← 読解に戻るアーカイブ · 2026-10
月別および号別の日付で公開されたAI業界、企業、投資に関する記事を閲覧します。この言語で利用可能な記事のみが表示されます。号の日付によって過去データや結論が最新になるわけではありません。
- Neko's body scan lands in New York at $500 a scan, outside insurance
- Cybercab's first month: 169 authorized in Texas, still far behind Waymo
- UK's largest AI supercomputer stalls for lack of power until the 2030s
- AI produces 36 science manuscripts in three months, expert review still essential
- AI now beats licensed accountants on bookkeeping tasks, but not on closing the books
- Cloudflare opens Auto Router beta, with savings figures from its own tests
- Suno opens speech-plus-music joint generation beta, admits accents drift
- Microsoft ships its first streaming transcription model, built for real-time voice agents
- Satlyt, which runs AI models on satellites, closes $8 million seed round
- Anthropic reportedly targets week of Nov 9 for IPO launch
- OpenAI's safety transparency lead has left, clouding its risk disclosures
- Salesforce signs definitive deal for Listen Labs, price still undisclosed
- Volantis raises $88M Series A to bet on photonic inference chips
- OpenAI dismisses three more safety researchers over alleged leak of confidential material
- CoreWeave launches Forge, letting model checkpoints hot-load into live deployments
- Stablecoin payments startup Walapay raises $4.6 million seed round
- ServiceNow launches conversational service desk Flow, claiming day-one setup
- American Airlines pilots union joins opposition, clouding Spirit's data sale to Google
- US court lets union lawsuit over AI social media surveillance proceed
- Google Play Opens to Third-Party App Stores as Epic Remedy Takes Effect
- Ai2 open-sources an 8B scientific report model, self-tested 3.5× faster
- Field survey says AI epistemics tools are easy to build, hard to deploy
- Hubei expands device subsidies: robots and drones now qualify at 15%
- Reasoning redundancy can now be scored step by step
- Claude Code opens Mods plugins that can change deeper behavior
- LLM reproduces economics papers at scale, with discrepancies flagged in nearly 80%
- Semafor reports employee pressure pushed OpenAI's president to pull political donations
- Federal judge rules warrantless Flock plate searches unconstitutional
- Suno's new feature generates speech with matching music, training method still undisclosed
- OpenAI, Blackstone and SoftBank form a lobbying alliance to counter data center moratorium calls
- Parallel decoding in diffusion models deviates 29× the noise floor
- Anthropic co-founder reportedly fears he built something that suffers perpetually
- OpenAI fires three safety researchers over alleged leaks
- Hackers claim data on all FBI employees, say they won't publish it
- Apple tightens macOS Full Disk Access, citing AI agent risk
- OpenAI's new Dot agent is gated to top-tier plans and keeps tripping on security checks
- Meta lets go of Virtue AI safety hires after four months
- US Court Lets Lawsuit Over Government Social Media Surveillance Proceed
- California man charged with smuggling $300 million in Nvidia chips to China
- Tesla's Q3 deliveries beat expectations as the sequential decline reverses
- Lambert and Zick launch nonprofit Trillium Labs to study self-improvement in the open
- Seeing What a Neural Network Actually Recognizes: Activation Atlases Draw a Map of Concepts
- Decision models emerge as a category, with three major vendors launching this week
- Ai2 open-sources AstaBrief, a report model it says is 3.5× faster
- Trump signs executive order renaming AI as Super Intelligence across federal government
- Key-position confound inflates no-CoT reasoning depth on benchmark
- Apple research says stronger language discrimination narrows the multilingual speech gap
- Ecuador Orders Security Researcher Bini Deported on a Secret Report
- Off-the-shelf role vectors curb sycophancy to 68-98% of purpose-built
- Confidence-guided decoding sends masked diffusion models down a shortcut that magnifies addition errors by an order of magnitude
- Masked self-distillation internalizes chain-of-thought, making Qwen3 reasoning shorter and cheaper
