1110 | New Model Kings and Clever Machines

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

Milo: And I'm Milo. It's a packed one today. The thread running through basically everything we're going to talk about is that the shape of computing is shifting underneath us — new models at the top of the leaderboard, new chips at the bottom of the power curve, and in between, a whole lot of people asking who they can actually trust with their data.

Mia: Yeah, and some of it is genuinely fun too. There's a paper airplane database in here somewhere. But let's start at the top, because Anthropic made the biggest move of the last day or so.

Milo: Right, so Anthropic launched two new models, Claude Fable 5.2 and Claude Mythos 5.2. And the headline claim is that these outperform Opus 5.5 on most benchmarks, while costing eight dollars per million input tokens and forty dollars per million output tokens. They also run about thirty percent faster.

Mia: That pricing is interesting in itself, because it's not cheap, but it's not the eye-watering tier either. And the flagship detail here is Fable. It scored 65 on the Artificial Analysis Intelligence Index, which is the highest score anyone has posted so far on that index.

Milo: And it has a million-token context window. The one honest caveat people kept circling back to — and I think this is the right caveat — is that Fable is very verbose. So you get the top score, but you may be paying in tokens and patience for that verbosity.

Mia: Which, if you're billed per output token at forty dollars a million, verbosity isn't a personality quirk, it's a line item. That's actually one of the unresolved questions around this launch. What we know is benchmarks, price, speed, context length, verbosity. What we don't know yet is real-world reliability. Benchmarks tell you how a model does on curated tests. They don't tell you whether it holds up over a full workday on your actual tasks.

Milo: And there was one really fun demonstration of what these models can do when they're pushed hard. Somebody had Fable 5.2 one-shot their nephew's beat 'em up game. One shot — meaning the model built it in a single prompt run. It generated roughly 175 images and 140 audio clips, and it bot-playtested all twelve stages of the game itself.

Mia: Twelve stages, playtested by the bot. That's the part I find remarkable, because playtesting is usually the human-in-the-loop bottleneck. If the model can iterate against its own game, that's a loop closing that used to require a person. It's a demo, obviously, a novelty project for a nephew, but it's the kind of demo that used to take a small team.

Milo: So that's the top of the market. But the same day's news shows the other end of the market moving just as fast, and in some ways more consequentially for most people.

Mia: Meta quietly shipped what might be the more economically interesting model. It's called Muse Ember, and it's their smallest model, and it matches Haiku 5.5's max score at roughly a third of the cost.

Milo: A third of the cost, and we're talking ten cents per million input tokens, fifty cents per million output tokens. Compare that to the eight and forty we just discussed and you can see the spread. There's a premium reasoning tier, and there's a commodity tier, and the commodity tier keeps getting better at the same price point.

Mia: Which raises the question a lot of people were asking: why would you ever pay premium prices if the cheap model matches the mid-tier model? And the honest answer is we don't fully know yet where the small models fall apart. That's the same unknown we had with Fable, mirrored at the low end.

Milo: And underneath the model layer, the hardware layer is churning. Google expanded its TPU portfolio with two new parts: the 8tx, which is focused on secure inference, and the 8i Max, which is built for massive context. And the context there is that this makes four TPU systems Google has shipped in one year.

Mia: Four systems in a year. That's a cadence that used to belong to phone makers. And notice the specialization — one chip for security-sensitive inference, one for enormous context windows. That maps directly onto what the model makers are doing. Fable has a million-token context; Google builds a chip for massive context. These companies are designing hardware for the workloads they expect.

Milo: So if you zoom out, the prediction that falls out of today's news is: cheaper, more specialized inference across the board. Cheaper models, specialized silicon, and — this is the part that connects to our next topic — specialized algorithms.

Mia: Because it's not only chips. There were three algorithmic results that all point the same direction: you can get speedups in software that people usually go buy hardware for.

Milo: The flashiest one is called Next-Concept Modeling. It's a coarse-to-fine approach using VQ-VAE for text generation, and the claim is up to 21 times faster than diffusion language models at matched scale.

