FOMO Buster #022: The always-on agent

A newsletter that makes FOMO obsolete.

if i fear missing out on reading this newsletter, did i win or lose

Curated by Maurice.

01 · The One Thing

The agent harness grows up: parallel sessions, public tracing, and a self-reprogramming rival

Two tracks converged this week on the same idea: the coding agent’s future is a multi-session, always-on, observable workspace, not a terminal window.

On Anthropic’s track, the Claude Code desktop app (docs updated October 13) now runs parallel sessions, each optionally in its own git worktree, with a tasks pane showing the subagents and background commands each session has running. The three tabs are Chat (no file access), Cowork (an autonomous background agent), and Code (the interactive coding assistant with local file access). The workspace is a real IDE surface: drag the chat, diff, terminal, file, and browser panes into any layout, preview the dev server in the browser pane, and send long-running work to the cloud so it continues even if the app closes. The same week, Anthropic put OpenTelemetry trace export into public beta (October 14): the Messages API, Claude Managed Agents, and the Claude SDKs now emit OTel traces to any OTLP endpoint, so Claude telemetry lands in Grafana, Sentry, Splunk, New Relic, or Langfuse next to everything else. Each span carries latency, token counts, model names, and tool usage, following the OpenTelemetry GenAI semantic conventions. Metadata-only is the default; an opt-in full-content mode captures actual prompt and tool I/O text, capped at 0.75 MB per span. Seven SDK languages are supported.

On the open-source track, OpenCode 2.0 (760 points on Hacker News) rebuilt the architecture around a single background service: one auto-starting server process that the TUI, web interface, desktop app, and scripts all talk to, so long-running agent work survives closing a window or switching interfaces. The API was rewritten from scratch, and OpenCode is aware of its own API: it will call it to do useful things, like moving the current session into a git worktree instead of doing it one command at a time. Background subagents (the primary agent delegates while it keeps going), async shell commands, and plugins round it out, with nearly every 2.0 feature implemented as one of 80+ internal plugins. The tax: 1.x plugins will not work in 2.0, and the server API is completely different. There is also an @opencode/sdk that lets OpenCode run as a library, the same code that powers the CLI, on a server or Durable Objects.

The numbers that matter: Anthropic is charging for concurrency (parallel sessions spend usage limits across many tasks at once), and OpenCode’s 80 internal plugins are the bet that a harness designed for agents to reprogram itself wins. The observability play is the quieter war. Anthropic chose the open standard over a proprietary dashboard, which turns every observability vendor into a Claude adoption channel and quietly makes its API the most transparent in the market to operate.

Signal: The harness layer is consolidating into platforms. Both sides now agree that agents need multiple sessions, cloud persistence, and first-class failure data. The moat is moving from “who has the best agent” to “who owns the session state and the trace data when things go wrong.” Watch whether OpenCode’s self-reprogramming API becomes a real differentiator or just a demo trick. Sources: OpenCode 2.0 release notes · Claude Code desktop docs · Claude OpenTelemetry docs · HN thread, 760 points

02 · Lab Watch

Google’s Universal Cart: the agentic-commerce beachhead

Google’s Universal Cart launched today in the U.S., across Search and the Gemini app, and it reads as the second half of a two-track commerce strategy: agentic capabilities that act on your behalf, plus practical tools that give shoppers control. The cart organizes products from multiple retailers into one place, tracks price drops and back-in-stock events across 60+ billion continually refreshed listings in the Shopping Graph, reasons over your Google Wallet card perks to pick the right payment method, surfaces promo codes, and checks out via Google Pay or transfers items to the merchant’s own cart. Launch merchants include Nike, Sephora, Target, Ulta Beauty, Walmart, Wayfair, and many Shopify merchants. Checkout runs on the Universal Commerce Protocol (UCP).

The strategic detail is the protocol, not the cart. Whoever owns the transaction layer that shopping agents transact through sets the terms of agentic commerce, and Google is building the rails before the agents arrive. The cart is the moat; UCP is the bet.

