0922 | Cancelled, Censored, and Watermarked: Tech's Quiet Power Moves

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

Milo: And I'm Milo. It's been a busy day on the feeds, and there's a weird thread running through a lot of it — software and hardware that people didn't ask for, didn't choose, or can't turn off. And the counter-story, which is people building things and serving users well.

Mia: Exactly. We've got a cancelled Apple operating system running in a browser emulator, and then Apple today shipping AI features that users can't fully disable. We've got censorship removal from open models, hidden watermarks, an app that detects Meta's smart glasses, forks and lockouts, radios without satellites, and a PDP-11 serving web pages from 1991.

Milo: Let's start with that opening one, because it's genuinely great. Copland — Apple's famous cancelled operating system — is running in the DingusPPC emulator in the browser. Specifically Copland D11E4, which is the last build, and it took eleven patches just to unlock it enough to boot.

Mia: And for people who don't know the history: Copland was Apple's big attempt in the mid-nineties to rebuild the Mac OS from scratch. It kept slipping, kept growing, and eventually Apple cancelled it and went out and bought NeXT instead — which is how Steve Jobs ended up back. So Copland is this symbol of a decision Apple made behind closed doors that killed years of work.

Milo: Right, and the emulator part is the fun detail. DingusPPC emulates old Power Mac hardware, and getting a prerelease OS like this running means the emulator is now accurate enough to handle something that was never even finished. Eleven patches to unlock the build — that's archaeology as much as programming.

Mia: But here's why it's on the show today, and why the discussion around it got interesting. People drew a straight line from Copland to Apple right now. Apple published documentation on how to disable or restrict Apple Intelligence on the Mac — which sounds like giving users control.

Milo: Except, and this is the part commenters kept hammering, when you turn it off, the local models aren't removed. Those are roughly eight to sixteen gigabytes of AI models sitting on your disk even with everything switched off. And there's a workaround circulating for macOS 27 that stops the system from even downloading the AI models, purely so people can save storage.

Mia: And it gets better, or worse, depending on your view. On macOS 27, even when you've disabled every Siri setting and every Apple Intelligence service, there's a process called "Siri AI.app" that keeps running. Commenters ran it down, checked their settings, checked again, and it's just there, spinning.

Milo: The word that kept coming up in the discussion was "crapification." That's user language for the feeling that a product you like keeps getting things added to it that you can't remove. And people were careful to distinguish two different complaints. One is resource use — disk space, memory, background processes. That's practical, you can measure it.

Mia: The other is autonomy. Several commenters said the model downloads themselves weren't the real issue — it's that the toggle says "off" and the machine doesn't behave like it's off. If you document how to disable something but the process keeps running, what does "disable" mean?

Milo: And that's where the Copland parallel came in. Someone in the thread basically said: Apple has always shipped software its users didn't ask for. Copland was built for years and nobody outside Cupertino could stop it. The difference today is that the software ships to your machine and stays after you say no.

Mia: To be fair to the other side — and there was one — a few people pushed back and said having the models pre-installed is a latency choice. If you later flip Apple Intelligence on, it works instantly. Local models, on-device processing, that's a privacy story too; your data isn't leaving the Mac.

Milo: That's true, and the documentation existing at all is something. Some commenters noted that a decade ago Apple wouldn't have published a page about disabling features at all. But the unresolved question everyone kept circling: why can't "disable" also mean "delete"? Eight to sixteen gigs on a base-model Mac is real storage. Nobody in the thread had a good answer for that one.

Mia: Okay, so from Apple's big expensive local models to the other end of the local AI spectrum. This is the story that honestly surprised me most today.

Milo: Let's do the hardware anchor first. The M5 Ultra Mac Studio with 256 gigabytes of unified memory is getting praised as basically the ideal machine for running local AI agents. The reasoning is that agents need context — lots of it — and unified memory lets you load big models that a discrete GPU simply can't hold.

