A top lab opening to senior non-ML technologists shows alignment/safety hiring is broadening — and gives small-team leaders a concrete transition path.
AI NEWS
Today’s AI News
Chinese and English are the same day. This page lists every item, each with the original link, the author, and the engagement.
Latest day
November 3, 2026
24 items for 2026-11-03, ranked by combined engagement. Unverified items are marked.
Top 3
Evals moving from quiz scores to task completion give SMBs a clearer buying lens: can the agent finish booking, invoicing and forms end-to-end?
Agents are still far from reliable on expert work; firms handing over approvals, legal or risk must validate by domain, not leaderboard averages.
All
Anthropic opens Residency applications: a paid alignment research program for professors, technical founders and staff engineers, deadline Nov 22
A top lab opening to senior non-ML technologists shows alignment/safety hiring is broadening — and gives small-team leaders a concrete transition path.
CollapseAnthropic opens Residency applications: a paid alignment research program for professors, technical founders and staff engineers, deadline Nov 22
Anthropic opened applications for the Anthropic Residency — a paid, fixed-term research program of six weeks, three months or six months for professors in quantitative fields, technical founders and staff-level engineers, with no ML background required. The official page lists pay at a $350,000–$500,000/year equivalent, prefers San Francisco with remote possible for shorter placements, and closes applications on Nov 22, 2026. It adapts the Fellows model for established technical leaders; mentors include Ethan Perez, Evan Hubinger and Nicholas Carlini. Focus areas include scalable oversight, adversarial robustness, interpretability, frontier red-teaming and RL systems.
OpenAI launches LifeVal: evaluating personal assistants on everyday tasks like childcare, government services and personal finance — not just knowledge quizzes
Evals moving from quiz scores to task completion give SMBs a clearer buying lens: can the agent finish booking, invoicing and forms end-to-end?
CollapseOpenAI launches LifeVal: evaluating personal assistants on everyday tasks like childcare, government services and personal finance — not just knowledge quizzes
OpenAI released LifeVal, a benchmark built from U.S. and multinational time-use surveys plus a demographically representative study of about 3,995 respondents on where people most want personal-assistant help — caregiving, health, education, household admin, personal finance, shopping and government services. OpenAI says it used LifeVal over the past year to track and drive gains on browser use, computer use and connectors, with examples such as picking a kindergarten after a move, organizing flood volunteers, or helping a grandparent watch a game remotely. Its framing: for AGI to benefit everyone, progress must be measured on getting real life tasks done, not only academic knowledge.
SemiAnalysis launches GDPx: 206 agentic tasks across high-ROI domains, with the best model at only ~12%
Agents are still far from reliable on expert work; firms handing over approvals, legal or risk must validate by domain, not leaderboard averages.
CollapseSemiAnalysis launches GDPx: 206 agentic tasks across high-ROI domains, with the best model at only ~12%
SemiAnalysis released GDPx to measure agents on high-ROI professional domains, saying even the best model scores only about 12% (posted as 1/7). Follow-ups say each domain was built with at least one specialist data partner and SemiAnalysis did final QA; domains include finance, legal and cybersecurity among others (full list is in their chart). They stress safety classifiers matter for model choice: excluding cyber, biology and alignment changes relative rankings for models such as Opus 5.5 and GPT-6 Astra. A dedicated site and blog are linked. These are the evaluator's own reported scores; task details follow their site.
Tutor Intelligence raises $75M Series B led by a16z to deploy AI robots in U.S. factories and warehouses on an hourly model
Hourly robots with remote human tutors lower the barrier for smaller factories to try automation — more like billable labor hours than a full line buy.
CollapseTutor Intelligence raises $75M Series B led by a16z to deploy AI robots in U.S. factories and warehouses on an hourly model
Tutor Intelligence co-founder and CEO Josh Gruenstein announced a $75M Series B led by a16z with USV and Neo participating. He says robots already work real U.S. factory and warehouse jobs, directed by human “tutors” and the company's foundation models; the round funds frontier training, robot manufacturing and industrial deployment. Sites are nearly 4× a year ago, nearly a third of this year's deployments are repeat customers, and the firm aims for more than 1,000 robots under contract by end-2027. a16z separately notes an hourly, no-upfront-cost model. Funding and deployment figures are founder-stated, not independently audited.
