1013 | AI Everywhere, Freezer Attacks, and a 50fps GT4

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

Milo: And I'm Milo. Today's feed is a good one — there's this thread running through almost everything: what happens when machines do work that used to be human, who pays for it, who profits from it, and where the new bottlenecks and fault lines show up. So let's just dive straight into the big one.

Mia: Yeah, let's talk about Rich Hickey's argument, because it set off a real debate. Hickey — the Clojure guy — says LLMs are not tools. They're automation. His reasoning: a tool is something you wield, something you control directly, something that gives you feedback as you use it. With an LLM, you don't have that kind of control or feedback loop, so calling it a tool is a category mistake. It's automation, full stop.

Milo: And there's a direct response to that from yyhh.org, which pushes back. Their counter is that an LLM can be a tool — if the human keeps a feedback loop and exercises judgment. In other words, the tool-ness isn't in the model, it's in how you use it. If you review, verify, and steer, you're still the one manipulating the work.

Mia: That disagreement is interesting because both sides actually agree the human matters — they just locate the difference differently. Hickey says the lack of inherent control disqualifies it from being a tool. The response says control is something you build around it. And honestly, the empirical evidence from the threads this week is murkier than either pure position.

Milo: Right, so take the Ars Technica study. They found that the speed gains from AI coding get absorbed by the bottleneck of human review. You generate code faster, but someone still has to read and verify it, and that's where the time goes. So the acceleration is real on the production side and then it hits a wall.

Mia: Which raises the obvious question — if review is the bottleneck, does review become the new job? Some commenters were basically asking whether the role shifts from writing code to judging code, and whether that's better or worse. There wasn't consensus on that. It's genuinely unresolved.

Milo: And then there's the firsthand account that puts a very human face on it. An engineer who got laid off wrote that models like Fable don't actually replace veterans — the veterans still outperform — but the models made him cheaper than his compensation. That's a brutal distinction. The technology didn't win on quality; it won on economics.

Mia: That reframes the whole "does AI replace developers" question. The answer from that account is: not by being better, but by changing what your salary is measured against. It's not capability displacement, it's price displacement.

Milo: And then you have Diomidis Spinellis taking the long view, going further than anyone else in this thread. His claim: AI will replace cognitive work generally, and thinking will become a leisure activity — like hunting became after domestication. Hunting went from survival necessity to hobby. He's saying cognition goes the same way.

Mia: That's the most extreme position in the discussion, and it's worth noting nobody in these threads fully endorsed it or fully demolished it. It sits there as the horizon case. If Hickey is right that LLMs are automation, Spinellis is asking: okay, automation of what scale? Cognition itself? And if the yyhh.org response is right that human feedback loops preserve tool-ness, then Spinellis's leisure-brain future is much further off.

Mia: The positions genuinely conflict and the thread doesn't resolve it.

Milo: One thing that strikes me — the productivity math is the missing piece everywhere. Nobody has real numbers on whether AI-assisted work nets out faster once you count review, debugging, and rework. The Ars study says the gains get absorbed, but absorbed entirely? Partially? Unknown.

Mia: Right. And what comes next on this is watching what companies actually do. If the laid-off engineer's story generalizes — cheaper-than-compensation dynamics — companies will act on that economics regardless of where the tool-versus-automation philosophy lands. Okay — staying in the money zone, let's talk about the AI financial machine, because it got loud this week.

Milo: So the headline item: New Constructs came out and called the Anthropic IPO the most ridiculous of 2026. Their math: they value the company at $150 billion, against a two-trillion-dollar ask. And the loss side — $42 billion in losses in 2025. That gap between what they'd pay and what's being asked is enormous by any measure.

Mia: A thirteen-fold gap between the analyst valuation and the ask. Commenters were picking apart what that means — is it bubble pricing, is it rational given optionality, nobody agreed. But the New Constructs framing was the anchor: ridiculous of the year.

Milo: Meanwhile Meta launches Muse Wonder, which they're calling a frontier model, with two features doing the heavy lifting: Active Memory and Active Focus. And they trained it at the Prometheus datacenter. That's the spend side of the same coin — frontier claims require frontier infrastructure, and Prometheus is the physical embodiment of that.

Mia: And here's the prediction that landed: Kevin Kelly said back in 2016 that by 2026, Google's main product would be AI, not search. And people are marking that as essentially correct now. A ten-year-old prediction that held.

Milo: That's a good prediction track record, and it cuts against the "this is all hype" reading. If the money is irrational, why did the ten-year structural prediction come true?

Mia: Though you can hold both: the shift to AI being real, and the specific IPO valuation being detached. Those aren't contradictory. The interesting unresolved question in the thread was how far money keeps flowing into AI infrastructure — whether the Prometheus-style spending keeps getting funded at these loss levels. And then there's a counterweight data point: GLM-Legal from Z.ai scored 19.2% all-pass on Harvey LAB, and reached parity with frontier models on LegalBench and PLawBench.

