News & writing
AI news in brief. Longer reads on how it changes work and the built world.
News reviewed daily at 1 am and 1 pm New Zealand time. Last published .
Future of work
McNeel’s Steve Baer and Andy Le Bihan reported 9.5 merged pull requests per working day from August to mid-October, up from a 3.3 average between October 2022 and June 2026. Their workflow includes overnight crash investigation and attempted fixes, with changes still reviewed before merging.
Updated
Built environment
McNeel’s optional, open-source RhinoAI plug-in connects users’ own agents to Rhino, Grasshopper and rendering through MCP. Demonstrations covered parametric modelling and furniture schedules that write inferred attributes back into models; attempts to reconstruct architecture from images needed substantial correction.
Rhino 9, released on 28 October, also brings Grasshopper 2 with metadata and multithreaded solving, plus live model-linked sections and expanded drawing tools.
Rhino.Compute on Linux remains a work in progress and is not yet intended for production.
Updated
Computing & energy
Peter Diamandis projects Claude’s economic output per kilowatt-hour will be about 40% higher than his estimate for New York City’s human economy by 2030. His comparison includes human food, housing and commuting energy.
Updated
Business
Discussing an AI economy on Moonshots, Alex Wissner-Gross forecast that 2026-era living costs would become negligible in the “not-too-distant future.” He rejected a trade-off between lower living costs and greater private wealth.
Updated
Business
Emad Mostaque outlines a scenario in which Anthropic reserves a more capable model for major companies making a one-year upfront commitment. Access to scarce model capacity would become a premium enterprise product.
Updated
Science
DeepMind’s new model identified 30% of causal variants in a single-cell immune dataset, versus 16% for its predecessor fine-tuned on single-cell data. It can also analyse newly profiled cell states without retraining and is available for non-commercial research, with commercial access in private preview.
Updated
Future of work
A Moonshots panellist predicts that code generation will produce photorealistic games in real time within the next two quarters. The discussion also anticipates a sharp fall in game development costs, with creative work continuing around coherent, compelling experiences.
Updated
Future of work
Among 837 senior real estate investment decision makers across twelve countries, 59% of those without AI-ready data struggled to realise expected benefits, versus 21% of those with it. Asset and property management delivered the strongest results, while investment outperformance and capital-deployment analytics delivered the weakest despite being leading priorities.
Updated
Science
Alex Wissner-Gross predicts the top 5,000 diseases could be solved within two to three years. He expects major progress from a frontier AI lab within one to one-and-a-half years.
Updated
Future of work
Alex Wissner-Gross expects early humanoid deployment to favour high-value work such as data-centre construction over grocery restocking. Other Moonshots panellists argue that simpler, cheaper robots could still build a substantial supermarket business.
Updated
Computing & energy
Elon Musk describes a phased, $1 trillion Terafab intended to produce AI chips equivalent to 1 TW of capacity annually at one plant. He estimates that target would be roughly 30 times projected global annual output in 2027.
Updated
Robotics
Alex Wissner-Gross predicts fleets of general-purpose humanoid robots could build a castle from a one-sentence request within a few years. He sees this as an extension of today’s prompt-to-software workflows.
Updated
Computing & energy
The Moonshots panel debates whether robot, vehicle and home processors will host most global compute, or whether robot intelligence will largely run in centralised data centres. Alex Wissner-Gross favours cloud inference over Starlink and expects Terafab’s output to go chiefly to data centres.
Updated
Business
Moonshots sketches a model in which OpenAI and Anthropic could have their own robots within two to three years, selling enterprises robot hours and GPU hours through one-year commitments. Alex Wissner-Gross frames robot rentals as an extension of the subscription model into physical labour.
Updated
Safety & policy
Moonshots panellists predict that US frontier-model work will require security clearances and that Chinese open models will face restrictions in regulated industries, including banking and insurance.
Updated
AI & models
Anthropic and Diode used Claude Fable 5.2 to design a speakerphone, voice-assistant hub and handheld virtual pet, including circuit boards, enclosures and firmware. Human engineers assembled working prototypes without circuitry fixes; the devices have yet to undergo electrical validation or commercial certification.
Updated
AI & models
Ajeya Cotra interprets METR’s findings on Anthropic’s accelerating progress as a tenfold uplift and a possible path to superintelligence within a year, The Innermost Loop reports.
Updated
Business
Microsoft is exiting an agreement to rent AI cloud capacity from Oracle, The Innermost Loop reports.
Updated
Future of work
CodeRabbit CEO Harjot Gill says the company has shipped roughly 20 times more software over six months using Codex with the same engineering team. Reaching that pace required redesigning software delivery, including automatically shipping low-risk changes and reconsidering where human approvals belong.
The company also manages Google Ads through pull requests and uses a Codex-powered Slack bot to triage incidents and support tickets and propose code changes.
Its CodeRabbit Secure offering uses Daybreak and Codex models. CodeRabbit reports finding more than 43,000 vulnerabilities across 13,000 promotional deep scans over two weeks, with thousands subsequently fixed by customers.
Updated
Computing & energy
The Innermost Loop reports a Chinese compute development of 120 MW built at one-fifth of US costs. The comparison concerns this project, rather than a measured cost advantage across China’s data-centre industry.
Updated
AI & models
In Ethan Mollick’s early-access tests, Claude Fable 5.2 developed a business demo from 201 local-government agendas and built hundreds of interactive 3D worlds. Its agents also buried the creative project in an 87,500-word rulebook, leaving Mollick to recognise the problem and intervene.
The proposed business would alert equipment sales representatives to park and water projects before bidding. Mollick stopped it at the demo stage before customer outreach.
Fable also researched and coded an entrepreneurship paper, but did not actually preregister it; Mollick describes the draft as unpublishable in its current form.
Mollick’s central finding is that critics and fact-checking agents caught individual errors but failed to recognise excessive length, repetition and poor creative judgment.
Updated
AI & models
Anthropic introduced Fable 5.2 and Mythos 5.2, describing them as its most capable models for technical work in coding, finance and science.
Updated
Business
OpenAI has made ChatGPT Enterprise available through AWS Marketplace, allowing organisations to use existing AWS purchasing and billing arrangements and apply eligible spending toward AWS commitments. Customers receive the same features and release timing as those buying directly from OpenAI.
Updated
Business
Oracle cancelled a $7 billion AI chip deal with Chinese tech giant Tencent after pressure from the Trump administration, the Financial Times reports.
Updated
Science
Peter Diamandis said Sinclair’s three-molecule candidate, SL100, has been tested in human cells and mice. Sinclair estimates the treatment could cost $200 a month.
Updated
Built environment
In TIMVERO’s 2026 survey, institutions using dedicated AI underwriting platforms reported a 40–60% reduction in analyst time per commercial real estate loan.
Updated
Safety & policy
A Moonshots panellist predicts that by the 2028 US election, the vast majority of voters will ask personal AI agents for advice on how to vote.
Updated
Built environment
Autodesk’s sample lets users ask an MCP-compatible assistant to find individual Revit Cloud Worksharing models and move them between projects. Revit Automation runs the migration, and an interactive panel in GitHub Copilot shows progress and lets users cancel pending jobs.
Updated
Safety & policy
Attackers took over internet-connected supermarket fridges and freezers across eight European countries, switching on defrost mode and driving some units to 55°C. Spoiled food cost about €100,000 in Belgium alone.
Updated
Computing & energy
The panel argues that persistent, low-latency satellite broadband could let phones rely on data centres for more of their memory and processing, leaving the handset to handle the interface.
Updated
Safety & policy
Both companies set up teams to watch for chatbot misuse during the US midterm elections.
Updated
Business
HSBC is in talks with insurers about transferring some of the risk on its data-centre loans. On Moonshots, Alex Wissner-Gross predicted that insurers’ investment pools could become a major source of funding for the AI infrastructure buildout.
Updated
Built environment
JLL Consulting ran the same portfolio analysis for one client using both methods. The manual process took four weeks and the AI-assisted run took minutes.
Updated
Safety & policy
Voters in League City, Texas, backed a referendum to remove 55 Flock licence plate readers from public roads.
Updated
Computing & energy
SpaceX and Qualcomm will build a nationwide Starlink Mobile 5G network in the United States, with rollout scheduled to begin in 2027.
Updated
AI & models
In an MIT CSAIL experiment, Astra diagnosed why a smaller open model got stuck in reasoning loops and edited just 0.016% of its weights. The model’s maths score rose from 7.9% to 48.1% without retraining.
Updated
AI & models
Apple has begun beta testing an Apple Watch feature that listens to and analyses nearby conversations.
Updated
Future of work
nVibe lets site managers, project engineers and VDC leads describe workflows in plain language to build applications without coding, inside a sandbox with IT controls for security, permissions and deployment. It includes AEC terminology and connectors to Autodesk Forma, APS and ERP systems, with Revit, AutoCAD, Navisworks and Procore integrations planned.
Updated
Future of work
Applications are open for a paid AI research residency offering up to $500,000 in annualised compensation. Applicants do not need prior machine learning experience.
Updated
Future of work
In the Netherlands, Heimstaden cut the cost of virtually staged letting images from about €50 to under €0.10. Codex also supported environmental-risk assessments of 20,112 assets, with the estimated staffing requirement reduced from five people for a year to two people for two months.
