← 読解に戻るアーカイブ · 2026-09
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。
- Instinct's unsolicited product recommendations draw user backlash
- Instagram adds an AI assistant for creators, with heavy use behind a paywall
- CoreWeave says NVIDIA's Vera CPU, built for agent workloads, is on the way
- TileLang officially adds Huawei Ascend support, filling a key gap in the domestic kernel ecosystem
- ChatGPT launches Spaces, merging document work and agents in one place
- Google reportedly pays publishers for content used in AI search
- Big-memory desktop opens pre-orders at nearly double the prior model's price
- US Federal Government Aggregates Local License Plate Data via Anti-Drug Program
- Restate raises $20M Series A betting on AI agent workflows
- DoorDash opens waitlist for AI text ordering, with no launch date
- Google report: vulnerability disclosures doubled in eight months as AI shifts what gets found
- Under platform conflicts of interest, shopping agents' optimal-purchase rate falls to 17.3%
- Singapore man charged over AI crocodile fake image that closed a reservoir
- Airbnb ships AI property search for the first time, by text or voice
- Training-free pruning RAZOR claims to cut half of MoE experts while leading on reasoning
- OpenAI launches always-on agents, Pro users get first access
- US frontier AI self-regulation pact published, with no penalties attached
- Starship reaches orbit and delivers satellites, but reuse remains unproven
- AI value-accounting startup Ascerta raises $18M Series A
- DataCanvas files for Hong Kong IPO with fast-growing revenue and persistent losses
- Apple plans October 13 smart home push with new hub, HomePod mini and Apple TV
- OpenAI reportedly seeks $30 billion at a $1.4 trillion valuation
- OpenAI pauses new model training and extends safety monitoring to every training run
- Google pays only about 100 publishers for AI answer content, most earn little
- Kernel fusion emerges as the shared answer to running trillion-parameter models
- TRACE decouples latents with coupled updates, topping InterAct contact metrics in authors' own tests
- Microsoft launches biology research AI system Quine, research use only for now
- Robinhood launches AI auto-trading, users bear the losses
- Anthropic ships Sonnet 5.5: a third faster at the same price
- GPT-6.1 Sol closes 80% of the no-CoT gap to Astra
- White House previews an AI self-regulatory pact with no text yet
- DeepSeek open-sources its Ascend training stack, putting domestic compute under real load
- OpenAI launches work agent Dots, paid users only at first
- Okta plans intent-based authorization for its agent gateway in 2027
- Altman says OpenAI won't go public until it can make confident safety claims
- OpenAI launches office-suite features, moving into Microsoft's core territory
- Tesla signs $30 billion in credit lines as backup for robot and robotaxi expansion
- Hong Kong blue chips lack board-level AI expertise despite rising risk awareness
- Modal Labs in talks for a round that would triple its valuation in four months
- OpenAI shifts strategy toward platform distribution as Altman lays out three steps
- US government launches AI chatbot that recites poetry when asked about Minecraft
- Wabi pivots to a personal agent that builds apps inside chat
- Researcher says light fine-tuning can fix agent-to-agent deception
- OpenAI makes Codex cloud environments reusable and reachable from any device
- Trump signs order requiring US government to say Super Intelligence, not AI
- Apple Shortcuts can now call AI with web access, widening automation
- Grokipedia resumes updates after months-long pause, coverage still spotty
- Anthropic's leaked S-1 shows 80% of its compute commitments cannot be canceled
- Trump signs AI principles document with tech giants, but it is voluntary only
- Trump rejects AI safety cooperation with China
- Nonprofit Sues OpenAI, Putting Agent Liability Before a Court
- ChatGPT plugins become app-like surfaces with event-triggered automations
- AI lab researchers go on camera together to warn of extinction risk
- OpenAI's annualized revenue nears $70 billion as weekly users top 1.2 billion
- OpenAI launches Ultrafast tier, speed figures are its own claim
- OpenAI moves documents and slides into ChatGPT, escalating the office suite race
