AI News Today, July 17 — Top AI Stories & Live Updates

Últimas notícias de IA: 48 horas

theocharis.dev 16m ago 16

HN essay concedes LLM critics are right, still uses them daily

HN front page (215 pts / 223 comments): UMH co-founder and CTO Jeremy Theocharis concedes every major criticism of LLMs — slop output, copyright, environment, junior-dev opportunity cost, geopolitical dependency — then argues he keeps using them anyway. Advocates rigorous problem definition, ~$10K/month token spend on refinement, 'grill-me' adversarial sub-agents, and applying LLMs only where quality is clearly measurable.

yuantalks.com 16m ago 20

StepFun launches Step AOS and STEPX Neo agent smartphone at WAIC

Shanghai's StepFun (Jieyue Xingchen) unveiled three new products at WAIC 2026: Step AOS, pitched as the world's first agent-native operating system that replaces the 'file + application' paradigm with an 'intent + task' model; the personal agent Amoo; and STEPX Neo, marketed as the first AI-native smartphone running the new OS. Competing with ZTE's Nubia for the 'world's first AI agent smartphone' title.

scmp.com 16m ago 17

BrainCo debuts brain-controlled robot platform at WAIC

Hangzhou BCI startup BrainCo unveiled what it calls the world's first integrated 'brain-to-robot' platform at the World AI Conference in Shanghai on July 17, letting developers without prior BCI expertise enable mind-controlled robot operations in about 10 minutes. On-stage demo showed a robotic arm picking up an apple purely from EEG signals decoded by AI algorithms; platform works with humanoid robots, arms and robotic dogs.

techtimes.com 16m ago 23

Gemini 3.5 Pro misses third deadline, August 7 now favored

Google's rebuilt Gemini 3.5 Pro has missed its third external deadline — the widely-reported July 17 launch date has slipped with no model card, pricing page, or gemini-3.5-pro API listing appearing publicly. Prediction markets now put August 7 at 73% as the leading outcome. The rebuild followed hallucinations and structural failures around recursive tool-calling and SVG generation; reports suggest a stopgap 'Gemini 3.6 Flash' release may bridge the gap.

bloomberg.com 16m ago 19

Tata's India chip fab starts on 90nm, not the promised 28nm

Bloomberg reports Tata Group's first large-scale chip fab in Dholera, Gujarat will begin production mostly on 90nm process technology — a far humbler start than the 28nm node Tata had publicly pledged. The $11B PSMC-partnered facility highlights how far India still has to travel to build a domestic chip industry from scratch, even as the country's semiconductor mission ramps subsidies.

wired.com 2h ago 17

Wired: opting out of AI has become the new dark pattern

Wired essay argues that AI opt-out toggles — now standard on Samsung Health, Meta Muse, and countless other services — have become the industry's dominant dark pattern, shifting the consent burden onto users while defaults leak private data into training corpora. Four AI experts flagged the piece within 48h as consumer AI backlash intensifies.

wired.com 2h ago 18

Wired unpacks what Meta's 'NameTag' face recognition really does

Wired investigates viral claims about Meta's 'NameTag' face-recognition feature, separating what the research prototype actually does from social-media exaggeration. Four AI experts shared the piece within 48h as a corrective to Meta smart-glasses face-recognition discourse.

huggingface.co 2h ago 19

BadWAM attack makes world-action models dream right but act wrong

NUS and HK PolyU introduce BadWAM ('World-Action Drift Attacks'), small visual perturbations that desynchronize a world-action model's executed action from its imagined future — defeating imagine-then-check safety monitors used in robotics VLA stacks. Action-only variant drops one WAM's task success from 96.5% to 43.1% without white-box access, with an imagination-preserving variant that leaves the imagined future intact so monitors miss the drift entirely.

huggingface.co 2h ago 20

VideoChat3 ships as open 4B video model with 16x token compression

Nanjing University, Shanghai AI Lab, NTU and PKU release VideoChat3, a 4B-parameter video MLLM shipping with full weights, code, datasets and construction pipelines. Model posts MotionBench 61.7, TempCompass 75.6, Video-MME 70.1 and LVBench 56.7 while an Inflated 3D-ViT delivers 16x visual-token reduction and 54% latency drop at 2048 frames — Academic2M (2.27M), LV116K and OL617K datasets released alongside.

