1113 | AI Shakeups: IPOs, Exits, and the Week in Tech

||下载

节目备注

From Anthropic's IPO filing and Jensen Huang's sudden exit to Gemini 4, GPU fingerprinting, Microsoft's big outage, Apple's on-device models, and more: this episode walks through the week's biggest AI and developer news, plus what it means for users, developers, and the industry.

Timeline

  • 00:00:04 Opening
  • 00:00:56 Anthropic's IPO and Jensen Huang's exit
  • 00:03:39 Google's Gemini 4 and Antigravity
  • 00:05:25 On-device, small, and restricted: where AI models run
  • 00:09:14 The Microsoft 365 outage
  • 00:11:22 Security week: PostgreSQL RCE, patches, and a spy probe
  • 00:13:56 Rust 1.100 and open-source agent tooling
  • 00:16:10 App review backlogs, DMCA takedowns, and open-source apps
  • 00:18:26 AI in hiring and finishing abandoned work
  • 00:20:53 Grid costs, dwarf planets, and mystery radio pulses
  • 00:22:40 Closing

Related links

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, everybody. I'm Mia.

Milo: And I'm Milo. It's been one of those news cycles where it feels like the ground is shifting under the whole industry at the same time — money changing hands at a scale we've never seen, leadership vacuums at the very top, and meanwhile, down in the trenches, people are still fighting about compiler features and embroidery software.

Mia: That last part is why we do this show. Today we've got Anthropic filing for an IPO and Jensen Huang stepping down at NVIDIA on the very same beat, Google pushing a new Gemini model with aggressive pricing, a Microsoft 365 outage that took out its own status page, a twenty-two-year-old PostgreSQL bug turned remote code execution, Rust hitting a milestone a decade in the making, and a bunch of smaller stories that all rhyme with each other in weird ways.

Milo: So let's not waste any time. The two stories everybody's talking about are the Anthropic IPO and the Jensen Huang resignation. And the fact that they landed in the same news cycle is doing a lot of work in how people are reading them.

Mia: Right. Let's start with Anthropic. They've formally filed for an IPO. Ticker ANT on the Nasdaq. The numbers in the filing: third quarter 2026 revenue of seventeen billion dollars, which is twelve times year over year. That is a staggering growth number. But there's also a net loss of twenty-three point three billion — and the filing notes most of that is non-cash, which matters a lot for how you interpret it.

Milo: Yeah, and that's exactly where the discussion splits. If you're bullish, you look at twelve-x revenue growth and say the loss is an accounting artifact of stock-based compensation and similar non-cash items, and the real story is a company scaling revenue faster than almost anything in history. If you're skeptical, you look at a twenty-three billion dollar loss against seventeen billion in revenue and ask what the unit economics actually look like once you strip out the accounting.

Milo: Trading starts November nineteenth, so we're going to get a public price discovery on that argument very quickly.

Mia: And that's what makes it interesting — for the first time there'll be a market-clearing price on a frontier AI lab, rather than private valuations negotiated behind closed doors.

Milo: Now, into that context, NVIDIA drops the other bomb. Jensen Huang resigned as President, CEO, and board member — effective immediately. No transition period announced in what we have, no successor named in the announcement itself, and crucially, no stated reason.

Mia: And the Hacker News thread on that was pure speculation, which is worth being honest about. Commenters were throwing out everything from health to internal dynamics to the state of the AI industry cycle — but nobody actually knows the cause.

Mia: The only thing the thread agreed on is that a sudden, same-day, no-notice exit at the company that supplies the compute underneath the entire AI boom is a big deal, and it lands at an especially strange moment, right as one of NVIDIA's biggest customers is going public.

Milo: So what happens next for both companies is genuinely unknown. We don't know why Jensen left, we don't know how the market will price ANT on the nineteenth, and we don't know whether these two events are connected or just a coincidence of timing. I want to resist the temptation to narrativize it too hard.

Mia: Fair. Let's hold that uncertainty and move to Google, which is very much not standing still. Google's Antigravity rolled out Gemini 4 Hydrogen, and also introduced an Orchestrator that pairs two models — Carbon and Hydrogen. The reported result is near-Carbon success rates, but with twenty-four percent less quota used and nineteen percent faster.