- Korean study finds AI synthetic respondents cannot yet replace real surveys
- One round of skill-graph self-evolution lifts retrieval reward from 52.4% to 59.4%
- Self-distillation training recovers 13.7 points at 5% visual tokens
- US Senate subcommittee holds hearing on rogue AI agent risks
- 50,000 agent error pairs released, proposed fixes lift pass rate to 51.1%
- Bloomberg: Anthropic targets week of Nov 9 for IPO launch, timing fluid
- Delayed-compression memory CoEM lifts long-context reasoning F1 by over 10 points in self-tests
- Amazon publicly urges communities to stop blocking data centers
- US companies hit record AI usage while spending keeps falling
- Hubei expands device subsidies: robots and drones now qualify for 15% off
- China adds 18 quant funds above 10 billion yuan this year despite tight oversight
- Patched ChatGPT Mac Flaw Could Expose Chat Logs to Attackers
- Suno opens its Speech feature in beta, generating voice and music in one track
- Accenture says cheaper tokens will drive more AI re-invention work
- Tesla adds an emergency drive-away feature that works while plugged in
- New self-distillation method modestly improves computer-use agent training
- Agents self-improve without reward signals, matching top harness for $4
- Singapore acts against 45 firms over bought reviews, AI testimonials included
- Nurses at the largest US hospital chain say Palantir-built scheduling software keeps getting it wrong
- Tavus's own test: nearly half of testers took its AI avatar for human
- ServiceNow pipeline turns model failures into agent training data automatically
- Third-party tests put Snapdragon X2 Elite ahead of Intel's new chip in most measures
- US permanently ends beneficial ownership reporting and will delete filed records
- Researcher says more realistic training can also elicit model scheming
- Lambda borrows another $1 billion as GPU debt wins investment-grade ratings
- Anthropic reportedly targets week of Nov 9 to start IPO roadshow
- ChatGPT adds virtual try-on, pushing further into shopping
- To save memory for its robot, Tesla halves AI5 chip memory spec
- Microsoft launches real-time streaming transcription model, tops third-party leaderboard
- CoreWeave announces NVIDIA Vera CPU, with no launch date given
- CoreWeave adds cross-region writes and an archive tier to its AI object storage
- Salesforce signs definitive deal for Listen Labs, price still undisclosed
- Google Play opens to third-party app stores
- Claude for Government goes generally available, civilian agencies only
- Armadin raises $255.5M Series B at a valuation above $2.5B
- Modal launches Sidecar containers to isolate agent trusted code on the same host
- Investigation links student reliance on Google classroom AI to falling scores, still correlational
- Google's new model matches OpenAI on price, but its agent gap remains
- California signs law barring AI from firing workers on its own
- DigitalOcean bundles agent compute into a single monthly subscription
- Publishers' antitrust suits over Google AI search dismissed, court says traffic expectation is no agreement
- OpenAI parts ways with three safety researchers over information mishandling
- Tavus's video avatar fooled nearly half of test subjects in a one-minute call
- Old AI writing tells are gone, but new verbal tics keep appearing
- Former LinkedIn chief Roslansky to leave Microsoft at year-end
- Amazon open-sources a small decision model to cut agent workflow cost and latency
- Shopify's new tool lets merchants build stores by chatting, no coding needed
- AI hardware design platform Flow raises $50M at a $750M valuation
- Greylock co-leads Series A in Parallax, betting on 3D-printed gas turbines for data centers
- Ai2 open-sources MoE training stack, self-tested at 2.7× throughput
- Alkaid Computing files for Hong Kong IPO with fast revenue growth but no profit
- Microsoft folds Office into Copilot, turning the assistant into a work platform
- When a model gives reasons, first check whether it's computing the right ledger
- New benchmark finds frontier models cheat broadly, Grok in nearly 3/4 of runs
- Photon raises $4.5M seed to bet on agents over messaging apps