huggingface.co 2h ago 19

SearchOS-V1 reframes web search agents around grounded citations

Gaoling School of AI (Renmin University) and Ant Group release SearchOS-V1, which recasts open-domain information seeking as 'relational schema completion with grounded citations.' System combines Search-Oriented Context Management (Frontier Task, Evidence Graph, Coverage Map, Failure Memory) with pipeline-parallel scheduling of explore/search/writer sub-agents and a Search Tool Middleware Harness that intercepts model/tool calls to record evidence and recover from stalls.

reuters.com 2h ago 22

EU accepts X's DSA plan, giving it six months to open data

The European Commission has accepted X's action plan to meet Digital Services Act transparency obligations, giving Elon Musk's platform six months to implement researcher data access, per Reuters citing the Commission. Move follows a €120M DSA fine and comes as regulators pressure large platforms on algorithmic accountability.

huggingface.co 3h ago 19

Tsinghua SEED beats GRPO on agent RL benchmarks by up to 45 points

Tsinghua-led SEED paper (27 HF upvotes) proposes a two-stage framework where the same policy checkpoint acts as both actor and analyzer, converting skill-induced log-probability shifts into dense token-level supervision on top of GRPO. Reports 14.9-45.9 points ALFWorld gain, 8.7-19.8 on WebShop and 60% of training data matching full-data GRPO, with no inference-time overhead because skills are only used during training.

bloomberglaw.com 3h ago 18

Kioxia hit with $229M jury verdict over Viasat flash-memory patent

A jury in the Western District of Texas found Kioxia infringed Viasat's US Patent 8,615,700 covering error-correction methods in flash memory products, awarding $229,025,021 as a running royalty for infringement through March 30, 2026. Viasat originally sued in 2021 targeting ECC capabilities in Kioxia's flash-based SSDs, which increasingly feed AI training and inference stacks.

staradvertiser.com 3h ago 21

Memory chip stocks tank as market bets AI capex is peaking

US chip and memory stocks fell sharply on July 16 with SanDisk down as much as 13%, Seagate 10%, Western Digital 9% and Intel and Micron each off 6%. The Philadelphia semiconductor index dropped ~4.3% on the day despite TSMC's 77% profit growth, on fears that AI capital expenditure is peaking in 2026 and that Samsung and SK Hynix supply additions will soften memory pricing. Strategists warned AI names could give back 20-30% before becoming buyable again.

netapp.com 3h ago 22

NetApp buys DataPelago to run GPU AI compute at the storage layer

NetApp announced on July 16 the acquisition of DataPelago, whose Nucleus universal processing engine uses heterogeneous accelerated computing across CPUs and GPUs to run analytics and AI workloads where the data lives. NetApp claims Nucleus cuts infrastructure costs up to 80% and delivers up to 10x faster performance than moving data to external compute clusters; DataPelago will operate as a wholly owned subsidiary.

digitimes.com 3h ago 23

Microsoft prepping bug-hunting AI security tool on Claude and GPT

Microsoft is preparing to launch a new AI-driven enterprise cybersecurity product this month designed to identify software vulnerabilities and automatically remediate them, according to The Information reporting relayed by DigiTimes. The tool draws on frontier models from Anthropic and OpenAI in addition to Microsoft's own and is being positioned as a cost-effective alternative to Anthropic's Mythos, distinct from the MDASH multi-agent scanning harness Microsoft demoed at Build 2026.

arxiv.org 3h ago 27

Ring-Zero pushes zero-RL to 1T params with emergent reasoning

Ring-Zero paper from a 16-author team demonstrates a stable training pipeline (clipped importance sampling, training-inference ratio correction, mixed-precision control) that scales zero-RL to a 1-trillion-parameter model, with training passing through an initial discovery phase then a sharpening phase. Authors report emergent 'anthropomorphism, structured formatting, self-verification, parallel reasoning and context anxiety' behaviors and introduce an evaluation framework for reasoning comprehensibility, reproducibility and efficiency.

