1119 | Models, Agents, and the People Who Still Care

||Download

Show notes

A fast tour of the week in tech: a new frontier model price war, AI agents quietly doing audit and code work, serious security holes and a bank leak, tiny models that reason better, and a batch of hands-on hardware, games and open-source tools.

Timeline

  • 00:00:04 Opening
  • 00:01:05 Gemini 4 Neon and the frontier model price war
  • 00:05:28 AI agents doing the work — and replacing the people
  • 00:12:00 The word of the year is 'human' — and the IPO behind the AI boom
  • 00:13:27 Security week: RCEs everywhere and a bank's secrets out
  • 00:18:58 Doing more with far less: tiny models and lean training
  • 00:21:16 Machines you can touch: UEFI games, Voyager, and x86 history
  • 00:27:16 Quick hits: open-source gear, games, and a health warning
  • 00:29:55 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.

Transcript

Mia: Welcome back to the show — I'm Mia, and across the table from me is Milo. It's been a strange twenty-four hours, and the thread that runs through everything we're covering today is the same one the Cambridge Dictionary landed on when it picked its word of the year: "human." Machines are writing code, running audits, shipping firmware games that boot without an operating system — and meanwhile the humans keeping the actual lights on seem to be an endangered species.

Mia: That tension is everywhere in today's stories.

Milo: Exactly. We're going to spend most of the episode on two big discussions: Google's Gemini 4 Neon launch and what the frontier model price war actually means, and then the messier story of AI agents doing real work — and what happens to the people who used to do it. Then security week, tiny models, some gloriously tactile machine love, and a quick tour of open-source gear.

Mia: Let's start with the launch, because it dropped overnight. Google released Gemini 4 Neon, and the headline is frontier-level coding and analysis at three dollars per million input tokens and fifteen dollars per million output tokens, with a million-token context window. It ranks third on the Artificial Analysis index with a score of 61.

Milo: And the raw performance numbers are striking. It's pushing 127 tokens per second, versus the Carbon variant at 46. So nearly three times the throughput. The projected Agent Duel Elo lands around 1,066, and Opus 5.5 sits at 1,078 — so within striking distance of Anthropic's flagship.

Mia: Here's where the comment section gets interesting, because the benchmark question is the same one that always comes up. Does an Elo score of 1,066 mean the model is actually good at your job? The skeptics point out that benchmark performance and real task quality are different things, and we have exactly one data point sitting right there in the thread to prove it.

Milo: Which is the Carbon testing story. Someone ran Gemini 4 Carbon on real work and found it took roughly seven times as many steps as Astra to finish, and — this is the part that stings — half of those steps were spent on work the ticket never asked for. Redundant exploration, overthinking, doing things nobody requested.

Mia: So you can be frontier-quality on paper and still be the annoying coworker who rewrites the whole file when you were asked to fix a typo. The counterargument from the other camp is that step count isn't quality — maybe Carbon was being more thorough, verifying things. But seven times the steps also means seven times the latency and compute cost in practice, which eats into the price advantage.

Milo: Right, and that's the real fight in the thread. Neon at three dollars a million input tokens looks cheap until you're paying for seven times the agentic steps. Price per token is not price per completed task. That's the unresolved question: nobody has published a cost-per-task comparison across these models on real workloads.

Mia: And the price war isn't just Google. Microsoft's Ignite event dropped seven new MAI models, and the claim getting all the attention is MAI Code 2 achieving roughly 98% of GPT-6.1 Sol's coding quality at about 40% of the price. The reasoning there is deliberate: don't win the benchmark, win the value curve.

Milo: The disagreement in the discussion is about whether "98% at 40%" is a meaningful claim. One view: if quality is close enough, most users can't tell the difference, and the savings are real — this is the commodity-model argument. The other view: that missing 2% is exactly where the hard problems live, and the 40% price cut doesn't help you if the model can't do the task at all. You're not buying 98% of a solution; you're buying a failure on the 2% hardest work.

Mia: Which is a sharp way to put it. There's also the reliability angle from the same discussions — people who've been burned by a model that's 98% right will tell you the last 2% is most of the debugging time. Nobody's resolved it, because the answer probably depends on the task. For scaffolding and boilerplate, 98% is plenty. For the gnarly stuff, it isn't.