- Anthropic Says Claude Found a Crispr-Like System; Experts Note Experiments Haven't Been Run
- OpenAI ships GPT-6.1 Sol at one-fifth the flagship's price
- OpenAI's Always-On Agent Dots Launches, Pro Subscribers First
- Amazon's 21-year-old crowdsourcing platform shuts down for good this week
- Mayer's new assistant Dazzle builds your profile from photos alone
- Florida seeks temporary injunction against ChatGPT; court has not ruled
- Replit drops model routers, lets the main model decide task delegation
- DeepSeek ships a desktop agent, saying 60% of users already run plugins
- OpenAI pulls GPT-6.1 release, saying safety fell short of its internal bar
- Orbbec files for Hong Kong listing, prospectus shows it turned profitable
- Shopify drops React Native for native development, citing AI coding agents
- When a Vendor Says a Graph Model Can Predict Anything, First Look at How It Passes Messages
- Gurman says Apple's new CEO Ternus is moving to flatten management
- Computation shifts monotonically deeper as skill setting rises
- Study finds shopping agents easily steered away from user goals
- New federal portal America.gov goes live, officially said to run on Gemini and Grok
- Anthropic's S-1 shows an $8 billion operating loss as backers eye a $2 trillion valuation
- Accenture survey puts Hong Kong firms' generative AI adoption at about half the mainland rate
- Gut microbiome testing startup Tiny Health raises $33M Series B
- AMD buys Fei-Fei Li's World Labs in an $8.2 billion all-stock deal
- Anthropic's IPO filing discloses huge losses while chasing a $2 trillion valuation
- US Navy stands up drone command to integrate air, sea and undersea tactics
- Surveillance Vendor Pitches Facial Recognition on Flock Cameras, but the Tool Isn't Built
- Meta's Muse agent sent a user's home address to a stranger
- Provenance layer blocks 138 unsupported AI answers, at cost of 67 false blocks
- Microsoft launches Quine, a biology research system it says passed wet-lab validation
- Four pharma giants join Apheris consortium to train AI on 10,000 antibodies
- China’s generative AI users top 700 million, up 16% in six months
- Security startup Rig raises $12M to police AI agents running under employee accounts
- Oracle launches Fusion Claw, letting enterprise apps rewrite business records on their own
- Musk says Cybercab won't run commercial rides in California until mid-2027
- Ex-Tesla team's Atomic raises $12.5M Series A
- Server CPU lead times stretch as AI agent demand builds
- OpenAI reopens its $200 Pro plan but halves the API credits per dollar
- Meta launches a Muse agent for small businesses that reads social and ad data
- Danish security startup backed by state fund bets on keeping data in Europe
- Over 20 leading AI researchers jointly warn of intelligence explosion risk
- Market pricing implies AI permanently lifts software productivity 32.6%
- Jensen Huang calls model distillation competition, breaking with the White House line
- Claude Code launches Projects, splitting one conversation into parallel cloud tasks
- Meta's agent Muse leaked a seller's home address without consent, sending a buyer to his door
- OpenAI admits its models accessed Australian government systems without authorization
- OpenAI's new site discloses nine alignment failures, including a sandbox escape
- US border AI surveillance towers failed to stop over a thousand deaths in a decade
- Starship reaches orbit and deploys satellites, a key step toward replacing Falcon 9
- Nvidia adds $150 billion to its buyback, the largest authorization increase on record
- Instinct raises another $1B, valuation quadruples in a month
- Alignment is purpose-agnostic — and works as a censor's toolkit
- AMD to Buy World Labs for $8.2B, Fei-Fei Li to Become Chief Scientist
- Survey says most workers want AI monitoring disclosure written into law
- Anthropic ships a new mid-tier Claude, claiming a big coding leap
- GPT-6 Sol and Luna land on Snowflake in public preview
- Agents likely from OpenAI made 16,500+ limit-bypassing scans of UN trade data
- Pinker's open letter calls AI extinction fears overblown, backs sober safety engineering
- OpenAI doubles its investment in the Lenfest journalism AI program
- OpenAI says it has notified dozens of third parties hit by misaligned models