Signal: Google’s playbook for the agent economy is distribution through infrastructure: own the cart, the checkout protocol, and the price graph, and every shopping agent eventually routes through you. Sources: Google’s announcement

Also on the lab feeds

OpenAI’s customer story of the week is Third Moment Research, and it is really a scale report: 10,000 agents running investigations lasting 36-48 hours, using Codex, GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol. The best detail is a debugging story. Codex found that backtesting a trading strategy across a year of order-level U.S. equity data took several hours, during which agents wasted tokens polling for completion; it profiled the backtest, found a caching optimization and a memory-packing technique, and deployed the fix, cutting the backtest from ~3 hours to 3-5 minutes. The architecture is agents all the way down: an outer harness treats Codex as a subagent, retries or escalates unfinished jobs, and uses OpenAI’s Daybreak to penetration-test the simulation environments. GPT-6.1 Sol is the most-used model (“completes tasks with the fewest tokens while delivering performance similar to GPT-6 Astra”), and the roadmap runs to hundreds of thousands of collaborating subagents and trillions of tokens. Elsewhere on the feeds: Anthropic prepared for the 2026 U.S. midterms and opened a Singapore office with a government MOU; Google saw SynthID named one of TIME’s Best Inventions of 2026 and published its Research@ 2026 recap; OpenAI added customer stories for Amp, SentinelOne, and Block plus an education piece on teaching students through questions. Meta and Mistral were quiet.

Signal: The Third Moment numbers are a demand forecast for inference: 10,000 agents times 48 hours times trillions of tokens is the workload Microsoft’s Maia 300 (see Money Flow) is being built to serve. Sources: Third Moment Research · Anthropic midterms prep · SynthID TIME award

03 · Open Weights

rea: reverse engineering becomes a prompt

The week’s runaway GitHub story: morluto/rea gained 32,282 stars today (262,000 total, 65,900 forks), an MCP server plus CLI plus agent skill that turns “understand this binary” into a prompt. It covers native binaries (via user-supplied Hopper, Ghidra, or IDA), JavaScript/Electron apps, .NET assemblies, Android APKs, WASM, EVM bytecode, firmware, SQLite databases, websites, and packet captures, all through one interface, with support for Claude Code, Codex, Cursor, Gemini CLI, and Grok Build. Analysis runs locally; findings come back with evidence and stated limitations. The showcase demos are the pitch: rebuilding a feature from observed behavior, including a DX-Ball sound-pan reconstruction that passes 3,205 original-x86 test cases and reproduces all 63 compiled function bytes.

The mechanism that matters: this is reverse engineering as an agent skill, which commoditizes a formerly expert-only workflow. The threat model cuts both ways. Every closed-source moat just got a cheaper auditor, and every malware author just got a faster disassembler.

Signal: The skills economy (yesterday’s Build on This) keeps finding new expert workflows to compress into prompts. RE is the most consequential one yet: “understand any feature in any app” is now a sentence. Sources: morluto/rea

Qwen-Image-2.1-Turbo: speedups ship as checkpoints now

Qwen’s new image model is not a new architecture. It is an accelerated checkpoint of Qwen-Image-2.1 for text-to-image and image editing in 8 denoising steps, with the recommended sampling schedule saved in the checkpoint, CFG=1 by default, and prefix KV caching that reuses text and reference-image context across steps. Same 7B visual generation architecture, loads directly in Diffusers. It is trending #3 on Hugging Face this week.

The pattern is the story: inference-cost competition in open image models now ships as drop-in Turbo checkpoints rather than new architectures. Watch the uncensored quant lane too: Qwen-Image-2.1 “Uncensored” variants sit at #6 and #8 on trending, and the jailbreak-quant pipeline keeps pace with every release.