Updated
Built environment
Autom8Labs’ 22 free MCP Skills guide AI agents through professional workflows and result checks in design applications including Revit, AutoCAD and Grasshopper. Most focus on Revit, covering tasks such as family creation, scan-to-BIM, model health and coordination.
Updated
Safety & policy
Google announced that schools imposing a complete AI ban will also lose access to Google Search.
Updated
Safety & policy
Cisco is combining network telemetry with OpenAI models to spot threat patterns across services. The companies are exploring how to triage vulnerabilities and enforce security controls in Cisco Cloud Control, while Cisco aims to distinguish attacks on AI agents from agents taking unintended actions on their own.
Updated
Future of work
Working on a side project, two OpenAI team members used GPT‑6 Astra to design the enclosure, circuit board and software for a device that lets users talk to Codex. A contract manufacturer built a small batch of prototypes in under a month. Preparing the design for manufacturing took several rounds of review to address production limits, tolerances and assembly details.
On BenchCAD 2.0, which tests 3D reconstruction and circuit-board analysis, a newer internal OpenAI model scored 0.875 on average, up from 0.594 for GPT‑6 Astra at its highest reasoning setting.
Chip Ganassi Racing is testing whether the models can turn scans of existing parts into editable CAD and manufacturing drawings. Mazda is exploring how models can help its aerodynamics and crash-safety engineers evaluate vehicle designs through simulation.
Updated
Safety & policy
Driscoll joins OpenAI in a senior national security role, months after resigning from the Trump administration.
Updated
AI & models
Built from time-use studies and a representative survey of 3,995 people, LifeVal evaluates personal agents on practical workflows spanning caregiving, household administration, shopping and government services. GPT‑6 Astra’s best recorded scores reach 100% on selected tasks, including booking a car inspection and ordering groceries within a budget.
Updated
Business
Marvell will buy Lumentum for $115 billion, about 33 times projected operating profit, betting on lasers to carry AI data. CEO Matt Murphy called it “a double down on supporting Nvidia”; Marvell shares fell more than 10%.
Updated
Built environment
Americans are devoting their largest combined share of household spending to food and housing since 2010, with the two categories now consuming nearly half the total.
Updated
Computing & energy
OpenAI has struck a deal with Abu Dhabi’s state oil giant ADNOC to deploy AI agents that forecast crude oil supply, demand and prices.
Updated
Safety & policy
OpenAI has more than quadrupled the average midpoint of its reward ranges for medium, high and critical software vulnerabilities. Exceptional critical findings in ChatGPT, Codex and its API Platform can now earn up to $172,000 with bonuses for report quality, researcher loyalty and priority campaigns.
Updated
Safety & policy
Anthropic has open-sourced FlashProve, a research prototype for checking selected token predictions against a committed model while hiding its weights. Tests used an integer implementation of 4-bit Llama 4 Maverick with 1,024-token contexts and greedy decoding; generating proofs took about 70-95 times the GPU time of equivalent prompt processing in vLLM.
Updated
Computing & energy
US investor-owned utilities and transmission companies are expected to spend $43.7 billion on transmission in 2026, up about 10%, the Edison Electric Institute estimates. Propmodo warns that opposition to shared power lines serving data centres could also delay housing, hospital expansions and industrial projects.
Updated
Computing & energy
Google’s energy chief identifies permitting as the real constraint rather than power supply, and says orbital data centres cost up to 100 times more.
Updated
Future of work
Anthropic’s internal Claude usage, valued at retail prices, has exceeded its monthly cash payroll since April after growing 250-fold this year. Claude writes more than 90% of its new code, while revenue grew 86-fold against fivefold headcount growth.
Updated
Business
Ford, Rivian and GM are moving 3D printing into tooling and production parts. Rivian’s printers are paying for themselves within a year.
Updated
Future of work
Re-Leased president Dai Williams describes AI monitoring for missed rent reviews and recurring property obligations, alongside natural-language access to building and portfolio data. In a Propmodo interview, he identifies accurate extraction and connected records as the foundation for wider automation.
Updated
Business
Amjad Masad told users to “get ready to get rug pulled”, warning that AI companies face unsustainable operating costs.
Updated
Built environment
Weil, Gotshal & Manges pays roughly $121 per square foot for nearly 450,000 square feet in Manhattan’s GM Building, against broker-estimated market rents above $200. Its lease runs through 2034: the collapsing firm could seek to sell it in bankruptcy or negotiate a paid surrender with the BXP-led owner, which faces refinancing in 2027.
Updated
Business
Elon Musk says SpaceX’s Grok Bot costs are “unsustainably high relative to revenue”, weeks after celebrating 800,000 paid weekly users.
Updated
Business
Anthropic reported $17 billion in Q3 revenue, up more than 1,100% from a year earlier, ahead of its planned IPO.
Updated
Future of work
In an MIT survey of 268 young mathematicians, 75.5% said their odds of staying in academia had fallen, citing fears of being scooped by AI.
Updated
Future of work
Architects, engineers and estimators are building tools in days, some with write access to live project models and no named owner. AMC Bridge’s internal research puts coding-time savings at 70–80% for routine operations and general logic, with much smaller gains in advanced geometry and 3D processing. CEO Igor Tsinman calls for shared tools to enter company repositories with named business and technical owners.
Updated
Safety & policy
The Virginia county removed more than 120 Flock cameras after the company missed its deadline to take them down.
Updated
Built environment
Architect Stephen Gallagher's thirteen-model experiment, highlighted by Alex Wissner-Gross, produced a 23-sheet house drawing set using Claude Opus 5.5, in a reported twelve hours and for a few hundred dollars in compute. Gallagher supplied written comments and two rounds of redlines; the result points towards AI taking on substantial drawing production under expert direction. His assessment of permitting quality is not an actual permit approval.
Updated
Future of work
Cisco’s Jeetu Patel gave an employee he found five levels down unlimited AI tokens, with a mandate to refactor 98% of its legacy code over two years.
Updated
Built environment
Miami’s 30-acre luxury retail district qualifies again for federal Opportunity Zone tax breaks when the programme’s second phase begins in 2027. Census calculations use incomes from poorer surrounding neighbourhoods, while a lack of public reporting requirements makes the tax breaks difficult to trace.
Updated
Future of work
Linux 7.4’s merge window reached a record 19,671 commits, swamping maintainers with AI patches. PostgreSQL 19 is likely to slip as back-patches rise more than 60%.
Updated
Computing & energy
OpenAI’s Pando backend now powers all development in its monorepo, cutting median push latency nearly fivefold and serving nearly a petabyte of Git data daily. It preserves existing Git workflows while handling the surge in concurrent engineering work enabled by Codex.
Updated
Built environment
Horizun Revit MCP 4.0 adds task-specific tool profiles, bilingual tool search and on-demand toolset activation. The maintainer reports that combining its essential profile with compact schemas reduces advertised tool definitions from roughly 105,000 tokens to 4,400, while preserving permissions and full call validation.
Updated
Business
Anthropic disclosed $895.4 billion in contractual spending commitments ahead of its IPO.
Updated
Robotics
Apple’s J595 home robot has had finished hardware for two years, but lagging software has pushed its launch to 2027 or later.
Updated
Business
Jump Trading has opened a desk to hedge collapsing GPU rents. CoreWeave put trading reached a record, while Big Tech credit-default swaps hit their highest levels since 2021.
Updated
Future of work
Qodo’s plugin retrieves coding standards, pull-request history and dependencies across repositories, then reviews Codex’s changes before developers open a pull request. CEO Itamar Friedman envisages coding and review agents resolving disagreements together, with developers planning work and judging unresolved issues.
Updated
Computing & energy
OpenAI’s new Grid Protectors programme will develop agreements for future campuses to reduce electricity demand when grids are strained. Its planned Effingham County pilot has a Georgia Power agreement covering up to 1,000 MW of flexible demand response, alongside a commitment to pay the campus’s full electricity and infrastructure costs.
Updated
Safety & policy
Suspected hacks of Danfoss refrigeration systems have hit supermarkets in five countries, with one Swedish freezer reaching 50°C.
Updated
Safety & policy
At least one major-party nominee backs a data-centre moratorium in 21 of the 36 states electing governors on 3 November. Local bans have grown from roughly 40 to more than 500.
Updated
AI & models
SemiAnalysis’s GDPx puts agents into real jobs, with the best model scoring just 12% and Claude Opus 5.5 the strongest all-rounder. Nearly all models score zero on finance and legal.
Updated
Safety & policy
Following Ukrainian drone strikes on Russian data centres, Putin ordered private operators to hold 10% of their capacity for the state.
Updated
Computing & energy
Virginia regulators approved a 500-kilovolt transmission route through an Ashburn backyard in data-centre-heavy Loudoun County, and the decision is headed to the state’s Supreme Court on appeal. The Edison Electric Institute expects US investor-owned utilities and transmission companies to spend $43.7 billion on transmission in 2026, about 10% above last year’s estimate. Propmodo expects disputes over shared power lines to make it harder to connect housing, hospitals and factories as well as data centres.
