1002 | Agents, Benchmarks, and Border Phones

Mia: Welcome back to the show, everybody. I'm Mia, and across from me as always is Milo. Today we've got a thread of stories that honestly all run through the same nerve: who's in control, of code, of data, of platforms, of the AI economy itself.

Milo: And that theme really does carry everything today, from coding agents getting a serious architecture layer, to phone searches at borders, to Google putting a price tag on free apps. It's an episode about control shifting around. Let's get into it.

Mia: Let's start with the coding agents, because two things shipped that are attempts to make agents actual infrastructure rather than demos. First, a paper: Context Language Models. The idea is you stop treating context as something the system manages around the model, and instead you let the model treat its own context as a file that it manages itself.

Milo: Right, and the claimed result is that this approach beats the state of the art in context management while using fewer FLOPs. And there's a second trick in there that I think is almost as interesting: suffix cache reuse. So when the model rewrites or edits that context file, the parts that didn't change can be reused from cache rather than reprocessed.

Mia: And that's the part that makes it practical, right? Because if the model is constantly rewriting its entire context, you'd expect the cost to explode. If the suffix cache lets you only pay for what actually changed, then the "fewer FLOPs" claim hangs together. The framing here is a real inversion, though. Everyone else is building smarter retrieval, smarter trimming, smarter summarization around the model. This says, no, give the model a file and let it decide what to keep, edit, and throw away.

Milo: And that inversion raises the obvious question, which the paper's authors are implicitly betting on: do models actually know what's relevant better than the scaffolding code does? If the answer is yes, a lot of context engineering work becomes redundant. If it's no, you get agents that confidently delete the wrong things from their own memory.

Milo: There's no consensus yet — it's one paper claiming benchmark wins — but the direction of travel is clear: move agency for context management inside the model.

Mia: Which brings us to the other side of the coin: the harnesses around the model. Pi 1.0 shipped this week, described as a hardened, minimal, extensible coding agent harness. Key pieces: Codemode, extension support, deferred tool loading, and a full-screen TUI as the default.

Milo: Deferred tool loading is worth pausing on, because it's the same philosophy as the context paper in a different layer. Don't load every tool upfront, load what's needed. The whole point of Pi 1.0 seems to be restraint. Minimal, hardened, extensible — it's not trying to be a kitchen sink, it's trying to be a solid base that other things get built on. Codemode as an extension mechanism suggests they want the agent's capabilities themselves to be programmable.

Mia: And Pi 1.0 isn't alone. There's also Pi Durable, an experimental framework for long-running durable agents. It's about fifteen thousand lines, it runs over pluggable storage backends — memory, SQLite, JSONL — and it provides a Node execution environment.

Milo: That's the piece the whole agent ecosystem is missing, honestly. Coding agents today run in a session, and if the session dies, the work is gone or corrupted. Durable execution means an agent that can run for days, survive restarts, and pick up where it left off. Fifteen thousand lines is small for what it's claiming, which is either elegant or early — and it's explicitly experimental, so probably both.

Mia: So the picture across the day: the model manages its own context as a file, the harness is minimal and loads tools lazily, and durability gives agents persistence across time. Everything's pointing at unsupervised, long-running agent work. Except — and this is the kicker — the community is deeply skeptical that we're there.

Milo: Right, and that skepticism showed up on the "Goalposts" poll site this week. It's an HN poll site that votes on whether AI has actually met past Hacker News challenges — the famous predictions and bets people made about what AI would be able to do. And the debate that came out of it was over unreliable LLM coding and the fact that there are still no unsupervised workflows in real use.

Mia: And that's the unresolved question hanging over everything we just described. You can ship hardened harnesses and durable execution all you want — if the model underneath still produces unreliable code, you can't put it in front of production without a human watching. The infrastructure is racing ahead of the reliability. Nobody in these discussions was claiming unsupervised workflows work today; the argument was more about how far away they are and whether the trajectory is even the right one.

Milo: There was also a fascinating counterpoint in this space: a Rust project called Bez, which is generating a browser engine from web specifications and tests. Which sounds like exactly the kind of unsupervised, long-horizon, ambitious agent work we're talking about.

Mia: But the commenters weren't buying it as a solved problem. The main pushback was that AI can't yet build things that are fast and well-architected. A browser engine is arguably the hardest kind of systems software there is — performance-critical, deeply spec'd, enormous. Generating one that's correct is one thing; generating one that's fast and well-architected is another thing entirely.