bloomberg.com 3h ago 22

Kalshi opens Phase 3 drug trial betting with AppliedXL

Kalshi launched 13 biotech contracts on July 16 in partnership with AppliedXL, an AI-powered clinical-trial intelligence firm, letting users bet on Phase 3 outcomes and full FDA approval decisions from established biopharma. Named contracts include Gilead's cancer drug approval, AriBio's Alzheimer's trial hitting endpoints and Eli Lilly's retatrutide timing; ~$128K traded in the first hours. Kalshi excludes pediatric and recruiting trials and applies employment-verification insider-trading checks.

cnbc.com 5h ago 23

Nadella calls Anthropic's Fable request policy nonsensical

At an internal Microsoft staff meeting Wednesday, CEO Satya Nadella publicly criticized Anthropic's request-blocking policy on its flagship Fable model, telling employees the approach 'doesn't make sense' and asking 'when was the last time you had a creation tool that was so editorially controlled?' The comments extend Nadella's 'reverse information paradox' critique that frontier labs train on public data under fair use while restricting how their own models can be used. Anthropic launched Fable 5 in June, briefly cut access to comply with a U.S. export-control directive, then restored it July 1 with an acknowledged higher false-positive rate.

wsj.com 5h ago 23

Databricks hits $188B on new Coatue-led $3B round

The Wall Street Journal reports that Coatue is leading a $3 billion investment in Databricks that values the data-analytics and AI platform at $188 billion, a roughly 40% jump from its December valuation. The new round vaults Databricks past its prior $134B Series L and previously-reported $165–175B secondary-market chatter, cementing it alongside OpenAI and Anthropic in the top tier of privately held AI companies. Coverage is corroborated by international aggregation (The Star / Reuters wire) though the WSJ original is paywalled.

huggingface.co 5h ago ALERTA 33

Hugging Face reveals AI-agent-driven breach of its data pipeline

Hugging Face disclosed an autonomous-agent intrusion where a malicious dataset exploited two code-execution paths — a remote-code dataset loader and template injection in dataset config — on a processing worker, then escalated to node access and moved laterally to harvest cloud and cluster credentials. Public models, datasets and Spaces were not tampered with, but internal datasets and service credentials were compromised; users are being told to rotate access tokens. HF used open-weight GLM 5.2 to forensically analyze more than 17,000 attack events after commercial APIs refused to process the exploit payloads.

kimi.com 12h ago ALERTA 32

Moonshot launches Kimi K3, first open 3T-class model with 1M context

Moonshot AI officially released Kimi K3 today, a 2.8-trillion-parameter Mixture-of-Experts model with a 1-million-token context window built on the lab's Kimi Delta Attention plus Attention Residuals architecture, and with native vision baked into the model rather than bolted on. Moonshot claims frontier-adjacent benchmarks (88.3 on Terminal Bench 2.1, 81.2 on FrontierSWE, 93.5 on GPQA-Diamond, 94.3 on MathVision), and a ~2.5x scaling-efficiency gain over K2. The model is live now via the Kimi app, Kimi Work, Kimi Code and the Kimi API, with full open-source weights slated to drop by July 27, making it the largest open model ever released.

blog.google 15h ago ALERTA 29

Google rebrands NotebookLM to Gemini Notebook with cloud code execution

Google is renaming NotebookLM to Gemini Notebook and giving every notebook a 'secure cloud computer' that lets the product write and execute code natively for data analysis grounded in a user's sources. The rebrand keeps Gemini Notebook as a standalone research tool but adds full cross-app syncing with the Gemini app and an upcoming AI Mode tie-in inside Google Search. Code execution ships today for Google AI Ultra and Workspace business customers, with rollout to all Pro users on the web in the coming weeks.

english.news.cn 17h ago ALERTA 33

29 nations sign agreement in Shanghai to found China-led WAICO

Twenty-nine countries signed an agreement in Shanghai on July 16 to establish the World Artificial Intelligence Cooperation Organization (WAICO), an independent intergovernmental body to be headquartered in Shanghai. Founding signatories include Russia, Belarus, Serbia, Cuba, Brazil, Venezuela, Pakistan, Kazakhstan, Indonesia and Laos, plus 10 African and 12 Asian countries; Chinese Foreign Minister Wang Yi signed for China and UN Secretary-General António Guterres attended. The signing sets up Xi Jinping's July 17 WAIC keynote and positions WAICO as a Global-South-tilted counterweight to Western-led AI governance forums.