Mia: And the key phrase there is "at matched scale." Speedup claims are cheap when you compare against something smaller or worse. Matched scale means this is a real architectural comparison, and 21x is not an incremental number. If it holds up outside the paper's conditions, it changes what diffusion LMs are for.

Milo: Then there's Flide, which is a different beast. It's an exact inference engine for Jev-style decision models, and it gets up to 9.5x faster than vLLM or SGLang by using exact state forking.

Mia: And "exact" is doing a lot of work in that sentence. Most inference speedups come from approximations — quantization, sampling tricks, speculative decoding. Exact forking means you're not trading accuracy for speed, you're restructuring the computation so the same answer comes out faster. That's the kind of speedup that doesn't degrade.

Milo: The third one is older-school but maybe the most broadly useful: e-means. It's a mini-batch k-means method that reaches the test error of Lloyd's algorithm at 25 epochs in just 1.6 to 2.4 epochs. It does that by using scaling laws and by reviving dead clusters.

Mia: K-means is everywhere — clustering, vector search, preprocessing. If you can get the same quality in a fifth of the epochs, that compounds across every pipeline that touches it. The mechanism is interesting too: dead clusters are a classic failure mode where centroids get stuck with no points assigned, and reviving them is a known idea applied systematically here.

Milo: So the pattern across all three: the speedups aren't coming only from silicon. Chips and algorithms are racing each other. And then there's a chip story that pushes this race all the way to the edge of the network.

Mia: DeepX taped out a chip called the DX-M2, an edge chip, and the claim is that it can run inference for 100-billion-parameter language models under five watts. The trick is Samsung's PIM — processing-in-memory — where compute happens inside the memory itself.

Milo: Five watts. For scale, a laptop CPU under load pulls tens of watts. If a 100-billion-parameter model really runs on five watts, that's phone-or-smaller class hardware running frontier-adjacent models locally.

Mia: And I want to flag the appropriate skepticism here: this is a tape-out, meaning the design is committed to fabrication, but it's a claim until silicon arrives. Samples are expected in March 2027, which is a long way off. Tape-outs fail, and claims made at tape-out time are the most optimistic version of the story.

Milo: But even with that caveat, it extends the same race we've been describing. Frontier models on datacenter chips, cheap small models, better algorithms, and now the attempt to push inference down to the edge where the data lives. Everything is pointing at "inference gets cheaper and closer to you."

Mia: Okay. Let's pivot, because there's a cluster of stories about the less shiny side of all this — what happens to people's behavior and expectations when AI is everywhere.

Milo: The first one is genuinely unsettling if you care about transparency. Someone analyzed 1,400 Claude Code sessions and found that Opus 5.5's reasoning changed sharply on November 3rd — without the model name changing.

Mia: That's the key detail. Not "a new model shipped." The same-named model behaved differently, sharply, on a specific date. Which raises the obvious question: what changed under the hood? Was it a silent update, a routing change, a re-tune? Users had no announcement to point to. They noticed because they were paying attention to their own sessions.

Milo: And it's a transparency problem even if the change was an improvement. If your tool's behavior shifts and the name doesn't, you can't build on it reliably. People designing workflows around a model need to know when the ground moves.

Mia: The second story is about AI changing behavior nobody chose. Somebody tracked cold emails on Hacker News and the numbers are stark. From 2013 to 2021, pitches were roughly zero percent of the cold emails people received. In 2026, it's 88 percent product pitches, at record volume — and the driver is LLMs.

Milo: That's not just spam getting worse. That's LLMs making personalized-sounding outreach so cheap that the entire signal of "someone took time to write to me" collapses. When writing to a specific person costs nothing, writing to a specific person means nothing.

Mia: And the recipients didn't choose this. The senders mostly didn't either — they're responding to incentives. When one person sends ten thousand AI-generated emails, everyone else has to too, just to compete. It's an arms race nobody voted for.

