0723 | Interview Take-Home Unmasks Supply-Chain Plot

Mia: Welcome to HackerNews Daily, a podcast from Bri. I'm Mia.

Milo: And I'm Milo. Today we're looking at a developer who uncovered something strange during a take-home interview project for a blockchain company linked to Rivos.

Mia: We'll also talk about a reminder that depending on any service, even volunteer-run ones, can end badly. And Henry Ndubuaku and Roman have built a system that post-trains the on-device model Gemma to do something pretty clever.

Milo: Plus, Terrence Tao tested whether ChatGPT could reason about a famous unsolved math problem. Let's get into it.

Mia: A developer who did a take-home interview project for a blockchain company called Rivos or Rivos Inc. recently shared a report that reads like a cyber-thriller. The interview process looked polished: a LinkedIn profile, a GitHub organization, even a working product website.

Milo: Right, and the task seemed normal on the surface. Candidates were told to clone a repository, run some local commands, and submit a pull request. But hidden inside a pre-commit Git hook was obfuscated malware, and it persisted even after you deleted the repository.

Mia: Here's how the persistence worked. The hook wrote encrypted command-and-control server details and an auth token to a separate Git global config temp directory. Then it created a hidden systemd user service called ld-unique.service with a randomized file name under ~/.config/systemd/user/.

Milo: And that service was set to run at boot. Every 3,600 seconds it executed an XOR-encrypted Base64 payload, which turned out to be a reverse-shell SSH client reaching out to the attacker's server.

Mia: So the whole thing was an elaborate supply-chain attack disguised as a job interview. In the Hacker News discussion, some people speculated that scams like this might be fully automated in what they called the agent era.

Milo: Others pointed out that North Korean threat actors have shown sophisticated engineering in this style of operation for a long time, possibly also leveraging large language models. One commenter said the attack actually resembled a CTF challenge.

Mia: Switching gears, a recent incident sparked a broader reminder: depending on any single service, even a volunteer-run one, means you can eventually lose access.

Milo: The specific trigger was Codeberg banning cryptocurrency projects. A commenter in the discussion said the takeaway is to make backup plans and self-host whatever you can. Their specific recommendations were Gitea, or Forgejo, which is a Gitea fork run by Codeberg itself.

Mia: There were also some more decentralized alternatives mentioned. One person brought up a Git option built on Nostr, where your public key represents sovereign project ownership, so it is not bound to any single entity.

Milo: And a third commenter pointed to Git over Reticulum. That option provides Git, issue tracking treated as work documents, and release management. They linked to the Reticulum network manual's Git section for the details.

Mia: Now here is an interesting approach to making on-device AI more reliable. Two creators, Henry Ndubuaku and Roman, built a system called Cactus Hybrid that post-trains the on-device model Gemma 4 E2B to self-assess its own accuracy.

Milo: The key is that every response gets a confidence score between zero and one. What is clever is how they generate that score. Instead of using common routing signals like text-based self-rating or token entropy heuristics, which they found were barely better than a coin flip, they tap hidden state signals across different layers in the small model.

Mia: The authors claim those internal signals carry meaningful self-awareness. In practice, a developer can accept the on-device answer when the confidence score is high and route the query to a larger cloud model when it is low.

Milo: And the numbers are compelling. By offloading only fifteen to thirty-five percent of queries to Gemini 3.1 Flash-Lite, the hybrid system matches Gemini 3.1 Flash-Lite on most listed benchmarks.

Mia: The routing fractions vary by task. ChartQA needs just fifteen to twenty percent offloaded. LibriSpeech is twenty-five to thirty percent. MMBench, GigaSpeech, and MMAU require thirty to thirty-five percent. And MMLU-Pro demands the most: forty-five to fifty-five percent of queries sent to the cloud model.

Milo: The creators say their motivation is the rising expense of frontier models, combined with the need for privacy and speed. A small on-device model is fast and private, but sometimes wrong, and this confidence mechanism gives you a practical way to decide when to trust it.

Mia: Finally, Fields mathematician Terrence Tao recently used a public ChatGPT conversation to explore whether the model could reason about a famous unsolved problem: the Jacobian Conjecture, and its known counterexamples over finite fields.

Milo: Tao fed the model a specific construction from the literature, an eight-by-eight matrix, and asked it to categorize the counterexample as a standard crypto-automorphism, a symplectic automorphism, a tame automorphism, or none of the above.

Mia: ChatGPT correctly analyzed the matrix, recognized the counterexample, and noted it does not fit the tame automorphism class. Tao described that classification as a casual observation, but still one worth making.