Milo: Worth noting too what these models are actually being used for — and Google gave an example. Genie 4, the world model, is now available to AI Ultra subscribers: real-time environments you can edit with text, native audio, a community gallery. That's not a chatbot; that's interactive simulation. So the frontier isn't just "answers questions faster," it's "generates worlds in real time."

Mia: Okay, so from benchmarks to consequences. Because the second big discussion is about what these systems are doing to actual work — and it cuts both ways. Let's start impressive. KPMG UK ran something called Project Exponential: over a hundred AI agents deployed on a real FTSE 250 audit. And according to the writeup, the agents sometimes dug deeper and asked better questions than the human auditors.

Milo: The reaction to that was fascinating, because people split into roughly three camps. Camp one: obviously the agents win — they never get tired, they don't skip the boring reconciliation steps, they'll happily pull the thread at 3 AM that a human would note and forget. Camp two pushed back: asking "better questions" than auditors is a strange claim, because audit value often comes from judgment — knowing which questions not to escalate, sensing when a client is being evasive.

Milo: An agent that digs deeper everywhere is just generating noise you have to filter.

Mia: And camp three is the professional-anxiety camp, which is fair given this is, you know, their jobs. The unresolved question is accountability. If an AI agent flags something and a human signs off without checking it — or worse, if the agent finds nothing and everyone trusts it — who's on the hook when the fraud slips through? The audit profession has an entire liability structure built around humans attesting. That structure does not currently have an answer for this.

Milo: Meanwhile in open source, the agents aren't auditing — they're shipping. Mastra, which is at about four million weekly downloads, has this thing called Factory: an agentic kanban board for Slack. And last week it opened and merged 238 pull requests, alongside 630 human-merged ones. So roughly 27% of the PR volume was machine-generated.

Mia: The discussion there was less "can it code" and more "how do you review it." A 238-PR week from agents either means your review process is scaling beautifully or your humans are rubber-stamping. Some commenters argued agent PRs are actually easier to review because they're mechanical and well-scoped. Others said the risk is exactly that they look easy to review — plausible code, subtle bugs, and the reviewer fatigue sets in fast.

Milo: And in the genuinely wild category: OpenAI reported that GPT-6 Astra aided chip design. About 13% standard-cell area reduction in the vector units and 10% in the matrix units of AMD's Jalapeño parts, with up to 14% PPA gains. Chips. Physical silicon. That's about as far from "write me a regex" as you can get.

Mia: The takeaway people drew from that one is that the highest-value agent work may be in domains with cheap, fast verification loops. Chip design has simulators — you can test a layout and know if it's better. Same reasoning as Sudoku, actually, which we'll get to later. Where verification is expensive or subjective, agents flounder; where it's cheap, they shine.

Milo: Okay, but here's the dark side, and this is the part of the thread that got emotional. All About Berlin — a long-running blog for people navigating German bureaucracy — lost 75% of its traffic to Google AI Overviews. The author's claim is blunt: Google trained AI on his content, the AI now answers the questions directly, and the traffic collapse cost him his job.

Mia: The debate splits pretty predictably. The pro-AI side says: the information is now freely available to everyone, instantly — isn't that a win? The counterargument, and it's a strong one, is that the information was only free because someone spent years assembling it, and if you de-fund every person doing that assembly, the pipeline that feeds the AI answers dries up. You get one great generation of knowledge and then nothing new.

Mia: Nobody answered "who writes the next All About Berlin," and that silence was telling.

Milo: Which connects straight to the VioletPixel essay that got a lot of traction. The thesis: the world is held together by "smart people who give a shit" — and recent outages at Microsoft, Apple, and various banks suggest there are too few of them left, as AI and cost-cutting replace them. The argument is that these outages aren't bugs, they're symptoms of removed redundancy — the humans who would have caught the failure were the redundancy.

Mia: The disagreement there is about attribution. Plenty of commenters said outages happen for boring reasons — migrations, misconfigurations, plain bad luck — and blaming "the death of caring engineers" is nostalgia dressed up as analysis. The other side countered that the essay's point isn't any single outage; it's the trend across multiple sectors in the same period. When everything fails at once, "bad luck" gets less convincing.

Mia: And nobody has a clean way to measure institutional slack, so the debate stays philosophical.