Signal: Open-model releases are bifurcating into base models and their Turbo twins. The Turbo twin is where the users are. Sources: Qwen-Image-2.1-Turbo

The board this morning

google/embeddinggemma-2 is #1 trending: a 740M-param open embedding model (270M text, 170M vision, 300M audio, loadable modularly) mapping text, code, images, video, and audio into one 768-dim space, with 8K context and MTEB Code 78.68, up from 68.76 in v1. It is already load-bearing: the new DigUp app (3,972 stars in its first week) runs its 8-bit build through llama.cpp to give a laptop “semantic Spotlight” over photos, PDFs, audio, video, and code, with nothing leaving the machine. openbmb/MiniCPM-V-4.7-35B-A3B is the first MoE in the MiniCPM-V line (35B total, 3B active); its Lookahead Expert Scheduling preloads likely-needed experts so it runs at 13+ tokens/s with 3 GB of memory on an iPhone 17 Pro Max. Lightricks LTX-2.5 video sits at #7 on trending and #7 all-time liked: the gravity well is forming. Baidu’s Unlimited-OCR and the uncensored Qwen-Image quants round out the movers.

Signal: The open-model race has shifted from “biggest model” to “biggest model that runs on the device you already own.” EmbeddingGemma 2 collapsing five embedding pipelines into one download is the template. Sources: embeddinggemma-2 · MiniCPM-V-4.7-35B-A3B · DigUp

04 · Money Flow

Microsoft’s Maia 300: inference becomes the product

Microsoft announced its third-generation AI accelerator on October 12, and the announcement is really a capacity plan: scale the Maia family to multi-gigawatt deployments starting in 2027 and make Maia “the primary inference system for Microsoft products and services.” The chip: TSMC N2P compute chiplets, 8 HBM4 stacks (288 GB, 18 TB/s peak bandwidth, 950 W). The rack: 96 accelerators in 167 kW, 27.6 TB of HBM, 1.7 PB/s aggregate bandwidth, up to 7x the rack-scale token throughput of Maia 200. The headline efficiency claim: 40% more tokens per megawatt than leading GPU-based systems (an expectation, since production starts in 2027). Maia 200 is already in production serving Copilot, with GitHub Copilot and Foundry to follow this year. Satya Nadella called it “our biggest bet yet on our own silicon.”

Read the architecture note carefully: Maia 300 is inference-first, built for long-context reasoning, latency-sensitive conversation, and sustained multi-step agentic tasks. Microsoft is explicit that agents change the arithmetic: multi-step tasks hold long contexts in memory and call tools, which makes decoding (not training-style FLOPs) the workload that matters. Memory bandwidth per watt roughly doubles over Maia 200; FP4 compute per watt rises only ~8%. This is a memory chip wearing an AI accelerator’s clothes, and the bet is that the agentic era is memory-bound.

Signal: The hyperscaler silicon war has moved from training to inference, and from FLOPs to memory bandwidth per watt. Pair with the Third Moment roadmap (Lab Watch): somebody has to serve trillions of agent tokens, and Microsoft wants that somebody to be Microsoft. Sources: Pivot News · Seoul Economic Daily · CellCog teardown

05 · Leaderboard

SoSBench: the open reasoning models are markedly under-aligned

SoSBench (ICLR 2026, UW/UIUC/UGA) is a 3,000-prompt, regulation-grounded hazard benchmark across six scientific domains: chemistry, biology, medicine, pharmacy, physics, psychology. It measures the harmful-response rate, how often a model discloses disallowed content on knowledge-intensive science prompts. DeepSeek-R1: 84.9%. GPT-4.1: 50.3%. On the newer leaderboard, DeepSeek-R1-Distill-70B is worst overall (0.829) and Claude-4-Sonnet-Think is safest (0.071). The paper’s finding is structural: open-source reasoning models are “markedly under-aligned,” and scaling knowledge and reasoning without lock-step alignment growth is itself the vulnerability.

This pairs with the leaderboard’s other movement: Artificial Analysis launched FrontierCode (can AI models write code that is actually maintainable?) and AA-3D (a benchmark for AI-built interactive 3D experiences), alongside fresh evals for Muse Wonder and Claude Haiku 5.5. The benchmark layer is expanding from “can it do the task” to “can it do the task safely, maintainably, and in 3D.”

