Anthropic Defines AI Risks and Meta Backtracks on Crawling

Anthropic Wants to Stop AI Models Only After 50 Deaths

Anthropic has submitted its own bill to U.S. lawmakers on Capitol Hill for regulating advanced AI systems. The draft would allow the U.S. government to stop the release or use of a model if it poses a foreseeable and imminent risk of killing or seriously injuring more than 50 people or causing over a billion dollars in damages. The threshold applies to three sources of risk: the creation or release of biological weapons, a system that escapes its developer’s control, and cyberattacks. A company spokesperson explains that the text reflects the recommendations of the company’s in-house Advanced AI Framework, which the firm has been promoting to federal and state politicians since June.

Under the draft, developers of frontier models would have to conduct safety tests and publish the results. Independent auditors would register with the U.S. Department of Commerce and evaluate the labs and their most advanced models at regular intervals; the companies would be allowed to pay for these audits, as long as the compensation does not depend on the outcome. If an auditor identifies an imminent risk, they must inform the Department of Commerce within 72 hours of completing the assessment. The department could then petition a federal judge in the District of Columbia to suspend the model. The draft explicitly limits this procedure: other departments, agencies, and officials would not be allowed to independently stop the development, deployment, or internal use of a model. This follows far-reaching export controls imposed months earlier by federal officials in a series of phone calls concerning two of Anthropic’s most advanced models.

In parallel, Anthropic announced it would give six auditors from the non-profit organization METR 'employee-like access,' including company-owned hardware and office space, as well as internal tools on par with its own risk assessment teams. According to the company, customer data, legally privileged information, and details about the model architecture are excluded. METR publishes its findings, and Anthropic cannot withhold anything, but it can request the redaction of sensitive sections. Dario Amodei had proposed embedding external auditors in a blog post in September. Third-party auditing is not yet mandatory: the bipartisan FRONTIER Act from July would require independent verification organizations, and OpenAI first signaled partial support for federal regulation in mid-September. The Trump administration has so far rejected broader oversight and is relying on voluntary commitments, but as of last week, it requires the reporting of security incidents after Claude sent a false murder tip to the police. The FTC also served OpenAI and Anthropic with requests for information regarding potential consumer harm. → Storyboard18, Politico, Reuters, Bloomberg Law

Synthszr Take: Less than twelve hours passed between the first report from Washington late on October 16 and the story in Indian business media the next morning at 9:51 a.m. local time. What went around the world was a draft without a committee and without a vote: a document a company handed to lawmakers. Nevertheless, the thresholds of 50 deaths and a billion dollars in damages are now in Singapore and Mumbai and are being cited from there long before anyone in Congress has voted on it. This speed is the real leverage: A phrase that circulates globally early enough becomes a reference point for regulators who still have to write their own text. In twelve hours, Anthropic generated more regulatory reach than most laws do in twelve months.

Meta Removes Clause for Crawling Partner Websites After Protests

Meta has removed a controversial passage from its commercial terms of service that would have allowed the company to automatically read the websites associated with business accounts and use their content to train its AI products. The wording was part of an update scheduled to take effect on October 30. It authorized Meta to use scrapers, bots, crawlers, and test access to collect, index, cache, and periodically retrieve the domains, pages, and URLs related to a business customer’s products and services. The second part stated that Meta could use any content from these domains to provide, develop, improve, research, and train its own products. Meta rejected this interpretation from the start: Head of Communications Andy Stone and spokesperson Daniel Roberts stated that nothing would change regarding the use of web crawlers. After the removal, Roberts added that the consolidated crawling terms had only applied to advertising and business products, and this was now clarified in the text itself; content providers could still signal their preferences via robots.txt. The background to the debate is the ongoing dispute between news organizations and OpenAI over whether training without a license qualifies as fair use. In parallel, Meta has begun to charge users for posting links on its platforms. → Ars Technica, The Verge

Synthszr Take: The reversal came within days, and only because a few lawyers and news outlets read a passage that would have taken effect on October 30 without any further action. For anyone with a business account, this means: the negotiating power lies in the window between the announcement and the effective date, which is often two weeks. Meta is removing the clause while simultaneously claiming that nothing ever changed anyway, which doesn’t add up. The next version will likely contain the same permission in more innocuous terms, distributed across supplementary conditions that no one reads in context. A tool that automatically compares and highlights changes to platform terms is now part of the essential toolkit for every publisher and creator who lives off of reach.

Torvalds Calls AI Disclosure Requirement 'Utterly Stupid'

The Linux maintainers are getting rid of the Assisted-by tag, which developers have been asked to use in their commit messages since January to indicate if they have used 'any advanced coding tool.' Sasha Levin, a kernel developer and Nvidia engineer, submitted the corresponding patch to the mailing list on October 9, citing a consensus from the Maintainers Summit; the discussion has since garnered over 50 replies. Jeff Layton of Red Hat had previously argued that the tag was not being consistently used anyway and effectively served as free advertising for providers of proprietary language models. Laurent Pinchart disagreed on October 13, wanting to abandon the requirement only in exchange for another disclosure mechanism, because rules set expectations, even if some individuals circumvent them. Linus Torvalds responded on the same day on the mailing list with the sentence that such rules were 'utterly stupid,' and sarcastically suggested that Pinchart should also stop using compilers and submit directly in assembly. → Techpresso

Synthszr Take: A commit tag is administrative work disguised as quality assurance. Torvalds' compiler jibe hits the sore spot of the kernel process: the patch is evaluated, not its origin story, and it was never any different before 2026. Still, Pinchart is right that written rules shape expectations, even towards people who circumvent them.