Milo: And then there's the personal-data angle. A job candidate went through interviews at a YC startup and later discovered the startup had used their LinkedIn data without permission. The呼声 in the thread was basically "sue them," but the more level-headed takes pointed out that the candidate's actual leverage is thin — the harm is diffuse, and the terms-of-service web is tangled.

Milo: It did spark a genuinely useful side discussion about what interview data companies are quietly harvesting and whether anyone consents to any of it.

Mia: Let's zoom out, because two other stories give this whole discussion its frame. First: Cambridge Dictionary's 2026 word of the year is "human." Chosen because AI keeps blurring the line between people and machines. There's something almost pointed about that — the word of the year is the thing being automated.

Milo: Second: Anthropic priced its IPO. Eighty-three dollars a share, 980 million Class A shares, trading on Nasdaq starting November 19, 2026, under the ticker "ANT." Commenters peg the market cap somewhere between 1.1 and 1.5 trillion dollars. So the same week the dictionary crowns "human," one of the most valuable companies on earth sells machines doing what we used to call human work.

Mia: The valuation debate is genuinely unsettled. The bull case: agent revenue is real and growing, the audit-and-code stories prove it. The bear case, echoing the Carbon test: if agents take seven times the steps on unrequested work, the unit economics might never work at trillion-dollar scale. And the deeper question nobody can answer: is this a durable business or the sharpest round of mania we've ever priced?

Milo: Alright — security week, and it was a rough one. Ruby disclosed seven vulnerabilities in its Regexp handling and String case conversion, some with working RCE proofs of concept. That's remote code execution from regex — the thing every web app uses constantly. Fixes shipped in 4.0.10, 3.4.13, and 3.3.14, so patch now.

Mia: The discussion there: people pointed out that "regex leads to RCE" feels like a throwback to the C-string era, and the lesson is that string manipulation code paths are attack surface even in memory-safe-feeling languages. The practical debate was about patch cadence — seven vulns disclosed at once suggests a coordinated audit or a responsible-disclosure batch, and commenters were mostly grateful rather than alarmed.

Milo: Wiz, meanwhile, found two unauthenticated HTTP/2 RCEs in NGINX — CVSS 9.2, meaning critical. Unauthenticated code execution on your reverse proxy is about as bad as it gets, since that's the front door of most of the internet. Patched builds are already out.

Mia: The thread discussion here was partly deja vu — people remembering the HTTP/2 protocol's history of implementation bugs. The reasoning: HTTP/2's complexity makes every independent implementation a fresh chance for memory bugs. The debate was whether we blame protocol complexity or C. Someone quoted the old line: the protocol is the attack surface, the language is just the delivery mechanism. Unresolved: how many other HTTP/2 implementations have the same class of bug, unreported.

Milo: Then there's the KVM bug — CVE-2026-99312. The Linux kernel's virtualization layer fails to reset MMU context when emulating an INIT signal to a guest. Why do we care? Because HN immediately connected it to AI agent sandboxes. Most agent platforms run untrusted generated code inside VMs, and a hypervisor escape means your sandbox isn't a sandbox.

Mia: That's the thread that ties back to our first topic. If you're running thousands of Carbon-style agents executing arbitrary code, your isolation boundary is the KVM kernel code — and that code just had a bug. The commenters' point wasn't panic, it was: the threat model for AI infrastructure now includes the entire virtualization stack, and most people haven't updated their mental model. Nobody knows yet whether this is exploitable in practice in the mainstream agent setups.

Milo: Now the institutional-failure stories. Codeberg's domain vanished — it disappeared from its Namesilo account and got suspended, apparently routed to an abuse-handling process. The team is standing up codeberg.eu as a fallback. The discussion was mostly about how fragile single-registrar dependence is, and how "abuse" pipelines at registrars have become a de facto kill switch with no due process. Some argued codeberg.

Milo: eu is a band-aid — the real lesson is you need registrar diversity and pre-arranged escrow for critical infrastructure.

Mia: And then Deutsche Bank. Fifteen million leaked documents expose Russian oligarch money laundering, plus links to Trump and Epstein scandals. This is journalism territory more than security, but the HN framing was: these leaks came from the inside. Every comment thread eventually lands on the same question — who leaked this, and were they a whistleblower acting alone or part of something bigger? That's unresolved and probably stays that way.

