> daily_signal(2026_07_09)

The AI agent Anthropic said would erase your job just reached your phone, and its own data says coding was never the real target.

PickBits Daily Signal · Thursday, July 9, 2026

By Mark Pickering · 8 min read · July 9, 2026

// tl;dr

Anthropic put Claude Cowork, its autonomous work agent, on the web and your phone, and released usage data showing coding is only 8.7% of what people do with it. Across 1.2 million sessions at more than 600,000 organizations, business-process work and content drafting dominated. The agent can run scheduled tasks with no device online, though Anthropic's own demo leaves the drafted email unsent.
California closed the enforcement gap that let a driverless car break traffic law with no one to cite. New DMV rules effective July 1 let officers issue a formal "Notice of AV Noncompliance" to a manufacturer, give emergency crews a two-minute window to clear an area of a robotaxi fleet, and lift the state's prior ban on autonomous heavy trucks.
Meta is rolling out a mandatory update that disables the camera on its Ray-Ban glasses when the recording light is tampered with or destroyed. Some owners moved from tape to paying third parties to kill the LED so they could film people covertly. Meta says it will also pull listings for those services and has left legal action on the table.
A Stanford team published a Nature study showing a language model can forecast the results of social-science experiments about as well as a panel of expert humans. Across 476 treatment effects and 105,165 participants, simulated results tracked real ones at r=0.85, and r=0.90 for studies published after the model was trained. The authors flag that it overestimates effect sizes and cannot replace a real experiment.

Anthropic put the work agent we said in May was coming for entry-level jobs directly onto your phone, then published the numbers that answer a question the industry keeps dodging: coding is 8.7% of the work, and the rest is the ordinary office desk most people are actually paid for. California only just closed a basic gap that let a driverless car break the law with nobody to answer for it, years after robotaxis started driving on public roads. Meta locked its camera in hardware only after people started paying to destroy the warning light. The last story runs the other direction: a large model pointed at a slow, expensive research problem instead of a person.

Today: an AI agent reaches your pocket to do office work instead of code, California closes the ticket nobody could write, Meta bricks a camera over a defeated warning light, and a language model forecasts human behavior about as well as the experts who do it for a living.

1. Anthropic put its work agent on your phone, and its own data says coding is not the point.

The tasks most at risk are the ones on an ordinary office desk, not an engineer's.

Anthropic expanded Claude Cowork, its agent that carries out multi-step knowledge work on its own, from a desktop-only app to the web and mobile, starting with Max subscribers at $100 a month and reaching other plans "in the coming weeks." The pitch: start a task at your desk, get updates on your phone, and let Cowork finish in the background even with no device online. Its example is a 6 a.m. Monday client-prep job in which the agent works through email threads, transcripts, and recent news, builds a briefing, and leaves a follow-up email drafted but unsent. Alongside the launch, Anthropic published data from 1.2 million anonymized sessions across more than 600,000 organizations: business-process automation made up 33.4% of use, content creation 16.4%, and software development just 8.7%.

Back in May, Anthropic's own chief executive said AI would erase half of entry-level white-collar jobs, and we noted he was also the one shipping the agents to do it. This is that shipping continuing, now in your pocket, with a usage split that cuts against the comfortable story. The tasks Cowork is actually used for, the reports and contracts and decks, are the everyday office work most people are paid for, not the engineering the coverage fixates on. Anthropic frames the same number as "everyday business work is on the rise," which is the sunny reading of a finding with a sharper edge: the near-term automation exposure is the ordinary desk, and the tool is built to act on its own and schedule work while you sleep. The unsent email in the demo is the tell that a human is still meant to be in the loop, for now.

TechCrunch report July 7 2026 Anthropic expands Claude Cowork autonomous work agent to web and mobile starting with Max subscribers at 100 dollars a month usage data from 1.2 million sessions across more than 600000 organizations shows business-process automation 33.4 percent content creation 16.4 percent software development only 8.7 percent agent runs scheduled tasks in the background leaves follow-up email drafted but unsent — techcrunch.com · July 7, 2026

Why this matters: If your day is reports, spreadsheets, contracts, or briefings, that is the work Anthropic is aiming its agent at, and its own numbers say that is where the automation is landing first, not on the engineers. The planning input here is concrete: the exposure map is 8.7% coding against roughly half the use in business-process and content work. Action this week: Run one bounded, low-stakes task through Cowork that you can fully verify, and keep a human on the approval step, because the agent's default is to act and even schedule work while you are away. If you manage a team or set AI policy, use the usage split as a guide to where review, data-access controls, and retraining actually need to go, and treat "AI is a developer tool" as the assumption to drop. Follow this one; the launch matters less than which desk jobs quietly get handed over next.

techcrunch.com: The coding-agent wars are spilling into the rest of the office with Claude Cowork (July 7, 2026)
venturebeat.com: Anthropic brings Claude Cowork to mobile and web as usage data shows most users aren't coding (July 7, 2026)
9to5mac.com: Anthropic is expanding Claude Cowork to mobile and web (July 7, 2026)

2. California closed the loophole that let a driverless car break the law with no one to cite.

Traffic law only ever covered human drivers, so a robot breaking it had no one to answer to.

