1115 | Leaks, Loopholes and a Caught Booster

||下载

节目备注

From Nvidia's circular chip financing and a caught Starship booster to Lovable apps leaking passports, Google's multi-agent results, a tax code proven wrong 138 times in Lean, and every version of Monkey Island side by side — a tour of what's new, broken, and rebuilt in tech.

Timeline

  • 00:00:04 Opening
  • 00:00:50 Who leaked your data, and who pays for being watched
  • 00:08:31 Shorter clocks, longer scams
  • 00:10:58 Agents doing science, and learning you can watch
  • 00:13:43 Browsers and toys built for agents
  • 00:15:56 Nvidia's circular money and a 753B on your desk
  • 00:19:30 Flight 15: booster caught, ship skipped
  • 00:20:46 Archived by law, attested by default
  • 00:22:31 From hand-assembled BASIC to a machine-checked tax code
  • 00:25:50 Monkey Island 13 times, and KDE on film
  • 00:26:46 Closing

This episode is produced by Bri. Bri uses advanced AI technology to turn the feeds you care about into podcasts made for listening. Contact us at hi@bri.so.

文字稿

Mia: Welcome back to the show, I'm Mia.

Milo: And I'm Milo. It's been one of those news cycles where the same thread keeps showing up in different stories — who has access to your data, who decides when things get disclosed, and who gets to build the tools everyone else has to live with.

Mia: Right, so today we've got a privacy stack that goes from vibe-coded backends leaking passports to an eight-hundred-million-euro fine against Meta to smart glasses that make bystanders the product. Then we'll swing through disclosure deadlines, research agents, hardware financing, a Starship catch, and a few stories about how software history gets preserved.

Milo: Let's just dive in, because the first one deserves a long look.

Mia: So, Crissy Field did an analysis of a hundred fifty-nine thousand Lovable projects — Lovable being one of those AI app builders where you describe what you want and it ships something. And of those, twenty-four thousand six hundred sixty-four Supabase sites were exposing data with no authentication at all.

Milo: That number is the whole story, honestly. One in six, roughly. And this isn't metadata leaking — we're talking passport details and over two hundred thousand messages sitting behind open endpoints.

Mia: And I think the reason it resonated is that it isn't a vulnerability in the classic sense. Supabase works the way it's designed to. Lovable works the way it's designed to. The failure happens in the gap between them, where a person who has never thought about row-level security asks for an app and gets a backend with a public door.

Milo: That's the "cost of letting anyone ship a backend" argument. There's a real tension in the comments here, because on one side you've got people saying this is the same story as every abstraction wave — most websites have always been insecure, we just didn't have a tool that let us count them at this scale. A hundred fifty-nine thousand projects is a census, not a sample.

Mia: And the other side says no, this is qualitatively different, because the abstraction is now doing the wiring for you. If a human sets up a database wrong, that's one site. If the generator's default path produces an unauthenticated Supabase client, that's tens of thousands of sites that fail identically.

Milo: Which raises the uncomfortable question: whose responsibility is it? Some people want Lovable to refuse to generate a Supabase integration without auth configured. Others say that's nannyware — the tool can't know your app is supposed to be private. And then there's the grimmest take, which is that it's nobody's job and the market will just punish the exposed ones, except the people punished are the app users, not the builders.

Mia: Right, the passengers pay for the pilot's mistakes. And what's still genuinely unknown is whether platform defaults change because of this. Nobody in the reporting says Lovable or Supabase has announced anything. The audit just documented the scale.

Milo: I'll say one thing though — an audit with named numbers is a lot harder to wave away than "people say vibe-coded apps are insecure." When the count is twenty-four thousand six hundred sixty-four, someone has to decide that number is acceptable or fix it.

Mia: Okay, stay in the privacy lane but go up a level in scale. Italy fined Meta about eight hundred eighty-seven million dollars — eight hundred five million euros — over Android camera-roll access.