Mia: So Bez becomes a great test case for the whole debate: it's exactly the kind of project that unsupervised agents supposedly enable, and the community response was essentially "watch what happens when quality bar is actually high."

Milo: And one more piece in this cluster that ties it to the real world: the OpenID Foundation put out a 2025 whitepaper on identity management for agentic AI. Authentication, authorization, delegated authority for agents.

Mia: This matters because the unsupervised workflows debate isn't just about code quality — it's about what an agent is allowed to do on your behalf. If an agent is going to run for days and take actions, who is it acting as? Whose credentials does it hold? Delegated authority is a genuinely unsolved problem, and a standards body taking it on is an admission that this is now infrastructure, not research novelty. Keep that in your head, because we'll come back to identity later in the show.

Milo: Speaking of performance infrastructure, let's shift to databases and compilers — a whole different kind of engineering speed race happening this week. First up: turbopuffer v3, and the headline is an architectural reversal. They demoted the ANN vector index from being the primary index to being a secondary index.

Mia: And the reason is beautifully practical. When your vector index is primary, every query type that isn't a vector search — attribute filters, full-text search, other query plans — has to be shoehorned through or rewritten around it. Make the vector index secondary, and attribute filtering, FTS, and everything else get their own first-class query plans. Vector search becomes one option among several instead of the axis everything rotates around.

Milo: Which, ironically, echoes the theme from our first segment. Everyone built infrastructure assuming one thing — vector similarity — was the primary use case. Then reality arrived with mixed workloads, and the architecture had to flip. Same story as context management: the scaffolding assumption broke.

Mia: Now, on the search side, ParadeDB closed an eight-times performance gap against PlanetScale's TIN. TIN is closed-source, so it's not available for people to run themselves. ParadeDB got there with BM25 optimizations and hit 81.9 queries per second on the StackExchange benchmark.

Milo: The interesting part here is less the number and more what it says about open source versus proprietary. Eight-x is not a rounding error — that's a genuinely large gap — and they closed it through what sounds like sustained algorithmic work on BM25 rather than a single breakthrough. And since TIN isn't open, users can't verify or fork the competition; ParadeDB's whole pitch is that you can.

Milo: Whether the benchmark is representative is the open question, as it always is with benchmarks — but StackExchange data is at least real-world-shaped text search, not synthetic vectors.

Mia: And in the same week, the Rust compiler picked up a 4.57 percent mean wall-time reduction over two months. And the breakdown is the interesting part: it came from algorithmic changes, LLVM 23, PGO on Clippy, and new Polonius work.

Milo: Polonius is the long-running effort to replace the borrow checker's dataflow algorithm, so seeing "new Polonius work" in a release note is a signal that's still alive and moving. And I love that PGO Clippy shows up in the list — profile-guided optimization applied to the linter — because that's the same theme again: not one heroic rewrite, but a pile of disciplined algorithmic and toolchain improvements that add up.

Milo: Four and a half percent mean wall-time across all compilations is a real quality-of-life improvement for a huge developer population.

Mia: There were also two smaller stories in the tooling world this week worth folding in. SvelteKit 3.0 released on October 1st, 2026 — config moves to vite.config.ts, the $lib alias becomes plain lib, and remote functions are still experimental but usable in production.

Milo: Usable in prod but still experimental is a very honest label, and remote functions are the feature everyone's watching because that's the server-calling-from-client ergonomics that frameworks are all converging on. And then there's the GitButler post arguing that Git 3.0's SHA-256 default is costly — specifically that rehashing breaks external hash references, and there's no old-to-new hash mapping provided.

Mia: That's a real gap. Every CI system, every commit-signing record, every bug tracker that references a SHA becomes stale the moment you rehash a repository. The complaint isn't that SHA-256 is wrong — it's that the migration is incomplete without a translation table. Different flavor of the same day: platforms changing their foundations and the downstream dependencies left holding the bag.

Milo: Which is actually a perfect segue to our next segment, because platform control and what happens to your stuff when the platform moves — that's the story of phones, borders, and cars.

Mia: Let's start with the one that got the most attention: a leaked video of Magnet Forensics' GrayKey tool bypassing Apple's 72-hour inactivity reboot.

Milo: For context: Apple introduced this behavior where after 72 hours of inactivity, an iPhone reboots into a more locked-down state — that's the after-first-unlock versus before-first-unload distinction, AFU versus BFU. The reboot is a security feature: once the phone reboots, some keys are no longer resident in memory, and extraction gets harder. The leaked video shows GrayKey preserving that AFU state, defeating the reboot, keeping the phone in the more extractable condition for police access.