Milo: Which brings us to the trust cluster, because there were three stories about tools and institutions asking for trust they may not have earned.

Mia: The big one is Google. They've reversed a 2023 promise: consumer Gemini can now route your personal Drive and email content to human annotators.

Milo: And "reversed a promise" is the operative phrase. In 2023 the assurance was that your consumer data wouldn't be reviewed by humans in that way. Now that routing exists. If you built habits on the 2023 assurance, that foundation just moved — which, notably, is the exact complaint people had about the silent model change we just discussed. The pattern repeats at a bigger company.

Mia: The second one is more technical but arguably more dangerous. ChatGPT Desktop forwards the user's SSH agent to remote machines. Which means a remote agent can authenticate using your local SSH credentials.

Milo: Let's unpack why people found that alarming. Your SSH agent holds the keys to your servers. Forwarding it is a real, legitimate technique — it's how you hop between machines without copying keys around. But if you forward your agent to a machine that's running an AI agent with broad permissions, you've just given a software agent the ability to authenticate as you to whatever those keys open.

Mia: Powerful and convenient, yes. But the blast radius of a prompt injection or a mistake on that remote machine is now your entire SSH footprint. The capability is fine; the default question — should this be easy to do? — is what people were arguing about.

Milo: And the third story in the trust cluster is the strangest. There's a lawsuit alleging that the Ethereum Foundation Mandate required employees to sign a "seppuku" pledge — a suicide pledge — or take compensation and leave.

Mia: These are allegations from a lawsuit, so we should be careful: they haven't been proven. But the claim itself, that an organization tied its employment terms to something that extreme, is the kind of thing that, if it holds up, says something about power dynamics inside institutions people trusted.

Milo: So the trust thread: a search company walks back a privacy promise, a desktop app hands your keys to remote machines, and a foundation faces allegations about coercion. Different scales, same underlying question — who is actually looking out for the user?

Mia: And that question has a surveillance-flavored answer in our next group of stories, which are literally about watching and being watched.

Milo: First, Someone Was Here. It's a Mac menu bar app that photographs users while they're away from their machine. It does face recognition locally, and phone alerts are a 4.99 dollar lifetime purchase.

Mia: The framing matters here. This isn't spyware — it's a tool for catching whoever uses your computer when you're not there. Your roommate snooping, a coworker, a thief. Local face recognition means the images don't need to leave your machine, which is the privacy-respecting architecture for a surveillance tool.

Milo: Though it's worth sitting with the irony: you install it to catch people watching you, and now your camera photographs anyone near your machine. The line between protection and surveillance is thin and depends entirely on who controls the data.

Mia: Then there's MeowDisplay, which is open-source and goes the other direction — it mirrors your Mac onto an iPhone or iPad. It has a Window view, Stage Peek, and Direct Touch, with latency around 14 to 17 milliseconds.

Milo: Under 20 milliseconds is the threshold where screen mirroring stops feeling like mirroring and starts feeling like the actual screen. Direct Touch means you're tapping the iPad and the Mac responds as if you touched it. For an open-source project to hit that number is a real engineering accomplishment.

Mia: And then the zoom-all-the-way-out version: Watched.earth. It shows every satellite that passed over any place on Earth in the last 7 days, with real Sentinel-2 and Landsat imagery.

Milo: So type in an address and you can see which observation satellites flew over it this week, and look at their actual photos. Public Earth-observation data has existed for years, but the interface turns it from an archive into something like a window. Suddenly anyone can ask "who was looking at this place, and what did they see?" — and get an answer.

Mia: Which loops right back to the surveillance theme. Satellites, cameras, face recognition. And in the middle of all that watching, there's a small story about making AI output look less like AI.

Milo: Right — STE80 Humanizer. It's an open-source agent skill, meaning something an AI agent can load and use, that removes 26 known AI writing tells and rewrites text 80 percent of the way toward ASD-STE100.