Milo: Explain the actual mechanism, because this one's subtle. It's not that Meta stole photos. It's that Android's photo permission was app-wide — one grant covers everything in your camera roll — and Meta's warning was generic. The regulator's position was that users couldn't give informed consent to something that broad.

Mia: And I think that's why this fine is interesting as a precedent rather than just a big number. It's not a claim of misuse. It's a claim that the permission model itself, plus a vague disclosure, doesn't constitute consent. If that logic holds, it's a challenge to the whole design of broad permissions.

Milo: The counterargument in the discussion is basically yours, but reversed: where does this end? Every permission on every phone is a bundle. Location is a bundle. Contacts are a bundle. If informed consent requires per-item granularity, the UX becomes unusable, and then people just tap through, which is arguably worse.

Mia: And the rebuttal to that is that the answer isn't granular prompts, it's better defaults and scoped APIs — which, notably, the mobile ecosystems have been slowly building, like limited photo selection. So the fine reads to some people as a regulator dragging the ecosystem toward the design it should have had anyway.

Milo: What's unresolved is whether anything changes for users. Eight hundred five million euros is real money but Meta can absorb it. The question everyone keeps coming back to: does the fine change the permission flow, or is it just a tax on the old one?

Mia: Which connects perfectly to the third story, because it's the same problem from the hardware side. The EFF published a warning about Meta and Google smart glasses, Apple Watch Audio Intelligence, and the Amazon Bee — and their framing is that these devices shift privacy burdens onto bystanders who cannot opt out.

Milo: That bystander framing is the key move. Everything we've talked about so far — the Lovable leaks, the Meta fine — is about people who at least clicked something. The person walking past your smart glasses never agreed to anything. There's no consent flow for being in the background of someone else's device.

Mia: And each product does it slightly differently. Glasses capture people around the wearer. Audio Intelligence on a watch — the idea of a device continuously processing ambient sound around you — that's bystanders again. The Amazon Bee is a screenless band that's always listening for your tasks, which means it's also listening to everyone near you.

Milo: The disagreement here is about whether this is fixable by design or structural. Some people argue there are technical answers — recording indicators, on-device processing, blur everyone else. Others say if the value proposition is passive, continuous capture, then any friction you add for bystanders undermines the product, so the incentive structure fights the fix.

Mia: And notice the pattern across all three of these stories. In each case, the party with the least power and least information bears the risk. The Lovable app user, the camera-roll owner who got a generic warning, the stranger on the sidewalk. Exposure is scaling with how much stuff we ship and wear, not with how carefully anyone designs.

Milo: There's actually one more story that belongs in this cluster, because it shows the funnel side of the same problem. CBC did an investigation and found about fifteen hundred postings for remote "travel jobs" on LinkedIn and Monster that were funneling applicants into something called Evolution Travel, which is an MLM that charges fees. Over a thousand of those ads have since been removed.

Mia: The reason it fits here: job boards are another place where the platform's permissions and defaults decide what people are exposed to. A listing that looks like employment but costs you money is a consent problem too — the disclosure exists, buried, but nobody reads it because the frame says "job."

Milo: And fifteen hundred across two mainstream boards tells you the moderation isn't keeping up with the volume. Same theme as the Lovable numbers — volume beats review.

Mia: So let's move from scams under pressure to a different kind of pressure — deliberate this time. Google's Project Zero changed its disclosure policy. As of now, it's "Full 90" — they publish ninety days after reporting. Then it tightens: sixty days starting February 2027, thirty days starting January 2028. And they removed patch soak time entirely.

Milo: For people who don't follow this world: Project Zero finds security bugs and tells the vendor privately, then discloses publicly after a deadline to force fixes. The soak time was extra padding — if the vendor patched, Project Zero would wait before publishing so more users could update. Removing it means: your patch ships, the clock doesn't extend.

Mia: And the trajectory is the striking part. It's not a one-time tightening, it's a ratchet with dates attached. Ninety to sixty to thirty. That reads like a message to vendors: the slack you're enjoying is scheduled to disappear.

