📅 November 11, 2026 · ⏱️ Read time: 5 min · 🔗 Issue No. 34
You're in the loop — a Chinese lab just shipped an open model that beats DeepSeek on the tasks banks actually run, and priced it at pennies. Meanwhile, a researcher poking at your iPhone surfaced a hidden list of 4,500+ phrases iOS quietly refuses to autocomplete — a small reminder that the models shaping your words aren't always the ones you can see.
Today: a $5.3B defense-AI megadeal, Trump's surprise turn on regulation, and why the "agentic web" is quietly draining publishers — plus a how-to for running open-weight models on your own machine, five fresh tools, and a prompt worth stealing.
🔁 The Loop
Open weights, a defense megadeal, and a policy turn
A new open-weight model opens up: the week's story is less about one launch than about who gets to self-host the frontier.
Tencent's Hy4 crashes the open-model party. Tencent shipped Hy4, a general model it says tops the open-source field on real financial work — scoring 3.04 to DeepSeek-V4.1-Flash's 2.97 and Kimi K3.1 Pro's 2.68 in a blind test graded by 213 finance pros across 579 tasks. It runs about $0.90 per million input tokens and $2.70 output — a fraction of frontier pricing — and cut its hallucination rate from 12.95% to 7.2% versus the preview. Weights and system cards are promised "in the coming days," with a cheaper Hy4 Flash close behind, so the real test is whether developers adopt it once they can self-host. See the benchmarks.
Shield AI buys its own eyes for $5.3B. Autonomy startup Shield AI agreed to acquire radar-maker Echodyne for $5.3 billion — $4.8B at close plus up to $500M in earnouts — and raised $2.5 billion alongside it, including a $1B Series H that values Shield at $28.8 billion, more than double its $12.7B mark in March. Echodyne's compact solid-state radars already ship into counter-drone systems in 40-plus countries; folding them into Shield's Hivemind autonomy stack is a bet that defense AI has to own its sensors. Read the deal.
Trump says he "doesn't mind guardrails." In a 60 Minutes interview aired November 8, the president said AI's risks "can be managed" and signaled he'd work with Democrats on safeguards — a notable softening — while warning that tight US rules can't leave the country in "a deep second position" behind China. No specific policy was named. What he said.
Going deeper: why China's open-weight models keep topping the cost-performance charts, and what Shield's raise says about the defense-AI gold rush.
🌊 Deep Current
The web is learning to answer itself — and publishers pay the tab
As AI agents read the web on our behalf, the clicks — and the ad money — stop flowing to the sites that wrote it.
The new middleman. A run of research this month put hard numbers on a shift you've probably felt: the open web is increasingly read by machines, not people. A Washington University in St. Louis study of 55,393 Google queries found AI Overviews now surface on 13.7% of searches — and that 11% of the claims they generate, roughly one in nine, aren't supported by the pages they cite.
Follow the clicks. Those summaries keep readers on Google. Pew has found people are about 40% less likely to click through when an AI answer sits on top, and buyers are reacting: in a Digiday survey of 18 media shops, several said they're shifting 10–30% of their Google search budgets toward Meta and TikTok, chasing the clicks search no longer sends.
When an agent does the browsing, the open web stops being a place people visit and becomes a database machines quietly strip-mine for answers — and the ad revenue drains out with the traffic.
The catch. Agents don't shop like people, either. A University of Toronto study cited by MIT Technology Review found AI agents are measurably less likely to recommend products carrying a "sponsored" tag — so the ad formats that funded the web may not survive the handoff to software that ignores them. Some publishers are adapting: Time now sells agent-targeted slots and publishes machine-readable versions of its articles.
The bottom line. If your business depends on people landing on a page, the ground is moving — not because any one model launched, but because the web's default reader is becoming an agent that summarizes, rarely clicks, and tunes out your ads. Watch whether "answer engines" start paying for the sources they lean on, or whether the open web's economics quietly reorganize around whoever owns the agent.
🛠️ The Workbench
Run an open-weight model on your own machine
With models like Hy4 and DeepSeek shipping open weights, you can run a capable model locally — private, free, and offline. Here's the five-minute version using Ollama.
Install Ollama (ollama.com) for Mac, Windows, or Linux — it's a single download that runs quietly in the background.
Open a terminal and run ollama pull llama3.2 (or another small model) to confirm everything works before you grab a bigger one.
Chat immediately with ollama run llama3.2 and type at the >>> prompt — no internet required once it's downloaded.
Want a nicer interface? Install Open WebUI, a browser front-end, and point it at Ollama's local endpoint for a ChatGPT-style window.
Match model size to your RAM: ~8B models run comfortably on 16GB; quantized builds trade a little quality for speed if you're tight.
Keep everything local — nothing you type leaves your machine, which is the whole point for sensitive work.
🗣️ Overheard
What the timeline's buzzing about
🎵 Suno, again: The Elvis Presley estate and Eminem's publisher Eight Mile Style both sued Suno this week, claiming its AI songs mimic the artists' voices and lyrics.
🎥 Per-viewer dubbing: A newly surfaced Microsoft patent describes rewriting a livestreamer's words — and lip movements — differently for each person watching.
📈 1.54M and climbing: DeepSeek's agent platform has reportedly hit 1.54 million daily active users, per Bloomberg.
🤐 The forbidden words: A researcher found iOS keeps a hidden list of 4,500+ phrases it won't autocomplete or swipe-suggest.
🧪 Bio-risk warning: Tencent's chief AI scientist Yao Shunyu warned AI could soon cause real-world harm in biology — harder to contain, he says, than cyber.
🔎 Fresh Finds
Five tools worth a look
🌐 OpenCode Browse: lets coding agents drive your real browser tabs and logins to finish web tasks for you.
✏️ openpencil: an open-source, AI-native vector design tool that turns text prompts into UI on a live canvas (already 5.7k stars).
🐘 Elephant-Goldfish: Google's free plugin where one agent drafts a plan and a fresh-context agent stress-tests it.
🔗 AXL: connects ChatGPT, Claude, or Cursor to 1,200+ sales and marketing actions over MCP.
🛡️ Hacktron: an AI security-research agent that hunts software vulnerabilities — OpenAI just tapped its researchers for external testing.
★ = sponsored placement, if any.
🧪 Prompt Lab
The "steelman, then stress-test" prompt
Borrowed from the Elephant-Goldfish idea above: make one model argue a plan at full strength, then make it attack its own work with fresh eyes.
You are two experts in one. First, as an optimistic planner, write the strongest 6-step plan to achieve: [GOAL]. Then switch roles to a skeptical reviewer with no attachment to the plan. List: - the 3 assumptions most likely to be wrong, - the single failure mode you'd bet on, and - the one change that would most reduce risk. Finish with a revised 6-step plan that survives your own critique.
Want an image to go with a write-up? Try this named-style prompt:
Modern gouache editorial illustration: a friendly rounded robot and a person sharing a desk, reviewing a glowing blueprint together; warm indigo-and-amber palette, soft paper-grain texture, generous negative space, no text.
⏪ Rewind
Yesterday’s most-opened link
In case you missed it: the study finding that roughly one in nine claims in Google's AI Overviews isn't backed by the sources it cites — readers couldn't stop forwarding that stat.
Stay in the loop — the InTheLoop team