Milo: That's the interesting engineering story, right? Instead of just shipping a better model, they're shipping a pairing where presumably one model handles something cheaper and the other handles the hard parts, and the end-to-end system beats the efficiency math of just running the expensive model alone. Near-Carbon success at less quota and more speed — that's a systems-level claim, not just a model claim.

Mia: And then there's the pricing, which is where it gets aggressive. Gemini 4 Hydrogen is one dollar fifty per million input tokens and seven fifty per million output tokens — and it's flat across a one million token context window. That last part is the part people should sit with. Flat pricing at a million context means no long-context surcharge, which has historically been where providers claw back margin.

Milo: So if you're a competitor setting frontier pricing, this puts real pressure on you. And on raw capability, at high thinking levels it scores fifty-one on the AA Intelligence Index. Whether that benchmark number translates to your workload is, as always, an open question — but the pricing structure is concrete and it's out there now.

Mia: Okay, from Google's cloud-scale models, let's zoom all the way in — because there's a cluster of stories this cycle that all point the same direction: models are getting smaller, more local, and in some cases, harder to trace.

Milo: First one: Apple. The Foundation Models framework shipped a native Swift API for on-device and cloud LLMs. Two things stand out — third-party models can plug in through a Language Model protocol, and there's something called Dynamic Profiles. So Apple is formalizing the interface layer for local inference, which means app developers can build on-device AI without shipping their own runtime.

Mia: That's the consumer-facing version of a trend. The other end of it is a Hacker News post people have been calling "smolmaxxing" — running models on hardware they absolutely were not meant for. The poster got a sixteen-billion-parameter mixture-of-experts model running at about five tokens per second on a Xiaomi TV box with roughly one gigabyte of usable RAM. And a 1.7-billion-parameter model at twenty-four to twenty-six tokens per second on a 2013 iPad Air.

Milo: A 2013 iPad Air. That machine predates most of modern machine learning tooling, and it's doing interactive-speed inference. Five tokens a second on a TV set-top box is slow for chat, but it proves the point: with aggressive quantization and MoE architectures where you only activate a fraction of parameters, the floor for "what can run this" has collapsed.

Mia: Now here's the flip side of that coin, and it's a serious one. Forensic weight analysis — people who study the actual published model weights like fingerprints — claims that models from DeepSeek, Moonshot, MiniMax, and Ant Group, produced since 2026, carry marks consistent with NVIDIA Blackwell GPUs. And Blackwell is export-restricted. So if that analysis holds, it raises a compliance question those companies would very much like to not have.

Milo: And we should be careful here — this is a claim from forensic analysis of weights, not a confession or a documented paper trail. The fingerprints allegedly in the weights are suggestive; whether they prove anything, whether there's an innocent explanation, whether regulators act on it — all unknown. But methodologically it's a fascinating development: the weights themselves are becoming evidence.

Mia: Which connects to the last piece of this cluster: LibreAGI. It's a private BitTorrent tracker where members preserve open-weight models that are likely to disappear. They run a signed public catalog and they enforce seed ratios.

Milo: The signed catalog is the detail that makes it credible rather than just a warez site — you can verify that the weights you download are the weights that were published. And seed-ratio enforcement solves the actual hard problem of archival torrents, which is keeping people seeding terabyte-scale files years later. The premise is sobering, though: the track is "these models might vanish." Companies delist or relicense open weights all the time.

Milo: So you've got the full spectrum now — Apple making local inference a first-class platform feature, hobbyists running models on TVs, forensics reading model weights like GPUs, and an archive building insurance against deletion.

Mia: All right — infrastructure story number two, and this one is less fun: Microsoft. On November twelfth, 2026, Microsoft 365 had a broad outage. Their own description: elevated load, because many servers were unexpectedly unavailable. Exchange, Teams, SharePoint, Copilot, Defender — all affected.

Milo: Note that Copilot and Defender being on that list is significant. Your AI assistant and your security tooling going down at the same time as your email — that's the whole modern stack failing as one unit.

Mia: And the Hacker News Ask thread confirmed widespread Exchange Online failures from people's firsthand experience. But the detail that became the center of the discussion: Microsoft's own status page was down. Returning a 502.

Milo: That's the part commenters seized on, understandably. The status page is supposed to be the thing that's isolated from the thing it monitors. If the status page and the service share fate, then during an incident you have no authoritative signal at all, and everyone falls back on Twitter and Hacker News threads — which is literally what happened here.