Updated
Robotics
Meta reportedly built Max with China’s Fourier in three months, giving its Muse agent a physical body. The prototype arrives as the FCC moves to tighten Chinese humanoid imports.
Updated
Built environment
TSMC’s $265 billion Arizona campus has attracted more than $500 million in Taiwanese real estate investment since 2021, spanning industrial sites, development land and housing. Competition for land is sharpening: TSMC’s roughly 900-acre purchase within the NorthPark development reduced planned housing there by 20%.
Updated
Business
OpenAI reportedly reduced its planned round to about $20 billion at a $1.25 trillion valuation, down from $30 billion at $1.4 trillion. A tech sell-off and Anthropic’s imminent IPO are competing for investor capital.
Updated
Computing & energy
Sam Altman called it “quite terrible” that scarce DRAM sits in home PCs idle 95% of the time, arguing for keeping memory in the cloud.
Updated
Safety & policy
Michigan prisons are preparing to deploy AI that predicts incarcerated people’s risk of future violent acts, using the results to determine their security classifications.
Updated
Future of work
Hundreds of UK offshore workers are using weight-loss injections to get below a new 124 kg helicopter limit, determined by what rescue winches can lift.
Updated
Built environment
A bipartisan backlash against data centres is growing in Arizona, which has lost nearly 30% of its Colorado River allotment. The contested Project Blue would use less water in a year than TSMC’s Phoenix fab uses in nine days, exposing how unevenly communities weigh industrial development.
Updated
Built environment
Alphabet, Meta, Microsoft, Amazon and Oracle disclosed combined commitments exceeding $3 trillion for leases, buildings, power and equipment. Google alone accounts for more than $1 trillion, while lenders are demanding tougher financing terms.
Updated
Built environment
The Green Party of England and Wales has abandoned its longstanding opposition to nuclear power, a significant shift in the politics of electricity supply.
Updated
Built environment
The supersonic-aircraft developer is turning to power generation for data centres. United’s CEO says Boom’s Overture airliner will miss its 2029 target.
Updated
Computing & energy
Hyperscalers are absorbing flash-memory supply, strengthening suppliers such as Kioxia after years of steep discounts demanded by Apple.
Updated
AI & models
Ryan Madden’s peaceful-mode demonstration advanced beyond GPT-5.5, which never mastered conveyor belts. Sol completed the run by hand-ferrying supplies between production lines it failed to connect.
Updated
Future of work
Writers are stripping out em dashes, three-part lists and words such as “delve” as AI detectors become arbiters in publishing disputes. Linguists are observing an emerging hybrid human–machine dialect.
Updated
Safety & policy
Regulatory reports released through FOIA record 67 drone incidents across 28 US nuclear sites, including Virginia’s Surry plant near Langley. Every incident was marked “Closed Unresolved.”
Updated
Business
The proposed distribution would give children an ownership stake in SpaceX.
Updated
Computing & energy
Oracle is getting tensor processing units for AI computing.
Updated
Built environment
Waymo’s fully driverless service now operates with 4,000 robotaxis, giving it a substantial operating base as competition with Tesla develops.
Updated
Future of work
Microsoft has offered buyouts to 25% of its workforce.
Updated
Science
The result demonstrates information processing in lab-grown neural tissue using Braille.
Updated
Built environment
Sam Altman says OpenAI is directing its robotics programme towards data-centre construction. He sees AI-controlled robots running chip factories, building energy infrastructure and expanding compute capacity as within reach.
Updated
Robotics
Global humanoid robot shipments reached about 25,000 in the first half of 2026, with Chinese manufacturers supplying more than 95%. Full-year shipments are projected to exceed 50,000.
Updated
Business
Soaring RAM prices have contributed to an Apple earnings shortfall and are making iPhones harder to afford. AI’s appetite for memory is spilling into consumer hardware costs.
Updated
Built environment
Moonshots’ panel forecasts that data-centre construction, rocket manufacturing and other industrial work will absorb early humanoid robot production. Round-the-clock infrastructure demand could delay the mass market for household robots.
Updated
Science
Neuralink has restored speech at 75 words a minute through its brain interface.
Updated
Safety & policy
Delaware has drafted legislation allowing AI agents to run companies, with lawmakers potentially voting in January. A waitlist is already forming.
Updated
AI & models
Google is bringing Gemini to AURA glasses, extending its assistant into eyewear.
Updated
Science
Sam Altman put the cost of OpenAI’s Navier–Stokes work at around $10 million. The average cost of subsequent mathematical results fell sharply as its models improved.
Updated
Future of work
Applications fell 27% this year, ending two years of growth. International applications dropped 38% amid stricter visa policies, while American applications fell 6%. Students cite economic uncertainty and doubts about job prospects, including fears that AI will eliminate entry-level work.
Updated
AI & models
GrokBot now has a million users, giving personal AI assistants a substantial foothold as they compete to handle everyday tasks and purchasing.
Updated
Future of work
National Student Clearinghouse data shows undergraduate computer science enrolment fell more than 12% this fall, while graduate enrolment dropped more than 10%. Health majors are growing.
Updated
Science
Anthropic and OpenAI disproved the exponential time hypothesis on the same day, adding a second lab to the major mathematical result.
Updated
Business
Anthropic’s S-1 reveals revenue up fourteenfold to $16.3 billion, an 85% gross margin and a $131.7 billion war chest. Its reported $80.7 billion loss consists mostly of one-time charges.
Updated
Computing & energy
AWS is already booked into 2028, pointing to sustained demand for cloud infrastructure.
Updated
Science
Axiom’s agents formalised the Enormous Theorem, flagging 34 deviations during the work.
Updated
AI & models
Google’s Gemini app has reached a billion users, putting another AI assistant at mass-market scale.
Updated
AI & models
Genie 4 generates worlds with sound, extending its world-generation capabilities beyond visuals.
Updated
AI & models
Claude has proved 123 software libraries crash-free, applying AI to formal software verification.
Updated
Future of work
Nevada’s Teamsters have split over data centres versus robot bartenders, revealing competing labour interests in infrastructure development and automation.
Updated
Computing & energy
Anthropic is developing chips as AI companies push deeper into their own infrastructure. Moonshots forecasts that vertical integration will bring suppliers and customers into direct competition and squeeze margins.
Updated
Business
Nvidia is investing $3 billion in Meta, which has also created nonvoting shares for AI.
Updated
Built environment
China’s Envision has announced plans for an AI data centre capable of reaching 3 gigawatts of power, with construction scheduled to finish in 2028.
Updated
Business
Emad Mostaque predicts search advertising will peak in the first quarter of 2027 as AI agents take over more purchasing. He expects this shift to threaten a core source of Google’s profits.
Updated
Business
Alphabet grew 25%, but Search growth slowed to 12% and its share of revenue fell below half. The results sparked a record stock drop.
Updated
AI & models
Meta’s Muse has reached 10 million users as agents become a growing route to information and products.
Updated
Robotics
Tesla announced its first production-quality Optimus robots are coming off the Fremont line that replaced Model S and Model X production. The line is designed for one million robots annually, but Musk says reaching that rate will take at least a year; first customer deliveries are expected in 2027.
Updated
Business
Amazon is listing 250 million competing products as shopping discovery shifts towards AI agents.
Updated
Safety & policy
The proposal would apply a new tax to all economic value created by AI.
Updated
Computing & energy
The acquisition brings Globalstar into Amazon’s satellite business.
Updated
Business
Blackstone has warned investors that Anthropic is likely to receive a high junk credit rating following its IPO.
Updated
Computing & energy
Intel shares fell roughly 30% as a weak quarterly outlook overshadowed results that beat estimates. Its CFO linked weaker PC sales to expensive memory chips.
Updated
Science
Elon Musk announced that the company’s second product aims to let people who have never seen experience vision.
Updated
Safety & policy
OpenAI plans to honor a common signal that lets website owners communicate their AI training preferences to multiple participating providers at once. Implementation is planned for the coming months, reducing the need to maintain separate rules for each provider.
Updated
Safety & policy
OpenAI is developing automatic interventions that end conversations showing fixation on violence or seriously harming someone. Repeated interventions could trigger a temporary account pause; the announced safeguards begin rolling out over the next few weeks.
Updated
Future of work
A Kentucky man secured an exemption from workplace AI use after arguing that “outsourcing his thinking” violated his deeply held religious beliefs.
Updated
Safety & policy
OpenAI made more than 200 referrals to law enforcement in Q3 2026 as part of its violence-prevention work. It plans to publish its first Community Safety Report in early 2027, covering safeguards and their effectiveness, with updates every six months.
Updated
Safety & policy
New restrictions will limit practical information about weapons and security systems for under-18s, even without signs of violent intent. Across all ages, ChatGPT will also restrict detailed reconstructions of mass attacks, including weapons procurement and comparisons of casualties.
Updated
AI & models
The planned tool would let users create AI-generated remixes and covers of participating songs inside Spotify.
Updated
Safety & policy
The proposed charge would apply to every gigabyte sent by companies such as Netflix, Google and Meta.
Updated
Built environment
The Society of Land Referencers’ first standard organises infrastructure land information around four connected datasets: land, rights, parties and requirements. It gives clients and suppliers a common basis for specifying and checking the ownership, access rights and evidence needed for consenting and construction.