Mia: So the security feature exists, and there's a commercial tool on the market that specifically neutralizes it. That's the whole story in one sentence. The unresolved questions are the ones you'd expect: how does the bypass work, who has access to these tools, and what does Apple do next. This is a cat-and-mouse game where every defensive feature gets a countermeasure, and the leaked video suggests the current round went to the extraction side.

Milo: And it collides directly with the second story: US Customs and Border Protection can search phones at the border without a warrant. Basic searches require no suspicion at all. Advanced searches require only reasonable suspicion — not a warrant.

Mia: Put those two together and you get the scenario people were clearly thinking about: a device at the border, a search that legally doesn't need a warrant, and a commercial tool that keeps the device in its most extractable state. The border search authority has been litigated and debated for years, but the practical landscape keeps shifting as extraction tools get more capable. The legal framework hasn't moved at the same speed as the technical capability.

Milo: Moving from pockets to cars: Northeastern ran a study called "Automatic Transmission" on connected vehicles. The numbers: 19 out of 21 connected vehicles sent traffic to third parties. And of the 30 companion apps they looked at, 7 sent personally identifiable information to trackers.

Mia: Nineteen out of twenty-one is nearly universal. Almost every connected car in the study was phoning home to someone other than the manufacturer. And the companion app numbers show the problem extends to your phone too — seven of thirty apps sending PII to trackers.

Mia: What's still unknown is the full destination list and what exactly is in those payloads — the study establishes that traffic goes to third parties, and the open question is how sensitive it is and whether drivers have any meaningful control over it.

Milo: And here's a lighter but genuinely strange item in the same privacy-under-pressure theme: security people found undocumented SDR capability — software-defined radio — in ESP32 chips. Raw IQ capture, frequencies from 2.2 to 2.7 gigahertz, plus 4.8 to 6.0 gigahertz on the C5 variant, at up to 80 megasamples per second.

Mia: This was found across multiple projects independently, and it's remarkable because it means a cheap commodity microcontroller that's already in millions of IoT devices has a hidden radio capability nobody documented. Eighty megasamples per second is serious bandwidth. The fun part is the dual reading: for hobbyists, free SDR hardware hiding in a three-dollar chip; for security people, undocumented radios in deployed devices are a question mark.

Mia: Both readings are true simultaneously, which is why it captured the community's imagination.

Milo: And ESP32s are, of course, everywhere in connected cars and IoT devices too, so it slots right into the same anxiety. Which brings us to the AI industry itself and the pressure it's under from regulators.

Mia: The big one: the FTC has opened an investigation into OpenAI, Anthropic, and other AI companies over potential product dangers. And notably, this comes after the Hugging Face hack incident.

Milo: The sequencing matters there. A security incident in the AI ecosystem — the Hugging Face hack — seems to have been a catalyst, and now the regulator is looking at potential product dangers across multiple companies at once, not one company in isolation. That breadth is the story. It's not a complaint about one chatbot; it's an inquiry into whether the products themselves pose dangers.

Mia: And running parallel to the enforcement side, there's the standards side — that OpenID Foundation whitepaper we flagged earlier. Authentication, authorization, delegated authority for agents. So you've got a regulator asking "are these products dangerous" and a standards body asking "how do agents hold identity at all." Those two efforts haven't met in the middle yet, and that's the unresolved part.

Mia: You can regulate the product and still have no answer for what happens when an agent with delegated credentials does something harmful. Identity for agents is the missing layer between the FTC's questions and the industry's roadmaps.

Milo: And then there's the industry building outward at the same time. OpenAI and Synopsys announced a partnership on GPT-Synopsys — a frontier AI model for chip design, built on Synopsys EDA tools, with revenue sharing between the parties.

Mia: Chip design is a fascinating domain for this because it's one where correctness is verifiable by simulation and fab yield — you can't bluff your way to a working silicon. Synopsys brings the toolchain, OpenAI brings the model, and revenue sharing suggests this is a product, not a research PR. It's also a signal of where "AI product" is heading: not chat, but domain-specific models embedded in professional toolchains.

Mia: Whether a frontier model can actually out-perform human chip designers on the hard parts remains to be seen — that's the open bet.