Milo: He then guided the model through type-two polynomial automorphisms with distinct eigenvalues, pressing it to relate symplectic structure to the monodromy of Hamiltonian flows. So he was really pushing the boundaries of what the model could handle.

Mia: The chat shows the model performing solid pattern-matching and algebraic manipulation under expert steering, but it is not generating novel mathematical insights on its own. Tao's own remarks in the thread frame the exercise as a test of current AI reasoning, not a breakthrough.

Mia: Microsoft is drawing a line under SMS and voice calls for multi-factor authentication. According to a tenant-administrator email shared on Hacker News, those two options will be removed for M365 tenants on February 1st, 2027.

Milo: That is a specific date quite far out — but there is an earlier milestone. The same notice says Microsoft will begin automatically enrolling users in passkeys starting September 1st, 2026.

Mia: Right. So the timeline is: automatic passkey enrollment next September, and then roughly five months later, SMS and voice codes go away entirely.

Milo: And that passkey push comes with a friction point the Hacker News discussion surfaced. Microsoft Authenticator passkeys cannot be used with third-party password managers like Bitwarden.

Mia: Which means users report they are keeping Microsoft Authenticator installed solely for that one login. They cannot consolidate into the manager they otherwise use everywhere else.

Milo: So you get the security benefit of passkeys, but you lose the cross-platform convenience that makes passkeys attractive in the first place — at least for this specific Microsoft implementation.

Mia: And if an organization has users who standardized on something other than Microsoft Authenticator, this change forces a split — one app for Microsoft, another for everything else.

Milo: The discussion framed it as a design choice that prioritizes the ecosystem over the user experience. Whether that changes before 2026 is an open question.

Mia: Moving to a different kind of design tension — a blog post on Fiddery titled "Businesses with ugly AI menu redesigns" argues that AI-generated signage and advertising is becoming a new signifier of low-effort, low-skill output.

Milo: The core idea is that when a customer sees obviously AI-generated menus or ads, it shapes how they perceive that business — and not in a positive direction.

Mia: One Hacker News commenter put it bluntly. They said they used to find chalked sandwich boards annoying, but now they would actually prefer one, because it signals a human made an effort.

Milo: That same commenter predicted businesses will start standing out by continuing to hire human effort — specifically for recognizably human output.

Mia: Another participant noted that this dynamic is not entirely new. Chalk-written signs and menus have long signaled a high-quality restaurant, while photos of food and ring-bound menus signaled the opposite end of the spectrum.

Milo: So the AI-generated menu becomes the new ring-bound menu — a shortcut that telegraphs something about the establishment before you even order.

Mia: There was also a playful aside in the thread. Someone floated the idea of a chalk-writing robot or printer — essentially automating the human-crafted look without the human.

Milo: And one person added that they actually know someone who earned a degree in letterprinting — so the craft itself is real, and people still train for it.

Mia: Now to a historical mystery more than four centuries old. Scientists have conducted DNA testing on the remains of Grand Duke Francesco I de' Medici and his wife — and they found evidence of malaria.

Milo: This is the couple whose deaths in the 1580s have been debated ever since. The rival theory has always been poisoning by the relative who was next in line for the Ducal throne.

Mia: The findings were reported by CNN. And the malaria result is significant — but it does not fully close the case.

Milo: That is the key point the researchers themselves emphasized. Study co-author Gisella Caccone, a senior research scientist at Yale, stated — and this is worth quoting — "We can say that they had malaria, we cannot say that they were not poisoned as well."

Mia: So the DNA evidence confirms the presence of the malaria parasite. It does not rule out the possibility that someone also administered poison around the same time.

Milo: Another researcher, identified only as Ochoa, put it this way: the genetic evidence "decreases the margin for speculation." But they acknowledged there is no guarantee Francesco was not also poisoned.

Mia: Decreases the margin for speculation — not eliminates it. That is a careful way of saying we now know more, but we still do not know everything.

Milo: Which after 400-plus years feels fitting. The science advances, the mystery narrows, but it does not disappear.

Mia: Our final story comes from an essay on Hacker News arguing that ethics is the skill that matters most right now.

Milo: And the pushback in the comments was immediate. One person wrote simply: "ethics doesn't pay." Three words, very direct.

Mia: A reply acknowledged that in a corrupt system, ethics can indeed work against that system. But the commenter insisted ethics is indispensable at a civilizational level — and traced that idea all the way back to Aristotle.

Milo: That same commenter made a broader argument. They said many relatively new institutions run like perpetual-motion machines with voracious appetites — constrained only by positive law, meaning whatever is technically legal.

Mia: And they argued we need to recognize how we have lost our bearings, and find what they called a cooling-off period to search for them again.