Mia: For people who don't know that standard: ASD-STE100 is Simplified Technical English, the writing standard used in aerospace documentation. Short sentences, controlled vocabulary, no ambiguity. So this skill isn't just "make it sound human" — it pushes text toward a formal, spec-like register.

Milo: And think about what that implies. There's now a market for tools whose entire job is making machine text pass as not-machine text. That exists because, as we said with the cold emails, the tell that something is AI-written is becoming a social signal people act on. Tools like this are the countermeasure to the countermeasure.

Mia: It also slots into the agent wave we opened with. Skills like this are how agent ecosystems mature — not just agents that do tasks, but agents that fix other agents' output. Okay, let's shift to something a bit more grounded: lean software. There was a whole cluster of projects whose shared idea is that software has gotten fat and it doesn't have to be.

Milo: The flagship comparison is DBDelve. It's an open-source database client written in Rust using GPUI. Launch time: 1.1 seconds. Idle memory: 133 megabytes. A 1,000-row query: 0.21 seconds. And the comparison everyone reached for is DBeaver: 5.2 seconds to launch, 606 megabytes idle, 1.96 seconds for the same query.

Mia: So roughly five times faster to launch, a quarter of the memory, ten times faster on the query. And the point isn't really "DBeaver bad" — DBeaver is a mature, feature-rich tool. The point is what the gap reveals: a lot of that weight is framework and abstraction tax, not functionality.

Milo: Next in the same family is CUI, an immediate-mode UI library for C99. About 170 components, no dependencies, and 0.1 to 0.25 milliseconds per frame.

Mia: A quarter of a millisecond per frame. And no dependencies in C99 is a genuine feat — most C UI work involves dragging in platform toolkits. Immediate mode is also a philosophy choice: instead of maintaining a tree of objects, you redraw state every frame. Simpler mental model, tiny footprint.

Milo: Then zebkit, which attacks the heavyweight champion of heavy software: Electron. zebkit embeds Chromium — Blink, V8, PDFium, FFmpeg — as a linkable library. Result: 34 percent smaller and 2x faster cold start than Electron.

Mia: The insight there is that Electron ships Chromium as a whole runtime around your app. zebkit says: you're already linking C libraries, so link the browser's engine too. Same web tech, less scaffolding.

Milo: And then two deep cuts. First, glibc malloc now batches brk syscalls — a thousand malloc-of-16-bytes calls resulted in only 2 brk calls. And malloc at 16 bytes is 22.7 times faster than calling sbrk directly at 16 bytes.

Mia: The lesson there is one of the oldest in systems programming: the syscall boundary is expensive, and batching across it wins. What's notable is that this optimization landed in glibc — the foundation of nearly every Linux program — so everyone benefits without changing a line of their own code.

Milo: And the deep-est cut: VM/370 CE. It's a community-improved, license-free version of IBM's last unrestricted VM/370, and it runs via Docker with Hercules emulation.

Mia: So people are running a 1970s mainframe hypervisor, in a container, on a modern laptop. It's nostalgia, but it's also a statement: this software was good enough that people maintain it, license-free, fifty years later. Lean software lasts.

Milo: Which connects to the next topic, because infrastructure — who owns it, what it costs — was itself a story today.

Mia: Payments Canada launched the Real-Time Rail. That enables 24/7 real-time clearing and settlement of payments nationwide.

Milo: The significance is the word "settlement." Real-time payment systems in various countries still batch settlement later, which creates risk. Actual real-time clearing and settlement means money genuinely moves at 3 a.m. on a Sunday, final, nationwide. That's a foundation everything else — payroll, instant transfers, new fintech products — can build on.

Mia: And the second story is about public data changing hands: USPS now sells official ZIP Code boundary polygons, starting at ten thousand dollars. Prior options are free approximations.