Milo: And that matters for the valuation argument, because it suggests frontier-adjacent capability may be cheaper to reach than the $2 trillion price implies. If a challenger matches frontier models on legal benchmarks, the moat story gets complicated. Nobody in the thread settled whether that validates or undermines the IPO ask — it just sits there as evidence on both sides.

Mia: Right. Okay, shift gears — from money to security, and this next one is genuinely strange. A cyberattack in France hit roughly 500 supermarkets through Danfoss software — that's refrigeration management software — and the attackers cranked the fridges up to 50 degrees Celsius.

Milo: Fifty degrees. That's not ransomware-locking-files, that's physically destroying inventory. Millions of euros in losses. And it wasn't France only — the same attacks on Danfoss fridges hit Europe broadly: France, Belgium, the Netherlands. And commenters flagged the precedent: this echoes an earlier attack on US military commissariats, same kind of target, same kind of tactic.

Mia: What's unknown — and commenters kept pressing on this — is who did it. No attribution in what we have. That changes how you read the whole thing. State actor? Extortion? Sabotage? Destroying food rather than encrypting it is an odd choice, because destruction is harder to monetize than ransom.

Milo: Right, and the discussion of why — was it meant as pressure, was it a demonstration, was it incompetence — was all speculation. The fact pattern is clear: ~500 stores, Danfoss as the vector, 50-degree fridges, millions lost, US commissariat precedent, no known attacker.

Mia: From infrastructure attacks to a bug bounty story that's almost funny. Google paid $90,000 for a remote code execution flaw at developers.google.com. The mechanics are the best part: a Flask cookie was signed with a secret — and the secret was contained in the cookie itself.

Milo: That's self-defeating cryptography. The signature is supposed to prove the cookie is trustworthy because only the server knows the secret. If the secret travels inside the cookie, the signature verifies nothing. Anyone could forge anything.

Mia: And commenters reading this one had that mix of "how did this ship" and "we've all done something adjacent." Signing a payload with a key embedded in the payload is the kind of mistake that survives because it works in every test — nothing fails until someone deliberately forges. Ninety thousand dollars says it was exploitable for real.

Milo: And then the supply-chain story, which is scarier in a quieter way. A commit in the Zod repository — version 4.6.6 — contained malware. The payload delivery is the novel bit: the malware fetched its payload via an Ethereum contract. Using a smart contract as the command channel.

Mia: That's clever because blockchain-based delivery is hard to take down and hard to monitor. The saving grace, and it's important to say clearly: it never reached npm. The commit existed in the repo, but the package registry was never poisoned, so end users were not exposed through installs.

Milo: Still, the discussion circled around what it means for trust in shared code. If a well-known repo's commit history can carry contract-fetching malware, the review surface for maintainers gets even worse. And that connects to — okay, let's move to the platforms-drawing-lines discussion, because it's the same anxiety showing up institutionally.

Mia: QGIS made a flat decision: a policy that prohibits totally — completely bans — contributions made with AI. Their stated reason: unsustainable pressure on maintainers. Not "label your AI contributions," not "AI allowed with review." A total ban.

Milo: And given what we just said about the Zod commit and review bottlenecks, that policy reads differently. If your maintainers are already the bottleneck — and remember Ars found human review is the bottleneck in AI coding generally — then a flood of AI-generated contributions with the same review cost is exactly the pressure QGIS says they can't absorb. The ban is a capacity decision as much as an ideological one.

Mia: Some commenters obviously debated whether a total ban is the right shape versus a review-based policy, but the reported reasoning from QGIS itself is the maintainer pressure. And on the other side of the ledger, Codex CLI version 0.157.0 — it now launches by default a persistent, self-updating daemon. It copies about 374 megabytes and updates itself hourly.

Milo: Hourly. That raises the question commenters were asking: what exactly is running on your machine, continuously, without you deciding each time? A self-updating persistent daemon is a meaningful shift in the software contract between a CLI tool and its user.

Mia: And then a reminder that platforms have fragility that has nothing to do with AI: GitHub had an incident on October 12th, 2026 — degraded Actions, Issues, projects, and notifications, mitigated gradually over time. Not a security event, just degradation across the core collaboration surfaces.

Milo: But it fits the theme, because if your workflow depends on Actions and Issues and the review pipeline, an outage in those is an outage in your ability to work. Three stories this hour: a project banning AI outright, a tool installing a self-updating daemon, and the central platform wobbling for a day. The next thing to watch is how the ecosystem reacts to AI-driven changes like these — whether more projects follow QGIS.

Mia: Okay, from policy to a genuine mystery in silicon. This one is delightful if you like documentation archaeology. There's an x86 instruction — vfmaddrnd231pd — it's an FMA instruction, fused multiply-add with rounding control. It's documented in Intel's SDM, the Software Developer's Manual, in the 2024 edition. And then it's absent from the 2026 edition.

Milo: A documented instruction vanishing from the manual. And the second half of the puzzle: on Haswell — which is old silicon at this point — executing it raises #UD, an undefined-opcode exception. So the hardware says "I don't know this instruction," while the 2024 documentation said it exists.