DeepSeek Updates Model, Touts Lower Resource Consumption

DeepSeek released an updated AI model over the weekend, describing it as faster and leaner. According to Bloomberg Technology, the update intensifies competition with Anthropic, OpenAI, and Chinese providers Moonshot Kimi and Z.ai. The terms 'faster' and 'leaner' are the company’s own statements, not the results of a documented comparison with the mentioned competitors. DeepSeek is thus positioning operational efficiency alongside pure model capability as a selling point. The announcement does not specify what exactly is changing in terms of memory requirements, computational effort, or end-customer price. → TechOrange 科技報橘

Synthszr Take: In this announcement, 'lean' is a price argument disguised as a technical one. DeepSeek names four competitors in the same breath—Anthropic, OpenAI, Moonshot Kimi, and Z.ai—thereby drawing them all into the same calculation: What is the cost of a response that doesn’t need to be corrected afterward? Back in August, DeepSeek kept token prices stable for V4 Pro while others were tweaking speed, and the pattern is now repeating itself with efficiency as a selling point.

AI Shopping Agents Force Retailers to Overhaul Access and Payment Authorization

According to a Wall Street Journal report, online retailers must redefine how they reach customers, authorize payments, and protect their customer relationships as AI shopping agents insert themselves between merchants and buyers. The responses vary: Amazon is blocking prominent external agents, Target is allowing approved providers, and Walmart is cooperating with AI providers to attract customers through them. Agents like Meta’s Muse and OpenAI’s dots are now selecting products, bidding in auctions, booking travel, and ordering groceries. This creates a visibility problem for retailers, as third-party agents often disclose little about whom they represent, and different models evaluate the same products differently. The optimization provider BrightEdge estimates that purchases via agents are increasing by 43 percent monthly and have roughly quadrupled since April, while only about three percent of the e-commerce companies it serves allow agents to complete a transaction. → Techpresso

Synthszr Take: Three percent of the shops managed by BrightEdge allow an agent to complete a purchase, while the volume of such purchases is growing by 43 percent per month. This gap can be closed in the backend. Product data must be structured in such a way that an external model can evaluate it correctly, and every order needs verifiable proof that the agent is acting on behalf of the buyer. Waller precisely identified the expensive part: The verification shifts from the payer to their authorization, and for this, most checkout systems simply don’t have a field yet.

F-22 Pilot Prompts Autonomous Drone to Target for the First Time

In October 2026, a US Air Force pilot guided Anduril’s unmanned YFQ-44A Fury in flight for the first time from the cockpit of an F-22 Raptor. According to a report from The War Zone, both aircraft took off from Edwards Air Force Base in California and met in the air, after which the pilot transmitted an updated mission plan to the drone via a tactical data link. Instead of remote-controlling every movement, he gave high-level tactical instructions that the onboard autonomy software interpreted and executed. The Air Force describes the human’s role in the cockpit as that of a “quarterback”; according to their statements, the demonstration included dynamic flight maneuvers and simulated mission tasks. The Fury is Anduril’s contribution to the Collaborative Combat Aircraft program, for which the YFQ-42A from General Atomics was also selected for the first stage. → TechOrange 科技報橘

Synthszr Take: The pilot formulates an intent, and onboard software translates the execution into concrete maneuvers. This shifts the accountability for decisions to a codebase that cannot testify in any court, and the term “quarterback” describes this situation more accurately than the Air Force probably likes. As long as a human is guiding one machine, oversight can still be claimed; with the stated program goal of several Fury drones per cockpit, this turns into rubber-stamping, because attention cannot be divided.

China Showcases AI-Controlled Interceptor Drones for Small Aircraft

At an anti-terror forum in Beijing, China presented low-cost, AI-controlled interceptor drones designed to neutralize small unmanned aerial vehicles. The systems were showcased at the Great Wall-2026 International Forum on Counterterrorism, which ran from October 13 to 15 and, according to the organizer, brought together more than 320 defense experts and delegates from 62 countries. The event was organized by the People’s Armed Police, the country’s paramilitary unit responsible for internal security. According to the exhibitors, the interceptor drones target the Low, Slow and Small category, i.e., slow, low-flying micro-devices; prices, speeds, operational ranges, or the principle of operation were not mentioned in the report. Also showcased were humanoid robots, robot dogs, unmanned ground vehicles, detection technology, and laser-based defense systems. → TechOrange 科技報橘

Synthszr Take: The technology that intercepts a drone in mid-air is largely the same as the technology that places a package on a balcony: sensor fusion, target recognition, millisecond-level control. The dividing line between civilian and military use in these systems lies in the software’s permissions, less in the bill of materials. The fact that delegates from 62 countries are in Beijing for three days discussing risks of misuse while the hardware is sitting on the exhibition table illustrates the situation quite well.