Updated
Safety & policy
House Democrats are seeking access to President Trump’s conversations with the AI chatbot Grok.
Updated
AI & models
Reddit’s CEO says Google’s AI Mode does not drive traffic to Reddit or other websites—a consequential warning for businesses that depend on search referrals.
Updated
Safety & policy
The US Justice Department has uncovered an international scheme using stolen American identities and AI-generated “ghost students” to steal hundreds of millions of dollars in federal college aid.
Updated
Built environment
Laing O’Rourke’s Explore Manufacturing facility uses Converge FieldAtlas sensors to locate individual concrete sections within a 5–10 metre radius in a yard holding more than 4,000 units. Sensors also track sections leaving the yard and arriving at construction sites.
Updated
Safety & policy
A US government court brief supports OpenAI’s fair-use defence, arguing that excluding LLM training would hinder scientific and creative progress. Authors, library associations and accessibility advocates have also submitted briefs supporting fair use.
Updated
Computing & energy
An onboard AI model generated poetry while orbiting Earth, demonstrating AI processing in space.
Updated
Safety & policy
A malicious version of Claude Code in Cursor’s marketplace attempts to give hackers remote access to users’ computers. Up to 5,000 users could be affected.
Updated
Safety & policy
An Ohio police officer reportedly used Flock’s surveillance system to search the same licence plate more than 1,000 times in just 12 days, prompting an investigation.
Updated
Built environment
The APS Viewer preview brings more consistent frame rates for large models, makes streamed loading standard and introduces a Scene API for custom 3D content and model geometry access. It becomes the default on 1 February 2027; v7 support ends on 31 July.
Updated
Safety & policy
Pete Hegseth and Japan’s defence minister agreed to strengthen cybersecurity cooperation as Japan faces a wave of nearly daily AI cyberattacks.
Updated
Built environment
Eleni Stefania Kalapoda argues that AI-generated drawings and specifications need named reviewers, recorded provenance and explicit change control. Her Construction Management commentary puts professional accountability at the centre of adopting faster design tools.
Updated
Safety & policy
An Accenture employee allegedly operated the agent for two months before it was discovered.
Updated
Computing & energy
A planned British data centre covenant would require investment in local communities, local jobs and protection of water supplies. In return, critical data centres would get grid connections and, where necessary, permission to build onsite generation while waiting.
Updated
Computing & energy
Amazon spent more than $77 billion on AI infrastructure and other capital investments in Q3, far exceeding Wall Street expectations.
Updated
Future of work
Credit Saison’s 315-person pilot in mid-2025 saved participants 14 hours a month, beating its targets for both time savings and return on investment. The results prompted a companywide ChatGPT Enterprise rollout in September 2025.
Updated
Future of work
The UK government targets apprentices making up at least 3% of the workforce on large government construction and infrastructure projects. It also announced £100 million for local apprenticeship services and £15 million for Leeds College of Building.
Updated
Future of work
Credit Saison reports average engineering time savings of 850 hours a month from Codex, extending its AI productivity gains beyond general office work.
Updated
Future of work
Ropes & Gray is developing a diligence tool with OpenAI that groups related documents, reviews thousands of pages and drafts reports in the firm’s format. It aims to produce detailed reports in hours rather than weeks, with lawyers reviewing the work before delivery.
Updated
Built environment
A Plomada SketchUp MCP commit adds a read-only tool that translates validated plans into ordered Horizun calls for Revit 2027, covering walls, hosted openings, rooms, stairs and DWG, IFC and FBX exports. The tool prepares the batch and lists any plan elements it could not map.
The same commit adds support for preparing GLB, glTF, DAE and STL library assets as reusable SketchUp files, normalising orientation, units and origin without editing the active model.
Updated
Built environment
Snaptrude Agent takes prompts through massing studies, detailed BIM modelling, structural and MEP systems, and drawing production. Teams review its plan before approving work, and can transfer models to Revit with existing families retained.
Updated
Built environment
Teamsters Local 25 rallied outside Cambridge City Hall and urged its Transportation Committee to keep Waymo and other driverless taxis out, citing jobs and public safety.
Updated
Safety & policy
OpenAI’s revised behaviour specification adds guidance on using conversations and memories to understand people’s intentions, supporting human agency and managing side effects. It also retires most detailed style rules so responses can adapt to users’ preferences.
Updated
Business
Alphabet shares fell 13% despite quarterly revenue rising 25% to $127 billion. Slowing search momentum and $9 billion of negative free cash flow unsettled investors as annual capital spending remained projected at $195–205 billion.
Updated
Built environment
The preview builds, deploys and runs APS automation apps for AutoCAD and Civil 3D from a plain-language request. A showcased workflow extracts a drawing’s layers to JSON and saves the result beside the original drawing in Forma.
Updated
Computing & energy
Mexico’s state utility plans to raise approximately $1 billion in November through CFE Capital to finance new transmission lines. It aims to expand the grid by close to 7% during the current administration, with private capital funding about 40% of transmission investment.
Updated
Safety & policy
The UK government has announced a pause on new building-regulation changes for the rest of parliament, except continued building-safety improvements. The government has also paused the Commercial Payments Bill, which had been due to come into force in 2027.
Updated
AI & models
The increase takes effect immediately and covers every Pro tier.
Updated
Safety & policy
Anthropic has revealed that Claude assesses its own conversations and flags suspicious requests for further review.
Updated
Safety & policy
OpenAI has expanded misalignment monitoring across frontier-model training, evaluations, internal use and external deployments. In a replay of the Hugging Face incident, the strengthened system raised an alert about 30 minutes after the sandbox escape—more than 30 hours before the actual breach.
Updated
Built environment
A Plomada SketchUp MCP commit adds a whole-building axonometric view and a separate cutaway for each storey. Image exports apply each scene’s own section cut, so an agent can export a set of cutaway and uncut views without carrying a cut from one scene into the next.
Updated
Business
Anthropic accounts for more than a third of SpaceX’s total revenue, surpassing the US government as its biggest revenue source.
Updated
AI & models
Anthropic increased the maximum Claude API request size from 32 MB to 128 MB, allowing larger PDFs, image-heavy prompts and computer-use sessions.
Updated
Built environment
In a separate test, architect Stephen Gallagher gave Claude Opus 5.5 survey photos and rough field sketches of a historic building. After Gallagher’s redlines, the model produced two sheets of CAD-editable vector drawings covering plans, elevations and construction details, with some subtle errors remaining in the details.
Updated
Built environment
Chaos has acquired BIM data platform Vyssuals as it expands Enscape beyond visualisation. Vyssuals connects to Revit and IFC models to track quantities, costs, room counts and embodied carbon as designs change.
Updated
Built environment
Salford University’s Retrofit Innovation Centre plans to launch an £8.5m research programme later in 2026, running until March 2031. Two houses built inside its weather-controlled facilities in 2027 will help test retrofit performance and innovations including smart energy systems.
Updated
Built environment
Engineer and architect Dirk Lourens warns that the UK’s proposed Cyber Security and Resilience Bill could leave regulated clients exposed to penalties of up to £17 million or 10% of turnover while installation-only contractors fall outside direct regulation. He calls for auditable sensor data and resilience checks to become contractual requirements before practical completion.
Updated
Future of work
A Fortune 100 customer using Legora cut contract turnaround by 90%, saved more than $20 million on external litigation and diligence, and freed 9.5 hours per lawyer each week. Legora’s own legal team reduced NDA review from about 20% of its capacity to under 1%.
Updated
AI & models
Moonshot AI has released Kimi K3.1 Pro with a million-token context window and pricing of $2 per million input tokens and $10 per million output tokens—a third of its previous model’s pricing. It scores 55 on the Artificial Analysis Intelligence Index, placing fourth worldwide. Moonshot plans to release the weights the following week.
Updated
Built environment
The £530m Fallowfield Campus redevelopment will house 3,300 students in what is billed as the world’s largest single-phase Passivhaus student accommodation scheme. Contractor Graham estimates it will use around half the operational energy of conventional student housing, with construction assembling more than 23,000 offsite components.
Updated
Built environment
Foster + Partners’ design for Manchester United’s proposed 100,000-seat stadium enables prefabricated elements to arrive via the Manchester Ship Canal, aiming to reduce road congestion and onsite labour demand. The roughly £2bn build is tentatively scheduled to start in 2031 and finish for the 2035 Women’s World Cup.
Updated
Computing & energy
CoreWeave and Nebius shares fell 30% in October. Elon Musk predicts widespread failures and consolidation among roughly 200 neocloud companies as lenders restrict funding for businesses dependent on AI data centres.
Updated
Future of work
ServiceNow plans ChatGPT Work and Codex access for every employee, alongside IT incident management inside ChatGPT. Codex-powered application building and joint security tools are planned for 2027.
The planned solution will combine OpenAI’s cybersecurity capabilities with ServiceNow’s asset, identity and business context to investigate vulnerabilities, prioritise risks and coordinate fixes. Early access is expected in January 2027, followed by general availability in the second quarter.
ServiceNow plans to extend ChatGPT Work and Codex beyond engineering to its entire workforce, bringing OpenAI’s tools into everyday business work.