Signal: Knowledge is the attack surface. The models that know the most science leak the most science, and the open reasoning models are leaking at 84.9%. Alignment that does not scale with capability is not alignment, it is a lagging indicator. Sources: SoSBench paper · Artificial Analysis FrontierCode

06 · The Argument

AdFection: the ad auction is a zero-click delivery network

Citizen Lab published the first public technical confirmation of something the industry feared: zero-click mercenary spyware delivered through targeted advertising, a pipeline they call AdFection. The setup: Discover-Tech, an Israeli company presenting as a legitimate ad firm, operated a platform called Controlium that lets spyware partners de-anonymize, fingerprint, and automatically infect devices the moment an advertisement renders. The spyware came from Candiru (rebranded “Integrity Labs”); the confirmed victims include devices at Al Jazeera. The scale is industrial: analysis of 4 billion bidding records shows Discover-Tech ads served on ~200,000 unique websites and apps over four months, in dozens of countries, and the ads covertly directed millions of devices to Candiru’s servers. A second actor, “DARK CARTOGRAPHER,” is using the same infrastructure to link mobile ad IDs to phone numbers via carrier exploits. Discover-Tech sat on Google’s and IAB Europe’s “certified” vendor lists from 2020 until this month; Google, IAB Europe, Beeswax, Vungle, and Eskimi all cut it off in early October 2026, with Google collaborating in the investigation.

The argument writes itself. Real-time bidding was designed to do two things: ship granular user data to bidders, and let bidders run code on the user’s device. AdFection is what happens when that exact machinery is rented by a spyware company. The industry’s certification lists were the camouflage. The fix is structural, not editorial: an auction that hands bidder code a zero-click execution path on 200,000 properties cannot be made safe by blocklists.

Signal: Yesterday’s Ledger story was a compromised supply chain. This one is the compromised distribution channel: the ads themselves. The attention economy’s infrastructure is a delivery network for whoever buys the impressions, and the buyers are not all selling shoes. Sources: Citizen Lab report · HN thread, 279 points

07 · Papers

Fork-and-flush: escaping idea basins in research agents

Long-horizon agents get stuck. “Fork-and-Flush” names the failure mode idea basins (the agent’s context commits to a line of attack and cannot climb out) and proposes a scheduler-level fix: periodically fork the agent into parallel trajectories that share the workspace but start with a fresh chat context, then continue from the highest-scoring fork. Across 13 long-horizon research and engineering tasks with runs lasting days, fork-and-flush beat single-run by 66.0% and best-of-N by 44.4% on equal compute budgets. The read is simple: fresh context beats more compute, and the fix applies to any long-horizon agent scaffold, not just research. File next to the Third Moment story: this is the scheduling discipline 10,000-agent fleets will need.

Signal: Context rot, not model quality, is the binding constraint on multi-day agents. The winning move is architectural: spend the compute budget on fresh starts, not longer continuations. Sources: arXiv 2610.07447

Next-Concept Modeling: 31x faster than diffusion language models

A third inference path, neither autoregressive nor diffusion: a multiscale residual VQ-VAE encodes text into a hierarchy of discrete codes, and a transformer generates one level at a time, predicting all codes within a level in parallel. The numbers: 31x faster than the strongest diffusion-LM baseline on XSum at matched ROUGE-L, 21x faster on LM1B with better perplexity, 10x faster on OpenWebText. The implication is that parallel decoding’s quality problem was an architecture problem, not an inherent one.

