Models, products and AI startups, near-duplicates collapsed to the most credible source
Meta ads promoted AI nudify apps with child abuse contentThe Tech Transparency Project reported that Meta failed to detect 332 ads containing child sexual abuse material (CSAM) this year, with the vast majority promoting China-made AI “nudify” apps. Some ads used photos of real minors — including a European royal press photo and a 14-year-old Instagram influencer — which TTP says were AI-animated into sexual videos.
Google updates WeatherNext to version 3Google released version 3 of its WeatherNext model, which now ingests some satellite weather data to shorten the lag between current conditions and new forecasts. The update is detailed in a white paper and aims to improve forecast timeliness while keeping lower computational cost.
Power users file class action against Anthropic over subscriptionsA group of Claude subscribers filed an expanded class action lawsuit against Anthropic, alleging deceptive advertising about the limits of its Max subscription tier. The suit was brought by attorneys Monica Vaca and Kati Daffan, both formerly at the FTC, and seeks legal accountability for Anthropic’s subscription claims.
Qwen releases Qwen-Drive-1.0Qwen released Qwen-Drive-1.0, a finetuned driving vision-language model that integrates 3D perception, VQA, and motion planning. The full Bf16 checkpoint is 9B and the Hugging Face repo links to a ~40-page technical report.
GPT-5.6 Sol helps run quantum experimentsOpenAI shows how an MIT researcher uses GPT-5.6 Sol together with Codex to autonomously run quantum computing experiments, analyze results, and calibrate qubits. The demonstration illustrates using GPT-5.6 Sol and Codex for experiment automation and analysis workflows.
nex-agi publishes Nex-N2.5-mini 35B modelA Reddit user posted about the model entry 'nex-agi/Nex-N2.5-mini - 35b' on r/LocalLLaMA. The post lists the model name and its 35B parameter size without further details.
Google Cloud, Accenture form FDE unit to train 1,000 engineersGoogle Cloud and Accenture created a joint unit called Accenture Gemini Enterprise Business Group to train up to 1,000 forward-deployed engineers to build custom AI applications on the Gemini Enterprise platform. The organization will sit under Accenture.
Ling-3.0-flash-VL releasedinclusionAI published Ling-3.0-flash-VL on Hugging Face, a multimodal model with native image and video understanding, a 1M-token context window, and 124B total parameters. It activates about 5.5B parameters per token and uses sparse MoE and a hybrid backbone for efficient long-context multimodal reasoning.
Patagonia emerges as potential AI data center siteReports say Argentina’s Patagonia is being eyed for large AI data centers due to cool climate, hydropower, wind energy, and available land. Pampa Energía plans up to 500 MW in Neuquén with initial contracts expected by end of 2026, Green Capital aims to start 300 MW in Chubut, and OpenAI announced an unsigned project with Sur Energy.
ASML wins TSMC, Samsung, Intel; Huawei pursues alternativesASML convinced TSMC, Samsung and Intel to switch to larger photomasks, which should boost throughput of its newest EUV machines by about 40%. Meanwhile Huawei, via equipment maker Yuliangsheng and its suppliers, is pursuing a domestic strategy to reduce dependence on Dutch lithography technology.
Chrome moves to two-week release scheduleChrome has shortened its release cycle from four weeks to two weeks and launched Chrome 153 for desktop, iOS and Android. Google says the change enables faster security patches, reduces the N-day patch window, and speeds iteration of AI-related features.
GPU GB-per-dollar and bandwidth guideA user posted a GPU comparison guide for the LocalLLaMA community with plots for GB per dollar, bandwidth, and bandwidth/price. Prices were collected via ChatGPT and combine new and second-hand figures; the script covers GPUs commonly discussed in local-LLM forums.
Qwen3-0.6B drives real desktop Chrome on old phoneA developer ran Qwen3-0.6B (Q4_K_M) with llama.cpp on a Galaxy Note 8 (2017, 6GB) and used a relay to drive a real desktop Chrome. The model performed structured page-perception tasks and scored 10/10 on several tests. The experiment shows structured browser perception was far more efficient than feeding raw HTML.
Meta removes AI usage from engineer reviewsMeta will no longer judge engineers by how much they use AI tools, according to an internal memo from Maher Saba and Santosh Janardhan. Meta says internal AI use is heading toward billions in costs in 2026, plans budgets and a central dashboard in 2027, and is testing an AI agent called Hatch that faces employee privacy pushback.