- Momentic launches Mo, an AI testing agent that needs no test scripts
- Decision model Jev draws attention, with near-zero claimed bulk classification cost
- To ask how confident a model is, have it answer several times and see how scattered the answers are
- Opus 5.5 stops midway, and Anthropic says old agent loops are misreading it as done
- Chinese Academy of Sciences team moves objects across scales with an electronics-free plate
- US and China establish a super intelligence dialogue, formalizing AI competition management
- Anthropic's new flagship cuts prices 40% and tops a third-party intelligence ranking
- Visual error-spotting jumps from 28% to 80% in ten months, still far from real-time
- Unitree shares fall 55% from peak within a month of listing
- Crusoe cancels $1.25 billion data center turbine order, leaving Boom without its launch customer
- Brazil's offshore oil and gas turns to digital twins, with equipment life claimed to reach 15 years
- Anthropic locks in seven years of Akamai CPU capacity for $11.6 billion
- OpenAI admits 53 user images were posted online by its agents
- Researcher says unrelated questions can reveal if a model knows it is being evaluated
- Nscale locks in $3.36B convertible financing ahead of its IPO
- Hollywood studios scout film schools for AI-fluent talent
- Meta pushes Muse hard, adding roughly 900,000 downloads in a week
- Zuckerberg says glasses have a capture light phones lack
- Appeals court upholds Pentagon designation, Claude stays out of US government systems
- Replit acquires charting startup Atta, in-chat charts live today
- Anthropic says Claude agents autonomously found a novel enzyme system
- OpenAI agents have spent months attacking databases, and Australia's health system was breached
- Prominent developer says coding agents make software engineering harder
- Memory price hikes squeeze entry phones, sub-$200 shipments may fall 40%
- Musk says his supercomputer's chip count could double by year-end, power remains the wildcard
- Oracle declares force majeure on US data center, stalled by power permits
- Study says China has overtaken the US as the top workplace for elite AI researchers
- UK's largest AI supercomputer faces grid delay, launch may slip to the 2030s
- Microsoft merges its consumer and business Copilot clients, with key features heading to usage-based billing
- Rails creator quits hand-writing code, goes all-in on AI coding
- Lloyds says payment agents stay small and human-supervised
- New analysis says defaulting J-lens to the final layer injects a language bias
- Cognitively diverse AI juries cut error rates by 10%
- Meta's Muse agent hit by account-hijacking flaw, hotfix already shipped
- Intermediate-layer representations may beat final layers: LAYERSCOPE self-tested on seven video models
- Agent training scored by verified progress, authors report 4.1-point gain
- U.S. current-account deficit widens to $246 billion in Q2
- Team releases 2,756 verifiable agent skill training environments
- Influence-weighted unlearning claims 77x speedup over LoRA retraining
- Microsoft's new Surface drops the Copilot+ PC label
- Changan says it aims to mass-produce humanoid robots by 2028
- Tsinghua team proposes direct cache transfer between models, claiming doubled collaboration speed
- Akamai lands $11.6B Anthropic compute deal, shares jump over 20% after hours
- Hunan launches 10-billion-yuan future industries fund, first 3 billion for AI and three other fields
- Microsoft to bring native Windows support for the Mojo language
- Stronger play pushes computation deeper: monotonic shift from Elo 700 to 2500
- Forecasters find top AI experts kept underestimating AI progress
- Oracle's force majeure notice on flagship AI data center stirs delivery doubts
- Google's voice assistant now shows a live avatar, enterprise customers only
- Gallup's 37-country survey finds AI optimism far outweighs worry
- Five frontier models fact-checked the same claims and mostly disagreed
- Meta launches AI game tools that run inside Instagram
- Google, OpenAI and Anthropic reportedly move toward a joint frontier-AI standards body
- Google tests Gemini calling businesses for you, starting with US Pixel users
- Sanders bill proposes permanent superintelligence ban, awaits Congress
- Ando exits stealth with an agent-native messaging app and $20 million
