> daily_signal(2026_07_09)
Anthropic just put its autonomous work agent in your pocket — Claude Cowork now runs tasks on web and phone, and its own data shows coding is only 8.7% of what people actually use it for.
PickBits Daily Signal · Thursday, July 9, 2026
1. Anthropic just put its autonomous work agent in your pocket — Claude Cowork now runs tasks on web and phone, and its own data shows coding is only 8.7% of what people actually use it for.
On July 7, 2026, Anthropic expanded Claude Cowork — its agent that carries out multi-step work tasks on its own — from a desktop-only app to web and mobile, starting with Max subscribers ($100/month) and reaching other plans 'in the coming weeks.' The pitch is device-independent, scheduled autonomy: kick off a task at your desk, get updates on your phone, and let Claude finish the job in the background even with your laptop closed. Anthropic released usage data from 1.2 million anonymized Cowork sessions across more than 600,000 organizations that undercuts the 'AI = coding' framing: software development was just 8.7% of activity, with the bulk going to ordinary office work — reports, spreadsheets, contracts, and content drafting.
Key fact: If your job involves the office work Cowork is aimed at — drafting reports, processing contracts, managing content, prepping briefings — try a bounded, low-stakes task first (one you can fully verify) and keep a human in the approval loop, because the agent's default is to act autonomously and even schedule work while you're away; Anthropic's own example leaves the follow-up email 'drafted but unsent' for a reason.
techcrunch.com · venturebeat.com · 9to5mac.com · primary source
2. California just closed the loophole that let a driverless car break the law with no one to cite — new DMV rules that took effect July 1 also give emergency crews a two-minute veto over robotaxi routes and open the door to autonomous freight trucks.
Because California traffic law only ever covered human 'drivers,' officers had no formal way to cite a driverless vehicle even after a clear violation — running a red light, failing to yield to a pedestrian, entering an active emergency scene. A package of DMV regulations finalized April 28, 2026 and effective July 1, 2026 closes that gap: officers now issue a 'Notice of AV Noncompliance' to the vehicle's manufacturer and the DMV, local emergency officials can order an AV fleet to clear a designated area within two minutes, and the state lifted its prior ban on autonomous heavy-duty trucks. The Robot Report's July 8, 2026 look at the rollout — reported through fleet-operations provider Guident, which runs autonomous shuttles in Florida — frames this as the new operating reality every AV company doing business in California now has to build for.
Key fact: If you live, drive, or walk in California, know that a driverless car breaking traffic law now has a real enforcement path — a 'Notice of AV Noncompliance' goes to the manufacturer and the DMV — but the consequence lands on the company, not a driver in the vehicle; keep reporting incidents to local police as you would with any traffic violation.
therobotreport.com · sfist.com · dmv.ca.gov · primary source
3. People were destroying the little light that warns you a Meta camera is on — so Meta's new mandatory update bricks the camera the moment the recording LED is tampered with.
Meta began rolling out a mandatory update on July 7-8, 2026 that disables the camera on its Ray-Ban Meta smart glasses whenever the device detects that the capture LED — the light meant to warn bystanders they're being filmed — has been physically blocked, modified, or destroyed. Earlier glasses would only pause the camera while the light was covered; the change targets people who moved from tape to 'sophisticated efforts to modify or destroy' the LED, including paying third parties to do it, specifically to record others without consent. Meta also says it will remove ads, posts, and Marketplace listings for services that defeat the light, and has left legal action against those sellers on the table.
Key fact: If you own Ray-Ban Meta smart glasses, install the mandatory update — it's not optional, and a device with a tampered or damaged capture LED will have its camera disabled; conversely, if you're around someone wearing camera glasses, the still-lit white LED remains your signal that recording may be active.
petapixel.com · 9to5google.com · primary source
4. Stanford researchers just published a Nature study showing a language model can forecast how a social-science experiment will turn out about as well as a panel of expert human forecasters — even for studies published after the model was trained.
In a study published in Nature, Stanford's AI for Public Benefit Lab reports that GPT-4 can predict the results of social-science experiments about as accurately as a pooled panel of human forecasters. The team built an archive of 70 preregistered, nationally representative US survey experiments — 476 treatment effects across 105,165 participants — and had the model simulate how representative samples of Americans would respond to each experimental condition, then inferred the treatment effect. Simulated effects correlated with the real ones at r=0.85 overall, and r=0.90 for unpublished studies that could not have been in the model's training data. The promise: a cheap, fast first-pass simulation that could help researchers screen and design experiments before spending months and thousands of dollars running them — with the authors flagging real limits (the model overestimated effect sizes) and misuse risks.
Key fact: If you design or fund social-science, behavioral, or survey research, treat this as a screening tool, not an oracle: an LLM simulation could cheaply triage which hypotheses or message variants are worth running a real experiment on, but the authors' own finding that the model overestimates effect sizes means it can't replace human subjects — use it to narrow the field before you spend the budget, then still run the study.
ai4pb.stanford.edu · nature.com · primary source