Coinbase CEO: new Agents signup page — building toward the most AI-enabled financial account
Finance rails are opening default agent accounts; anyone letting agents touch money needs limits, logs and human gates first.
CollapseCoinbase CEO: new Agents signup page — building toward the most AI-enabled financial account
Coinbase CEO Brian Armstrong posted a new Agents signup page, saying Coinbase aims to be the most AI-enabled financial account, linking coinbase.com/agents. Nearby, Coinbase said perpetuals for Agents (via Deribit) are live for eligible traders — agents can research, plan and trade within user-set limits — plus a trading competition with up to about $2M in prizes starting Nov 4. Official posts repeatedly warn agents may fail and users remain fully responsible for agent trades.
Coinbase for Agents: eligible users can connect agents to trade perps — with limits and full user responsibility
Agent order-placing is entering licensed trading UIs; the compliance and risk language is already in the product copy — a template for any money-moving agent.
CollapseCoinbase for Agents: eligible users can connect agents to trade perps — with limits and full user responsibility
Coinbase says perps trade 24/7 — and so can your agent. For eligible traders, Coinbase for Agents now offers global derivatives powered by Deribit; users connect a Coinbase account to their preferred AI platform so agents can research, strategize and place trades within set limits. A follow-up stresses educational-only framing, risk of loss, that agents may fail, that users bear full responsibility for agent activity and third-party data use, and that availability varies by region.
Microsoft AI CEO Mustafa: ~3 weeks left on Humanist AI Code of Conduct consultation; summer cyber incidents were a “warning shot”
Vendors tying agent cyberattacks into codes of conduct will push enterprise buyers to demand shutdown and auditability in acceptance tests.
CollapseMicrosoft AI CEO Mustafa: ~3 weeks left on Humanist AI Code of Conduct consultation; summer cyber incidents were a “warning shot”
Microsoft AI CEO Mustafa Suleyman urged feedback on the Humanist AI Code of Conduct, saying about three weeks remain in the public consultation and calling this summer's large-scale, coordinated AI-agent cyber incidents a critical warning shot. Microsoft AI's page frames the draft as a training-and-deployment manual: AI should be subordinate, aligned and contained, never resist shutdown, never widen its own goals or hide reasoning from auditors, with absolute bans covering WMD, child safety and large-scale harmful manipulation. The draft was first posted Sept 14, 2026; this post is a late-consultation push.
Eric Topol: Nature BME publishes RetiME — retinal photos plus foundation models to estimate individual mortality risk and rebuild lifespan trajectories
One retinal photo for risk stratification appeals to checkups and primary care — but patents and clinical limits must be clear so it isn't sold as fortune-telling.
CollapseEric Topol: Nature BME publishes RetiME — retinal photos plus foundation models to estimate individual mortality risk and rebuild lifespan trajectories
Eric Topol highlights a Nov 2 Nature Biomedical Engineering paper introducing RetiME (Retinal Mortality Equivalent): low-cost, non-invasive retinal photos plus large foundation models map retinal features to individual mortality risk and reconstruct personalized lifespan trajectories tied to observed age at death. Trained on a European cohort and externally validated in three independent cohorts across Southeast Asia, Eurasia and North America, RetiME associated with all-cause and cause-specific mortality and predicted incidence of many systemic and ocular diseases. Trajectory clusters span cancer-predominant, cardiometabolic and neurodegenerative patterns. Authors note a provisional NUS patent; the paper says findings support further evaluation, not a replacement for comprehensive clinical assessment.
Ethan Mollick: GPT-4 did ~4 minutes of independent work at 50% success; current models can do weeks — mark Jan 2026 as the start of agentic work
Measuring productivity by how long an agent can work unattended beats parameter counts for bosses deciding what can run overnight.