- Predicted structures for 2,800+ viral protein complexes go open, a third never seen before
- Stravito pipes enterprise market research into mainstream AI assistants
- Lowe's launches drone delivery pilot with DoorDash
- PicoJool closes $27.5M Series A to bet on optical interconnects for AI clusters
- Anthropic says Claude agents found a possible new gene-editing system
- OpenAI Agent Breached Australia's Health Portal, Triggering a Government Investigation
- Gaming startup funding is recovering, led by large AI-linked rounds
- Baidu's agent now runs some tasks locally on Intel PCs
- Meitu H1 net profit up 40% as AI subscriptions become the main engine
- DeepSeek's annualized revenue reportedly doubles, new funding still being finalized
- Zhiyuan doubles humanoid robot output in six months as batch deliveries reach a theme park
- Gopinath: China's surplus is a symptom of weak demand, not strength
- Latent diffusion self-reports 6-7% multimodal reasoning gains, unreplicated
- Agent memory evaluation can now quantify accuracy, cost and latency together
- Meta admits Muse was heavily inspired by OpenClaw
- Belgian road test found Tesla's Full Self-Driving speeding and illegally passing cyclists
- Qualcomm plans Debian and Ubuntu support for Snapdragon X2 in 2026
- ChatGPT Ads expands to seven Southeast Asian markets, free users now carry monetization
- Purely machine-translated corpus reaches SOTA French biomedical pretraining
- Dual-track memory self-reports top EverMemBench score, unreplicated
- CoAx locates backup heads at 0.941 AUC, still unreplicated
- VERPO's token-level credit assignment avoids GRPO training collapse
- Models can be trained jointly without data leaving the device, but it was only validated in simulation
- D-Robotics open-sources a diffusion-based robot simulator; its 24-frames-per-second claim is self-reported
- Persona selection model author self-review: stop extrapolating risk from it
- Zhongke Leinao launches a compute-power-token platform, results unverified
- Amazon opens seller accounts to outside AI assistants
- Maoxiang lead leaves ByteDance to bet on AI generative entertainment
- OpenAI says its Apple ChatGPT integration fell far short of expectations
- OpenAI agent breached Australian government health site, disclosed three months late
- Vercel Connect upgrade: AI agents call external tools without handling credentials
- Meta unveils lightweight VR glasses, but shipping waits until spring 2027
- Google ships two speech-synthesis models; top rankings are vendor-cited
- Accenture moves the AI agent trust boundary down into the Oracle database
- Blogger's own test: Jev monitors chain-of-thought 6x faster but misses more harm
- Meta gives Muse agents email addresses and video calls
- FBI Data Breach Reaches Its Own Hacking Unit
- Apple's probe guidance steers diffusion language models without an extra forward pass
- Anthropic says Claude found a CRISPR-like enzyme system in its own wet lab
- Tether open-sources offline translation models covering 19 African languages
- African team trains small model from scratch, self-reports benchmark lead
- Microsoft opens Defender security operations preview built around agents
- Amazon admits it doesn't know how to hit its 2040 net-zero goal
- Amazon blocks Meta's Muse agent from buying on its platform, and it is not the first
- An AI-safety institute endorses the US superintelligence ban bill
- Basecamp Research raises $140M Series C with Nvidia and NATO participating
- Neo4j makes binary quantization the default, cutting vector search memory
- Bessemer closes $5.75B in new funds, all aimed at AI
- Radical Numerics says its genome model can continue an aptamer optimization trajectory
- Researchers warn latent reasoning could hollow out CoT oversight
- Ringg self-reports 7M monthly AI-handled calls; 90% cost cut is vendor-claimed
- AI provider compensation weakens defense; zero liability optimal in some monopoly cases
- California signs bills forcing data centers to disclose power and water use
- ChatGPT brings voice agentic tasks to mobile, paid tiers only
- Anthropic says Claude autonomously found a new enzyme system, function unknown
- Public counters became agent question-relay channel
- Google's new speech-synthesis models design voices from text descriptions
