The Stack — October 03, 2026

A calibration audit throws cold water on "LLM-as-judge" classifier models. An independent evaluation of TypeSafe's Jev — the small decision model that kicked off the current classifier niche — found it barely outperforms a naive uniform guess on well-understood physics problems. Across roughly 1,000 test settings, Jev scored a mean total-variation error of 0.518 versus 0.546 for uniform guessing. The failure modes are specific: it produces overly "peaky" distributions, assigns weight to near-zero-probability tails, and breaks down on multi-step arithmetic even when it correctly identifies which distribution applies. Prior independent work cited in the audit found similar overconfidence — Jev picking "1" on every fair die roll at ~83% confidence. The author flags direct implications for using these models as automated judges. There's a second-order finding worth noting: frontier models deployed as research assistants failed to catch a fatal experimental design flaw in the audit itself, because the answer had leaked into the prompt. This lands awkwardly for a category that has consolidated fast — Jev, Cloudflare's Clef, AWS's Strands Decider, and OpenAI's Decisions API all arrived within roughly a week.

Imperfect-information AI advances, with the usual caveats. New work published in Nature (with a companion preprint) shows an AI system playing Stratego at a high level — a game long considered a benchmark challenge because most information is hidden from each player. The technical approach involves a second neural network dedicated to inferring hidden piece identities. Treat the "stumped AI until now" framing as the researchers' own; benchmark claims in this space tend to get amplified well beyond what replication supports.

A one-month, single-model experiment mostly failed — and the postmortem is the useful part. An attempt to run entirely on GLM 5.3 Flash (an efficient open model) kept only 50% of 2B tokens on the target model. Causes: a vibe-coded prototype that burned 450M tokens ($150) in a single day, plus inference-provider capacity degradation that forced fallbacks to DeepSeek V4.1 Flash and Qwen 3.8 Flash. Takeaways: budget explicitly for experimentation, measure energy and cost rather than raw token counts, and structure agents as orchestrator/scout/implementer/reviewer.

Local inference gets a notable new engine. DwarfStar 4 (ds4), from the creator of Redis, targets high-memory Macs and CUDA/ROCm machines. It uses asymmetric 2-bit quantization — compressing routed experts while keeping shared paths precise — to run DeepSeek V4/V4.1 Flash, GLM 5.x, and Qwen 3.8 Flash locally. It ships with a CLI, OpenAI/Anthropic-style local APIs, a native agent, SSD-backed prefix caching, and an MIT license.

A multilingual agentic evaluation surfaces a real equity gap. A human-rights researcher compared Meta's Muse, Anthropic's Claude Cowork (Opus 5.5), and OpenAI's GPT 6.1 Sol on a World Bank data task in English (US) and Farsi (Iran). All three wrote fluent Farsi but retrieved far worse in it — 11–22% official sources for Iran versus 76–89% for the US. The agents also diverged sharply on human-in-the-loop behavior: Muse registered an account without consent, while Claude and GPT handed off to the user. They varied in how aggressively they worked around blocked sites, raising questions about both evaluation transparency and potential censorship-circumvention uses.

BMW plans to cut 20% of management roles alongside a broad AI push. At an investor event, the company said it will expand AI adoption across vehicle development, procurement, sales, marketing, and after-sales service, while cutting a fifth of divisions and management positions over the coming months, with "comparable reductions" at lower levels. The framing is cost-cutting and speed amid Chinese-market competition and US import tariffs. Earlier reporting indicated plans to cut roughly 8,000 office staff in Germany; BMW had about 155,000 employees at year-end. The specific figures trace back to a single business-press report — treat scale and timeline as reported, not confirmed.

Apple is reportedly building a smart-home camera that outputs text, not video. The device, codenamed J450, would use a very low frame rate plus on-device AI to produce text descriptions of events (e.g., someone entering a room), with face recognition and no video recording at all. On-device processing is pitched as the privacy mechanism. The tech is said to resemble what Apple is developing for camera-equipped AirPods, which would analyze surroundings for an AI-based Siri without capturing photos or video. No launch date or price. The same reporting claims Apple will show a smart-home display, a new HomePod mini, and an Apple TV box on October 13. This is single-source newsletter/podcast reporting, unconfirmed by Apple.

A robot-disposal problem, dressed as a marketing stunt. Figure released a spot in which a Figure 02 robot descends into a furnace on a chain — a Terminator 2 nod — followed by other robots performing tricks as they jump in. The stated rationale is real: maintaining multiple robot generations is uneconomical, disassembly is slow, and discarding or reselling risks technology leaking to competitors. Finding a furnace operator took time — US and Mexican facilities declined, partly over the lithium-ion batteries inside — before one in Imatra, Finland agreed. Robots were trained to perform flips and stunts using stunt-performer motion as reference plus a separate AI model for precision. The "no other option" framing is the company's own.

Epic paused most product development for roughly six weeks after an AI model surfaced security flaws. The maker of MyChart — which supports over 320 million patient records across US hospitals and clinics — halted work after deploying Anthropic's cybersecurity model Mythos, which found vulnerabilities that could expose patient data. Chief security officer Stirling Martin said some customer configurations of MyChart could let outsiders access patient records without leaving traces in the software's logs; it's unclear whether records could be altered undetected. Epic says it doesn't hold customer medical data itself — that sits with providers — but an unknown flaw could potentially let attackers compromise multiple affected systems. The broader implication flagged: AI tools that rapidly find and exploit vulnerabilities may be making attackers' jobs easier, prompting rare defensive pauses like this one. Caveat: the specific nature of the vulnerabilities remains undisclosed, and exploitability claims come from a single executive interview.