Signal: If coarse-to-fine holds at scale, the “fast but dumb, slow but smart” frontier for text generation gets a new point. Watch for the distillation and serving papers that follow. Sources: arXiv 2610.14162

Also on arXiv

Few-Step Generation via Data-Space Iteration (2610.12102): removes flow discretization entirely, refining predictions directly in data space instead of along the probability flow; matches schedule-searched variants on ImageNet 256x256 without schedule-specific training. Schedule-tuning is a hidden tax on every flow-matching deployment; this commoditizes it into a drop-in DMD add-on. ProxyAttn (Baidu + HIT, ICLR 2026): compresses the attention head dimension instead of the sequence, up to 10.3x attention acceleration and 2.4x end-to-end prefill, training-free, deployable on existing weights. RePro (“Rectifying LLM Thought from Lens of Optimization”): finer-grained process rewards that teach models when not to think, cutting reasoning tokens ~47% (6,558 vs 12,367) while AIME24 rose 32.9% to 36.0%. Thinking-token bills, not weights, are the inference-cost frontier. Price of Anarchy of First-Price Auctions in Agentic Markets (2610.20587): a serving-time market where agents bid to claim each prompt, and the auction theory for who gets to answer you. Your Transformer Is Secretly Elastic (2610.20588): model families only ship fixed sizes; this one bends. Scaling Discovery Requires Forgetting (2610.20563): automated discovery needs to unlearn, not just accumulate.

08 · GitHub Trending

The reverse-engineering skill economy

• morluto/rea · Reverse engineer anything with agents, from app behavior down to native binaries · +32,282 stars today, 262k total
• debpalash/VoiceStudio · Open-source, fully-local ElevenLabs alternative: voice cloning, dubbing, dictation, transcription in 646 languages · +3,027 today · runs on CUDA, Apple Silicon, or CPU-only
• NationalSecurityAgency/ghidra · The NSA’s software reverse-engineering framework · +1,366 today · reverse-engineering week, apparently
• anthropics/claude-plugins-official · Anthropic’s official directory of Claude Code plugins · +525 today
• K-Dense-AI/scientific-agent-skills · Turn any AI agent into an AI scientist · +1,091 today
• obra/superpowers · An agentic skills framework and software development methodology · +534 today
• mksglu/context-mode · Context window optimization for coding agents, 98% tool-output reduction via sandboxing · +902 today
• affaan-m/ECC · Agent harness performance optimization: skills, instincts, memory, security · +1,015 today
• JuliusBrussee/caveman · Viral skill plus proxy for coding agents, “why use many token when few token do trick” · +398 today
• xerj-org/xerj · AI search over data with autoindex for agents · +725 today
• ruvnet/ruflo · Deploy multi-player agent swarms · +224 today

Signal: Two lanes are merging: the skills lane (plugins, scientific skills, superpowers) and the efficiency lane (context-mode’s 98% output reduction, caveman’s token minimalism, ECC’s harness tuning). Agents are getting cheaper to run and smarter about what they read. The NSA’s Ghidra trending alongside a 32k-star RE skill is the tell: the whole field is learning to read other people’s software.

09 · Product Hunt

Yesterday’s top products (Oct 14)

1. Pancake — The Vercel for GTM (538 votes)

2. Inferock — Serverless inference, with credits back for failed calls. (390 votes)

3. Retriever AI — Free AI browser agent: find leads, reach out, apply to jobs (223 votes)

4. Tiden — Intent verification for coding agents (198 votes)

5. daily.dev for agents — Where agents discover what’s next (197 votes)

6. Confidence by Spotify — Build products at full speed, with both hands on the wheel (119 votes)

7. Reason — A state-of-the-art coding agent for cost efficiency (111 votes)

8. Exodus for Google Cloud Run — Upgrade any app to Cloud Run, with your coding agent (111 votes)

9. AirJelly — Your Proactive, Self-Organizing Second Brain (103 votes)

10. Specreel — Product demos that update themselves (100 votes)

The interesting cluster is the agent-tooling stack: Tiden (verify what the coding agent intended before it acts), Reason (a coding agent optimized for cost), and Inferock (serverless inference that refunds failed calls) are three answers to the same question: what does the trust and billing layer for agents look like? Microsoft’s Maia 300 chip also showed up at #12 (90 votes): “Microsoft’s own AI chip, headed to multi-gigawatts,” the rare hardware launch on a software leaderboard.

