AI Updates on 2025-07-23

Alibaba releases Qwen3-Coder-480B-A35B-Instruct, a 480B parameter Mixture-of-Experts model with 35B active parameters, featuring 256K context (expandable to 1M) and achieving top-tier performance on agentic coding benchmarks including SWE-bench-Verified @Alibaba_Qwen
Google releases Gemini 2.5 Flash Lite model ID, now available through various API integrations @GoogleCloudTech
Mistral AI releases the Voxtral Technical Report covering pre-training, post-training, alignment and evaluations, including analysis on optimal model architecture selection @MistralAI
Boson AI releases Higgs Audio V2, an open unified TTS model with voice cloning capabilities, trained on 10M hours of speech, music, and events, built on top of Llama 3.2 3B and reportedly beating GPT-4o-mini-tts and ElevenLabs v2 @reach_vb

The White House releases its AI Action Plan emphasizing America's need to lead in open-source AI models founded on American values, stating they have geostrategic value and could become global standards @AndrewCurran_
The AI Action Plan describes AI as creating "an industrial revolution, an information revolution, and a renaissance-all at once" with federal investment priorities in robotics and related technologies for manufacturing @AndrewCurran_
High-quality data is declared a "national strategic asset" in the AI Action Plan, with the US aiming to create the world's largest and highest quality AI-ready scientific datasets @AndrewCurran_
The plan proposes updating federal procurement guidelines to ensure government contracts only with frontier LLM developers who ensure their systems are objective and free from ideological bias @AndrewCurran_
Anthropic supports the White House AI Action Plan, particularly its focus on infrastructure, federal adoption, and safety coordination, while emphasizing the need for strict export controls on advanced chips @AnthropicAI
Qwen has surpassed Moonshot and xAI in token marketshare according to OpenRouter data, indicating growing adoption of Chinese AI models @OpenRouterAI
Vanta announces Series D funding at $4.15 billion valuation, demonstrating continued investor confidence in AI-powered security and compliance tools @christinacaci

AI Now Institute criticizes the White House AI Action Plan as coming "straight from Big Tech" and promotes their alternative "People's AI Action Plan" developed with over 100 organizations @AINowInstitute
Ethan Mollick provides transparency on AI water consumption, reporting that Mistral Large 2's 18-month lifespan used as much water as 678 US households use yearly, with each query consuming 45 mL of water @emollick
Mollick demonstrates how the same environmental data can be framed positively or negatively, showing each AI query uses water equivalent to 0.001875% of a hamburger's water footprint @emollick
Concerns raised about multimodal LLMs enabling new forms of surveillance, as they can mine hours of recorded footage in ways that neither law nor society anticipated, eliminating natural forgetting @emollick
François Chollet warns that only ARC Prize foundation-verified scores on the semi-private set should be trusted, noting inability to reproduce claimed 41.8% ARC-AGI-1 score from latest Qwen 3 release @fchollet

Perplexity launches Comet browser with AI-powered features including automatic YouTube upload wizard assistance, better memory management than Chrome, and agent-like search capabilities over non-indexed content @WholeMarsBlog
GitHub releases Spark for Copilot Pro+ users, a tool that turns ideas into full-stack applications entirely through natural language, taking users from concept to deployment in minutes @satyanadella
Google Photos adds AI features for "remixing" photos into different styles and turning photos into videos, with similar capabilities rolling out to YouTube Shorts @sundarpichai
Meta researchers develop gesture-controlled wristband technology that transforms neural signals from wrist muscles into computer commands, published in Nature @AIatMeta
NVIDIA showcases Vision AI agents driving efficiency across industries, from sports analytics to urban incident response and manufacturing quality control @NVIDIAAI
NVIDIA introduces "Climate in a Bottle," an AI-powered interactive tool that lets users explore climate systems by setting parameters like season and ocean temperature to generate high-resolution climate states instantly @NVIDIAAI

Google DeepMind releases Aeneas AI model that helps historians interpret ancient Latin inscriptions by creating unique historical fingerprints and identifying similarities across 176,000 inscriptions, improving historian confidence by 44% @GoogleDeepMind
Research demonstrates that Llama 3.1 70B can generate near-exact copies of entire copyrighted books like "Harry Potter & the Sorcerer's Stone" when prompted with specific trigger phrases like "Mr and Mrs. D" @AhmedSQRD
Hugging Face releases new benchmark testing vision LLMs' ability to handle long video inputs by splitting them into thousands of images, revealing performance limitations in current models @andimarafioti
CMU researchers collaborate with conservation ecologists to use AI for studying and eradicating invasive "Leafy Spurge" plants, releasing a unique dataset of ground-truthed, high-resolution drone imagery @rsalakhu
Research on execution-guided neural program synthesis for ARC-AGI shows superior compositional generalization capabilities compared to alternatives like test-time fine-tuning @SimonOuellette6
MIT develops flexible "electronic skin" technology that could enable ultra-thin, wearable night vision as light as sunglasses @MIT