advisories — plus more than fifty other CVEs — which JFrog believes are what they call LLM slop. Except for one. The concerning part is the track record: NVD quickly flagged these as critical, and CISA's branch agreed.

Milo: But when JFrog's researchers actually verified them, the claims fell apart. Cited code didn't exist in those versions or pointed at unrelated logic. Proof-of-concept payloads didn't trigger crashes. None of the CVEs appear on SQLite's official advisory page. And when they grouped all the advisories into one file, the scanner flagged them as AI-generated. So how did they test the specifics?

Mia: They cloned the official SQLite repository, checked out three different versions, built them in isolated containers, and ran the PoC SQL under AddressSanitizer. For one claimed use-after-free, the cited function didn't even exist in that older version — it was added in a later commit. For another advisory citing a patch in 3.51.3, a diff between 3.51.2 and 3.51.3 shows no changes to the file it named. Another cited a pair of line numbers that turn out to be a comment and a memory allocation.

Milo: Even the scoring wobbled — Red Hat initially gave one of these a 10.0 critical rating, then downgraded it to 7.6. And the Hacker News commenters zeroed in on the practical pain: organizations that have mandated patching policies based on CVE scores could get burned chasing a vulnerability that doesn't actually exist.

Mia: Let's bring the AI productivity question home with a piece by Bjorn Roche, an engineering leader in New York, titled "The AI Productivity Gap." His argument is refreshingly honest: AI has already improved team productivity and will only get better — but building production features still takes almost as long as it used to. Because writing new code isn't where most of a senior engineer's time goes, and AI hasn't made the hard part any easier: figuring out what code to write in the first place.

Milo: He even notes that AI-written product requirements docs or ticket descriptions sometimes take longer to review than human-written ones, because AI writing can be overly detailed and harder to distill. He lays out notional before-and-after day breakdowns — and the numbers are striking. A senior developer saves about an hour and a quarter a day, roughly fifteen percent, even assuming AI makes coding three times faster. That's because their writing-new-code time already drops from an hour and a half to half an hour.

Mia: The junior developer gains more — about two hours a day, roughly twenty-five percent — because juniors spend more of their day coding, which is exactly the part AI boosts most. And Roche calls out the irony: leaders say "we only hire senior engineers because AI does the work of juniors now," when actually juniors are the ones positioned to gain the most.

Mia: Finally, a contrarian take on working with AI — from Ankur Sethi on his lab notebook blog. He still uses coding assistants, but with a deliberate slowdown for personal projects. His agent instructions are strict: never create, edit, move, rename, or delete project files unless explicitly asked. Show every proposed edit in the chat — because he types every single edit into his editor by hand. And don't run any commands that modify files or change the repository state.

Milo: That sounds like giving up the whole point of an agent, but his reasoning is careful. He says he's not ten times faster — he's probably only twice as fast, but he gains a much deeper understanding. Manually typing the code builds a mental model, surfaces hallucinations and bad design choices, lets him refactor as he goes, and builds a spatial map of his own codebase.

Mia: He contrasts that with reviewing AI-generated pull requests — which he describes as poring over hundreds of lines of overly-defensive, badly-commented, subtly incorrect code. And his worry is bigger than one project: he fears the industry is taking on what he calls cognitive debt, and that soon we won't understand how large parts of our digital infrastructure are put together. He calls shipping software you don't completely understand professional malpractice. A strong stance — and one worth holding onto as we see how the rest of this conversation plays out.

Mia: Andy Pavlo, the well-known database professor at Carnegie Mellon, is joining ClickHouse to set up and lead a brand new research group called ClickHouse Labs. In a blog post he wrote for the company, Pavlo explains his history — he joined Carnegie Mellon's computer science department back in 2013, and he says he's been following ClickHouse since it was open-sourced in June 2016. In fact, he initially thought the project had to be vaporware, because it shipped features that at the time existed only in a handful of closed-source commercial database systems — things like vectorized query execution built on C++ and SIMD instructions.

Milo: That's a strong endorsement of how ambitious the project was from day one. So what is the new research group actually meant to do?

Mia: ClickHouse Labs wants to be a top-tier industry research organization focused on databases, working closely with ClickHouse's engineers, customers, and partners rather than running in isolation. It'll also team up with ClickHouse's PostgreSQL group to help make the company's managed service a market leader in performance and reliability. Pavlo points to IBM Research and Microsoft Research as inspiration, and he says an immediate focus is clearing a backlog of optimizations the engineering team explored but hasn't yet validated and shipped to production. He also notes that ClickHouse engineers have published detailed implementation articles since the company was founded in 2021, and there's a 2024 database research paper on ClickHouse's core architecture that Pavlo actually assigns to his CMU students.

Milo: And readers of the Hacker News discussion had their own hopes. One commenter asked whether Pavlo's CMU lecture series would continue, possibly in a sponsored format from ClickHouse — and Pavlo replied that the seminars will go on, with a new series starting next month and an announcement due this week.

Mia: Over in the Rust world, the project's goals repository includes an accepted proposal for the 2026–2027 cycle called “Immobile types and guaranteed destructors,” championed by a developer known as LCNR, with language team support from Jack Huey. The core idea is to introduce new traits — like Move and Forget — that make explicit exactly what operations a type allows. That would let a type opt out of being moved, or out of being simply forgotten through the language's mem::forget function.

Milo: So basically, the language would let developers say clearly what a type is allowed and not allowed to do, instead of certain operations kind of happening implicitly. Why does that matter in practice?

Mia: The write-up says it follows the precedent of the earlier work on the Sized trait hierarchy, and it would be validated through minimal prototypes, formal RFCs, and real-world testing in the Linux kernel — with goals like scoped spawning, async drop, and pin-by-default behavior. In the discussion, one commenter said the change would make recursive async functions easier to write and let async functions borrow rather than clone from their outer scope. But another commenter pushed back on the recursion piece, arguing that the compiler's synthesized Future type can't contain an instance of itself without boxing — because that would be an infinite-size type — and deciding where exactly to cut with Box::pin isn't trivial.

Milo: So what's the real win, in their view?

Mia: They argued the bigger payoff is scoped tasks and structured concurrency. Scoped threads already guarantee they'll terminate before the spawning function returns, so they can safely share stack references. Futures, by contrast, carry no guarantee they'll run to completion and are more prone to cancellation. Another commenter described the proposal as “things you'd expect to work already,” noting that tokio::spawn is awkward and the Pin wrapper is unergonomic. And one more added that no other language offers this without drawbacks like garbage collection pressure or heap allocation, with C++ coroutines coming off as much more awkward by comparison.

Mia: And that's our show. Ten big advances in math and computer science, plus AI's limits on real-world coding, what expertise does to LLMs, and a whole lot more.

Milo: We covered wind and solar passing fossil fuels in Germany for the first time, twenty years of Pandoc, and even how to dodge cognitive debt when AI writes your code.

Mia: Thanks for listening. If you enjoyed the episode, share it with a friend who loves a good deep dive — and we'll see you next time.