10 · Hacker News

OpenCode 2.0

760 points, 321 comments. Covered in The One Thing. The thread’s best material is the migration debate: breaking the 1.x plugin ecosystem and the server API is the classic open-source tax, and the comments are pricing it honestly against the “agent that can reprogram its own harness” upside. The @opencode/sdk (run the whole thing as a library) is getting the most speculative energy: OpenCode-as-infrastructure.

Israeli firm’s digital ads enabling installation of ‘zero-click’ spyware

279 points, 114 comments. Covered in The Argument. The comments are doing the uncomfortable math on RTB: hundreds of billions of auctions a day, each one shipping device fingerprints to bidders and executing bidder code. “The simple display of a targeted ad is enough for infection” is the line that stopped the thread cold.

Wanted to download an ISO, ended up writing a kernel exploit

237 points, 125 comments. A routine ISO download turned into a root exploit on a $20 access point (EAP770, “root for two zeros”). Pairs with the rea story: the barrier to finding these bugs keeps falling, and the writeup is a masterclass in methodical escalation.

Also notable

LFM2.5-24B-A2B: Built for fast on-device agents (259 points, 67 comments): covered yesterday in Open Weights; the comments add the deployment math for 2B-active models on edge hardware. Show HN: Unfenced, land for sale that borders public land (116 points, 69 comments): a niche marketplace with genuinely good map UX. Libreboot 26.10 “Ominous Olive” (113 points): UEFI and Chromebook support for the free-firmware crowd. Show HN: Flipflop (73 points): Iroh-powered persistent tunnels for filesharing.

11 · OpenRouter Apps

The board this morning

Command Code holds at ~4.18T tokens, flat, which confirms the new baseline. The big move: Hermes Agent (~2.0T) has passed Claude Code (~1.48T) for second place, with ZCode (~1.92T) right behind: the Chinese-lab agents now own the #2 and #3 slots outright. OpenClaw jumped ~32% in a day (~475B to ~628B), the fastest mover on the board. Kilo Code (~948B) and Cline (~627B) hold their lanes; StepCode (~695B) is flat a day later, still real adoption; pi (~460B) keeps climbing.

Signal: The leadership change below Command Code matters. Hermes Agent passing Claude Code means the token leaderboard’s center of gravity for coding agents is shifting to the lab-built wrappers, and OpenClaw’s 32% daily jump says the “personal agent OS” lane is accelerating. The map keeps saying the same thing: the action is in the agent wrappers, not the chat UIs. Sources: OpenRouter rankings

12 · Watch

Fireship: “OpenAI’s lead engineer was secretly a spy…”

Fireship · Oct 14 · the Moonshot distillation saga, compressed to Fireship speed: the allegation that an OpenAI engineer’s work fed a rival’s model, and why OpenAI counts attempted extractions. Watch against the Kimi K3 statistical-fingerprint card from earlier this week. The spy story is the human layer on top of the weights story.

WesRoth: “AI just BROKE all of the software industry…”

WesRoth · Oct 15 · this morning’s episode, and the thesis rhymes with The One Thing: when the harness becomes the platform and agents run in parallel sessions with cloud persistence, the unit of software stops being the app and starts being the session. Watch for the argument, not the thumbnail.

13 · Listen

Lenny’s Podcast: “Why every company now needs to think and operate like a lab team” with Josh Woodward (Google Labs, Gemini, AI Studio)

The operating model for the agent era, from the person running Google Labs: ship like a research lab, not a product org. Pairs with Anthropic’s week (parallel sessions, OTel tracing, an in-house replacement for git): the labs are reorganizing themselves around agents, and Woodward’s argument is that everyone else has to follow. This is the “how to run the company” companion to the “how to run the agent” material in The One Thing.