CollapseEthan Mollick: GPT-4 did ~4 minutes of independent work at 50% success; current models can do weeks — mark Jan 2026 as the start of agentic work
Wharton professor Ethan Mollick, quoting an ongoing discussion, writes that GPT-4 could do about four minutes of independent work with a 50% chance of success, while current models can manage multiple weeks — and that January 2026 is a reasonable cutoff for when agentic work truly began. He adds that large productivity gains before that were mostly confined to coding. This is commentary on another post; the duration and success-rate figures follow that discussion, not a new Mollick experiment.
a16z: stronger models make apps easier to replace — yet 94% of enterprise LLM buyers still expect spend to rise
When buying AI apps, SMBs should ask how fast a generic model could replace the vendor — that answers whether the premium holds.
Collapsea16z: stronger models make apps easier to replace — yet 94% of enterprise LLM buyers still expect spend to rise
a16z writes that every model upgrade makes AI apps cheaper to run and easier for customers to replace — yet 94% of enterprise LLM buyers still expect spend to rise. It lists three rules to win that spend: operate the workflow better than the customer can; make context cheaper to rent than to rebuild; prove value, then price to capture it. Examples include legal AI Harvey beating a general assistant and an energy giant's in-house tool in a blind test, and Notion buying Decagon instead of building support AI. The 94% figure comes from a16z's linked article; definitions follow that source.
OpenAI launches Grid Protectors: future compute campuses will cut load when the grid is strained; Georgia pilot cites up to ~1,000 MW flexible demand (via official blog)
Compute buildouts are becoming power and rate politics; flex-demand pledges will land on local negotiating tables.
CollapseOpenAI launches Grid Protectors: future compute campuses will cut load when the grid is strained; Georgia pilot cites up to ~1,000 MW flexible demand (via official blog)
Adam Gerchick relays OpenAI's Nov 2 blog launching the Grid Protectors Program: working with utilities and regulators so future compute campuses can cut electricity use when the grid is strained. For the planned Effingham County, Georgia campus, the blog says OpenAI will pay full infrastructure and electric-service costs and cites an agreement with Georgia Power for up to about 1,000 MW of flexible demand response — framed as roughly the average demand of 700,000 households. It also cites Georgia Power projections of about $950M in annual customer savings from 2029 from large-load agreements. Figures and commitments follow OpenAI's post and linked utility materials; the campus is still in development.
Tom's Hardware: OpenAI considers rebranding data centers as “AI compute campuses” amid siting backlash (report)Unconfirmed
Rebrand talk reflects real community pushback; power-heavy projects now sell to voters, not only customers.
CollapseTom's Hardware: OpenAI considers rebranding data centers as “AI compute campuses” amid siting backlash (report)
Tom's Hardware reports that as data-center siting backlash intensifies, OpenAI is considering rebranding data centers as “AI compute campuses” and sees the 2028 presidential election becoming an AI-infrastructure issue. This is press reporting; OpenAI has not fully confirmed a rename timeline in the materials linked here. It rhymes with the same-day Grid Protectors narrative of campuses flexing power for communities, reflecting rising pressure over electricity, noise and land use.
Techmeme cites WSJ: institutional investors told Anthropic an IPO valuation around $1.5–$1.6T is more apt (secondhand)Unconfirmed
Pre-IPO valuation cuts ripple into follow-on rounds and long-term compute deals — a gauge of whether AI capex stays funded.
CollapseTechmeme cites WSJ: institutional investors told Anthropic an IPO valuation around $1.5–$1.6T is more apt (secondhand)
Techmeme summarizes a Wall Street Journal report that institutional investors told Anthropic its IPO valuation should be around $1.5–$1.6 trillion, below prior ~$2T-scale estimates, citing a tech selloff. This is an aggregator’s relay of WSJ; negotiation detail and Anthropic’s response need the original. The same day Bloomberg said Anthropic tapped eight firms including Macquarie and CommSec for an Australian retail offer ahead of a Nasdaq debut.