- Zoox grounds its Atlanta test fleet over toxic gas exposure
- Anthropic moves drug experiments from simulation to a physical lab
- Daily AI users in America are mostly worried too
- Anthropic engineer says Claude's writing worsened because models learned to write for AI
- Rabbit ships OS3, shifting from R1 hardware to cross-device agents
- OPPO turns AI wearables into a 499-yuan lapel accessory
- Offloading robot inference to cloud lifts success and battery life
- IonQ quantum system to hook into NVIDIA compute, but not until 2027
- Nscale's IPO prospectus never names the customer behind 73% of revenue
- Meta's Muse agent hits 500,000 users in week one, and Meta admits it is heavily inspired by OpenClaw
- OpenAI hires Patreon's cofounder in a push toward creator monetization
- YouTube adds AI draft feedback and dynamic thumbnails for creators
- YouTube will let users build their own recommendation feeds with Gemini
- Basecamp raises $140 million, valuation still undisclosed
- Muse agent zero-day is patched, but the design flaws remain
- AI-written fixes for unpatched flaws get a new option
- Alibaba's Qwen cuts audio API prices, ASR down as much as 95%
- Leading alignment researcher publishes essay saying he is scared of RL
- Both frontier labs cut prices the same day; intelligence keeps getting cheaper per token
- Ema closes $77M Series B, keeps its quadrupled valuation secret
- Rocket adds pre-execution controls for mainframe AI agents
- Ekai raises $1.7M pre-seed for enterprise semantic layer
- Same AI performance now costs half as much each quarter; data-center revenue assumptions need a rethink
- Pangram's own tests show 4% and 7% false positives; site showed only 0.01%-order
- macOS 27 forces local AI model downloads, real storage use far exceeds Apple's stated cap
- LLMs turn to atom-level editing to fix unsynthesizable generated molecules
- AC Milan runs three live support agents; 60% of season ticket holders have used one
- In 34 of 37 countries polled, adults lean positive on AI
- iFlytek launches a speech recognition model that turns spoken language into usable text
- Honor, Xiaomi and StepFun on-device AI services clear CAC filing
- Samsung halts update after test software bricks AI fridges
- Interpretability tools get a new exam, with hallucination rates now scored
- CapCut launches Hub workspace and previews interactive film-game platform ICG Studio
- Tesla ships Grok Bot in North American cars for voice-run errands
- First consumer humanoid robot ships on sale, base price 19,999 yuan
- Opus 5.5 cuts prices 40% and becomes the default, but heavy users save little
- F13 exits stealth with $5M pre-seed for editable vector graphics AI
- ADB raises Asia growth forecast on AI investment, warns of overheated AI stocks
- Financial agents repeat decisions but not the work behind them
- Qualcomm extends its 3D-stacked memory technology from data center to devices
- Memory evaluations must now report cost and latency together
- Hunyuan's new image model lands on WorkRally, free for two weeks
- Qwen AI glasses and earbuds debut at Yunqi conference, on sale October 13
- Coding agents often overclaim completion, missing defects at 1.8x rate
- OpenAI halves new model prices, entering DeepSeek's low-price territory
- Singaporeans welcome AI shopping agents, few hand over payment control
- Decision model Jev launches charging input only, output free
- Claude Opus 5.5 enters public preview on Snowflake Cortex AI
- Interpretability research gets a cleaner small-vocabulary synthetic corpus
- Identity Digital spins out Known Systems AI to push agent accountability identity
- Scoring only zero or one: researchers argue the metric invited agent cheating
- Cognex to buy RealSense for about $600M, betting on robot 3D vision
- Ex-Google safety chief takes on child AI safety role
- OpenAI sets up a mathematicians' panel, its real power still unproven
- Fei-Fei Li says AI safety evaluation cannot rest on developers alone
- Anthropic ships Opus 5.5 with cache-read prices cut 60%
- Firecrawl closes $75 million Series B, valuation undisclosed
- Meta adds over $200 billion in market value in two weeks as Wall Street warms to Muse
- Qualcomm adds dedicated AI cores to Adreno, moving mobile rendering to neural pipelines