Mia: Which means somewhere between 2024 and 2026, either Intel un-documenting something that never really worked, or removing support for something that did. The mismatch between hardware behavior and documentation is what has engineers scratching their heads — the thread was mostly people doing exactly that: scratching heads. There's no explanation in what we have. It's an open puzzle, and honestly a fun one.

Milo: A nice palate cleanser. And staying in the "neat engineering" zone, there was a cluster of projects this week that all share a vibe: doing impressive things with unfashionable or underestimated technology. Let's start with the wildest one — Gran Turismo 4, running in a browser, at 50 frames per second.

Mia: And the key technical detail: it's not an emulator. The PS2 code was recompiled statically to WebAssembly. That's a fundamentally different approach — translate the game's compiled code once, ahead of time, into WebAssembly, instead of interpreting or JIT-emulating the PS2 at runtime. And it uses WebGPU for rendering.

Milo: Static recompilation of a whole PS2 game is a massive undertaking, and hitting 50fps — which is the game's native framerate, by the way — in a browser with no emulator is the proof it worked. Commenters were mostly in awe and picking apart how the approach differs from emulation in terms of fidelity and effort.

Mia: Next: w1700k-sdr. Someone turned a Gemtek Wi-Fi 7 router into a software-defined radio receiver — 320 megasamples per second, across the 2.4, 5, and 6 gigahertz bands.

Milo: That's repurposing consumer silicon for something it was never sold as. The radios in those routers are capable of wideband sampling, and this project exposes that as an SDR. Cheap wideband spectrum monitoring out of hardware you can find lying around.

Mia: Then the minimalist extreme: marketyour.games, a commercial web application written in C, served via CGI on OpenBSD. The stack is the BCHS stack — that's OpenBSD, C, httpd... actually the canonical pile — with SQLite behind it. And it's a single binary.

Milo: A commercial product on a stack most people associate with the nineties, and the argument implicit in it is that CGI plus a single C binary is comprehensible end to end. One process, one binary, no framework layers. The trade-offs — memory safety being the obvious one — were part of the discussion, but the simplicity was the draw.

Mia: And in the same family: zt, an adaptation of Templ to Zig. Type-safe HTML templates, escaped by default, transpiled at build time. So the safety is enforced by the language and the escaping is the default, not an opt-in.

Milo: That's the same design instinct as the BCHS app: move correctness to compile time and keep the runtime simple. Different language, same philosophy. Okay — that cluster flows naturally into culture and the physical world, because the next stories are also about old ideas getting new life.

Mia: Citroën is resurrecting the 2CV — the classic French people's car — as a small retro electric city car, around £13,000. It was presented at the Paris Motor Show, with sales targeted for 2028.

Milo: £13,000 for a new EV city car is a notable price point in a market where EVs skew expensive, and the retro angle is doing real work — the 2CV is one of the most recognizable car designs ever. Whether the 2028 target holds is the open question.

Mia: And on the energy side: balcony solar — one to three panels that plug into a regular outlet — is now legalized in California and nine other US states. This has been a normal thing in Germany for years, and the US is finally catching up on the regulatory side.

Milo: It matters because it drops the barrier to solar participation from "own a roof, permit an installation" to "hang a panel and plug it in." Renters can do it. That's a genuine widening of who can generate power.

Mia: And then the geopolitical story, which is stark. Yandex bot traffic — from AS13238 — dropped 99% after drone attacks on data centers. And users reported fewer pro-Russia posts afterward.

Milo: Ninety-nine percent is not a dip, that's a collapse. And the correlation users drew — bot infrastructure destroyed, and the character of the discourse changing — connects physical infrastructure to information infrastructure. Data centers are not just hosting; in this case they apparently were amplification.

Mia: With the usual caveat that correlation isn't attribution — we know the drones hit and we know the traffic fell, and we know what users reported. The mechanism between those facts is the part that remains inference. But it's a striking data point about how physical and digital infrastructure are the same infrastructure now.

Milo: Which is a good bridge to the last cluster — mobile and ecosystem odds and ends. The state of Linux phones: conditions are improving. Better hardware like the RK3588S, Qualcomm SoCs getting support, Phosh maturing as the shell.

Mia: But the report's core claim is that hardware isn't the real barrier anymore. The barrier is companies and the ecosystem. Apps, services, vendor buy-in. The phone works; the world around the phone doesn't fully.

Milo: That's the same shape as several stories today, right? The marketyour.games single binary works. The SDR router works. GT4 in the browser works. The hard part is never the clever technical core — it's the ecosystem and the economics around it.

Mia: And that's honestly the thread I'd pull through the whole episode. Hickey versus the tool argument, the review bottleneck, the maintainer burnout, the $2 trillion ask, the fridge attack, the poisoned commit, the daemon on your machine — every one of them is about the layer around the technology deciding whether it works or hurts.

Milo: Well said. That's the show for today. Thanks for listening — we'll be back with the next one.

Mia: See you then.