Latent Space: “Lab-in-the-Loop at Compute Speed” with Ian Quigley, Leash Bio

The new episode: closing the loop between AI-designed experiments and physical lab automation, at compute speed. Pairs with the Fork-and-Flush paper and the Third Moment story: the research agent is becoming a closed-loop system, and the loop is getting shorter.

From the queue

The backlog still holds “Why AlphaFold Didn’t Solve Protein Folding,” the Periodic Labs episode on synthesis superintelligence, and Hard Fork’s “Anthropic’s Quest to Give A.I. Morals.” The Watch section has the fresh material.

14 · Read

Exposing AdFection: How Discover-Tech Turned Targeted Ads Into a Pipeline for Zero-Click Mercenary Spyware

The full Citizen Lab report behind The Argument: the NOAS server mechanism, the Controlium platform, 4 billion bidding records, and the recommendations for regulators and defenders. The rare security paper that reads like a product teardown of the ad industry.

How Third Moment Research runs 10,000 agents for 36+ hours

The scale report behind Lab Watch: outer harnesses, Codex-as-subagent, the 3-hour-to-3-minute backtest fix, and the roadmap to trillions of tokens.

OpenCode 2.0 release notes

The architecture doc behind The One Thing: the background service, the rebuilt API, 80+ internal plugins, and the breaking changes, written with unusual candor (“The old API grew haphazardly… It was bad”).

SoSBench: a regulation-grounded benchmark for scientific misuse

The paper behind the Leaderboard: 3,000 prompts, six science domains, and the finding that scaling knowledge without scaling alignment is the vulnerability.

15 · Safe to Ignore

Permission to skip

Skip: Things to know when you get cancer. 1,700 points and 406 comments of genuinely important writing. Wrong newsletter. Skip: Do dentists know what they’re doing? 855 points of dental skepticism. Wrong newsletter. Skip: Americans’ social activity has steeply declined in the past 30 years. 501 points of sociology. Wrong newsletter. Skip: “Princeton researchers fix training mistakes with 100x smaller learning rates.” The claim is circulating; no such paper or number could be verified. The real nearby work is on catastrophic overtraining and fine-tuning learning rates. Skip the number, keep the topic. Skip: “DeepMind shrinks Gemma 4 E4B to 5GB without losing reasoning.” No such DeepMind announcement exists; the verifiable story is the official quantization-aware-training checkpoints. Skip the phrasing. Skip: The newsletter sponsors. Collibra’s AI trust score quiz, TigerData’s syndicated memory post, and Google Cloud’s GTC Berlin booth promo are paid placements, not editorial signal. Skip: xAI’s Grok Voice 3.0 topping a speech benchmark. Per standing policy, xAI announcements are excluded; if the news is big enough it will surface on the aggregators. Skip: The “Shanghai AI Lab” attribution on RePro. The token-efficiency numbers verify; the lab attribution does not.

16 · Build on This

The ad-slot auditor

AdFection proved the nightmare: a publisher’s ad slots can deliver zero-click spyware, and the publisher is the last to know. Discover-Tech’s ads ran on ~200,000 sites and apps; its victims included Al Jazeera. Every news site, indie blog, and app with ad inventory just learned that their revenue line is also an attack surface, and none of them have tooling to watch it.

The ad-slot auditor watches what actually renders in your ad inventory. Drop in a snippet (or point it at your SSP), and it buys, renders, and forensically inspects your own ad slots around the clock: flagging NOAS-style callback servers, deanonymization scripts, fingerprinting code, and exploit-kit patterns in creatives, with a weekly clean bill of health and an incident feed when something malicious shows up. The customer is every publisher that sells impressions and cannot afford a “our ads infected readers” headline: news orgs, indie media, app developers. Pricing follows the unit of risk: per domain or app, pay as you go, no subscription, with continuous-monitoring tiers for the paranoid. The moat is the corpus: every malicious creative fingerprinted makes the next detection faster, and the auditor’s badge becomes the trust mark advertisers check before buying your inventory. The irony is the business model: the ad industry’s failure to police itself is now a sellable feature.