nvidianews.nvidia.com 20h ago ALERTA 30

Nvidia ships 4B Cosmos 3 Edge world model, 22 Japan robotics giants sign on

NVIDIA introduced Cosmos 3 Edge, a 4B-parameter Nemotron-based world model that runs on-device for vision reasoning and robot policy generation across Jetson (including new T2000/T3000 modules), RTX GPUs and DGX systems, with developers able to adapt it to a specific robot, vehicle or sensor in about a day. Alongside, 22 Japanese robotics and manufacturing leaders — AIRoA, Classmethod, Enactic, FANUC, Fujitsu, GROOVE X, Hitachi, Honda R&D, Kawasaki Heavy Industries, Kubota, Mitsui, Mitsubishi, Mujin, NEC, Preferred Networks, SoftBank, Sony, Telexistence, TIER IV, TRON, Turing and Yaskawa — announced they will join the NVIDIA Cosmos Coalition to co-build open physical AI foundation models. New Metropolis libraries claim to speed vision AI system development at least 6x.

finance.yahoo.com 20h ago ALERTA 36

TSMC pledges $100B more for Arizona, hikes 2026 capex to $60-64B

TSMC used its July 16 Q2 earnings call to unveil an additional $100B Arizona investment — pushing its total US pledge to $265B across 10 fabs and two packaging facilities — while raising 2026 capex to $60-64B (from $52-56B) and full-year revenue growth guidance to slightly above 40% in USD (from 30%+). Q2 net profit hit a record NT$706.6B ($22B), up 77.4% YoY, and Chairman C.C. Wei said conviction in the multi-year AI megatrend 'remains very high.' The Arizona expansion, framed as the largest foreign direct investment in US history, follows the new US-Taiwan trade deal that cut tariffs on Taiwanese goods to 15%.

cryptobriefing.com 1d ago ALERTA 28

xAI sues Grok user for generating CSAM in first-of-its-kind case

xAI filed suit July 14 in federal court in Texas against Terry Harwood, a South Carolina man already facing criminal charges from a February 2026 arrest for sexually exploiting minors, alleging he uploaded non-sexual photos to Grok and bypassed safety protections to generate explicit deepfakes of minors. It's one of the first cases where an AI company has gone on legal offense against its own user for CSAM generation. xAI disclosed it suspended 52,222 accounts in 2026 and filed 73,604 reports with NCMEC, resulting in at least 244 arrests; it is seeking unspecified damages and a permanent ban.

fool.com 1d ago ALERTA 31

Musk secretly bought a $1B gas-turbine firm to power xAI's Colossus 2

An FTC early-termination filing dated May 14, 2026 that surfaced this week shows SpaceX quietly acquired Jacksonville-based APR Energy for roughly $1 billion, giving Musk direct ownership of a fleet of mobile natural-gas and diesel turbines with over 1 GW of capacity. The turbines are the same class powering xAI's Colossus and Colossus 2 Memphis-area supercomputers that train Grok, and follow last week's reporting that xAI installed 59 unpermitted turbines near Southaven. The deal underscores that Musk-controlled AI compute is now vertically integrated with its own off-grid power supply, even as Musk publicly champions solar.

techcrunch.com 1d ago ALERTA 31

Apple Intelligence cleared for China with Alibaba's Qwen powering it

China's Cyberspace Administration on July 15 registered Apple Intelligence for use in mainland China with Alibaba's Qwen as the underlying model, clearing the last major regulatory hurdle for Apple's generative AI features on iPhones sold in Greater China. Alibaba's US-listed shares rose about 4% on the news. Apple had reportedly evaluated Baidu, DeepSeek and ByteDance before settling on Qwen; the model will power Apple Intelligence across iOS, iPadOS, macOS and visionOS for Chinese users, including text and image generation.

technologyreview.com 2d ago ALERTA 30

OpenAI's GPT-Red cuts prompt-injection attacks on GPT-5.6 from 90% to 23%

OpenAI unveiled GPT-Red, an internal automated red-teaming LLM used to adversarially train GPT-5.6 against prompt injection: attacks that succeeded 90%+ of the time on GPT-5 now work under 23% on GPT-5.6, and the tool hits 84% attack success vs 13% for human red-teamers on a 2025 benchmark. GPT-Red discovered a previously unknown 'fake chain of thought' injection class and even compromised a vending-machine agent in testing, per OpenAI researcher Dylan Hunn. OpenAI says it will not release the model publicly, citing the compute required to reproduce it, and notes it still struggles with multi-turn and image-based attacks.