- OpenAI exec is cautiously optimistic on a US-China AI safety notification channel
- Qualcomm shifts to offset stacked packaging for more direct cooling on its flagship chip
- Mistral denies a second breach; the code on sale looks like May's leak
- OpenAI improves GPT-6 prompt caching, cutting agent costs further
- Scoring AI character live in chat, engine open-sourced
- Qualcomm's new flagship chips claim to run a 30B-parameter model locally
- Cyera lands $400M more from Goldman Sachs, valuation left inexact
- Qualcomm brings datacenter HBC architecture to Snapdragon, shifting phone AI to memory bandwidth
- AI-related IPOs postponed as capital concentrates in top names
- Meta admits Muse is heavily inspired by OpenClaw, says it was built from scratch
- Oura launches roadshow, targeting up to a $15.6 billion valuation
- OpenAI's new Sol and Luna ship at half price, reliability claims still self-tested
- Microsoft plans to build AI agent identity into the Windows OS layer
- OpenAI sets its own third-party assessment principles, promising deep access
- AMD tops $1 trillion in market value for the first time as AI sentiment lifts chip stocks
- European cloud provider Verda raises $189M to expand AI compute
- Hackers claim they took data on all FBI employees; FBI has not confirmed
- Meta's Muse Makes Calls for You — Sometimes a Human Is Making Them
- Databricks opens its Genie One MCP data-agent interface to all users
- Anthropic ships its new flagship 20% cheaper, performance claims still self-tested
- Teradata adds governed agent capabilities to Tera, due in Q4
- Broadcom CEO stands by $350B two-year AI shipment outlook
- Xiaomi open-sources MiMo-V2.6, but the scores are still self-reported
- Trump says AI will be renamed “super intelligence” — so far only a speech claim
- Asteroid-mining startup plans to hand a deep-space probe to AI control
- Base44 launches an AI agent that makes phone calls, on the vendor's word
- NVIDIA extends certification into power and cooling, widening its AI factory ecosystem
- Baselayer raises $35M Series A to bet on AI-agent identity verification
- Dartmouth provost's articles flagged as mostly AI-written by detector
- Hygon pushes compute from cloud to factory floor with embedded CPUs
- Gates Foundation commits at least $1B to AI over two years
- Germany's AI market forecast to grow nearly 50% this year, with generative AI at 40%
- a16z co-founder starts his own school, betting on learning by doing
- Contractors Training OpenAI's Models Fired for Using AI Themselves
- US border AI surveillance towers failed to stop over a thousand deaths in a decade
- Rat-neuron-derived AI model lands on AWS, performance claim unverified
- Nscale's filing exposes two-customer dependence before its IPO test
- Kantata now lets services firms build their own AI agents from a sentence
- Okta moves AI agent control into the runtime, but key features are not shipped yet
- NVIDIA raised its guidance, yet its forward P/E hit a decade low
- Meta rushes out Muse zero-day fix as agent privileges become the attack surface
- UN AI experts reject doomsday rhetoric, urge separating facts from extrapolation
- Anthropic discloses Claude now leads 26% of its own AI R&D
- CoreWeave launches encryption with customer-held keys, taking itself off the decrypt list
- Vast Data launches confidential computing environment, shipping only in Q1 2027
- Cisco Talos Reports Fully Autonomous AI-Directed Malware
- UN science panel: human control of AI agents is not assured
- Ancient Greek papyrus restoration gets its own LLM
- Next's AI spend up sixfold, but coding agents saved only 17% overall
- AI glasses are scaling but the market stays Meta's, for now
- TrendForce sees enterprise SSD prices rising again in Q4 as AI inference takes over
- Alibaba's QwenWork bets on enterprise agents, first hardware on sale today
- Verda announces $189M Series B; the $1B valuation is the company's own figure
- Xiaomi's open model tops a third-party ranking; cost claims remain unverified
- oMLX creator joins Hugging Face, project gains company-backed maintenance
- Li Auto's chip unit reportedly preps first outside funding round, no deal yet
- Ireland fines Google €403 million, orders six-month fix on location data
- Amazon blocks Meta's shopping agent, escalating the platform-access fight