Anthropic Halves Junior Hires, Agents Outnumber Workforce 6:1

Anthropic has published an analysis of its own operational data, which shows a significant shift in hiring behavior and work organization within the company. The study by Nathan Wilmers, Regina Catipon, Maxim Massenkoff, and Peter McCrory is based on internal figures on employment, AI usage, calendar and communication data, as well as 17 employee interviews. Key metric: In August, the ratio of interactive work hours from AI agents to the number of employees was 6 to 1. In parallel, the share of junior positions in technical departments has dropped from 21 to 14 percent since the beginning of 2025, and the junior share of new hires fell from 27 to 9 percent. One-third of ML managers and one-fifth of software managers stated in an internal survey that they are hiring less because of AI. → Anthropic

Synthszr Take: Six agent hours per employee, and next to that is the second figure: half a workweek just to read what the agents deliver. The ratio thus primarily measures how much output is no longer being reviewed by anyone. The fact that people at Anthropic are starting to build visibility and monitoring tools for their own agent pipelines is an admission that the control layer is missing and is being retrofitted by hand. Ironically, the junior roles, whose share of technical new hires has dropped from 27 to 9 percent, were previously the layer that proofread others' results and built up their judgment in the process.

Rokt Reduces 15 Career Ladders to Four, Betting on AI Generalists

The e-commerce technology provider Rokt has reduced its internal career paths from 15 to four because AI tools are blurring the lines between specialized disciplines. The Deep View reports this in a weekend edition produced in partnership with Rokt, based on an interview with Chief Operating Officer Bill Barton. According to Barton, all functions within the company, from legal and finance to marketing, now work with AI tools. This makes it easier for developers to approach non-technical topics, and non-technical employees can build things themselves. According to Barton, Rokt has specifically trained all employees in AI and has also brought non-technicians into hackathons and internal competitions. → The Deep View

Synthszr Take: Eleven eliminated career ladders are, in truth, eleven eliminated promises of advancement, and this hits new professionals harder than any round of layoffs. The intermediate steps that are now being eliminated were the places where one built up judgment through thousandfold repetition: Junior Analyst, Junior Designer, Junior Developer. Barton’s list of tenacity and humility sounds like a recruiting brochure, but it describes a real recruitment problem, because there’s no certificate for attitude and no curriculum yet for directing agents.

Google’s Willow Chip Reduces Quantum Errors 40-Fold

In a paper published in Science on Thursday, Google Quantum AI demonstrated that so-called magic states can be generated on the Willow processor with an error rate approximately 40 times lower than without the new method. Magic states are single-use resources: they are needed to execute Non-Clifford-Gatter, the operations without which a quantum computer cannot run truly useful programs, and they are consumed after use. According to Google, their production consumes up to 95 percent of a machine’s total capacity in some estimates. The team led by Emma Rosenfeld and Craig Gidney reports a logical fidelity of 99.99 percent for the generated T-states.

The established method dates back to 2004, when Sergei Bravyi and Alexei Kitaev described magic state distillation: many noisy states are distilled into a few clean ones, which requires extensive qubit factories that do nothing else. The cultivation method proposed by Google researchers in 2024 works like a filter instead. A coarse state is injected into a small qubit array, undergoes a series of checks, any attempt with a detected error is discarded, and the survivors continue to grow. Computationally, the proposal at the time promised that relevant programs could be run with about one-tenth of the previously expected number of qubits; the current experiment is intended to provide practical proof.

Willow uses superconducting circuits and, according to the publication, achieves error rates on the order of previous experiments with ion-based qubits, but faster according to Rosenfeld, which would reduce the total runtime. Isaac Kim of UC Davis calls the approach a way to make fault-tolerant quantum computing significantly cheaper. Dominik Hangleiter of ETH Zurich sees consequences primarily for medium-sized machines with a few hundred logical qubits. Zhenyu Cai of Imperial College London classifies the experiment as a significant step but points to open questions. The demonstration has so far only worked on a small scale.

In the cryptocurrency scene, the finding is being discussed under the keyword Q-Day. Cultivation is one of the assumptions behind Google’s 2025 estimate that RSA-2048 could be cracked with less than a million noisy qubits. However, breaking encryption would require magic states with error rates close to one in ten billion, several orders of magnitude below the value now measured. → Superintelligence, Google, New Scientist, Decrypt, The Block

Synthszr Take: There are a good two years between the 2024 preprint and the Science paper, which is an astonishingly short path to real hardware for an idea that previously only existed in simulations. The distillation method by Bravyi and Kitaev took significantly longer from 2004 to reach robust demonstrations on chips. The pace has changed: theorists and chip specialists at Google are working on the same artifact, eliminating the detour through external labs, funding applications, and replication rounds. This also explains why the 40-fold error reduction is already being published, even though it was achieved with small qubit arrays and cracking RSA would require an error rate of one in ten billion. When theory and manufacturing are under one roof, the half-life of research advantages shrinks: the next paper proposal from Santa Barbara could be on a wafer as early as 2028.