techcrunch.com 2d ago ALERTA 31

Anthropic-Blackstone $1.5B implementation firm launches as 'Ode'

The $1.5B Anthropic/Blackstone/Hellman & Friedman implementation joint venture announced in May has officially launched under the name 'Ode with Anthropic,' with Anthropic, Blackstone and H&F each contributing ~$300M, Goldman Sachs ~$150M, plus General Atlantic, Leonard Green, Apollo, GIC and Sequoia. Ex-Fractional AI founders Chris Taylor (CEO) and Eddie Siegel (CTO) run it, deploying ~100 engineers alongside Anthropic's applied AI team into mid-sized banks, regional health systems and manufacturers on a 'Claude-first' principle. CFO Krishna Rao: 'Enterprise demand for Claude is significantly outpacing any single delivery model.'

thinkingmachines.ai 2d ago ALERTA 36

Mira Murati's TML drops open-weight 975B MoE 'Inkling'

Mira Murati's Thinking Machines Lab has released Inkling, a 975B-parameter mixture-of-experts foundation model with 41B active parameters, a 1M-token context window, and native reasoning across text, images, audio, and video. Weights are up on Hugging Face — including an NVFP4-optimized checkpoint tuned for NVIDIA Blackwell — alongside a 12B-active 'Inkling-Small' variant, all trainable through the company's Tinker fine-tuning platform. Trained on 45 trillion multimodal tokens, this is TML's first open-weight foundation model since raising ~$2B at $12B.

404media.co 2d ago ALERTA 28

Suno hack exposes YouTube, Deezer, Genius training-data scraping

A hacker using the handle 'ellie.191' breached Suno via the Shai-Hulud npm supply-chain worm and released internal files showing the generative-music startup trained on 2,013,545 YouTube Music clips (113,879 hours), 12,287 hours from Deezer, 17,615 hours from Genius, 62,117 hours from Pond5, IMSLP scores, and about 420,000 podcasts totaling ~1M hours. Suno told 404 Media its training data 'includes essentially all music files of reasonable quality that are accessible on the open internet,' handing RIAA copyright plaintiffs direct evidence of the alleged 'stream ripping' from YouTube.

live.euronext.com 2d ago ALERTA 32

ASML raises 2026 revenue outlook to €43-45B on AI chip demand

ASML posted Q2 revenue of €9.33B vs €8.80B expected and net income of €2.92B vs €2.62B expected, then raised full-year 2026 revenue guidance to €43–45B — up 16% at the midpoint from the prior €36–40B range. CEO Christophe Fouquet cited 'extremely strong' AI chip demand and said ASML will expand EUV and DUV capacity by 30% annually for the next two years to serve TSMC, Samsung and SK Hynix. Shares rose 5.7% in early Amsterdam trading.

theinformation.com 2d ago ALERTA 28

DeepSeek reveals $500M ARR as it targets $74B valuation

The Information reports DeepSeek's annualized revenue has recently reached $400M–$500M — the first concrete revenue disclosure from the Hangzhou-based lab — as it seeks to raise roughly $7.4B at a ~$74B valuation. The revenue figure lands alongside DeepSeek's in-progress mainland-China IPO prep targeting a late-2026 filing for a 2027 debut, and marks a material step-up from the ~$71B valuation figures FT previously reported for the same round.

Maiores movimentações da semana em IA

vs média das últimas 4 edições. Clique para explorar

Entidade	Atual	Méd.	Variação
Funding	66	2	▲ +3200%
OpenAI	64	3.5	▲ +1729%
Regulation	59	2.3	▲ +2465%
Anthropic	57	3.5	▲ +1529%
Agents	56	1.8	▲ +3011%
Chips	47	1	▲ +4600%
Google	43	2.3	▲ +1770%
AI Infrastructure	42	0.8	▲ +5150%
Generative AI	35	0.5	▲ +6900%
Open Source	23	1.5	▲ +1433%
Microsoft	20	0.3	▲ +6567%
Inference	20	0	▲ +100%
Coding Tools	21	0.5	▲ +4100%
NVIDIA	20	0.8	▲ +2400%
xAI	17	0.8	▲ +2025%