Mia: The caveat people kept repeating: an RTX 5090 setup is still faster. But to get comparable capability you're looking at a system that costs twelve thousand two hundred ninety-nine dollars. The Mac Studio is expensive too, but it's a different league of expensive, and for agent workloads — where you're often memory-bound rather than compute-bound — a lot of people said it's the sweet spot.

Milo: Now the opposite end, and this is the fun one. A project called mini-AGI trained a continuous-learning model completely from scratch — on a laptop RTX 3070 with eight gigabytes of VRAM. The author was upfront that it's still toy-level, but the techniques are interesting: a dynamic mixture-of-experts architecture, training on a single stream of data.

Mia: The discussion there split into two camps. One camp said this is the most interesting direction in the space — continuous learning is what current big models fundamentally can't do, they're frozen after training. If someone can make learning-from-experience work at small scale, that's a real research contribution, not a benchmark increment.

Milo: The skeptics said, look, it's toy-level by the author's own admission, and small-scale experiments historically don't scale — the dynamics change completely when you go up orders of magnitude. But even the skeptics conceded the constraint is the point: doing anything interesting in eight gigabytes forces efficiency tricks that might matter later.

Mia: And in between the 3070 laptop and the 256-gig Mac Studio, there's Kev — a family of tiny decision models, built on Qwen3.5, in zero-point-eight, four, and nine billion parameter sizes. The framing is that these are "Jev-style" decision models: not chatbots, not generators. They classify, they route, they verify.

Milo: That division of labor argument got a lot of agreement. The idea is you don't need a giant model to decide "which of these three tools should handle this request" or "does this output pass the check." Small, fast, cheap models doing judgment work, big models doing generation work. Several people said this is how their production setups actually look already.

Mia: Which brings us to the deployability piece, because capability is only half the equation. Cloudflare's Python Workers are now generally available. That means Python frameworks and AI libraries run natively in the Workers runtime — no glue JavaScript layer, no translating your Python into JS shims.

Milo: People who've fought with those glue layers were genuinely relieved. If you're doing AI inference or ML-adjacent work, your whole stack is Python, and until now edge platforms made you wrap it. Now it runs where it runs. The synthesis comment that stuck with me: capability was already there, deployability was the bottleneck, and this removes a chunk of it.

Mia: All right, buckle up, because the next topic is spicy. Who controls what your model says — and do the companies quietly change their minds after you've paid?

Milo: Start with the open side. There's a tool called Heretic that automatically removes refusal behavior from LLMs — the technique is called abliteration. You install it with pip install heretic-llm, point it at a model, and it strips out the censorship.

Mia: And what made the discussion substantive rather than just "neat tool" was the author's argument about measurement. He's pushing for standard metrics — how many refusals does the model still make, and KLD, which measures how much the abliterated model's output distribution diverges from the original. The point being: if you're going to modify a model's behavior, you should be able to quantify what changed.

Milo: That resonated because it's exactly what the closed providers don't give you. Which brings us to the two data points people kept citing. First: Fable 5 — Anthropic's Claude — became permanent in subscription plans, and independent measurements show thinking tokens dropped drastically in August. Not announced, measured, after the fact, by people outside the company.

Mia: Second: xAI shipped Grok 4.7 with a bigger base model at the same price — two dollars and six dollars per million tokens — scoring forty-six point three percent on CursorBench. Decent on paper. But the Hacker News reaction was a wall of complaints about laziness — users posting screenshots of responses that were just "Done!" with no work shown.

Milo: So both companies, in different ways, shipped changes that users experienced as silent downgrades or silent behavior shifts. And the commenters connected it directly to Heretic. The argument went: with open weights, behavior changes are visible and measurable — you can run the metrics. With closed models, you find out from independent measurements or from your own frustration.

Mia: To be fair, the pushback: some people said the thinking-token drop might be a legitimate efficiency improvement — fewer wasted tokens, same quality, cheaper to serve. And a "Done!" response isn't necessarily wrong, it might just be terse. The counter-counter was: then say so. Announce the change. Let users choose. The lack of communication is the actual complaint, more than the change itself.

