> daily_signal(2026_10_22)
Washington ordered a foreign AI model pulled out of federal systems, Chicago took the delivery robots to court over the faces they scanned, and two labs bet the next edge is an AI that does less.
PickBits Daily Signal · Thursday, October 22, 2026
// tl;dr
- The Department of Homeland Security ordered every civilian federal agency to remove DeepSeek, and the order reaches any third-party product built on it. Issued under the 2018 Federal Acquisition Supply Chain Security Act, it cites data transfer to China, PII collection, and critical iOS flaws. Agencies have 60 days to request a written exception, and it covers civilian agencies only.
- Two Cook County class actions say Coco's and Serve's sidewalk delivery robots scanned faces and hands under Illinois' biometric law, then kept the data to train AI. BIPA sets $1,000 to $5,000 per scan with no need to prove harm. Both companies deny collecting identifying biometrics.
- Cloudflare launched Clairic, a "System One" model built only to read; it returns typed values, not prose. In Cloudflare's own test, a pipeline on Clairic ran 49% cheaper and about 3x faster than an all-LLM version, billed on input only. Independent accuracy benchmarks aren't out yet.
- Microsoft Research and MIT open-sourced probabilistic bias correction, a layer that sharpens the hard 2-to-6-week weather forecast instead of replacing the models agencies already run. Published in Nature under an MIT license; gains are probabilistic and concentrated in the tropics.
The one that actually moves something this week is the DeepSeek order. We've tracked those fights all month — China's own export controls on the 17th, the open-source release just yesterday — and this time it's Washington flipping it around, ordering the model pulled out of every civilian federal agency and treating a frontier model the way it would a banned hardware vendor.
The other three are quieter, and they land closer to home. The sidewalk-robot fight we watched turn into an arrest last week is now about what the robots were recording and who kept it. And Cloudflare and a Microsoft and MIT team each shipped something deliberately smaller. Cloudflare's is a model that only reads and never writes; the MIT group's is a correction layer bolted onto forecasts we already run, just to make them less wrong.
Homeland Security pulled DeepSeek from federal systems the same week Chicago sued the delivery robots and Cloudflare shipped an AI built only to read.
1. Homeland Security ordered every civilian federal agency to remove DeepSeek from its systems.
The clause that matters reaches past DeepSeek's own apps to any third-party product quietly built on it.
The Department of Homeland Security has issued Exclusion and Removal Order #2026-01 (docket DHS-2026-0628), filed for public inspection on October 19 and set for Federal Register publication the next day, on the recommendation of the Federal Acquisition Security Council under the 2018 Federal Acquisition Supply Chain Security Act. In plain terms: civilian executive-branch agencies must take Hangzhou DeepSeek's products out of their systems, and the removal requirement extends beyond DeepSeek's own apps to third-party products and services built on DeepSeek implementations — the vendor tool in a stack that quietly calls those models. Agencies have 60 days to apply, in writing, for an exception; otherwise the exclusion takes effect immediately.
DHS's stated reasons read like a threat model: Chinese laws that let the state reach into a private company's infrastructure, personal data collected at account creation, the risk that prompts carry sensitive information to China, and critical flaws already found in DeepSeek's iOS app. Washington is treating a foreign frontier model the way it treats a compromised hardware vendor, and using a 2018 supply-chain law to do it. This covers civilian agencies only, not states and not private companies. The order doesn't decide anyone else's policy; it makes the federal government's own call, in public and on the record.
Why this matters: If you run IT or procurement, the order quietly redraws what counts as a DeepSeek dependency — not just the app your staff installed, but the vendor product three layers down that calls those models and never told you. My own read is that the "built on it" clause is the whole story here: a ban on an app is easy to satisfy, while a ban on everything built on the app is a software-inventory problem most organizations can't answer today.