Milo: So regulators are circling, standards bodies are writing papers, and the industry is signing chip deals. Let's talk about the economics, because the money story this week is brutal in both directions.

Mia: Start with the losers. Web development education is collapsing under GenAI. Course creators are reporting revenue drops of more than fifty percent. And 2ality, the long-running JavaScript blog by Dr. Axel Rauschmayer, is taking its books offline due to AI crawler traffic.

Milo: Those are two different failure modes and worth distinguishing. Course creators are seeing demand evaporate — people are getting answers from models instead of buying structured learning. 2ality's problem is different: it's the cost side. AI crawlers hammering the site made hosting the books untenable. So one is demand destruction, the other is cost destruction, and both hit the same community — the people who write the reference material that, ironically, the models were likely trained on.

Mia: That's the point a lot of people made: there's a real question about whether the knowledge pipeline that produced these models can survive the models. If writing a JavaScript book no longer pays and hosting one costs money, who writes the next generation of training data and reference material? Nobody has an answer, and the fifty-percent revenue drops are the leading indicator.

Milo: On the winners-and-costs side: Micron's CEO said memory supply will be "much tighter" in 2027 and 2028. Over 75 percent of 2027 output is already committed, and he said there's no line of sight to balance.

Mia: Seventy-five percent of next year's output committed already, with no line of sight to balance — that's the CEO of a memory manufacturer saying the shortage is structural, not cyclical. This connects straight back to the AI capex story: data centers eating the memory supply. And it means memory prices flowing into everything — phones, cars, laptops. If you were wondering why this matters for regular people, that's the transmission mechanism.

Milo: Now the taxpayer angle. A New York Times piece reported that Meta claims R&D tax credits on Nvidia AI chips for its data centers, trimming nearly four billion dollars off its tax bill last year.

Mia: Nearly four billion. The question people raise is whether buying compute hardware counts as "research" in the way tax credits were designed for. That's a live policy question, not a settled one — whether those credits are being applied the way legislators intended or whether the definition of R&D has quietly expanded to include capital expenditure. Watch for that debate to get louder.

Milo: And in the same week, an FT op-ed called the idea of an AI sovereign wealth fund "techno-imperialism." The HN reaction was sharp, and largely against the piece — commenters pointed out it was ignoring existing sovereign wealth funds like ADIA or Temasek.

Mia: That's the counterargument that carried the discussion: sovereign wealth funds already exist, they're already massive, they're already run by governments — the UAE's ADIA, Singapore's Temasek. So calling an AI-specific version "techno-imperialism" without engaging with the fact that the institution is decades old struck a lot of readers as a framing problem rather than an analysis.

Mia: The genuinely new question underneath — should AI compute and capability be treated as a strategic national asset the way oil reserves are — remains open, and both sides of that argument are defensible. But the op-ed didn't win the room.

Milo: Which all feeds one picture: enormous capital flowing in, memory supply tightening, education revenue collapsing, tax policy straining to keep up. An economy reorganizing around AI faster than its institutions can respond.

Mia: And that reorganization is exactly what our next segment is about — platform control, and what happens when the gatekeepers move the goalposts.

Milo: Three stories. First: a developer publicly criticized Google's Android Developer Verification Program. The specifics: a $25 fee, identity checks, and this applies even if you're distributing free APKs.

Mia: That last clause is the crux. The fee is trivial for a business — the issue is the principle. Identity verification and a paid registration for the right to distribute software for free, to users who sideload. It's Google inserting itself into every distribution path for Android software, and the developer's argument is that this converts an open platform into a permissioned one, one small step at a time.

Milo: Second story: Google will end ChromeOS updates in mid-2034, breaking its ten-year Chromebook update promise, amid a transition to something called Googlebook OS.

Mia: The ten-year promise was a major selling point — it's why schools and enterprises bought Chromebooks. Ending updates in 2034 means hardware bought under that promise ages out early. And the deeper story is the platform consolidation: ChromeOS and Android merging into one OS, with the Chromebook promise as collateral damage. If you bought devices on the strength of that commitment, the mid-2034 date is the moment you re-plan.

Milo: Third: Techrights is claiming Red Hat is being phased out following the IBM acquisition — pointing to badge changes to IBM and departures of people. Commenters pushed back and debated, and a big thread was about NixOS as a migration path.

Mia: The NixOS debate is the interesting part. If Red Hat's identity dissolves into IBM, where do people who valued it go? The NixOS argument is that declarative, reproducible system configuration removes the thing you actually needed Red Hat for — the enterprise-grade, well-managed distribution. The counterarguments are about support contracts and certification ecosystems that NixOS doesn't replicate.