Milo: The debate mirrors the debate we just had about Meta. One camp says shorter clocks mean users get fixes faster, full stop — vendors will always use whatever time you give them, so give them less.

Mia: The other camp says thirty days is going to break somewhere. Complex patches — think a bug in a kernel, or something in a dependency tree — can need coordinated fixes across many parties. If the deadline doesn't move for patch complexity, you either get rushed patches or you get vendors shipping partial fixes to make the date. Which one is safer is genuinely not obvious.

Milo: And there's a middle position worth naming: the soak time removal is arguably the more aggressive change than the day counts. Soak time was protecting users who hadn't updated. Removing it assumes patch availability equals patch adoption, which for most of the world's devices isn't true.

Mia: That's the real unknown — will the thirty-day window hold when patches are complex, and does removing soak time actually cost end users anything? Project Zero has bet no. We'll find out.

Milo: Nice segue though, because the next cluster is also about pressure and rigor — but on the frontier side. Google's Antigravity, with a system called /teamwork, is a multi-agent setup that reportedly produced new results in math, chemistry, kernel work, and verification, and the tools are now available externally.

Mia: The word doing heavy lifting there is "new results." Not benchmarks, not code review — actual research output across four pretty different domains. And they're letting people outside Google use the tools, which is how you find out whether the claims generalize.

Milo: That's the big open question — how general are /teamwork's results really? A system that shines in one domain can be a demo in another. Having it external is the honest test, because researchers outside the company have no incentive to be generous.

Mia: And here's the counterweight story, which I love. Somebody rebuilt the DeepSeek V3 architecture — the real ideas, multi-head latent attention, DeepSeekMoE with the aux-loss-free balancing, multi-token prediction — in about a hundred ninety lines of pure JAX. Thirty-six point two million parameters, trained on a MacBook Air.

Milo: Which is the whole point made compact: the architecture is teachable. You don't need the fleet to understand MLA or the MoE balancing trick. A tiny faithful reimplementation does more for comprehension than any blog post, because you can poke it.

Mia: And the third piece of the triptych: Aatish Bhatia's interactive site where you watch reinforcement learning happen — a four-neuron little "spark" learning to seek light over about two thousand trials. You literally see the learning curve as behavior.

Milo: So in one day we've got frontier claims at the top and tiny, public, reproducible teaching at the bottom, and both are advancing the same field. There's a comment-thread argument here about which matters more, and I think the honest answer is they check each other. The MacBook reimplementation is a replica you can audit; the frontier results are claims you can't yet. The teaching site makes the mechanism legible to people who'll never read a paper.

Mia: What I'd flag as unresolved: "new results" in chemistry and kernels is a very strong claim, and the external availability is the thing that will either validate it or quietly shrink it. Watch what independent people do with /teamwork.

Milo: And there's a bridge here to our next pair, because agents doing work need tools built for them — and two projects this cycle did exactly that. First, a macOS WebKit browser designed for coding agents. The features tell you everything: no focus stealing, playwright-cli-style shell commands, shared profiles, idle tabs get suspended.

Mia: Every one of those is a complaint about running a browser under agent control. Focus stealing is the big one — an agent-controlled browser grabbing your window mid-typing is miserable. Shell-command-style browser control means the agent drives it the way it drives everything else. Shared profiles mean the agent sees what you see. Suspending idle tabs means the agent doesn't eat your RAM with forty background pages.

Milo: And then terminactory, from the other direction — a terminal-native idle factory game written in Rust with ratatui, that runs alongside AI agents and receives messages from them via Claude Code hooks. So your agent can send updates to your game while it works.

Mia: It sounds like a toy — it is a toy — but it's pointing at something real: the terminal is becoming a shared surface where the human and the agent coexist. Work in one pane, the game reacting to your agent's messages in another.