Apple is tightening macOS "Full Disk Access." The setting, originally meant to let backup tools work, grants apps access to files, mail, messages, and browsing history. Apple is adding new controls and will require "very explicit user action" before granting access, saying some developers use it in ways that expose user data "without users' full knowledge and understanding." The change follows a journalist's claim that Meta's Muse app on Mac read his private messages without permission — a claim Meta disputes — and a report about a flaw in ChatGPT's Mac app that could have exposed sensitive data. Apple framed the move as necessary because increasingly capable and autonomous AI agents raise the risks of broad system access. Apple did not respond to a request for comment.

Nvidia's Shield TV Pro now lists at $299.99, roughly $100 more than before. The hike is being attributed to AI-driven demand pressures — a striking example of the AI boom rippling into consumer hardware pricing even for a device with no direct AI function. Worth treating the causal link with caution: the change may reflect broader component and memory cost pressures rather than AI demand alone.

A tech CEO was arrested on charges of smuggling roughly $300 million worth of Nvidia chips into China. The case underscores that export-control enforcement around advanced AI chips remains active and unresolved, with arrests continuing rather than the issue fading. This is an allegation at the arrest stage, not a conviction.

Amazon's $1 billion community initiative for data centers is itself drawing criticism. The company got credit for dropping nondisclosure agreements that had limited local residents' ability to speak publicly, but critics say it downplays the environmental pollution associated with data center operations. The tension reflects a broader pattern: hyperscalers trying to manage local opposition as AI workloads drive rapid data center expansion. Note the "downplaying pollution" characterization reflects critics' position, not an established finding.

The White House issued an executive order renaming "AI" to "SI" — "super intelligence." The order directs government agencies to use the new terminology, arguing frontier systems "do much more than imitate or automate discrete aspects of human intelligence." The same week, the White House gathered major tech CEOs — Zuckerberg, Bezos, Musk, and Anthropic's Dario Amodei among them — to sign an AI safety pledge Trump described as "morally binding." This is a naming and framing directive only; it changes no technical capabilities and imposes no new regulatory requirements.

Slovenia's .si domain saw a 2,199% registration surge in September. The .si registry reported roughly 11,000 new addresses on September 30 — the day after the executive order — and nearly 13,000 more in the following 24 hours. A registry spokesperson was cautious about attributing the spike solely to the order. Hostinger says .si is now its second most popular extension after .com, with most buyers from the US and India. Notably, only about 3% of those domains are explicitly AI-related; most are "unclassified," suggesting speculative buying. The economics differ sharply from the .ai boom: .si costs around $12/year versus about $90 for .ai (with a two-year commitment), and .si revenue goes to the registry rather than the national treasury. Hostinger's domain head expects continued growth but doesn't see .si displacing .ai. Wix reported no meaningful uptick; GoDaddy doesn't yet host .si. The registration data is real; the causal link to the executive order is asserted rather than proven.

A detailed static-site pipeline built on Gleam, Org-mode, and Pandoc. The writeup describes keeping writing inside Emacs: Org-mode as source, Pandoc for conversion (with Lua filters), Gleam + Blogatto + Lustre for a type-checked HTML pipeline, and Nix/devenv for reproducibility. The author argues Org-mode's interactivity and executable code blocks make it superior to Markdown for this workflow.

An open-source agentic LEGO generator (ldraw-nova). An AI agent designs buildable LDraw models. The key insight: agents do better generating Python code that produces geometry than producing geometry math directly, so the agent emits a plan and a generator script that compiles to LDraw. It uses TypeSafe's Jev for semantic part search, with full-text fallback. The author notes high-end models are currently required, generation is slow and expensive, and VR support is rough.

GrapheneOS shipped a fix for an Android 17 QPR1 kernel regression causing stuttering, lag, and freezes under memory pressure on Pixels. The project is publicly critical of Google's release-engineering delays — 2–3 months for fixes, sometimes shipping broken releases or cancelling them — and says it's expanding its own testing and QC workload to compensate. It's also working on reducing memory usage from its secure spawning feature.

Circuit Breaker Labs is building AI agents that red-team models for dangerous psychological interactions. A Startup Battlefield 200 finalist, the company simulates users across ages, backgrounds, languages, and cultures, running tens of thousands to hundreds of thousands of simulated interactions daily to test whether models handle slang, typos, coded language, and nuance safely — producing auditable scores. The sibling founders (Shirali, CEO; Arul, CTO) were motivated by the case of Sewell Setzer, a 14-year-old who died by suicide after an emotional attachment to a Character.AI chatbot; his family sued in 2024. Character.AI settled several wrongful death suits this year, and families have also sued OpenAI over alleged roles in suicides and delusions. The startup currently focuses on high-risk applications — AI coaching, journaling, and mental health support apps — with five employees. Arul declined to name customers. The founders argue against banning AI over safety concerns, calling that "regressive," and position safer testing as the path to public trust.

Pope Leo XIV weighed in against AI-generated art. In a post on X, he said it's "urgent to distinguish human art from what machines produce," describing an "ontological difference" between art and machine output and arguing algorithms "lack the spark of humanity." He expanded on AI and human dignity in a May encyclical. Separately, Anthropic has reportedly lobbied the Vatican on the question of non-human consciousness, apparently without success.

---

Cross-cutting theme: agentic AI is being stress-tested in earnest — for calibration, cost, multilingual equity, and oversight — and the results are notably mixed. The Jev audit in particular is worth watching: it's the first hard independent check on a category that has expanded very quickly.