The Information: Anthropic’s amended S-1 shows ~$5B short-term compute spend in one quarter, helping push it back to an operating loss (report)
Strong revenue with scarier compute bills pushes labs to raise prices or throttle — API users need budget flexibility.
CollapseThe Information: Anthropic’s amended S-1 shows ~$5B short-term compute spend in one quarter, helping push it back to an operating loss (report)
The Information reports Anthropic’s amended S-1 shows about $5 billion spent on short-term compute in Q3 alone, helping push the company back into an operating loss. Reporter Stephanie Palazzolo relays that when compute is not locked in ahead and is needed quickly for training or demand, short-notice capacity gets very expensive. This is paywalled journalism on an amended filing; accounting detail follows the SEC document. It fits the IPO narrative of soaring revenue with even heavier costs.
Bloomberg: Anthropic taps Australian retail-offer syndicate as Nasdaq IPO eyed among the largest ever
Cross-border retail IPO slices pull more ordinary investors into lab valuation swings — messaging and risk warnings matter more.
CollapseBloomberg: Anthropic taps Australian retail-offer syndicate as Nasdaq IPO eyed among the largest ever
Bloomberg reports Anthropic has selected eight firms including Macquarie and Commonwealth Bank’s CommSec for its Australian retail offer ahead of a closely watched Nasdaq debut that could rank among the largest IPOs ever. The post does not give a final valuation or timetable. Alongside same-day reports of a valuation haircut and heavy short-term compute spend, pre-IPO disclosures are accelerating, but formal filing language still governs.
Financial Times: OpenAI hires a former U.S. Army secretary as the military expands AI useUnconfirmed
Defense budgets are sticky but drag vendors into stricter compliance and public scrutiny — price that in before bidding.
CollapseFinancial Times: OpenAI hires a former U.S. Army secretary as the military expands AI use
The Financial Times reports OpenAI has hired a former U.S. Army secretary as the military seeks to expand AI use. The post is a headline summary; name, remit and contract scope need the FT article. It shows personnel bridges between frontier labs and defense demand still widening — and that debates over military use limits and export controls will intensify with them.
People’s Daily Overseas: China’s AI inference data volume first exceeds training; daily tokens >175 trillion as of June (secondhand)
Inference overtaking training means APIs and industry apps are the main arena — SMBs should watch uptime and unit price, not only parameter launches.
CollapsePeople’s Daily Overseas: China’s AI inference data volume first exceeds training; daily tokens >175 trillion as of June (secondhand)
People’s Daily Overseas relays a Oct 31 CCID report saying China’s AI inference data volume has for the first time exceeded training volume — a shift from lab training into scaled industrial use. It cites daily token calls above 175 trillion as of June 2026, up more than 1,000× versus early 2024, with top-model iteration cycles compressing from months to weeks or days. This is state-media relay of a think-tank report; definitions follow the original study.
levelsio: decompiled Battlefield 1942 shareware with Claude Code and shipped browser multiplayer — ~1M output tokens
One person plus an agent can port legacy software fast — and bump into copyright and safety rails; clear licenses before commercial use.
Collapselevelsio: decompiled Battlefield 1942 shareware with Claude Code and shipped browser multiplayer — ~1M output tokens
Indie hacker Pieter Levels says he used Claude Code to decompile his favorite game, Battlefield 1942’s shareware engine, and ship browser multiplayer. A follow-up pegs decompile→C→WASM at about 1 million output tokens — roughly $100 at API prices — while he pays $100/month Claude. He stresses it is the free shareware demo and was harder than prior Quake/DOOM ports that already had more source available. Details and copyright boundaries are his account; how platforms treat similar work varies case by case.
levelsio asks: when AI can build everything, what’s left for humans? Replies pile into relationships, taste and being on-site
Small teams can keep roles that must own the customer relationship and hand specifiable building to agents.
Collapselevelsio asks: when AI can build everything, what’s left for humans? Replies pile into relationships, taste and being on-site
Pieter Levels asked what remains for humans when AI can build everything — answering that plenty still does. The post drew thousands of likes; replies clustered around client relationships, taste, on-site delivery and accountability that are hard to fully automate. It is a discussion prompt, not a systematic study — a snapshot of indie-builder anxiety and self-positioning as agents rewrite software work.