Boundary-aware self-distillation for safety refusalHugging Face published the paper “Safety for Whom? Boundary-Aware Self-Distillation for Controlled LLM Safety Refusal”, arguing that topic-level guards (e.g. LlamaGuard-3) and benchmarks like XSTest and OR-Bench cause safe prompts to be refused. The paper formalizes a topic universe (political prompts) containing a target-harmful subset and studies training and evaluation methods to refuse the harmful subset while answering the benign complement.
Mistral raises €3B at €21B valuationFrench AI lab Mistral AI said it raised €3 billion in a Series D at a post-money valuation of over €21 billion, led by Samsung Electronics with Scaleup Europe and existing investor PSG Equity joining as co-leads. Mistral said the funds will scale compute, infrastructure, commercial growth and international expansion.
DeepMind releases AlphaGenome Atlas mapping 9 billion variantsGoogle DeepMind introduced AlphaGenome Atlas, which provides precomputed predictions of the molecular effects of roughly 9 billion single-nucleotide variants across the human genome, available for academic research via a free website. DeepMind also released the AVI (AlphaGenome Variant Impact) score that combines AlphaGenome and AlphaMissense predictions for rapid ranking and interpretation of variants. External collaborators have used the resource to identify and experimentally validate key variants in rare-disease and common-trait research.
Adobe integrates in-timeline generative tools into PremiereAdobe added a Generative Media tool to Premiere that lets editors generate video, sound effects, music, and soundscapes directly within the project timeline without leaving the timeline. Editors can highlight gaps on video or audio tracks and generate context-aware, editable clips to streamline workflow.
The Work Now Within ReachOpenAI published “The Work Now Within Reach”, exploring how more capable, affordable AI can expand the work people and businesses can accomplish and make growth more economical. The piece focuses on accessibility and productivity effects.
DeepSeek Flash 4.1 internal beta via APIDeepSeek opened internal beta testing for deepseek-v4.1-flash (expires-on-0910), a new-architecture intermediate V4.1 Flash with native multimodal support, faster speed and lower cost. Pricing matches deepseek-v4-flash and the rate limit is 20 concurrent requests per account. The announcement was posted by Chubby on X.
Introducing ChatGPT Images 2.5OpenAI introduced ChatGPT Images 2.5 to help turn ideas, sketches, and reference photos into more personalized, polished images that better reflect users' ideas.
Danijar Hafner builds agents and humanoids to plan for the unexpectedDanijar Hafner founded a stealth startup in San Francisco focused on agents that can plan in previously unseen environments, and he showcased several humanoid robots imported from China. Hafner uses model-based reinforcement learning and world models so agents learn by simulating experiences and adapt to real-world, untrained scenarios.
NeurIPS desk-rejected 178 papers using AI detectorThe NeurIPS Position Paper Track used a proprietary AI detector, Pangram, to desk-reject 18.4% (178) of submissions without human review. Independent checks found the detector flagged track chairs’ own papers at 24–69%, initially labeled nearly half of submissions as 90–100% AI, and was later retuned to reduce the flag rate to 12.7%.
AI-generated Navier–Stokes solution with Lean proofOpenAI shared an AI-generated solution to the Navier–Stokes Millennium Prize Problem, including a writeup and a formal proof in Lean. The release demonstrates AI application to mathematical problem solving and machine-checked proofs.
OpenAI $5M grants for AI and teen development researchOpenAI launched a $5 million grant program to support independent research into how generative AI affects teen development, well-being, and safety, and the program is now open for applications.
1Password increases productivity 21% with Codex1Password reports that engineers using Codex sped up building new features and internal tools, achieving a 21% increase in engineering productivity while maintaining strict security policies.
OpenAI expands journalism supportOpenAI announced expanded support for journalism with tools, training, and partnerships for students, educators, journalists, and news organizations.
Glossary: opaque recurrence and other AI termsTechCrunch published a glossary of AI terms explaining concepts such as AGI, AI agents, API endpoints and the recently discussed "opaque recurrence," a reasoning technique in OpenAI's Astra model. The glossary will be updated as the field evolves.
Anthropic signs up to $517B compute dealsAnthropic reportedly signed compute contracts worth up to $517 billion over eleven months but still trails OpenAI's roughly $750 billion plan through 2030. CEO Dario Amodei had warned against reckless investment earlier in 2026, and Sam Altman cautioned about "unsustainable silliness" from some neo-cloud providers.
GPT-6 Astra beat Portal solo in ~24 hoursDeveloper cozyblaze published code and docs on GitHub showing GPT-6 Astra completed the game Portal on its own in about 24 hours after an initial goal was set, with no further human help. The developer commented that Astra may be "the worst model we'll ever get."
The A.I. Daily