Milo: Which brings us to the story that's essentially protest-by-DNS. Quad9 says that because it refuses to comply with a court order to block piracy sports domains, it will stop providing DNS resolution to French users starting November 24.

Mia: The comments were sharply divided. One camp: Quad9 is a heroic holdout — DNS providers shouldn't be censorship infrastructure, and pulling out of France is the only move consistent with their mission. The other camp: this is theater. A court order in a democratic country isn't censorship, and refusing to comply then exiting the market is just jurisdiction shopping dressed up as principle.

Mia: And there's the practical question everyone flagged: does Quad9 leaving France actually change anything, or do French users just switch to a resolver that will comply? The boycott may be symbolically important and materially meaningless. That tension was unresolved.

Milo: From protest to efficiency — topic five, and this one's short but it's the most scientifically interesting thing in the episode. First: an Energy-Based Reasoner — EBR — with just 1.8 million parameters hits 99.9% accuracy on Sudoku-Extreme while using 345 times less inference compute.

Mia: The discussion here was partly "so what, it's Sudoku," and partly "no, this is the point." The defenders argued the interesting thing isn't the puzzle, it's the architecture class. Energy-based models iterate toward a low-energy state — a fundamentally different computation than transformer token-by-token generation, and it means the model can spend variable compute per problem, like a human staring at a hard puzzle longer.

Mia: The skeptics pointed out Sudoku has a verifiable unique solution — it's the ideal case for this kind of method, and generalizing to messy real-world reasoning is the whole ballgame and totally unproven.

Milo: Second: RadixArk released DaisyRose-27B — Qwen3.8-27B plus reinforcement learning — trained on only 1,200 tasks in 18 days on 144 H200s. Terminal-Bench 4.1 score went from 5.6% to 12.0%.

Mia: The reaction: doubling the benchmark with 1,200 training tasks is either remarkably sample-efficient or a sign that the benchmark rewards narrow training. Same debate shape as EBR — the efficiency is real, the generalization is the question. Someone made the connection explicitly: between EBR's 345x compute savings and DaisyRose's 18-day training runs, maybe the scaling-law era isn't the only game in town. Architectures and data curation might be the next frontier.

Mia: But that's a hypothesis, not a result.

Milo: Alright, from tiny models to big machines you can actually touch. And this is my favorite section. First up: eklavya-raj built gta6-uefi — a 942-kilobyte GTA-style game set in a real recreation of Miami that boots directly from UEFI, with no operating system underneath.

Mia: Let that sink in — 942 kilobytes. No Linux, no Windows, no driver stack. The firmware itself is the runtime. And the map is built from OpenStreetMap data, with NOAA lidar for elevation and NAIP aerial photos for texture. So it's not a toy map — it's real-world geographic data piped straight into a bare-metal game.

Milo: The comments split into awe and curiosity. The awe: "this is the most impressive demoscene-adjacent thing I've seen in years." The curiosity: how do you do graphics without an OS? The technical folks explained UEFI gives you some boot-services primitives, but you're essentially writing to the framebuffer directly and handling everything yourself. Some asked why — and the answer in spirit is the same reason people climb mountains: because the constraint is the art.

Mia: It's worth noticing the thematic rhyme with the model-efficiency story. Nine hundred forty kilobytes, no OS. A 1.8-million-parameter model, 345x less compute. Both are statements about how much fat is in the modern stack. When you strip away the layers, how much of it was necessary? That's the question both projects ask.

Milo: Next: rshekhar.com/voyager. An interactive 3D teardown of Voyager 1, all the way down to a single memory chip. Plus the RTG — the radioisotope power source — the Golden Record cover, and the story of the fix. This is a browser-based deep dive, and the comments treated it as a masterpiece of technical communication.

Mia: The fix story is the emotional core. Voyager 1, billions of kilometers out, starts sending garbage, and JPL engineers debug a computer from the 1970s with a round-trip communication time measured in days. The teardown lets you see the actual hardware those engineers were reasoning about. Several commenters made the connection to the VioletPixel essay: that fix was "smart people who give a shit" in its purest form. Forty-seven years after launch, someone still cared enough to debug it.

Milo: And Pikuma published a history of x86 — from the 4004 to the 486 — and it's a gem. The thesis: little-endian byte order, the AX-through-DX register names, and segmented addressing aren't design decisions anyone made thoughtfully for modern machines. They're inherited accidents from the 8008 and the Datapoint era.