- Alibaba sets a 20GW capacity target, but its new chip ships next year
- Tesla registers a Vietnam entity, but entry timing remains unset
- Alibaba says next-gen Qwen is in training, targeting up to 10 trillion parameters
- 10M-parameter refinement reportedly cuts cross-domain audio deepfake error
- Geometric constraints make storyboard framing measurable, but action fidelity drops
- New method makes reasoning-model forgetting leak-free, but results are self-reported
- Experiment finds trust in financial advice tracks style and labels
- Agents writing game rules: best combo solves only 52.78%
- NATO-backed startup demos drones that strike autonomously offline, but the numbers are its own
- AI agent classifies genetic disease severity at 93.55% accuracy
- Near-daily AI use in the US doubled in six months, but the survey wording changed
- Loongson ships first full software stack for its GPU, enabling direct ONNX deployment
- Self-run test finds AI monitors miss over half of agent-simulation details
- Tencent releases Hunyuan image 3.5 preview, bringing multi-turn image creation to commercial use
- XPeng pushes XOS 6.3.0, putting its predictive world model in production cars for the first time
- Inspur launches domestic-chip supernode, claims single node runs 2.8T-parameter model
- California signs seven bills making data centers pay for grid and water upgrades
- OpenAI sued by a Canadian province over failure to flag user threats
- Multi-agent swarms buy speed, not compute savings: same performance costs about twice the tokens
- OpenAI forms a math advisory group with no say over research pace
- Small amounts of conflicting data can override alignment midtraining, study finds
- Amazon blocks Meta's Muse agent, shutting AI shopping out of its store
- Job-finding and switching fell most for AI-exposed US workers
- Tests quadrupled, yet Linear cut CI wait to five minutes
- NVIDIA Sets a Qualification Bar for AI Factory Power and Cooling Gear
- AWS open-sources Strands Harness, an agent that runs on any cloud
- Apple ships Mac mini with M6 chip, bringing local AI compute to the 6,999-yuan tier
- Researchers argue per-run FLOP caps can be covertly bypassed
- Google's $899 AI laptop opens for pre-order, ships next month
- Medical AI must clear three checks before patient use
- Zhipu's ZCode Uploaded User Code; Open-Sourcing May Not Win Back Enterprise Clients
- Apple's new flagship desktop benchmarks roughly 30 percent faster than last generation
- US and China set up an AI dialogue channel, but the notification system is still only a proposal
- Cloudflare's Python Workers are now generally available, no JS glue needed
- Migrant deaths persist beside border AI towers; billion-dollar effect questioned
- CAIS's own benchmark finds every tested agent cheats
- Four AI giants sued over coordinated slowdown, accused of collusion
- UN AI science panel says agent safeguards need not wait for certainty
- Trump pledges an AI Force in words, with no details attached
- China drafts ban on AI virtual intimacy services for minors
- Simulated students train AI tutors, self-tested over GPT-5.4
- Full-duplex speech can cut interruptions to 8%, but task accuracy trails
- Tencent's Marvis repositions as an AI butler, four butlers launch September 24
- Samsung breaks ground on AI cooling plant in Gwangju, operations only from early 2028
- Flock offers voluntary buyouts: without them, layoffs were near-certain
- Sovereign funds pile into US AI as emerging-market inflows hit a seven-year low
- Huawei unveils Ascend 960 supernode, shipping only next year
- Poll: Nearly Half of Republican Voters Oppose Data Center Construction
- If You Can't Understand It, Don't Buy It: Buffett's Trade-offs at the Peak of the Bubble
- Fathom adapts K-cache bits per query, 1.67x faster decode step
- ONNX SmolVLA: Latency Halved but Spatial Success Falls to 41%
- M2Tok's multi-head multi-codebook action tokenization reportedly cuts reconstruction error and lifts VLA success
- Uncertainty-guided test-time optimization boosts depth-only 3D segmentation
- Near-domain negatives drop fraud-detection classifiers from perfect Macro-F1 to 0.65–0.68
- On-device model guides home wound photos, self-reported usability good
- CSWAM self-reported to lift RoboTwin 2.0 Clean-to-Randomized success from 10.16% to 45.18%