Milo: And people were genuinely conflicted about this. On one hand, ZIP Codes were never designed as geographic areas — they're mail routes — so official polygons are a real product that someone has to maintain accurately. On the other hand, ZIP Code data is foundational to demographics, insurance, epidemiology, voting analysis. Gating the official version at ten thousand dollars means a lot of civic and academic work runs on approximations.

Mia: It's part of a pattern with the payments story: infrastructure exists, but access to it is increasingly metered. The rails are open to banks; the maps are open to whoever pays.

Milo: Next cluster: hardware you can touch — solar, RAM, and silicon carbide.

Mia: Start with the nice one. Magenta semi-transparent CdTe solar panels let red and blue light reach crops underneath. And in the reported test, broccoli yield stayed steady under the panels.

Milo: That's the agrivoltaics idea — farms that generate power and grow food — but the transparency detail is what makes it work. Regular panels shade too much. Wavelength-selective semi-transparent panels pass the parts of the spectrum plants actually use. Steady broccoli yield means the panels cost the farm nothing in output and gain it a power plant.

Mia: Then the painful one for anyone who builds things: the Raspberry Pi 5 with 16 gigabytes of RAM is now 385 dollars — an 80-dollar increase.

Milo: And the reason is a memory shortage with a long horizon. Micron sees the RAM shortage lasting to 2028. SK Hynix sees it to around 2030.

Mia: That timeline is remarkable. Memory prices are cyclical, but multi-year horizons from the manufacturers themselves suggest this is demand-driven at a structural level — everyone building AI datacenters is buying the same DRAM that goes into hobbyist boards. The AI boom is taxing the price of a Raspberry Pi.

Milo: And the third item is Infineon, which presented the world's first functional power devices on 300-millimeter silicon carbide wafers at electronica.

Mia: The number to know is 300 millimeters. Silicon carbide has historically been made on smaller wafers, which limits output per batch. Moving to 300mm — the standard size for commodity silicon — means more chips per wafer and, eventually, cheaper power electronics. That matters for EVs, chargers, grid gear — anywhere efficiency and heat matter.

Milo: And notice the thread: silicon carbide for power, DRAM scarcity from AI demand, solar over crops. The physical layer of computing is being rebuilt right now, and it's competing with food, cars, and everything else for materials.

Mia: Okay. Last topic, and it's the human one. We'll close on it.

Milo: The Economist published an argument that humans will never again create a project quite like Wikipedia.

Mia: And the reasoning behind that claim is worth restating, because it's not about Wikipedia being irreplaceable technically. It's about the conditions that produced it — a massive voluntary, non-monetized, global collaboration at a moment when the internet still felt like a commons. Those conditions, arguably, don't exist anymore. Attention is monetized, contribution is platformed, and — tying back to our cold email story — authentic human effort is now the scarcest resource online.

Milo: And against that slightly melancholy claim, the source we found most charming: Foldable Flight. It's the world's largest paper airplane database — 109 planes, each with step-by-step folding video tutorials.

Mia: A hundred and nine paper airplanes, lovingly documented, by humans, for free, for fun. It's either a counterexample to the Economist's thesis or a poignant last specimen of it.

Milo: And the closing thought came from a Children of Men discussion. People were debating the ending — the fog, the ship, the ambiguity of whether the boat really came. And someone grounded the whole dystopia in data: the film's premise is global infertility, but in reality the global birthrate is still above 2.0, though it's declining widely.

Mia: That's the right note to end on. Children of Men works because it takes a real trend — declining fertility, eroding institutions, people losing faith in the future — and imagines it run to its limit. The trend is real; the catastrophe isn't here. Whether the Wikipedia-style projects, the voluntary commons, the things we build together for free — whether those survive the transition is genuinely unresolved.

Milo: And that's honestly the theme of the whole episode. Faster models, cheaper inference, cleverer algorithms, leaner software — the capability curve is all going one direction. The open questions are the human ones: reliability we can't yet verify, trust that keeps getting walked back, and collaboration we may not be able to rebuild.

Mia: We'll be back with more tomorrow. Take care, everybody.

Milo: See you next time.