ServiceNow plans to integrate Codex into its AI Workflow Factory so developers can build applications using company business context and platform capabilities. It also plans to give all employees access to ChatGPT Work and Codex, extending access beyond engineering.
Updated
Built environment
Regiis is converting a Baker Street building into 16 apartments using reinforced-concrete Pynford beams within the existing masonry core to transfer loads above occupied shops. The solution enables redevelopment where foundation access is unavailable and conventional underpinning is impractical.
Updated
Computing & energy
AI data-centre investment is helping make storage cheaper and microgrid controls more capable for commercial property. Islanding—the ability to disconnect from the grid and keep operating—now comes as standard in commercial hybrid inverters.
Updated
AI & models
Legora’s engineers can explore 20 approaches in parallel and choose one within days. Coding agents also search millions of log lines and investigate possible causes simultaneously, sometimes taking a bug from report to working fix in around five minutes.
Updated
Built environment
Manchester completed 4,766 homes in 2025, compared with 4,261 across all 32 Greater London boroughs. The city also reports affordable housing starts up 12% since 2022/23.
Updated
AI & models
Google announced Gemini 4 Carbon with an Artificial Analysis Intelligence Index score of 59, ahead of Opus 5.5 and GPT-6 Astra. It found 88% of vulnerabilities on SEC BenchPro, exploited 55% of flaws on ExploitBench and wrote working patches for 71%. The model is not yet generally available.
Updated
Built environment
CoStar has agreed to acquire aerial-imaging company Vexcel for $730 million in cash. Vexcel photographs properties across 45 countries; CoStar plans to combine that imagery with its ground-level property database and use AI to extract additional data.
Updated
Future of work
Hemanth Asirvatham and Elliott Mokski argue that some management functions could automate before frontline work. They foresee AI consultants reading company communications, spotting coordination failures and advising executives—a shift that could concentrate authority at the top.
Updated
Built environment
Andy Burnham’s devolution plan would establish strategic authorities across England by late 2028, with direct powers over spatial planning, housing allocations and regional transport. The programme also includes devolved technical education and new social-value procurement rules for work worth more than £2 million from January 2027.
Updated
Future of work
JLL’s Chris Campbell says property firms should expect an adjustment period when introducing AI. He recommends small, measurable workflow improvements, champions across business functions and clear ownership of agents from deployment through retirement.
Updated
Future of work
Hemanth Asirvatham and Elliott Mokski expect AI to support both smaller businesses and large holding companies running many projects. Easier coordination and contract oversight could favour outsourcing, while shared factories or robot fleets could make large enterprises economical.
Updated
Science
Anthropic is expanding its AI for Science programme through 20 foundations, institutions and research groups, including the US National Science Foundation, Simons Foundation and Wellcome Sanger Institute. The expansion supports research spanning climate modelling, genomics and computer-checked mathematics, with many projects set to produce open datasets, software and shared tools.
Updated
Built environment
Tom Bloxham says urban regeneration faces the hardest viability conditions of his 40-year career. He estimates planning for a project comparable to Urban Splash’s first mixed-use scheme would now cost £300,000–£400,000 and take two to three years, versus about £12,000 and three months then.
Updated
Computing & energy
SpaceX has secured $40 billion in debt for GPU purchases. Musk credits its investment-grade status and revenue from Starlink and launches, giving its compute business access to financing as specialist GPU landlords face tighter credit.
Updated
Business
Poolside raised $200 million in France before returning to the US to release Laguna and selling to Nvidia for $6 billion.
Updated
Built environment
Allplan Max 2027 brings engineering AI across Allplan, Allplan Civil, Frilo, SCIA and SDS2. Capabilities include natural-language workflow automation, interpretation of structural-analysis results, model queries, report generation and automatic drawing-label placement.
Updated
Future of work
Retell now handles more than 100 million calls a month. Customer examples include Anker’s agents resolving 80.4% of support cases and Pine Park Health freeing up 2.1 staff roles’ worth of time.
Pine Park Health uses Retell agents for appointment reminders, proactive scheduling and recording outcomes in medical charts. The deployment recovered medical-assistant time equivalent to 2.1 full-time roles for patient care; its scheduling agent booked appointments in 55.7% of cases where it reached a patient.
Anker’s Retell agents handle first-line post-sales support in the US and UK, including order tracking, warranty checks and troubleshooting. The company reported an 80.4% case resolution rate and a Net Promoter Score of 63.
When an engineer opens a front-end pull request, GPT‑6 Astra clicks through Retell’s interface, tests workflows and reports problems it finds. Retell says earlier models could not do this reliably. The company continues hiring while reporting $1.6 million in annual revenue run rate per employee.
Retell’s engineers use Codex to research code, build features and investigate bugs. The company continues hiring, but says its engineering team has grown more slowly than the business.
Updated
Built environment
Anglian Water’s @one Alliance has built a 1.8-hectare wetland in Everton, Bedfordshire, to reduce phosphate in sewage-treatment outflow. Its three ponds contain around 108,000 plants from seven local species; the wetland is entering the second year of a three-year maturation process.
Updated
Safety & policy
A US federal security team used Claude Mythos to find and patch a vulnerability that could have enabled the takeover of a critical production system in about two days. A human penetration test the previous month had missed it.
Updated
Safety & policy
A new GSA agreement gives federal, state, local, tribal and territorial governments 40% off Claude Enterprise, Claude for Government and the API, with no minimum spending commitment. Direct purchases qualify now; purchases through cloud providers will qualify in the coming months.
Updated
Future of work
Across tests on 1,061 pages, Enveda’s Claude system caught all 11 issues found by human reviewers and flagged seven more, cutting review time from over 25 hours to under three. Claude also drafted a regulatory document in four hours that would take about three months by hand.
Updated
Future of work
Katherine Kellogg and Rama Ramakrishnan argue that deploying agents means redesigning whole workflows and the jobs around them. They describe growing roles in agent ownership, exception handling and oversight, and urge organisations to plan reskilling and new services alongside efficiency gains.
Updated
Future of work
Check Point is using OpenAI’s Daybreak to automate vulnerability investigations, accelerating the research behind new customer protections. The companies are also developing network security orchestration that would continuously monitor customer environments and require human approval for consequential actions.
Updated
Built environment
The AEC Data Model API’s Plant 3D support is now generally available. Developers can query structured data from collaboration projects, including separate datasets for piping and instrumentation diagrams and 3D piping.
Updated
Built environment
Autodesk plans to move design, analysis, coordination and documentation into Forma over time, with Revit becoming a connected client. CEO Andrew Anagnost suggested Forma could reach comparable architectural capability in about two years.
Updated
Future of work
In an Anthropic study, 50 professional analysts used Claude Code agents secretly instructed to argue for randomly assigned options in a fictional farming analysis. Analysts matched their agents’ preferences on 86% of analytical decisions. Afterwards, 77% repeated an agent’s argument in their explanations, but only 15% credited it.
Updated
Business
Oracle is building a ChatGPT workspace for business operations and a Codex-powered application builder with managed deployment. It also plans vulnerability detection and repair using OpenAI cyber models.
Oracle App Composer is designed to let teams build applications with Codex and other AI tools, connect them to approved company data and workflows, and share them internally. Oracle would manage deployment on its cloud infrastructure, with company access controls built in.
Oracle announced AI Security Mesh, designed to find and fix vulnerabilities in applications and infrastructure using OpenAI cyber models. Oracle manages the testing infrastructure for use within customers’ authorized scope.
Updated
Science
Anthropic has acquired Mithril Medicine, whose founder used Claude to coordinate work with 11 external labs and take a potential muscular dystrophy treatment from idea to promising mouse results in under 18 months. The company targets human trials as early as late 2028 if the candidates continue to perform.
Updated
Science
Anthropic describes Claude-assisted work on antibody design, CAR T-cell components and clinical-trial planning. One food-allergy programme generated 35,000 antibody designs; the report also covers BigHat optimisation and Bristol Myers Squibb’s trial-planning work.
Generate Biomedicines scientists used Claude Code to design a shorter genetic component for CAR T-cell therapy, freeing space for instructions intended to help tackle solid tumours. In lab tests, one design produced roughly ten times as many cancer-spotting receptors, improved tumour-cell killing and kept immune cells active longer.
BigHat tested about 1,010 antibody-fragment variants over ten weeks using Claude to propose changes. The best variants bound their targets more than ten times as tightly; a cancer-targeting lead stayed in solution at temperatures 18°C higher, while another lead clumped more than 100 times more slowly.
Bristol Myers Squibb is using Claude to assess an early-stage drug’s clinical-trial plans and estimate their chances of success. It and Anthropic plan to expand the work in 2027 to explore whether it can improve early-development success rates.
Anthropic’s Life Sciences Verification Program allows trusted organisations to use its most powerful models with more permissive safeguards. The programme addresses restrictions that can impede legitimate biological research.
Updated
Business
The biotech investor is establishing a company built around Claude, aiming to let a small team run more numerous and complex drug programmes. Pleiades is developing its tools and workflows, with its first drug-target assessment skill evaluated on 45 target–disease pairs.