Action this week: Inventory the embedded case, not just the obvious one — map which vendor products in your stack are built on or call DeepSeek models, because that buried dependency is exactly what the order targets. If you're federal-adjacent, expect a DeepSeek-free attestation to start appearing in contracts within weeks, and map your model dependencies before an RFP forces you to. If you're fully private, read this as the federal risk assessment it is and set your own policy deliberately, because nobody is deciding it for you yet.
techradar.com: More US government departments ban controversial AI model DeepSeek over security risks (October 20, 2026)
the-decoder.com: China protests the US DeepSeek ban while blocking ChatGPT, Claude, and Gemini at home (October 2026)
federalregister.gov: Department of Homeland Security agency documents
2. Chicago sued the delivery robots over what their cameras recorded on the sidewalk.
Two Cook County class actions say Coco's and Serve's bots measured people's faces and hands under Illinois' biometric law, then kept the data to train AI.
Two proposed class actions filed October 9 in Cook County accuse Coco Robotics and Serve Robotics of violating the Biometric Information Privacy Act, Illinois' and the country's strongest biometric statute. The complaints say the robots' cameras and LiDAR measured the faces and hands of customers unlocking their orders, and scanned other pedestrians who never signed up for anything. The named plaintiffs are Mitchell Breidinger, who got a DoorDash order from a Coco bot on the Near North Side, and Barry Schain, whose Uber Eats order arrived by a Serve bot in Lincoln Park; both say the biometric data was retained to train AI and shared with third parties. Coco says it hasn't collected biometric data in Chicago, and Serve says its systems aren't designed to identify anyone.
What gives this teeth is the statute, not the outrage. BIPA sets damages of $1,000 per negligent and $5,000 per intentional or reckless violation, and the part that makes Illinois the venue is that it doesn't require you to prove you were harmed. This is the sidewalk-robot fight moving a step: just last week we covered a woman arrested in Philadelphia for blocking one of these bots, and that argument was about obstruction and jobs. This one is about biometrics, and the venue is no accident: Illinois is where a biometric claim actually carries a price.
Why this matters: If a little delivery bot rolled past you in Chicago this fall, two lawsuits now say it may have measured your face and hands, and that a scan of a stranger on the sidewalk is a biometric record somebody kept. What strikes me is that the pedestrian who never ordered anything has the least say here and, under BIPA, maybe the clearest claim of all.
Action this week: Watch whether a judge certifies the class if you're in Illinois, because that's the moment this goes from a filing to a bill. If you deploy cameras or LiDAR anywhere public — a kiosk, a robot, a smart-city sensor — treat face and hand geometry as regulated biometrics now: get written consent and publish a retention schedule before you collect, not after you're served. And if you're outside Illinois, notice the gap this exposes, because the same robot doing the same thing in a state without a BIPA-style law draws no comparable penalty.
chicago.suntimes.com: Lawsuits accuse Chicago delivery robot companies of collecting biometric data (October 16, 2026)
wbez.org: Lawsuits accuse Chicago delivery robot companies of collecting biometric data (October 16, 2026)
thestar.com.my: Two class-action lawsuits filed against food-delivery robot companies for biometric privacy violations (October 18, 2026)
3. Cloudflare launched an AI built only to read, not to write.
Clairic returns typed values instead of prose, and Cloudflare's own test ran a pipeline 49% cheaper and about 3x faster than the all-LLM version.
How much of the inference bill is a large language model doing work that doesn't need a language model? Cloudflare's answer this week is Clairic, which it calls the first of a new "System One" class of extraction models, hosted on Workers AI. Instead of generating prose, it returns typed values — numbers, dates, lists — straight out of text and images, with a 65,536-token context window and vision support, priced at $0.10 per million input tokens and billed on input only because it emits no output tokens.
The headline claim is Cloudflare's own: a three-step customer-claims pipeline built on Clairic ran 49% cheaper and about 3.2x faster than the same pipeline on a general model, answering in under a second. Hold those at arm's length, because they are the vendor's figures, independent accuracy benchmarks aren't out, and an extraction model that returns a confident wrong number is worse than a slow correct one. And the interesting part isn't the discount. The inference-cost story we've tracked all quarter has been a price war, with DeepSeek repeatedly slashing what it charges per token. Clairic is a different move: stop routing dull structured work through a model built to generate. Cloudflare says it will open the weights later, without giving a date.
Why this matters: If you run AI in production, Clairic forces a buying question: how much of what you pay a general model for is really just extraction, pulling one field out of a document, and would a reader-only model do it cheaper and faster. When I have watched teams chase inference savings, the win almost never comes from a cheaper token; it comes from noticing a workload that was mis-specified as "generation" in the first place.