Milo: The reason this matters, I think, is the shift from agent ergonomics as a hack to agent ergonomics as a product surface. Until now, if you wanted an agent to use a browser, you bolted Playwright onto a headless Chrome and hoped. Now someone's designing the browser around the agent's needs from the start.

Mia: And the skeptic position is fair: browsers for agents might stay niche. Humans still want to browse, and dual-purpose interfaces are hard. But the terminactory angle suggests a future where the agent isn't just using your tools, it's present in your environment — and tools will be designed for that presence.

Milo: Which leads naturally to the infrastructure layer underneath — and this next pair is the spiciest of the day. The Wall Street Journal reported that Nvidia used circular financing deals in mid-2026 to absorb TSMC wafer oversupply and keep GPU prices high. And — this is the part that made jaws drop — the report says Jensen Huang resigned over affairs.

Mia: Let me unpack "circular financing" for listeners, because the term is doing a lot. The allegation is that money flows in a loop: Nvidia invests in or funds entities that then buy Nvidia's product, so demand for GPUs is partly built on Nvidia's own capital. If the report is right, that's a way of absorbing excess wafer supply from TSMC while keeping prices propped up.

Milo: The comparison everyone reaches for is vendor financing from past cycles — when suppliers fund their own customers to keep volume going, and it works spectacularly until it doesn't, because the demand was never independent. The open question the piece leaves us with is exactly that: how durable is circular-financed demand? If the loop is what's holding up orders, a wobble anywhere in the loop propagates.

Mia: And the resignation is a separate, human bombshell — the report attributes it to affairs, which is a very different kind of story from the financing one, and it landed in the same report, which is part of why the discussion got so heated. Leadership churn at the company that supplies nearly everyone's compute, on top of questions about demand quality — people connected those dots whether or not they should be connected.

Milo: Careful not to overreach there — the report gives its reasons, and we should state it as reported, not speculate beyond it. What's fair to discuss is the interaction: whoever leads Nvidia next inherits both the circular-financing question and the supply relationships.

Mia: Now the second half of this pair is almost a rebuttal in spirit. NVIDIA's own blog detailed serving GLM-5.3 — a 753-billion-parameter MoE in NVFP4 format, roughly 464 gigabytes — across two DGX Stations connected over ConnectX-8 RDMA, with vLLM tensor-parallel-two, peaking at 607.8 tokens per second.

Milo: Do the arithmetic out loud with me: that model does not fit on one desk machine. 464 gigs of weights means you need two high-end workstations tied together with high-speed networking to serve it — this is a rack squeezed onto a desk.

Mia: And the reason it's in the same conversation as the financing story is exactly the plan's line: financing shapes supply, and supply shapes what developers can run. The circular financing question is about whether demand for all this hardware is real. The GLM-5.3 demo is about what the hardware can do when you have it. Those are related but not the same claim — the model demonstrably runs; the question is whether the market for running it is organic.

Milo: One connection more: rockets face similar all-in hardware gambles, which brings us to Starship. Flight 15 — SpaceX successfully caught the booster. The ship, the upper stage, was skipped, reportedly due to insufficient fuel.

Mia: So the catch is now — what, four, five times in a row? It's becoming routine, which is genuinely historic when you think about it: catching a skyscraper-sized booster out of the air is the boring part now.

Milo: And the ship is where the fight is. Insufficient fuel on the upper stage suggests the margins problem — range, trajectory, propellant budget — is the thing still biting. The unknown here is the exact anomaly chain behind the skip; "insufficient fuel" is reported as the reason but the causal story of how it got there isn't public.

Mia: The contrast is instructive though. Reuse of the first stage went from miracle to routine. The upper stage, which has to survive reentry and land precisely, is still where the engineering risk lives. Same program, two very different maturity levels.

Milo: And that's the bridge to our next cluster, because it's about things ending and things becoming routine in a different sense — infrastructure and law. IRS Direct File — the government's free tax filing system — its open-source repo is now archived and unmaintained. Development stopped in January. And the code is only public because federal law requires it, not because anyone's shepherding the project.