Mia: The stated cause — elevated load as many servers unexpectedly became unavailable — is doing a lot of vague lifting. "Many servers unexpectedly unavailable" doesn't tell you whether it was a config push, a capacity event, a network change, something else. The root cause is unknown, and the open question hanging over the whole thing is whether the consolidation of email, chat, files, AI, and security into one connected platform has made these outages structurally more painful than they used to be.

Milo: Right — when your outage domain includes your incident-response tooling, resilience stops being a per-service property and becomes a platform-design question. One to watch.

Mia: Let's shift to security, because this week PostgreSQL had a genuinely remarkable story. Mehmet Ince took a twenty-two-year-old bug — a type-confusion bug in PostgreSQL — and turned it into remote code execution. It's catalogued as CVE-2026-6460. One important caveat: an authenticated user is required, so it's not a worm-in-waiting; you need valid credentials first.

Milo: But twenty-two years. That code has been sitting in the codebase since roughly when PostgreSQL was a very different project, and it survived two decades of scrutiny before someone chained it all the way to RCE. It's a good reminder that "old and heavily audited" is not the same as "safe" — it just means the bug is patient.

Mia: And PostgreSQL's response was comprehensive. They shipped releases 18.7, 17.12, 16.16, 15.20, and 14.25, plus a 19 RC2, fixing fifty-two security CVEs and over a hundred and forty bugs in one sweep. And they also drew a line: PostgreSQL 14 is now end-of-life. So if you're on 14, that was your last patch, and it's time to plan an upgrade.

Milo: Practically, if you run Postgres in production, this week was your patch-and-plan week. Fifty-two CVEs across the supported branches is a lot of surface to be exposed on.

Mia: Now, from database security to a geopolitical one that's also a security story. Singapore has criticized the United States for obstructing its probe into Megaspeed CEO Alice Huang, who's alleged to have ties to a Chinese spy. And on the US side, Senators Wyden and Representative Neal have opened an investigation.

Milo: The friction point is interesting: Singapore is complaining that US obstruction is getting in the way of its own investigation. So you've got two jurisdictions that are nominally aligned, tripping over each other in a probe involving a tech CEO and alleged espionage. It's still unfolding — we don't know what the investigations will find or how the intergovernmental friction resolves.

Mia: Let's pivot to something more uplifting: tooling. Rust 1.100 is out, and the headline is that the never type is finally stable — after ten years and five failed attempts. For anyone who hasn't followed this saga, the never type is the type of expressions that never produce a value — it's foundational for certain kinds of API design around things that abort or loop forever — and it's been one of the longest-running unfinished features in the language.

Milo: Five failed stabilization attempts tells you how subtle it is. It touches inference and type-system corners where small design decisions ripple everywhere. Getting it over the line in 1.100 — a version number people will remember — feels like a deliberate closing of a chapter. But that's not all in the release: custom allocators, a DropGuard API, and cargo dependency cooldowns.

Mia: The dependency cooldowns are worth a beat, because they're a supply-chain feature. The idea is you can impose a delay before newly-published dependency versions are picked up, which protects against a freshly-compromised release propagating to your build in the first minutes. It's Rust acknowledging that the supply chain is part of the threat model.

Milo: And there's a nice thematic pairing here with the other item in this bucket: AgentRun, from GREP.AI — a YC F26 company — which they open-sourced. It's a managed agents platform: self-improvement, durable execution, and a control plane that works on any ACP harness.

Mia: Durable execution is the key phrase. Agents that fail midway, get restarted, resume — that's the same reliability discipline that languages like Rust bake into the type system, applied to long-running agent workflows. Both stories are about infrastructure for software that has to survive over long timescales, whether that's a ten-year stabilization saga or an agent run that spans hours.

Milo: Okay, next cluster: gatekeepers under strain. Start with app review. AI-generated app submissions are flooding Apple's review process. And it's not just Apple — Steam's review delays have grown from three to five days to weeks.

Mia: The mechanism is obvious once you say it: if generating an app costs almost nothing, submission volume goes up until the review pipeline saturates. The reviewer time is the scarce resource, and the cost of consuming it has collapsed. Apple hasn't announced a fix in what we have here; the delays are just the observable symptom.