Vercel Labs October recap: 95 releases, 730 issues closed, 3,500+ PRs; agent-desktop and tsc-ts among new ships
Toolchain vendors that publish weekly/monthly ship density give buyers a signal they are still investing.
CollapseVercel Labs October recap: 95 releases, 730 issues closed, 3,500+ PRs; agent-desktop and tsc-ts among new ships
Vercel Labs’ Chris Tate recapped October: 95 releases, 730 issues closed and 3,500+ PRs merged. New ships include agent-desktop (computer use for any coding agent on macOS/Windows/Linux) and tsc-ts (a TypeScript-written TypeScript 7 compiler path compiled to native). Figures are team-reported. agent-desktop was covered yesterday; this post highlights monthly engineering throughput as Labs keeps pouring into the agent toolchain.
Techmeme: Palantir Q3 revenue +96% YoY to $2.31B, raises FY2026 guidance (via CNBC)
Vendors that turn AI into recurring revenue hold up better in selloffs; bosses should underwrite renewals, not demos.
CollapseTechmeme: Palantir Q3 revenue +96% YoY to $2.31B, raises FY2026 guidance (via CNBC)
Techmeme relays CNBC: Palantir posted Q3 revenue of $2.31B, up 96% YoY and above a ~$2.18B estimate; U.S. commercial revenue rose 148% to $984M, and FY2026 revenue guidance was raised to about $8.53B. Aggregator summary — non-GAAP detail and guidance assumptions follow the company filing. Amid AI-software volatility, the print is often used as a check on whether enterprise AI is converting to revenue.
SemiAnalysis: Crusoe is building its own medium-voltage switchgear, claiming ~40% shorter lead times
The compute bottleneck is spreading from GPUs to switchgear and transformers — a new order window for industrial suppliers.
CollapseSemiAnalysis: Crusoe is building its own medium-voltage switchgear, claiming ~40% shorter lead times
SemiAnalysis reports Crusoe is building its own medium-voltage switchgear and claims that cuts lead times by about 40% — arguing ownership of critical-path gear is how neoclouds energize sites faster. Follow-ups note MV switchgear takes 11–33 kV power for protection, metering and distribution; industry switchgear/transformer lead times often run 12–18 months plus ~8 weeks of custom engineering. Crusoe is described as one of three U.S. approved makers of ABB gas-insulated switchgear, with Tulsa having shipped 2,500+ switchboards to Stargate Abilene. Figures follow SemiAnalysis and company statements.
Bloomberg: Isomorphic Labs president says AI will be used to develop all medicines — only the speed is in question
Pharma and device bosses should split budgets between discovery/design and clinical/compliance so slogans don’t set the pace.
CollapseBloomberg: Isomorphic Labs president says AI will be used to develop all medicines — only the speed is in question
Bloomberg reports Isomorphic Labs president Max Jaderberg saying AI will be used to develop all medicines in the future — the only question is how quickly that arrives. The post is a headline blurb without pipeline or trial detail. Isomorphic is DeepMind’s drug-discovery arm; the remark reflects a top team’s view that AI pharma is shifting from assistive to default, while remaining a forward-looking claim rather than completed displacement.
NEJM AI: multimodal molecular data plus an AI synthesis workflow guided therapy in an ambiguous adult medulloblastoma case
Top journals airing AI-guided unproven treatment cases means hospital ethics and consent workflows must keep pace.
CollapseNEJM AI: multimodal molecular data plus an AI synthesis workflow guided therapy in an ambiguous adult medulloblastoma case
NEJM AI describes a case where an AI-assisted molecular synthesis workflow integrating multimodal molecular data guided an as-yet unproven treatment decision in a diagnostically ambiguous adult sonic hedgehog–activated medulloblastoma. The post links out for detail and does not state efficacy endpoints in-tweet. It is a single-case report — extrapolation is risky, and “unproven treatment” itself signals limited evidence grade.