Milo: Then another reply landed harder. It pointed out that child labor pays. Sex trafficking pays. Slavery pays. And then it asked: what kind of person does that make someone?

Mia: That question flips the whole framing. The original claim was "ethics doesn't pay" — and the counter is that plenty of terrible things pay, but that is not how we decide what to do.

Milo: So the thread becomes a kind of live debate between a transactional view of ethics and a foundational one — and neither side really concedes.

Mia: So there's a fascinating Hacker News comment from a user named sroerick who actually built something inspired by that malleable computing article. They describe an interpreted Lisp that runs inside a web server and stores the AST — the abstract syntax tree — directly in Postgres.

Milo: Right, and the practical upshot is that agents use functions in a REPL instead of calling an API. So instead of the agent hitting an endpoint and getting JSON back, it's interacting with the program's own internals.

Mia: Exactly. And when agents encounter an issue, they can write new views and actually improve the program on the fly. The author calls the result pretty amazing, though they're honest about it — they mention what they call vibeslop jank and compare it to a little bit of jank you'd find in Emacs.

Milo: Which is an interesting comparison, because Emacs itself is famously malleable. You can reach into almost anything and change it while it's running. That's the tradition they're channeling.

Mia: And then another commenter, __d, suggests there might be an interesting middle ground here. Not fully malleable software that demands full programming skill, but something closer to the Unix tool model.

Milo: That's the idea of prebuilt utilities you compose with a simple syntax — pipes, redirects, small focused tools. You get some of the flexibility of malleable computing without requiring everyone to be a full programmer. It's a really practical counterpoint.

Mia: Switching gears — Mitchell Hashimoto argues that every programmer should understand SIMD. That's single instruction, multiple data — the kind of parallel processing inside a single CPU core. And his core point is that auto-vectorization with just the -O3 compiler flag is not as effective as many people hope.

Milo: And commenters on the discussion really flesh this out. One key insight is that the best SIMD optimizations often require changing your data layout — specifically, switching from array-of-structs to struct-of-arrays. That's a fundamental restructuring of how you store data in memory, and the compiler can't do that for you.

Mia: There's also a real limitation around floating-point math. Floating-point operations are non-commutative — the order matters — which means the compiler can't safely rearrange them for vectorization without potentially changing your results.

Milo: Right, and you can push things further with flags like -march=native or -march=x86-64-v3, which tell the compiler it can use newer instruction sets. There's also function multiversioning with something called target_clones, which lets you ship multiple versions of a function optimized for different processors.

Mia: But even with all those flags turned on, the practical reality — as commenters point out — is that non-integer code generally does not autovectorize in what's called normal-math mode. So if you want SIMD speed for floating-point work, you're probably going to have to write it yourself.

Mia: John C. Dvorak has died at age 80. He was born in 1946, and for a lot of people in tech, his voice and his writing were a part of the landscape for decades.

Milo: Yeah, early audience encounters with his work came through Byte magazine in the 1980s — that was a major publication — and through his regular appearances on ZDTV, which later became TechTV. He had this very distinctive style.

Mia: There's this remarkable story about Wikipedia and Dvorak that's been circulating. For decades, a Wikipedia article listed his birth year incorrectly, and the site refused to let him change it.

Milo: Right, and Dvorak wrote about this dispute at length. His argument was that Wikipedia had become the antithesis of the entire concept of a wiki — the idea that anyone can edit and correct information was, in his experience, not how it actually worked in practice.

Mia: It's one of those stories where the irony really stings — the man himself couldn't correct his own birth year on the platform that's supposed to be the encyclopedia anyone can edit.

Mia: Apple released Safari Technology Preview 248, with release notes published on the WebKit blog. And one Hacker News commenter noted the team has been on a roll with massive fixes and improvements, pointing to last month's announcement of over 500 fixes.

Milo: And someone fired back with a quip — that Apple had finally used some of its trillions to buy one Claude Pro subscription. A bit of AI humor there.

Mia: But alongside the jokes, there's a real question people are asking: when do fixes in Safari Technology Preview actually make it to stable Safari? It's not always clear.

Milo: One user in the discussion is tracking a specific WebKit bug that's already been fixed in STP but hasn't landed in the stable release yet. They note it will land eventually — maybe months later — but Apple doesn't preannounce its plans, and there's no known way to track which WebKit commits end up in which specific Safari releases.

Mia: So the cadence is that STP gives you a window into what's coming, but the timeline from preview to production remains opaque. If you're waiting on a particular fix, you just have to keep checking.

Mia: Let's talk about Reddit and AI licensing. The company already has deals in place with OpenAI and Google, giving those companies access to Reddit's data for training. But observers are pointing out that Reddit also seems to be working to keep other AI companies out.