Mia: That last detail is the melancholy one. Lots of government software is open by law, not by culture. Once the team stops, the repo becomes a museum piece — legally public, practically dead. It's a caution for anyone who thought government-built software was self-sustaining.

Milo: Meanwhile, on the opposite end of the health spectrum: PyPI redesigned its project pages, live now, with a new Security tab showing provenance states — full, none, or warnings — and per-file attestations.

Mia: Provenance — meaning cryptographic evidence of how a package was built and by whom — moving into the default page layout is a quiet milestone. It's not a setting you opt into; it's the UX. You land on a package page and the security posture is just there, next to the readme.

Milo: And the pairing is almost a parable. One public-interest software experiment ends; supply-chain transparency becomes ordinary. The unknown on the Direct File side is whether it resumes elsewhere — another office, another administration, a state-level version. Nobody knows.

Mia: Okay, last two topics, and they're the historical ones — which is a nice way to land. First: Lexifina formalized US Title 26 — that's the entire federal tax code — in Lean, the proof assistant. And they found a hundred thirty-eight real drafting faults. Including rediscovered known loopholes and novel ones — one example is a SALT joint-filing quirk.

Milo: Sit with what that means. The tax code is text. It was written by humans, amended thousands of times, and nobody has ever machine-checked it. When you formalize it — encode the definitions and rules so a proof assistant can check them — the inconsistencies and contradictions fall out. A hundred thirty-eight faults in the actual law.

Mia: And the loopholes are the fun part, because search over a formalized corpus can find combinations a human drafter would never notice — like a SALT interaction that only bites in joint filings. Machine-checking human law catches both the errors and the exploits.

Milo: Now zoom way out, because there's a family tree making the rounds: a hundred four software job titles, from the 1940s to now. And the observation that got the most discussion: DevOps, SRE, and platform engineering are the same work renamed, era after era. The newest entry is AI Agent/Context Engineer, dated roughly 2025.

Mia: There's a cynical reading and a generous reading of that tree. Cynical: the industry reinvents titles to reset salary expectations. Generous: each rename reflects a real shift in where the work sits — the tooling genuinely changes even if the underlying responsibility, keeping systems running, doesn't.

Milo: And the newest entry raises the same question both ways. Is "context engineer" a real new discipline — the skill of shaping what an agent sees being genuinely novel — or is it prompt writing with a promotion? That's the unresolved one: which titles on this tree survive the AI era?

Mia: And here's the perfect origin point for the whole arc: Wozniak wrote Apple 1 Integer BASIC on paper. Hand-assembling 6502 machine code. No assembler. Before the computer even existed.

Milo: Before the machine, the code. In his head, on paper, converting mnemonics to hex by hand. And that's the through-line of this entire conversation, if there is one: rigor keeps changing shape but never disappears. Woz hand-assembling, the tax code being formalized in Lean, PyPI adding attestations, Project Zero ratcheting down disclosure clocks — it's all the same impulse at different scales and eras.

Mia: Let's close on the lighter note. There's a site showing all thirteen versions of The Secret of Monkey Island side by side — eight editions, with twelve identical scenes for comparison — captured in ScummVM and Hatari. You can watch the remaster lineage decade by decade, pixel by pixel.

Milo: And alongside it, CultRepo released a documentary about KDE, covering the free-software desktop community from 1996 onward, and it got a very warm reception. Both of these are communities preserving and retelling their own history — the game through its versions, the desktop through its people.

Mia: The open question there is a nice one to end on: which of today's projects will get this treatment in twenty years? Which of the tools we talked about today — the agent browsers, the vibe-coded apps — will someone line up side by side?

Milo: Whichever ones survive long enough to have a version history worth studying. That's the show — Mia?

Mia: Thanks for listening, everybody. We'll be back tomorrow with whatever the internet dug up overnight.