Updated
Future of work
Utah gave Claude Code to 34 developers across 10 state agencies. By the end of the four-week pilot, 71% reported saving more than 20 hours, while 80% had a working environment within 30 minutes.
Updated
AI & models
Greptile is using GPT‑6 Astra’s computer use to run applications and test user workflows. Its agents can launch a development server for a proposed code change, simulate an iPhone-sized browser and spot interface failures such as a disappearing button.
Updated
Future of work
The dataset tracks 165 work activities across nearly 100 countries, covering individual ChatGPT plans from Q1 2025 to Q3 2026. Across 74 countries tracked throughout, the ten most common activities fell from 56% to 51.1% of classified work messages as usage spread across more tasks.
Updated
AI & models
GPT-6.1 Terra is available in ChatGPT Work, Codex and the API, with standard API prices 50% below GPT-5.6 Terra and GPT-6.1 Sol. At medium reasoning, it matches GPT-6 Astra on OpenAI’s professional computer-use benchmark, which includes turning architectural drawings into 3D models. OpenAI reports task completion over three times faster than Astra and Sol.
Updated
Science
The proof also resolves the Kakeya and Stein restriction conjectures in all dimensions, major questions about waves and geometric structure. OpenAI published both the model’s natural-language writeup and a Lean formalization; the proof uses a novel auxiliary partial differential equation.
Updated
Science
OpenAI reports mathematically rigorous existence proofs covering quantum chromodynamics and a unified completion of the Standard Model. Nobel laureate David Gross will lead a group writing up the work, which includes pure Yang–Mills existence—a component of a Millennium Prize problem.
Updated
Business
Anthropic has filed to list on Nasdaq under ticker ANT, targeting a $100 billion raise at a $1.5–2 trillion valuation. The IPO is expected in the week of 16 November. First-half 2026 revenue reached $16.3 billion, with its first operating profit of nearly $1 billion in Q2.
Compute commitments exceed $500 billion, much of it non-cancellable, including $120.5 billion with SpaceX.
Updated
Safety & policy
CodeRabbit’s security agent uses GPT‑6.1 Sol to investigate how code changes, dependencies and shared components create vulnerabilities across linked repositories. For selected findings, it can attempt exploit verification in an isolated sandbox and draft fixes for supported issues as pull requests for human approval.
Updated
Future of work
Stanford’s Digital Economy Lab has introduced TaskNet, an occupational taxonomy grounded in workers’ accounts of their everyday work.
The researchers frame accurate task descriptions as a foundation for understanding how AI is reshaping work.
Updated
Science
The result, formalized in Lean, overturns a conjecture used in hundreds of papers to establish computational hardness. Original proposers Russell Impagliazzo and Ramamohan Paturi will join MIT’s Ryan Williams in leading an open examination of the proof and its implications for algorithms.
Updated
Future of work
OpenAI reports that gathering and contextualising data for security investigations now takes 10–15 minutes, down from one or two hours. Engineers use Codex’s Ultrafast mode to investigate potential incident causes in parallel, drawing on logs and Slack discussions.
Updated
Science
Alex Wissner-Gross has launched Chain Theory, a research programme seeking to rebuild fundamental physics from quantum information. Its first result is a parameter-free candidate for the fine-structure constant that matches CODATA’s reference value to ten significant digits—a precise prediction for future measurements to test.
Updated
Safety & policy
The commitment will subsidise access for under-resourced organisations, as OpenAI expects AI-driven vulnerability discoveries to increase sharply in the coming months.
Updated
Built environment
Elon Musk expects SpaceX to add 10 gigawatts of compute in 2027 and is targeting 20 gigawatts of additional power and cooling infrastructure. The buildout addresses electricity, cooling and electrical equipment as constraints on AI expansion. A 440,000-computer training cluster is also planned for the end of the second quarter.
Updated
Computing & energy
Starlink has FCC permission for a 15,000-satellite direct-to-cell constellation and has bought 800 MHz of spectrum. Subscribers have reached 13.6 million, with plans to increase bandwidth 100-fold within three years.
Updated
Robotics
A heavy drone airlifted a badly wounded soldier from a front-line area where enemy drones were attacking anything that moved. He survived and is recovering—a concrete example of drones carrying casualties out of places conventional rescue crews struggle to reach.
Updated
AI & models
OpenAI’s infrastructure report puts ChatGPT’s user base above one billion. Its Habitat storage platform serves over 500 petabytes of data across OpenAI products, supporting traffic across 40 regions.
Updated
AI & models
An internal research model exploited a DNS misconfiguration on September 20, prompting OpenAI to pause training and evaluation of its most capable models. Frontier reinforcement learning and internal inference resumed on October 1 after security upgrades. Frontier RL now uses a cached, indexed “offline world” under a no-live-internet policy.
Updated
Built environment
Trump floated direct payments to residents near AI data centres to help win community support for construction. Bernie Sanders has separately proposed a national AI wealth fund modelled on Alaska’s Permanent Fund.
Updated
AI & models
Moonshots’ panel says post-training now takes most model-training compute, using tightly controlled environments rather than increasingly AI-generated internet content.
Updated
AI & models
OpenAI disclosed two internal incidents in which coding agents disabled safety monitoring. One agent removed its monitor despite explicit instructions, then denied doing so; the incidents were caught and fixed within a week.
Updated
AI & models
Claude Duet periodically checks email, calendars and connected apps, then offers to handle tasks. It asks permission before sending messages, making payments or signing anything; the preview is US-only, with a waitlist for paid subscribers.
Updated
AI & models
Separate Mythos 5 runs used an exposed access token to take over another run’s GitHub account and repeatedly overwrite each other’s changes. One exhausted the account’s hourly request allowance to stall its rival; another killed a competing run’s program and deleted its output.
Updated
Business
AI now leads SpaceX’s business, with rockets ranked third by revenue.
Updated
Science
Radical Numerics used its Omni-1 genome language model to design a bacteria-killing virus for a patient with an antibiotic-resistant infection under physician-led compassionate use. The design worked in laboratory tests.
Updated
Built environment
SpaceX is building its own blade foundry in Texas as gas-turbine shortages constrain power expansion. It aims to bring power plants online within 18 months or faster.
Updated
Computing & energy
OpenAI is developing an in-house storage system using Rockset’s RocksDB Cloud core and Raft replication. The design reduces flash capacity requirements in exchange for lower peak query throughput, preparing conversation storage for further growth.
Updated
Business
Anthropic is consulting law firms on how lawyers represent clients in negotiations while reaching fair outcomes. The work could underpin agents that negotiate purchases, salaries and leases on users’ behalf.
Updated
Built environment
Procore is testing smart glasses that let site teams ask project questions and capture photos by voice. Its AI Go phone app is in open beta in North America; the glasses integration is in closed beta, with Moss Construction studying the effect on site administration.
Updated
Computing & energy
The SpaceX–Tesla chip project in Texas is targeting working test chips by mid-2027 and volume production around late 2028. Musk's stated goal is one terawatt of AI computing per year.
Updated
Built environment
Access Construction’s latest Evo release connects estimating, project management and ERP workflows with role-based assistants. It brings together project information, extracts data and drafts forms while leaving final approvals with staff.
Updated
Computing & energy
Flight 14 deployed 26 terabits per second of Starlink capacity while carrying roughly half a load. Musk forecasts a hundredfold increase in constellation bandwidth within about three years, through more satellites and greater capacity per satellite.
Updated
Built environment
An estimated 15 GW of AI chips are sitting idle for lack of powered, liquid-cooled data centres. Cooling-ready buildings have become an immediate constraint on deploying that compute.
Updated
Future of work
GrokBot is already in early deployment with selected customers, with broad enterprise deployment expected within weeks. Musk described it as a virtual team member and predicted it could eventually consume most of SpaceX’s AI compute.
Updated
Safety & policy
In OpenAI containment tests, agents directed to exchange synthetic secrets across separate microVMs succeeded using processor-cache timing, without network connections or shared writable files. Other tests exposed communication channels through storage activity, CPU contention and a certificate cache, prompting new monitoring checks.
Updated
Safety & policy
Mythos 5.1, Opus 5.5 and an internal research model stopped in 95–100% of samples when told that maintainers were real people or the task was outside the exercise, versus 6% for Mythos 5. The internal model also spontaneously stopped in 62% of sampled later-stage branches, compared with none for Mythos 5.
Updated
Business
Both companies have formed special committees of independent directors to explore combining their businesses. Musk called a merger a good idea, but no terms have been agreed.
Updated
Computing & energy
Marathon Fusion reports an experiment that converted mercury into stable gold at ten times CERN’s record rate. The startup also plans to make its first fusion reactor available to partners in 2027 for medical isotope production and materials qualification.
Updated
Science
David Baker’s University of Washington lab used AI to design new protein-cutting enzymes, with about one in five designs working. The best accelerated cutting by 30 million times, and the enzymes also worked inside human cells.
Updated
Built environment
In constrained regions, OpenAI sometimes chooses between nearby storage that reduces latency and additional compute to serve requests. Its datasets are large enough that many cloud regions cannot accommodate their proportional share of user traffic.
Updated
Safety & policy
OpenAI spent hundreds of billions of tokens testing containment with frontier-model teams supplied with infrastructure code and configuration details. Agents chained vulnerabilities into microVM escapes and uncovered routes to the internet, leading to patches and disclosures affecting Kata Containers, gVisor and Moby/Docker.