Action this week: Audit one pipeline for the structured-extraction steps — invoice totals, dates, field parsing — that don't need a chatbot, and benchmark one of them on your own documents, not Cloudflare's demo. Get the accuracy question answered in writing before you switch, because the cost and latency wins are the vendor's and the error cost is yours. And watch for the promised open-weights release and whether other "System One" models follow, because if specialized readers hold up, routing everything through one big model gets harder to justify on the invoice.
blog.cloudflare.com: Introducing Clairic, our first System One extraction model (October 19, 2026)
unite.ai: Cloudflare releases Clairic, its first System One extraction model (October 2026)
ppc.land: Cloudflare's Clairic model cuts AI pipeline cost 49% in its own test (October 2026)
PickBits Daily Signal is free. If it lands in your inbox every day and it is worth something to you, the best way to support it is to forward it to someone who would read it. Subscribe today!
4. Microsoft and MIT open-sourced an AI that sharpens the hardest stretch of the weather forecast.
Probabilistic bias correction cleans up the 2-to-6-week window farmers and disaster crews plan around and forecasters have always found hardest.
A farmer deciding when to plant, or a crew pre-positioning for a storm, both make the call in the two-to-six-week window, and that's exactly where forecast skill has always collapsed. A Microsoft Research and MIT team, publishing in Nature, has released an open-source method called probabilistic bias correction that sits on top of existing forecast systems and corrects their systematic errors in that sub-seasonal range, rather than replacing the models agencies already run. Judah Cohen at MIT and Lester Mackey at Microsoft Research report it roughly doubles the skill of an AI-based forecaster and improves traditional physical models by more than 90%.
The code is on GitHub under an MIT license, archived on Zenodo, and model-agnostic, so any weather service can apply it to the forecasts it already issues rather than standing up a new system. The caveat the authors keep up front, and I'll keep it too, is that the gains are probabilistic and concentrated in the tropics, thin in Europe. This is sharper odds, not a six-week certainty. But it's the hopeful one in today's lineup. There's no black box to trust here: it's open code any weather service can read and run, and all it does is make the forecasts they already issue less wrong in the window that matters.
Why this matters: Weeks of warning before a dry spell or a storm is the difference between a harvest planned and a harvest lost, and now that warning is a free, open tool any weather service can run rather than a product only a rich agency can buy. What gets me is how unglamorous the win is: it doesn't replace anyone's forecast, it just quietly makes the one they already issue less wrong in the window that matters.
Action this week: Send this to anyone you know who runs sub-seasonal outlooks — a utility, a food-security group, an ag planner — because weeks of lead time is exactly what lets them move before the weather does. Read the honest limit first, though: the gains are concentrated in the tropics and modest in Europe, so treat it as sharper odds for planning, not a new certainty. And watch whether national weather services actually adopt it, because an open method only helps once someone runs it.
sciencemediacentre.es: Artificial-intelligence method improves weather forecasts six weeks in advance (October 2026)
latestly.com: Microsoft AI tool can predict weather up to 6 weeks ahead (October 2026)
letsdatascience.com: Microsoft releases probabilistic bias correction tool (October 2026)
» What to watch this week
- Whether the first DeepSeek-free attestations show up in federal contracts, and whether the removal order's "built on it" clause gets tested against a real vendor product. The easy part of the order is pulling an app; the hard part is proving nothing you bought is quietly built on the model underneath.
- Whether a Cook County judge certifies a class in the Coco and Serve suits. Class certification is the step that turns per-scan damages from a theory into a number, and it's the moment this stops being a filing and starts being a bill.
- Whether independent benchmarks back Cloudflare's Clairic numbers, and whether other labs ship their own reader-only extraction models. The 49%-cheaper claim is the vendor's; the category only matters if a "System One" model holds up on someone else's documents.
- Whether any national weather service actually puts probabilistic bias correction into routine operations. Getting a method from a Nature paper into an operational forecast is a bigger hurdle than publishing it, and adoption, not the result, is the real test of whether it helps.
Tomorrow's signal lands here.