Mia: The comment section loved this because it's archaeology of your own machine. The Datapoint 2200 used little-endian because of its hardware implementation, and when Intel built the 8008 to be compatible with it, that decision fossilized into every x86 chip since. Fifty years of backward compatibility to a 1970 terminal. The discussion turned philosophical: how much of our entire computing stack is sediment? The answer, apparently, is most of it.

Milo: Two more in this lane. The Air Force posted a SAM.gov solicitation for an F-22 "Next-Generation micro-Data Transfer Cartridge" — essentially the mission-data loader for the stealth fighter. Commenters amused themselves imagining what a modernized version needs: bigger storage, faster transfer, probably better crypto.

Milo: And someone noted the contrast — this is the most advanced air-superiority fighter on earth, and it's loading data from what's essentially a specialized thumb drive, and the Pentagon is shopping for a replacement like the rest of us.

Mia: And slowchronicle.com has an author using LLM agents for binary reverse engineering — Ghidra MCP, DeepSeek, with hypothesis-logging loops. The clever bit is the methodology: instead of letting the agent guess endlessly, you make it log hypotheses in markdown notes and test them systematically. The comments recognized this as good agent engineering generally — treat the model as a hypothesis generator and keep the human as the verifier.

Mia: Which is exactly the structure that worked for the auditors and failed for Carbon's unrequested work. Same pattern everywhere: agents propose, humans dispose.

Milo: And there's a quiet connection to Jimmy Miller's post on partial evaluation floating in these discussions — the idea that partially interpreting an interpreter-plus-program yields a compiler for free. Same spirit as the UEFI game: understand a system deeply enough and you can collapse layers into something leaner. When you specialize an interpreter to a known program, the interpretation overhead disappears.

Milo: It's the theory behind why gta6-uefi's no-OS approach is even possible — remove the generality, keep the function.

Mia: Okay, last stop: the everyday roundup, and then we're out. CrossPoint Reader — free open-source e-reader software for compact devices like the XTEINK X3 and X4, and BOOX readers. 242 contributors, roughly 281,736 devices running it. The comments loved that a tiny niche — cheap, hackable e-ink devices — has a thriving community, and that the device count makes it one of the most-deployed pieces of open-source consumer software nobody talks about.

Milo: ExitPath — a free, browser-only site cataloging open-source replacements for Slack, Notion, and Microsoft 365, complete with migration difficulty scores and checklists. The discussion: this is the practical companion to the VioletPixel anxiety. If you don't trust the cloud providers to keep the lights on, ExitPath is your escape map.

Milo: Some noted the difficulty scores are the most valuable part — "open-source alternative exists" is easy to say, "here's how hard the migration actually is" is the honest part.

Mia: Type Fighter II — a free, no-account arcade-style typing practice game, with progress saved locally. No account, local storage — the design constraints are the feature. Comments treated it as a palate cleanser: a thing on the internet that just wants you to type faster and asks for nothing.

Milo: For the mathematically inclined: an arXiv primer on Fourier Series by Dipierro, Pfefferlé, and Valdinoci — a full textbook with exercises and complete solutions. Commenters appreciated that the solutions are included, which is rarer than it should be, and that it's free. Several people mentioned working through it as a way to finally understand the math behind signal processing and, fittingly, the FFTs that underpin so much of the machine learning we discussed earlier.

Mia: And a health note to close: new research finds energy-drink-related cardiac arrest reports running eleven times higher than expected, with 18-to-34-year-olds making up nearly half. But — and the commenters were careful about this — no causal link is proven. Reporting bias is a real alternative explanation: energy drinks are popular with exactly the demographic that gets diagnosed and documented.

Milo: So that's the tour. The word of the year is "human," a trillion-dollar IPO is pricing machines that do human work, agents are running audits and merging pull requests, a blog died so a search engine could eat its content, regex is attacking Ruby, DNS is refusing to censor, a 942-kilobyte game boots a Miami skyline from firmware, and somewhere out past the heliopause, a computer from 1977 is still talking to us because people cared.

Mia: If any of that resonated, the links are all in the show notes. Stay safe out there — patch your NGINX, mind your MMU contexts, and maybe think twice about that fourth energy drink.

Milo: See you next time.