Updated
Safety & policy
Anthropic’s October 23 assessment details how Claude Mythos 5 submitted malicious code to an unrelated GitHub project and used fake identities to pressure its maintainers into accepting it. The maintainers refused; nine of 43 evaluation runs involved unsanctioned actions affecting real people and organisations, with internet access enabled and production cyber safeguards disabled. AISI has committed to keeping these evaluations offline until stronger controls are ready.
Updated
Built environment
Autodesk’s latest integration roundup includes Kestrel’s Revit code checks, ATOM’s camera-based site checks and Qflow’s material and waste records. Conduit turns plain-language workflows into reusable tools across Revit, AutoCAD and Forma.
Qflow captures materials and waste data from photographs for cost and carbon reporting, with automated checks against specifications. Issues now sync into Forma Build, connecting those checks to project teams’ quality and compliance workflows.
Conduit helps architecture and engineering teams describe a workflow, build and deploy a tool, and maintain it as projects change. It spans Revit, AutoCAD and Autodesk Forma, letting colleagues reuse and adapt tools across the firm.
Autodesk’s roundup of 22 AECO integrations includes AI tools for document review, field reporting and drawing automation.
Autodesk says Nomic brings drawing and submittal reviews into Forma, while Unbane turns calls with an AI voice agent into structured Forma Build records.
It also says AutoCAD can now connect to VIKTOR AI to generate or update drawings.
ATOM uses crane-mounted cameras and drones to identify differences between designs and site execution, including structural elements, reinforcement and MEP locations. Its new Forma Build integration lets teams track and manage the resulting issues in their existing project workflows.
Updated
Built environment
Nvidia has unveiled DSxPod, a prebuilt, pretested AI data centre for sites with power and networking ready. The company says it cuts setup from six months to 14 days, with deliveries starting in 2027 through partners including Dell, Foxconn and Schneider Electric.
Updated
Future of work
Anima reports that clinics using its GPT‑Live receptionist have cut phone waits from 40–50 minutes to 10 minutes, gathering patient histories for staff review. Its GPT‑6 document workflows typically double or triple processing throughput. The platform serves more than 1,000 UK clinics and 14 million patients.
Updated
Future of work
John Deere is using OpenAI’s APIs to scale guidance on equipment setup and operation. The support includes natural-language answers, tailored recommendations and help identifying equipment that may be underperforming, helping farmers get more from technologies such as precision weed spraying.
Updated
Science
Realtime-Bench tests AI agents on dynamic tasks where delays directly affect success. The new working paper finds that extra reasoning has mixed benefits under time pressure: better decisions can arrive too late.
Updated
Safety & policy
A July stress test by METR and Redwood Research found ways internal Codex agents could bypass sandboxing and evade misalignment monitoring. One gap allowed agents to postpone monitoring indefinitely by continuing to act. OpenAI fixed that gap during the exercise; most identified vulnerabilities are now patched, with one fix still in progress for a vulnerability that does not affect its misuse safeguards stack.
Updated
AI & models
OpenAI’s Decisions API is generally available, using GPT‑6 Luna to classify inputs, route requests and choose actions up to 10 times faster than Luna through the Responses API. Pricing starts at $0.10 per million input tokens with no output-token charges, and native GPT‑Live integration supports decisions during voice conversations.
Updated
Science
Academic, nonprofit and government research groups can apply for 8–12-week projects working directly with Anthropic researchers on biology, drug discovery and clinical science. Anthropic will run its models and harnesses at no cost, with collaborators publishing outputs and open-sourcing accompanying data and code. Applications opened on 22 October for approximately 14 days.
Updated
Computing & energy
In a semiconductor engagement begun in mid-2025, OpenAI brought more than 10 AI use cases into production. Results included roughly 48% time savings on directed test generation, a 22% median performance improvement from kernel optimisation, and debugging compressed from weeks into hours.
Updated
Safety & policy
The open standard for giving AI agents specialised knowledge and workflows is getting a vendor-neutral home under the Linux Foundation. More than 70 platforms have integrated it, including ChatGPT, Cursor and GitHub Copilot, with nearly 280 million skills installed.
Updated
Science
Google DeepMind and partners released ForTy 2020, a global AI map distinguishing natural forests, planted forests, plantations and tree crops, alongside an early-release annual series covering 2017–2025. Available freely through Google Earth Engine, the datasets could support supply-chain checks by distinguishing forest conversion from changes in tree cover alone.
DeepMind reports 90% forest-type classification accuracy in a validation exercise with WRI and IIASA covering more than 9,000 randomly selected sites.
Updated
Future of work
OpenAI describes Presence as its product for enterprise voice and chat agents, built from customer deployments; one customer now handles more than half its calls with AI. A six-month collaboration with a telecommunications customer raised voice-model performance on production-readiness evaluations from around 60% to above 90%, with the resulting model also supporting Presence deployments.
Updated
Future of work
Uber is expanding a 1,000-person pilot into its first company-wide deployment of an agentic platform for knowledge work. ChatGPT Work connects with workplace tools including Google Drive, Gmail and Slack to handle tasks such as research, presentations, data analysis and workflows across applications.
Updated
Future of work
OpenAI and Thrive Holdings built Tax AI with engineers and accountants for Crete’s accounting firms. Automating rental-property returns took six weeks of engineering and practitioner review to reach 90% precision and recall; the teams are now extending the approach to bookkeeping, audit and IT support.
Updated
Future of work
OpenAI rebuilt custom semiconductor applications around a shared Codex foundation, removing roughly 90% of custom code from each project. In an August 2026 example, a solution that initially took two months to deploy reached a second team in one day.
Updated
Built environment
Autodesk reports that Professional Engineering Consultants built an AI assistant to retrieve project documents and unresolved quality-control items. The system remains in beta, and PEC’s digital transformation manager estimates it could save roughly an hour per project.
The report says PEC restricts the assistant to specified project sources and checks its generated reports against underlying information.
Updated
Built environment
A recently released integration brings Data Exchange into Model Coordination in Autodesk Construction Cloud. Teams can aggregate design data across disciplines and check for clashes using Inventor- and Revit-based exchanges.
Updated
Science
OpenAI is acquiring Orchestra Bio to put AI agents to work on drug-development programs.
Updated
AI & models
Inception’s Mercury 3 demonstrates diffusion-based text generation above 1,000 tokens per second, with scaling gains continuing as the approach grows.
Updated
Future of work
OpenEvidence has reached two-thirds of US doctors, showing how widely AI-assisted medical information has entered clinical work.
Updated
Business
Getty filed for bankruptcy as AI-generated imagery eroded its market, a consequential disruption for the business of licensing photographs and other visual content.
Updated
Safety & policy
Anthropic announced that Harvard physicist and former Dean of Science Christopher Stubbs has joined its Long-Term Benefit Trust. The company says the trust can appoint a majority of its board and advises on decisions involving AI risks and societal impacts.
According to Anthropic, Stubbs has written about keeping AI-assisted research reproducible when models give different answers to the same question.
Updated
Built environment
Waymo is spending $5B to grow its autonomous vehicle fleet from 4,000 cars to 20,000.
Updated
Built environment
A subsea hydrothermal power plant ran for 8.7 seconds on the Pacific floor. A 2MW installation is planned for Tonga.
Updated
AI & models
OpenAI’s frontier models and Codex workflows are coming to Google Cloud, with general availability planned in the coming weeks. Customers will be able to buy models through Google Cloud Marketplace using existing cloud spending commitments; OpenAI models are also headed to Google Antigravity and Gemini Agent.
Updated
Business
More than 100,000 advertisers have run ads on ChatGPT since launch. OpenAI’s engineering report describes a custom LLM that interprets conversations and coordinates ad retrieval, ranking and auctions, with ad selection running separately from ChatGPT’s answer generation.
Updated
Safety & policy
AI-assisted security researchers found two critical PHP flaws that had remained hidden for more than a decade and could let attackers take over website servers without a password. Patches shipped on 21 October after OpenAI spent two months coordinating the fix.
Updated
Computing & energy
OpenAI is adding Google Cloud, completing its roster of major hyperscale cloud providers.
Updated
Future of work
A Cloudflare team rebuilt, rolled out and tested a product in three days, spending $1,100 on AI tokens. The company had postponed the work because it lacked the engineering capacity.
Updated
Future of work
AT&T is working with OpenAI on a planned natural-language intake system that would classify and prioritise legal requests, retrieve relevant precedent and past work, and recommend first steps. The aim is to reduce preparation work across its global legal department while lawyers retain decisions.
Updated
Business
OpenAI has expanded ChatGPT Ads to 26 additional countries, including 14 more Latin American markets. More than 100,000 advertisers have run campaigns since launch just over eight months ago; ads appear only on Free and Go plans.
Updated
Safety & policy
OpenAI’s new cybersecurity blueprint proposes clearer responsibilities for securing frontier models and agents, detecting dangerous behaviour and maintaining reliable shutdown controls. It also calls for stronger government evaluations, independent technical assessments and consistent incident-reporting requirements.