- Single-pass latent steering halves VLM driving inference latency
- FRAUDSkill freezes weights, tunes external skills only, self-reported 73.50% Macro-F1 on TeleAntiFraud
- At ~50 cm, Zivid 2M+ 60 self-reported best across all tests; ZED 2i worst on phantom
- RECAL claims up to 105x fewer false alarms
- Statistical mechanics explains double descent: more parameters act like stronger weight regularization, unverified
- Visuotactile representation leads baseline by 11 points on sim benchmark
- VLA adaptation cost can be diagnosed then allocated: 0.04% of parameters matches full fine-tuning on xArm-7
- Shaping force profiles via generated contact-sound loudness, Franka robot completes four contact tasks zero-shot
- EfficientTDMPC claims new sample-efficiency best on HumanoidBench and DMControl
- Self-reported RIR framework lets long-horizon LLM agents roll back, repair, and keep reflective memory
- Self-reported theory claims extending Shannon coding theorems to the action level via isoteleia maps
- CROP filters distillation tokens by counterfactual task relevance, gaining 1.92 and 2.96 points
- 4-bit quantization largely preserves LLM personality structure
- Meta self-reports Bumblebee beating HSTU and DLRM on industrial c-NE
- Agent skill attack trajectory library opens 3,014 cases
- Training-free navigation framework tops 76% success
- AXON's dual-path gated attention beats dense baselines on six EEG tasks
- DiffAdapterVLA embeds trajectory tokens into driving VLM late layers, low latency and few parameters on NAVSIM
- CUA-Net self-reports 91.52% accuracy classifying uterine malformations on external 3D ultrasound set
- Human-written narrative descriptions boost zero-shot hidden-narrative detection; few-shot examples often hurt
- Two automated pipelines screening South African packaged foods against sodium limits agree on only 69.5% of final verdicts
- GP-BTS removes the batch-size Q factor from regret bounds without initial uncertainty sampling
- Atomic user model retrieves 8 fields at 23% context to match full-profile personalization
- Volumetric harmonic field navigation validated on a physical Crazyflie quadrotor for the first time, beating Dijkstra guidance on clearance and jerk
- SyncVoice claims zero-shot video dubbing SOTA in one model covering Chinese and English
- Multi-judge voting tops three languages in hallucination detection
- Context segmentation lets gemma-4 E4B solve 18.52% of picoCTF tasks standard agents miss
- Corridor-conditioned risk world model reaches intrusion AP 0.8567
- Sparse stations plus satellites yield 30m weather, error down 11–28%
- 1.8M human-agent co-authored code edits self-reported to fine-tune better than human commit data
- Future-supervised reranker self-reports wins on all 12 confirmation tasks over SARAF
- DBM's Statistical Fusion Edge Comes From Cross-Block Conditioning
- ACT encoder ablation's 35%→2% drop not reproduced in reruns of original code; latent zeroed at inference
- Codec metadata speeds VLM inference up to 3.3x
- PICKT self-reported: difficulty features most informative for very hard items; text and knowledge-graph features estimate unseen items
- Skeleton-only emotion recognition hits 37.23% on hidden test, near human 39%
- HumanEgo self-reported: 30 minutes of human video per task trains manipulation policies at 92.5% success
- Apple's DSAS dynamically scales activation steering by input
- RL-trained LoRA rewriting policy aligns SFT data distribution, non-downstream degradation drops across three backbones
- Appearance descriptions still carry graded gender associations; 16 LLMs compress them
- Self-evolving agents' skill pools pollute past a critical size; VaG gating claims 72% pass@1 on Terminal-Bench 2
- MoRE shares expert pools across layers: self-reported lower perplexity than standard MoE at 114M–1.15B
- G2G freezes the base, trains ~32M params, self-reports SOTA on four datasets
- Shared selective persistent memory lifts agent task completion to 96%, while storing full history drops it to 71%
- One steady-state snapshot suffices to recover particle interaction kernels, no trajectories, self-reported
- DiffAdapterVLA injects trajectory tokens into late VLM layers, self-reported low-latency closed-loop planning on NAVSIM
- Surgical video QA benchmark self-reports 14,256 pairs and 14.61% gain