Milo: And there was genuine disagreement about whether open tooling can really force transparency. Optimists said every independent measurement project, every abliteration tool with standard metrics, raises the cost of silent changes. Pessimists said most users will never measure anything, and the providers know it.

Mia: That theme — invisible things shaping what you see — carries right into the next topic. Watermarks, AI-sounding text, and a Bluetooth spy detector.

Milo: Two watermark stories. SynthID, Google's system, encodes 136 bits into a 512 by 512 image — and those bits can include IDs linked to identity. Separately there's Spymark, a hidden watermark that tracks users. On paper both are provenance technology. In the discussion, the split was: is this invisible provenance, or covert surveillance?

Mia: The 136-bits-including-identity detail is what bothered people. Provenance saying "this image was AI-generated" is one thing. Encoding who generated it, or tracking where an image travels, is another. A few commenters noted the irony: the same technique that fights misinformation is indistinguishable from a tracking pixel you can't opt out of, because by definition you can't see it.

Milo: And there's no easy resolution. People asked: would you accept a visible watermark? Most said yes, reluctantly. Would you accept an invisible one tied to your identity? Almost nobody. But invisible is what makes it effective against abuse. That tension is genuinely unresolved.

Mia: The text side of this is actually more encouraging. There's a survey where 78 percent of developers said they stop reading articles that look AI-generated, and 98 percent prefer the author's own writing. Which sounds grim for AI content, but the takeaways in the discussion were more interesting than the numbers.

Milo: Right — the conclusion people landed on is that intentional writing is gaining value. The tells of AI-generated prose are now so recognizable that writing which is obviously, personally human stands out more, not less. Weird specific opinions, first-hand experience, actual voice — those became premium goods precisely because they can't be mass-produced.

Mia: Which connects to a piece about the "Tetris effect" — the way algorithmic feeds hijack attention, and the argument for an intentional internet: blogs, RSS, choosing what you read instead of being fed it. People shared their own versions of the Tetris effect — that feeling where your brain keeps playing the feed even after you've put the phone down.

Milo: And the practical advice thread was basically a resurgence-of-RSS thread. Several commenters said they'd rebuilt their reading around feeds and newsletters specifically to escape the algorithm, and the developer survey numbers back it up: when people can choose, they choose humans.

Mia: Now the surveillance side, and this app has the best name of the day: ZuckOff. A free app by Polish developer Pawel Szydlowski that detects Meta smart glasses via Bluetooth. Five thousand downloads in the App Store on launch day.

Milo: How it works: it identifies the Bluetooth signatures of Ray-Ban Meta glasses, Oakley Meta glasses, and Snap's specs. Everything runs on-device — nothing leaves your phone. The use case people cited immediately: bars, gyms, private spaces where someone could be recording you and you'd never know, because the glasses look like normal glasses.

Mia: But credit where due — the author himself is upfront about the limits. Detection isn't guaranteed — Bluetooth can be off or the signature can vary — and it's not immune to spoofing. Someone could potentially fake or mask a signature. So it's a tripwire, not a guarantee.

Milo: The discussion around it went in two directions. One: this is a market response to a real gap — Meta isn't giving bystanders any signaling equivalent of a camera's red dot, so someone built a workaround. The other: it's arms-race stuff. If detection apps get popular, glasses firmware evolves, signatures change, and the tool is chasing a moving target.

Mia: The philosophical point a few people made: we have social norms around visible cameras — you can see one, you can ask about it. We have no norms yet for a camera that looks like eyewear. Tools like this are society improvising norms before the law catches up.

Milo: Speaking of who checks whom — next up is oversight, and it starts with some very big names. An independent advisory group of mathematicians at the Institute for Advanced Study — Terence Tao, Timothy Gowers, Edward Witten — will advise AI companies on how they disclose mathematical results.

Mia: The significance people kept pointing at: these aren't consultants for hire, this is arguably the most respected collection of living mathematicians, and they're positioning themselves outside the companies. When a model claims to have proved something, the current state of affairs is that verification is expensive and the companies have incentive to overclaim.