Milo: And then there's a different kind of gatekeeping: Nintendo sent four DMCA notices to GitHub, taking down the Switch emulators Yuzu and Citron Neo, plus a hundred and forty-four repos of mod chip firmware. That's a sweeping action — it's not just emulators, it's the firmware ecosystem around them.

Mia: Meanwhile, on the more constructive side of open source: ExcalidrawZ, a native SwiftUI client for Excalidraw covering macOS, iOS, and iPadOS. Technically it looks like a nice piece of work — native, cross-platform. But on Hacker News it took criticism for borrowing the Excalidraw brand name. The pushback being that the name implies an endorsement or affiliation that isn't there.

Milo: And then the counter-example that ties this whole cluster together: ZigZag. It's a free, open-source embroidery application, written in Rust, with MCP support, and the author built it with ChatGPT. It exists because embroidery machines like the Brother Skitch PP1 normally require subscription software — and one person with an AI assistant built a replacement and gave it away.

Mia: So within one cluster you have AI-generated apps overwhelming a gatekeeper, a rights-holder using takedowns to enforce control, a naming dispute over borrowed trust, and open source with AI assistance simply routing around the paywall entirely. The common thread is that the old control points are all under pressure from different directions.

Milo: Which brings us to the AI-in-work cluster, and this one's got a philosophical edge. Jane Street has added LLM-focused interview rounds — described as less code-writing, more design and verification. Importantly, most of their rounds remain traditional, no-LLM coding. So it's an addition, not a replacement.

Mia: And that design detail — testing whether you can direct and verify an LLM rather than whether you can type the code — is a live hypothesis about what the job now is. Meanwhile the Quake 4 story puts that hypothesis into practice in the wild: the Quake 4 expansion was abandoned, and modders used AI to finish it. The result is OpenQ4, version 0.141, and it's now playable.

Milo: And that kicked off a real debate on Hacker News about crediting AI-assisted work. The disagreement, roughly: is this a triumph of community perseverance where AI was a tool like a compiler, or does finishing abandoned work with AI change what authorship and credit mean? Commenters were genuinely split, and the thread didn't resolve it. What's inarguable is that a piece of abandoned software is playable again.

Mia: And the bridge to the last item in this cluster is exactly that question of verification. There's an essay by Pron making the case that correctness is fundamentally hard — that results from computability theory, like the halting problem, and from complexity theory put complete understanding of programs at odds with programs themselves. You cannot, in general, fully verify what arbitrary code does.

Milo: That's the theoretical grounding under the whole cluster. Jane Street adding verification-focused interview rounds is the industry conceding that in an LLM era, the human's job is the part that can't be delegated — judgment about whether the output is right. And Pron's argument explains why that job doesn't get easier, because correctness was never fully achievable by humans either.

Milo: Whether AI-assisted development makes verification easier or just generates more things to verify — that's an unresolved question the debate circles without answering.

Mia: Let's close with the grab bag that turned out to be more coherent than it sounds. First, money and infrastructure: a Cornell study published in Nature Energy found that in 2025, data centers paid only thirty-eight percent of eight and a quarter billion dollars in grid infrastructure costs. The rest fell on households.

Milo: That's the cost-shifting debate in one number. The grid upgrades that data centers necessitate are funded mostly by ratepayers, not by the companies driving the demand. As AI buildout accelerates, expect that thirty-eight percent to become a political number, not just an accounting one.

Mia: Then out to the solar system. The Rubin Observatory found a new dwarf planet candidate, 2026 ED138. It orbits at forty-six AU with a period of about two hundred eighty-one years. And here's the haunting detail: it may have been captured on photographic plates back in 1937 — meaning it might have been sitting in an archive for ninety years, undetected, waiting for a survey with Rubin's sensitivity.

Milo: And finally, something genuinely mysterious. Researchers at Curtin and ICRAR linked an unusual red dwarf–white dwarf system to radio pulses that repeat every eight minutes. This is a long-period radio transient, and the origin is still unknown. A white dwarf paired with a red dwarf, blinking on a clock — the mechanism is an open question, and it's exactly the kind of thing Rubin-era astronomy is going to find more of.

Mia: And that's the show. From IPO filings to embroidery software to an eighty-one-year-old photograph of a dwarf planet — everything this week was, in some way, about who pays for infrastructure, who controls it, and who gets to verify it.

Milo: We'll be back tomorrow. Thanks for listening.

Mia: See you then.