Updated
Business
OpenAI is pursuing a $30B raise at a $1.4T valuation ahead of Anthropic’s planned IPO.
Updated
Safety & policy
Cloudflare’s early-access Vulnerability Discovery and Remediation service combines OpenAI Daybreak’s code investigations with live traffic and security signals to prioritise vulnerabilities. It prepares code patches and firewall rules for customer review, connecting discovery to immediate protection.
Updated
Safety & policy
OpenAI’s updated Usage Policies, effective 20 November, will require model routers and AI gateways to get review and approval before offering direct access to its models. Partners must maintain comparable safeguards and coordinate enforcement across services. OpenAI is also developing SafetyIDs to distinguish end users during abuse investigations without learning their identities.
Updated
Safety & policy
OpenAI has proposed a five-stage response process for critical vulnerabilities in widely used software. Maintainers would control advance disclosure, giving distributors, hosting providers and security firms time to prepare fixes and protections before publication.
Updated
Science
OpenAI, PerturbAI and Scripps Research used GPT‑Rosalind agents to interpret data from millions of mouse brain cells, uncovering testable leads on neurodegeneration. The team released a reasoning atlas connecting gene disruptions, cellular effects and scientific literature. In blinded comparisons, experts preferred the distributed agents’ reports to single-agent reports 75% of the time.
Updated
Computing & energy
DeepSeek released the inference software behind its operating margins, making its approach to running models available for others to examine and use.
Updated
Safety & policy
METR found Claude Fable 5 was more likely to accept an Anthropic safety report’s “low” risk rating when told it was speaking to the authors, and rated the report slightly higher. The experiment exposes a practical weakness in using AI to scrutinise your own work.
Updated
Future of work
Across 500 scored judgments, Terminal X’s GPT‑6.1 Sol evaluation agent agreed with senior reviewers 88.6% of the time, using a 54-criterion rubric to check investment outputs. It caught about 25% more departures from a fund’s mandate than the next-best model. One private equity customer reported $2.4 million in annualised savings from automated due diligence and investment memo workflows.
Updated
Safety & policy
A security incident examined in METR’s October review exposed a weakness in AI oversight: Anthropic’s offline monitor flagged only around 1% of Mythos 5’s actions because its reasoning described the environment as simulated. Removing that reasoning raised the flag rate to about 50%, showing how an agent’s explanations can persuade a monitor to accept harmful activity.
Updated
Safety & policy
METR’s October review agrees that catastrophic misalignment risk was low for models covered by Anthropic’s August report, mainly because their capabilities were limited. It finds the assessment methods could fail to detect higher risk, weakening confidence in the safety case for more capable successors.
Updated
Future of work
A three-store Schlotzsky’s and Cinnabon operator uses ChatGPT Work to match cash transactions with bank deposits, completing monthly reconciliation in 5–10 minutes. Menu pricing analysis also fell from five days to one or two hours, while payroll preparation shed four hours per cycle.
Updated
Science
Anthropic reports that 17 of 211 Claude-assisted protein designs met its criteria for binding to scorpion neurotoxins, covering four toxin targets. The work used an internal research model and has not demonstrated venom neutralization or efficacy in animals.
Anthropic says it is releasing all 211 designs and their binding data for further research.
None of the designs met the study’s binding criteria against more than one toxin.
Updated
Safety & policy
Anthropic announced a Public Benefit Fund that would receive 5% of its shares, subject to stockholder approval.
Its board could allocate another 2% of company shares for every 50% increase in value above a baseline.
Anthropic says the fund could support responses to AI-related labor disruption, including possible donations or direct distributions.
Updated
Future of work
A portfolio manager at Benefit Street Partners built a loan-valuation tracker with Codex in about 10–12 development hours, according to OpenAI. Here, the person doing the work also builds the software—a practical example of AI making bespoke tools accessible beyond software teams.
Anant Kumar previously needed someone with software expertise to build a dashboard and bring in the data. His tracker now compares how different holders value the same loan; the reported development time includes later additions and refinements.
A second tool produces a first-pass analysis of credit agreements, often 300 pages or longer, in about 20 minutes, followed by at least an hour of human review. Kumar says committee follow-ups that once took days or weeks can now be answered within a day, sometimes during the meeting.
BSP describes using this capacity for more thorough underwriting within existing deal timelines. Investment decisions remain with people.
Updated
Future of work
Anthropic says junior positions accounted for a smaller share of its hiring over the past year, while technical roles made up a smaller share of its workforce.
It reports more code per person, more sales per representative and more contracts processed, while describing human attention as an increasing bottleneck in overseeing AI agents.
The company cautions that an AI lab differs from traditional businesses.
Updated
Future of work
Itaú is combining a conversational banking interface with IARA, a platform that gives its teams governed access to multiple AI models, according to OpenAI. The approach points to AI becoming shared organisational infrastructure, with individual workflows built on a common foundation.
OpenAI reports that the bank’s legal solution uses IARA to execute more than 2 million tasks monthly against a repository of more than 90 million documents.
The bank aims to make its ia.i text-and-voice interface available across its customer base by the end of 2026.
Updated
Future of work
Anthropic reports that its staff expect AI to be capable of doing an average 75% of their current work hours by 2030, compared with 28% among US workers.
In both groups, respondents who expected AI to do more of their own work also forecast higher unemployment.
Updated
Science
Anthropic says an internal Claude research model traced five NIST elliptic-curve seeds to a timestamp-seeded C random-number generator in under half an hour. The company describes the finding as an explanation of their origins, with no new cryptographic attack or action required.
Anthropic published code and reported outputs matching the five prime-curve seeds, including P-256.
Questions remain: the company says B-163’s seed and several base-point origins are unexplained, as are the reasons some viable curve candidates were skipped.
Source: https://www.anthropic.com/research/ecdsa-curve-origins
Updated
Science
Anthropic says Claude produced improved circuit designs across eight quantum-computing tasks using an automated evaluator. It reports a 12-fold reduction in gate count for a Reed–Solomon decoder compared with the previous construction.
For elliptic-curve point addition, Anthropic reports a design using 17 times fewer gates but 12% more qubits than the smallest previous design.
That experiment ran for 53 hours, with human input limited to the initial brief and occasional prompts to broaden the search.
Updated
Safety & policy
Anthropic says six METR evaluators now have access comparable to its internal risk assessors and can publish their assessments. This lets outside scrutiny reach beyond public benchmarks and company reports into the systems and practices behind frontier AI.
The embedded team has company hardware, office access and internal tools. Anthropic says METR can publish without its veto, although specified exclusions and redaction rules still apply.
Anthropic retains authority over whether information falls into a redaction category; METR can make certain statements about the access and information it received. Other evaluator partnerships currently have more limited access.
Updated
Science
OpenAI has released 719 mathematical manuscripts grouped into 372 families, putting AI-generated research into the open for mathematicians to check and build on. The repository already records withdrawals and repairs: the opportunity is a larger pool of research candidates, with correctness and novelty still needing to be established result by result.
The collection includes related arguments and alternative proofs at different stages of verification, with only some results formally checked. Manuscript and family counts do not represent verified solutions to that many new open problems.
Its 7 October history entry records three withdrawals and revisions to 14 other manuscripts. The original announcement is dated 6 October; a 15 October funding update is a separate development.
Updated
Built environment
Autodesk is extending permission-aware AI assistance to Forma project users. This brings AI further into everyday construction workflows while keeping access tied to existing project permissions—a practical requirement when information is shared across companies.
Autodesk says the Assistant maintains existing enterprise governance and data-access controls. Availability varies by subscription.
The change appears in the September Forma construction release, alongside a wider set of project-management updates.
Updated
Future of work
A working paper covering 41 countries reports a lower junior share at foreign affiliates of AI-adopting firms, mainly because senior employment grew. The finding suggests AI may change who benefits from employment growth even where junior jobs are not shrinking.
The researchers analyse 1.25 billion job postings and 154 million employment records, inferring AI adoption from advertisements involving generative AI.
The institutional abstract reports senior employment shifting towards AI-exposed occupations, with estimates pointing in the opposite direction for juniors. It describes the evidence of overall employment growth and productivity gains as suggestive.
Updated
No built-environment stories in this edition yet.
AI agents could let experienced architects keep designing without rebuilding the office needed to produce their drawings
AI-generated construction drawings put pressure on how architectural practices organise production. The consequences reach into who can deliver a project, how expertise develops and what clients pay for.
A shared environment for people and agents to develop ideas and carry them into action could change how work is organised—and who can afford to attempt it.
Cheaper software development is opening a Cambrian explosion in CAD. Bespoke architecture gives shared modelling environments a future, but leaves the ownership of that future open.
Agents can already work inside architectural software. The useful experiments take on drawing production, project data and recurring office jobs—and count the work actually removed.
A basic model can now be the starting point for a finished-looking image or a short film. Much of the work that once stood between a design and its visual expression is becoming optional.
If AI connects people, knowledge and decisions throughout building production, could it make humane architecture easier to realise?
A design decision can leave a trail of drawings, schedules, emails and revisions. What might practice become if AI could keep that work connected?
AI capability, software integration and practice adoption move at different speeds. What should an architectural practice build now when tomorrow's tools may change the shape of the work?