Milo: Commenters who work in math were careful, though. They noted that machine-generated proofs are getting good enough that checking them is itself becoming a bottleneck — a model produces something plausible, and verifying it takes expert time nobody budgeted for. An outside advisory body doesn't solve verification, but it creates a channel where "we think this is wrong" can actually reach the company.

Mia: The open questions were: how much teeth will it have? Advising is not auditing. What happens when a company disagrees with Tao? Nobody expects a binding mechanism, so the group's influence depends entirely on the companies wanting the credibility.

Milo: Related, in spirit: Tim Dettmers announced an Open Source Week, and his argument is that academia is being reborn — but differently. His line is that a research unit should be an ecosystem, not a paper. Meaning: not "here's a PDF and a benchmark," but code, models, data, tools, a living thing people can build on.

Mia: That hit home for a lot of people who've watched AI research move from universities to labs. The argument was that the lab model optimizes for proprietary advantage, while the ecosystem model optimizes for the field moving forward. Dettmers betting that open research units can compete for talent and impact again.

Milo: And here's the proof that openness can be a feature, not a cost: Xiaomi's MiMo-V2.6, an open-weight, multi-modal model series positioned as low-cost. What people praised most wasn't the benchmarks — it was the benchmark charts being transparent. Legible methodology, no mystery.

Mia: Which in the current climate, after everything we just talked about — silent thinking-token cuts, "Done!" responses — transparency itself becomes a competitive advantage. Open weights let people verify; open benchmarks let people trust.

Milo: One more tool in that verification bucket: Transformer Explainer, an interactive visualization that runs an actual GPT-2 — the 124 million parameter model — live in your browser, and lets you watch the transformer machinery work step by step. People called it the best teaching tool they've seen for the architecture, because you're inspecting the real thing, not a diagram.

Mia: Which is a good bridge, actually, because inspecting the real thing is exactly what the next topic is about. Sometimes the upstream maintainers say no — and then what?

Milo: Story one: Noodle Gallery. It's a soft fork of Immich, the self-hosted photo manager. Not an independent alternative — a soft fork, meaning it stays close to upstream and adds features on top. The features it adds were rejected upstream, and the rejection reason was that they were "vibecoded."

Mia: And the Noodle author denies that label, which is where the discussion got interesting. Because "vibecoded" is a loaded word right now — it implies AI-generated code nobody fully understands, unreviewed, risky. The author's position is these are real features, thoughtfully built. Upstream's position is they don't fit the project.

Milo: The commenters split in a familiar way. Maintainers sided with maintainers: a project's scope is the maintainer's call, and a soft fork is the correct, peaceful resolution — everybody gets what they want. Users sided with users: when the thing you want is reasonable and the rejection reason is an aesthetic judgment about how the code was written, a fork is justified.

Mia: The genuinely new part of the debate was the "vibecoded" rejection itself. Several people said that's a slippery slope — judging contributions by their suspected authorship rather than their quality sets a weird precedent, especially when AI-assisted code is everywhere now. Others said maintainers can absolutely decline code for any reason, including "I can't comfortably maintain this," which is a legitimate maintenance concern even if the word used was rude.

Milo: Story two, similar shape, different domain: Raspberry Pi is blocking users from swapping the RAM chips on the board, citing support problems. The criticism writes itself — this is a company whose entire brand is user-modifiable, education-friendly computing.

Mia: The defenders pointed out the real-world version: if people solder in unsupported RAM and then file support tickets, that costs the foundation time and money, and the board has tolerances. The critics said: then mark the warranty void, don't block the hardware. Others noted the irony that the Raspberry Pi of 2012 would have been impossible if the 2025 version locked everything down.

Milo: Story three, the cautionary tale, and it's Bryan Cantrill's — the Oxide founder, formerly of Sun. His telling: Sun failed because it got bored of managing the business. The company that defined server computing stopped caring about its customers' actual experience, and the proof is a startup that evaluated Sun and chose Dell instead.

Mia: And Cantrill founded Oxide specifically to not repeat that — engineering-first, treating the customer's experience as the product. People in the thread who'd lived through the Sun era confirmed the pattern: brilliant engineering, contemptuous account management, death by neglect.