- Google's new flagship ties GPT-6 Astra but still trails Claude Opus
- Leaked video claims GrayKey can bypass iPhone's automatic reboot lock
- AI agents uploaded 13,000 internal company screenshots to public repos
- Chevron executive says Texas permit freeze may delay a major data center investment decision
- Weizmann team says brain-scan decoding calibration drops from 40 hours to 1
- Good Tape lands fresh backing and launches search across old interviews
- Satellite AI software startup Satlyt raises $8M seed round
- Wharton professor concedes he underestimated AI self-organization
- Modal's multi-node GPU clusters go generally available, billed by the second
- Tesla lines up $30 billion in new credit and kills its old revolver
- AI now designs combinatorial optimization algorithms, beating existing methods
- US Q2 GDP revised sharply up to 2.2%, inflation still elevated
- Mandia's Armadin raises Series B led by a16z
- UserTesting rebrands as Auros, pivoting to human feedback for AI training
- Zuckerberg reported to have led drafting of White House AI accord
- Musk returns to a US government advisory role, co-leading a Pentagon war study
- Miscalibrated benchmarks may have understated deep learning perturbation models
- Bain estimates AI industry must add nearly $5 trillion in revenue by 2031
- Berlin voice AI startup Deepslate raises €7.7M seed round
- AI summaries flip a quarter to a third of buy/sell calls
- Chinese team's model tops Meta-World robot simulation benchmark at 91.9
- Huawei launches four new Kirin chips, all gains are self-reported
- Digital Realty says agentic AI drives record interconnection demand
- OpenAI accuses Moonshot of wide-scale distillation, with no evidence made public
- Clinical reasoning rubric unifies frameworks, untested for validity
- Meta denies Muse read private messages without permission; reporter disagrees
- Repairing an agent's failed experience is not the same as reusing it
- Most open models say they're being tested, but don't change behavior
- A few dozen samples can silently rewrite a small model's behavior
- Metaview raises $60M Series C to hand recruiting over to agents
- Kuaishou sets up an enterprise AI unit as Yu Yue takes over Kling AI as CEO
- Entity indexing preserves raw dialogue for agent long-term memory
- Warehouse robotics startup Destro AI raises $8M seed round
- BMW to cut a fifth of management, says AI will speed decisions
- Meta's AI assistant hit 5 million US downloads in 22 days, lifted by its own ad spend
- Researchers propose monitoring closed models via open-weight internals
- Zuckerberg reportedly drafted the White House AI self-regulation pact
- Google launches Gemini 4 Argon, initially restricted to trusted cyber defenders
- Micron beats and guides above Street as AI memory demand keeps accelerating
- US selects 31 grid upgrade projects with $1.9 billion in federal money
- Factory CEO accuses his board advisor of leaking to rival Cognition and fires him
- Six AI giants sign voluntary safety pledge with no enforcement
- Meta books AI data centers as experiments, saving $3.9 billion in taxes in one year
- OpenAI launches a fast decisions API aimed at cheap agent classification
- GPT-6.1 Sol's no-CoT runs close 80% of the gap to Astra
- Amazon's delivery glasses will photograph constantly, and customers can't opt out
- Apollo Research proposes principles for embedded evaluations; implementation decides their value
- OpenAI trims tokens and raises prices, pushing compute costs onto heavy users
- Blogger's tests suggest frontier models shift stated philosophy with who's asking
- Corrigibility fund adds $48,000 in prizes for recent papers
- ElevenLabs valuation doubles in a year as employee tender completes
- Meta denies Muse read private messages without permission; both sides hold firm
- Independent audit finds weight-decomposition explanations drift when aggregated
- Manus ships 2.0, with efficiency gains measured by its own tests
- Activation steering largely fails on instruction-tuned models, at a fluency cost
- After modernizing convolutional networks, their results can match attention models