Milo: And that push to lock down access is raising a pretty sharp question in the discussion. One commenter flags a striking number: Reddit's gross margin sits at 91.5 percent. That is enormous, and it puts the whole licensing strategy in a different light.

Mia: Right. Because the content that powers those deals, the posts, the threads, the years of conversation, all of that comes from users. And the question surfacing in the discussion is whether Reddit has any plan to compensate the people who actually create that value.

Milo: Several participants in the thread are pretty direct here. They say the only compensation contributors appear to get comes through the ads users are allowed to see, and through small virtual stickers. That is the full extent of what comes back to the people writing the posts.

Mia: So you have a company with a ninety-one-and-a-half percent gross margin, selling access to user-generated content, and the contributors get ads and stickers. That is the shape of the debate happening right now.

Milo: Switching gears, there's a project called Hologram that's getting attention. It makes Elixir run in the browser.

Mia: And for people working in the Elixir ecosystem, this is a big deal. One commenter on Hacker News who uses Phoenix regularly says even at its current stage, Hologram offers a very promising developer experience, with a reasonable roadmap and clear intent.

Milo: The interesting part is what this unlocks. Local-first Elixir apps would cover one area that Phoenix does not handle out of the box today. Instead of writing glue code and error-prone wiring, you'd get the same result with just a few declarations.

Mia: So it fills a gap in the framework. And that matters because of how developers are making decisions now. A separate commenter notes that framework choice today is governed by fundamental capabilities more than by existing engineer expertise. It's less about what your team already knows, more about what the tool can actually do.

Milo: That same commenter mentions they're building an agent orchestrator in Elixir and Phoenix, which gives you a sense of where people see the language heading. Hologram running in the browser adds another reason to place that bet.

Mia: Now to a different story. A new survey from Redfin shows that most Americans say not in my backyard when it comes to AI data centers near their homes.

Milo: And the pushback isn't abstract. One Hacker News commenter puts it bluntly, they are awful to live near. Another cites a study from Arizona State University that found negative effects within twelve hundred feet of a data center.

Mia: But there is a counterpoint worth hearing. Another commenter says hyperscale data centers simply don't go on residential plots. These facilities require large commercial or industrial parcels with major utility connections. The siting logic doesn't really put them right next to houses in most cases.

Milo: Still, the debate has a real front line. A separate comment points to a data center in Northern Virginia, described as the front line of this whole fight, and it runs onsite power generation around the clock. That's generators running twenty-four seven, which gives you a sense of the noise, the emissions, and the local tension.

Mia: So you have a national survey saying most Americans oppose these facilities nearby, local testimony saying they're terrible to live beside, and at least one site running constant onsite power. That's the conflict playing out as AI infrastructure expands.

Mia: And that brings us to the end of our line-up. Let's hit the high points one more time.

Milo: We opened with a take-home interview project that went sideways. A developer built it out, then discovered the company it was for appears linked to Rivos or Rivos Inc. — and may not even exist. A cautionary tale about doing your homework before you do their homework.

Mia: From there, a reminder we all needed: any service can vanish, even the volunteer-run ones. You can end up holding the bag, so plan accordingly.

Milo: Henry Ndubuaku and Roman showed off Cactus Hybrid, a system that post-trains on-device models like Gemini Nano to get them navigating mobile UIs on their own — no cloud needed.

Mia: Terrence Tao did what Terrence Tao does. He opened a public ChatGPT chat and tested whether the model could reason through a math problem with him, live.

Milo: On the enterprise side, Microsoft is pulling SMS and voice as two-factor options for M365 starting February 2027. If you're still on those, the clock's ticking.

Mia: We looked at a Fiddery piece arguing that AI-generated signage and ads are producing what they called "ugly AI menu redesigns." The aesthetics of automation got a grilling.

Milo: DNA testing on the remains of Grand Duke Francesco I de' Medici and his wife turned up evidence worth a second look — Renaissance intrigue meets modern science.

Mia: And on Hacker News, an essay made the case that ethics is the skill that matters most right now. One commenter fired back bluntly — let's just say not everyone was on board.

Milo: A handful of quick hits: an interpreted Lisp that captures malleable-computing ideas, Mitchell Hashimoto arguing every programmer should understand SIMD, and a remembrance of John C. Dvorak, who passed at 80. His Wikipedia birth year was wrong for decades — he reportedly found the whole thing amusing.

Mia: Apple dropped Safari Technology Preview 248, Reddit's licensing deals with OpenAI and Google have observers asking who's being kept out, Hologram brings Elixir to the browser, and a Redfin survey says most Americans oppose AI data centers near their homes.

Milo: That is the show. Back with you next time — until then, keep questioning the inputs.