Milo: And the positive counterexample that closes the loop: Linear. They cut their CI wait times from over six minutes to about five, and halved the runner time per test — while their test suites nearly quadrupled. That's the discipline Cantrill is preaching, applied to a developer tool: serve the user even as your own complexity grows.

Mia: There's even a tie back to topic two — Fable 5's forced permanence in subscription plans got pushback in the form of independent measurement. That's users holding a vendor accountable when the vendor says no. Forks, measurements, workarounds — all ways of answering "the upstream said no."

Milo: Okay, infrastructure. Who owns it, and what happens when it breaks. Three very different stories.

Mia: Start with the hopeful one. Rhizomatica's HERMES is an open-source shortwave radio system that sends encrypted data over HF radio. Twenty watts of power, links of four hundred to six hundred kilometers. No satellites, no telecom carriers, no cables.

Milo: The communities this targets are the ones the network literally doesn't reach — remote and Indigenous communities, per Rhizomatica's whole mission. Twenty watts is roughly a bright light bulb. The discussion here was part admiration, part technical Q&A in spirit: HF is slow and noisy, but it is genuinely independent infrastructure. When everything else fails, a wire antenna and a radio still work.

Mia: Then the dark mirror of that: the npm package mathmain, which hid a remote-access implant. And the activation mechanism is genuinely clever — the malicious code only detonates when the matrix from an LU solve serves as the decryption key. In other words, it stays dormant in ordinary use, and triggers only in a specific mathematical context that the attacker chooses.

Milo: Commenters called it a supply-chain sleeper with a cryptographic tripwire. The scary implication: traditional scanning looks for obviously malicious behavior, but this looks like a math library doing math until one specific input arrives. Defenders said it shows static analysis alone can't catch everything; you need provenance and trust in who published the package.

Mia: And the third story is what happens when the centralized option fails: the FAA grounded flights at busy East Coast airports after a fiber-optic cut broke communications. One physical cable, cut, and air traffic control communications fail across multiple major airports.

Milo: The thread connected all three instantly: HERMES builds redundancy out of twenty watts and the atmosphere. The npm attack exploits trust in centralized distribution. The FAA outage shows what happens when critical systems run through single physical threads. Redundancy, trust, and physical concentration — three different failure modes of the same question: who owns your infrastructure?

Mia: Let's end warm. Preservation and play — three stories about old things still teaching us.

Milo: First, a restored PDP-11/83 — early eighties minicomputer — running a Mentec M11 processor and 2.11BSD, and it is serving a live web page right now. And the detail that made the engineers happy: it needed a custom kernel, because the M11 uses IEEE floating point instead of DEC's own format, which 2.11BSD historically assumed.

Mia: So someone had to reconcile a 1980s operating system with a processor that made a different floating-point choice. That's preservation as active engineering, not glass-case curation. And people in the thread just delighted in it — a machine older than most of its readers, on the modern web.

Milo: Second, staying in the theme of inspecting the real thing — Transformer Explainer again, which we mentioned earlier, is exactly this philosophy applied to modern AI. Don't read a blog post about attention; run GPT-2 in your browser and watch it.

Mia: And third, Grim Fandango. A document of the game's puzzle designs from 1996 surfaced, and the reaction was pure celebration — LucasArts' style, the art, the music. People who never played it in 1996 said the remaster's eased controls finally let them experience it properly.

Milo: The observation that tied the three together: old machines, old models, old games — all of them are more legible than their modern counterparts. A PDP-11 you can fully understand. A 124-million-parameter model you can watch. A puzzle document that shows its design reasoning. We preserve things partly for nostalgia, but mostly because understanding how something works starts with being able to see all of it.

Mia: Which, honestly, loops all the way back to where we started — Copland running in a browser, Apple Intelligence that won't turn off. The more opaque our systems get, the more we treasure the ones we can see into.

Milo: That's the show. Thanks for listening — we'll see you tomorrow.

Mia: Goodnight, everyone.