Mia: Nobody was claiming the transition is done — the claims come from one outlet and the evidence is badges and departures — but the migration conversation itself shows people are hedging.

Milo: Tie the three together and the pattern is unmistakable: every independent path — sideloaded APKs, Chromebook longevity, an independent enterprise distro — is converging toward gatekeeper control, and the community's answer in each case is some version of an open exit. Which is a theme we'll see again in our last segment, because the open-source world kept busy this week.

Mia: Let's do the community round with some depth, starting with the two perennial October threads: Ask HN, Who is hiring? and Who wants to be hired?, both for October 2026. The rules are unchanged and worth stating because they keep the threads usable: employers post location and remote status; job-seekers post their location, remote preference, technologies, résumé, and email. Recruiting firms and job boards are prohibited, and it's one post per company.

Milo: Those constraints are the whole design — no middlemen, no spam, direct contact between the two sides. And they run monthly, so they're a recurring pulse on the market. Combined with the education-collapse story from earlier, the hiring threads are where you'd look for evidence of what skills are actually in demand versus what the courses are teaching.

Mia: Also this month: the sixteenth RacketCon, October 3rd and 4th, 2026, in Oakland. The talk lineup includes Typed Racket inference, Pille — a low-level Rhombus — Herbie, and effect handlers.

Milo: That's a strong program for a language community. Typed Racket inference is about getting static guarantees without annotation burden. Herbie is the long-running floating-point accuracy improver — automatically rewriting math expressions to be more numerically stable. And effect handlers as a talk in the Racket ecosystem makes sense given how much algebraic effects theory has been flowing into programming languages lately.

Milo: Pille and low-level Rhombus point at Racket moving down the stack, which is a notable direction for a language famous for being high-level and teachable.

Mia: Next: StreetComplete's iOS port is in public beta, built with Kotlin Multiplatform and Compose Multiplatform sharing one codebase.

Milo: For anyone who doesn't know StreetComplete — it's the app that gamifies OpenStreetMap contribution by asking you simple survey questions about what you see on the street. An iOS port matters because the app was Android-only, and KMM plus Compose Multiplatform let them share the logic and UI across both platforms from one codebase. That's a real-world proof point for cross-platform Kotlin, from an app with meaningful complexity, not a todo list.

Mia: A couple of open-source oddments to round out. papero: a lightweight, open-source PDF extractor. No ML, CPU-only, outputs to Markdown, JSON, Excel, or Word, and it preserves tables, formulas, and bounding boxes.

Milo: And honesty in the discussion: one user reported the output was badly wrong on their documents. Which is the perennial PDF problem — PDF is a layout format, not a semantic one, so extraction is always lossy somewhere. A no-ML, CPU-only extractor is attractive for cost and privacy reasons, and preserving tables and bounding boxes is genuinely hard engineering. But the firsthand report of wrong output is the counterweight: works great until it doesn't, and you may not know when it doesn't.

Mia: And one more, which loops back to our very first topic: Cloudflare announced K2, serverless event streaming built on R2, for durable ordered log streams. And the origin detail is the good part — it was originally built as the ingestion layer for Basin Pipelines.

Milo: Durable, again. The same word from the agent framework. Infrastructure is being rebuilt everywhere on the assumption that things run a long time and must survive failures. And the "built for ourselves first" origin story is the classic Cloudflare pattern: solve your internal problem at scale, then productize it.

Mia: So let's land the episode. The connective tissue today: control is moving. Agents are getting context control and durable execution, but the community still isn't ready to hand them the keys unsupervised. Databases and compilers are being re-architected for speed. Phones, cars, and borders are tilting toward whoever holds the extraction tools. Regulators are circling AI companies while standards bodies write the identity rulebook.

Mia: The AI economy is enriching compute suppliers while starving the people who write the manuals.

Milo: And platform gatekeepers are tightening their grip while open exits — NixOS, KMM, open-source search, community threads — keep getting built anyway. The unresolved questions we're watching: does the model-managed context approach hold up outside a paper? Can agents cross the reliability threshold into unsupervised work? Will agent identity standards arrive before agents are everywhere? And who writes the next generation of knowledge if writing it no longer pays?

Mia: That's all the time we have. Thanks for listening — I'm Mia.

Milo: And I'm Milo. We'll see you next time.