California's DMV finalized a package of autonomous-vehicle regulations on April 28, 2026, effective July 1, 2026, closing a long-standing enforcement gap: because state traffic law only ever applied to human "drivers," officers had no formal way to cite a driverless car even after a clear violation like running a red light, failing to yield to a pedestrian, or entering an active emergency scene. Under the new rules, officers instead issue a "Notice of AV Noncompliance" to the DMV and the vehicle's manufacturer, and companies must turn over vehicle data and telemetry within 72 hours, or 24 if the violation is flagged as a public-safety threat. The rules also create "emergency geofencing directives": local emergency officials can transmit a digital order requiring a manufacturer to route its entire fleet out of a designated area within two minutes, and operators must maintain a two-way emergency-response link with a 30-second response time. Separately, the regulations lift California's prior ban on autonomous heavy-duty trucks, letting manufacturers apply to test and eventually deploy them commercially after a phased regime the DMV puts at 500,000 miles of testing per phase. DMV Director Steve Gordon: "California continues to lead the nation in the development and adoption of AV technology."

This is a genuinely new thread for us, and the underlying problem was ordinary enough that it showed up on a local news broadcast months before Sacramento wrote a rule for it: a Phoenix officer pulling over an erratic Waymo, a bystander saying somebody needs to be held accountable, and police confirming a citation was technically possible but something they rarely bothered with. California's fix is paperwork, not new technology: a formal notice and a data deadline, nothing mounted on the car itself. In the same package, the state also lifted its ban on autonomous heavy-duty trucks, betting that a paperwork requirement is enough oversight for vehicles that now weigh ten times as much.

The Robot Report piece July 8 2026 California DMV autonomous vehicle regulations finalized April 28 2026 effective July 1 2026 officers issue Notice of AV Noncompliance to manufacturer and DMV companies must turn over telemetry within 72 hours or 24 for safety-flagged violations emergency geofencing directives give local officials a two-minute fleet clearance window lifts ban on autonomous heavy-duty trucks DMV Director Steve Gordon quote California continues to lead the nation in AV technology — therobotreport.com · July 8, 2026

Why this matters: If you live, drive, or walk in California, a driverless car breaking the law near you now creates a real paper trail, even though the fine or fix lands on the manufacturer, not a driver in the seat. Action this week: Keep reporting AV incidents to local police exactly as you would a human driver; that report is what now triggers the manufacturer's Notice of AV Noncompliance. If you build or operate an AV or robotics fleet doing business in California, audit your compliance stack now: telemetry that can turn around within 72 hours, or 24 for safety-flagged violations, a system that can execute a two-minute emergency-geofencing directive, a 30-second-response communications link, and clearing the 500,000-mile-per-phase testing bar before heavy-duty deployment.

therobotreport.com: Tickets, geofences, and 1M miles: the new reality of California AV compliance (July 8, 2026)
sfist.com: Authorities can fine driverless cars, deploy emergency geofencing starting July 1 (April 29, 2026)
dmv.ca.gov: New autonomous vehicle regulations strengthen oversight and enforcement, authorize trucks and transit

3. Meta is bricking the camera on its glasses when you defeat the recording light.

The little white LED is being promoted from a courtesy to a hardware lock.

Meta began rolling out a mandatory update that disables the camera on its Ray-Ban Meta smart glasses when the device detects that the capture LED, the white light that tells nearby people the camera is recording, has been blocked, modified, or destroyed. Earlier glasses only paused the camera while the light was covered; the new logic targets owners who escalated from tape to "sophisticated efforts to modify or destroy the capture LED," some paying third parties to alter the hardware, specifically to film people without their knowledge. Meta says the change begins with its second-generation glasses, that it is working to remove ads, posts, and Marketplace listings for services that defeat the light, and that legal action against those sellers is on the table. The move follows documented non-consensual recording, including a case in which a woman was filmed without consent in a clip that drew more than a million views.

Back on June 27 we covered Meta's cheaper, non-Ray-Ban glasses reaching a mass-market $299 the same week the first code of conduct for face-cameras appeared, after people kept getting caught filming strangers. This is the next turn of that same story, and a real improvement: a hardware lockout beats a light someone can quietly tape over. It took people physically destroying the LED to get Meta to build it, and the lit light is still the only thing telling a bystander a camera might be on.

PetaPixel report July 8 2026 Meta rolling out a mandatory update that disables the camera on Ray-Ban Meta smart glasses if the capture LED recording light is blocked modified or destroyed beginning with second generation glasses responding to owners who moved from tape to paying third parties to destroy the LED to film people without consent Meta to remove Marketplace listings for LED-defeat services and pursue legal action — petapixel.com · July 8, 2026

Why this matters: If you own Ray-Ban Meta glasses the update is mandatory, and a device with a tampered or damaged light will have its camera disabled; if you are near someone wearing a pair, the lit white LED remains your only signal that recording may be active. Enough people now own camera glasses that Meta had to answer the consent question with hardware instead of trusting owners to be polite about a light. Action this week: Install the update rather than putting it off, and if you are weighing camera glasses for work, events, or public spaces, note that defeating the light can now void the camera and expose you to the legal action Meta says is on the table. Build recording consent into how you use the device instead of assuming a discreet capture is fine, and if you set policy for an office, a clinic, or a venue, decide now where wearable cameras stand before one shows up on someone's face in your building.

petapixel.com: If users conceal the recording light on smart glasses, Meta says it will disable the camera (July 8, 2026)
9to5google.com: Meta Ray-Ban smart glasses get a privacy-light camera update (July 7, 2026)

4. A language model forecast a social-science experiment about as well as expert humans.

Point the same technology at slow, expensive research instead of your attention.

Researchers at Stanford's AI for Public Benefit Lab published a study in Nature, "Large Language Models Can Predict the Results of Social Science Experiments," finding that a model can estimate experimental outcomes about as accurately as a panel of human forecasters. The team assembled 70 preregistered, nationally representative US survey experiments covering 476 treatment effects and 105,165 participants, then had GPT-4 simulate how representative samples of Americans would respond to each condition and inferred the treatment effect from the difference. The simulated effects correlated with the real ones at r=0.85, matching pooled human-forecaster accuracy; for unpublished studies whose results could not have been in the model's training data, the correlation rose to r=0.90, evidence the model was predicting rather than recalling.

We've followed this research-acceleration thread before, from the MIRA diagnostic AI in Nature to Meta's non-invasive brain-to-text work, and an earlier study already reproduced human survey and voting responses at high correspondence. The Stanford result goes further: the correlation held at r=0.90 even for studies published after the model's training cutoff, real evidence it was predicting, not remembering. That's a cheap first pass, screening a hundred ideas in an afternoon so the real budget goes to the handful worth running. The authors are upfront about the catch: the model overestimates how large effects are, a simulated result is not a finding, and the same tool that speeds honest research can just as easily manufacture plausible-looking "evidence" for a claim nobody tested.

Stanford AI for Public Benefit Lab project page 2026 Large Language Models Can Predict the Results of Social Science Experiments published in Nature GPT-4 simulated how representative samples of Americans would respond across 70 preregistered experiments 476 treatment effects 105165 participants simulated effects correlated with real ones at r 0.85 and r 0.90 for unpublished post-training-cutoff studies matching human forecaster accuracy authors caution the model overestimates effect sizes — ai4pb.stanford.edu · July 8, 2026

Why this matters: If you design or fund social-science, behavioral, or survey research, this is a screening tool, not an oracle: a simulation can cheaply triage which hypotheses or message variants are worth a real experiment, but the finding that the model overestimates effect sizes means it cannot stand in for real subjects. It saves study time and money instead of using real subjects to test claims no one plans to actually run. Action this week: Use it to narrow the field before you spend the budget, then still run the study. And if you work in policy, marketing, or communication and are tempted to skip testing because "the AI already predicted it," read the misuse caveats first and insist on the actual preregistered experiment before acting on a simulated result. Follow the thread; the next question is which fields adopt this as a first pass and whether the caveats travel with it.

ai4pb.stanford.edu: Predicting the results of social science experiments using large language models (2026)
nature.com: Large language models can predict the results of social science experiments (2026)

» What to watch this week

How fast Claude Cowork reaches plans below Max, and whether the office-work exposure shows up in real staffing decisions. The usage split is the tell; watch whether the resellers building on these agents pitch them as coding tools or, more honestly, as the thing coming for the reports-and-contracts desk.
Whether California's two-minute emergency-geofencing directive gets invoked for real, and whether other states copy the Notice-of-AV-Noncompliance model. A rule only proves itself the first time an emergency crew actually uses it to clear a fleet; watch whether other AV markets adopt the same paper trail instead of writing their own from scratch.
Whether Meta's LED lockout holds up in practice and whether the Marketplace-listing removals and threatened legal action materialize. A hardware consent signal is only as good as its enforcement; the question is whether the defeat-services market simply moves off Meta's own platforms.
Which fields pick up LLM experiment-forecasting as a first pass, and whether the overestimation caveat survives the adoption. The risk is not the tool but the shortcut: a simulated result quoted as a finding once the r=0.85 headline outruns the authors' own limits.

Tomorrow's signal lands here.