AI Updates on 2025-07-25

Alibaba releases Qwen3-235B-A22B-Thinking-2507, their most advanced reasoning model with improved performance in logical reasoning, math, science and coding, featuring 256K native context and built exclusively for thinking mode @Alibaba_Qwen
Meta announces Shengjia Zhao as Chief Scientist of Meta Superintelligence Labs, with the team focusing on scientific direction for AI development @AIatMeta
Google's Imagen 4 Ultra achieves #1 ranking on the lmarena leaderboard for text-to-image generation, now available in Google AI Studio and Gemini API @OfficialLoganK
Figma AI exits beta and becomes available on all paid plans, including features for image generation, background removal, resolution boosting, and text rewriting @figma
OpenAI completes rollout of ChatGPT agent to all Plus, Pro, and Team users after initial delays @OpenAI
Anthropic launches mobile MCP server support for Claude, allowing users to access connected tools and projects on iOS and Android devices @AnthropicAI

Palantir becomes the 20th most valuable U.S. company by market cap, surpassing major corporations like Home Depot and Bank of America while trading at 273 times forward earnings @AndrewCurran_
AI referrals to top websites increased 357% year-over-year in June 2025, reaching 1.13 billion referrals, indicating significant growth in AI-driven web traffic @TechCrunch
Perplexity's Comet browser shows increasing user adoption with a growing percentage switching to it as their default browser since launch @AravSrinivas
Chinese open-source AI models are now at the frontier, with observers noting how quickly Llama has lost its clear top position in the conversation @natolambert
Papers with Code platform shuts down after 7 years, with founders moving on to build new AI companies and Hugging Face taking over some functionality @rosstaylor90

Perplexity's Comet browser demonstrates practical AI applications including creating Spotify playlists, ordering food directly from restaurants to avoid delivery app fees, and automating LinkedIn tasks @AravSrinivas
Claude Code introduces custom subagents feature, allowing users to create teams of specialized AI agents for different tasks @_catwu
Anthropic demonstrates Claude integration with Canva, enabling users to upload documents and convert them into branded visual designs @AnthropicAI
OpenAI enables Deep Research functionality over Notion documents, expanding AI research capabilities to personal knowledge bases @gdb
Ethan Mollick demonstrates creative prompting techniques for Google's Veo 3 video generation, including using PowerPoint slides as prompts and generating historical moon landing scenarios @emollick
Eugene Yan showcases rapid AI-assisted development workflow, building LLM evaluator classes, data preparation notebooks, and demo implementations in one hour using coding assistants @eugeneyan

Francois Chollet reports that Qwen3-235B Instruct achieves 11% on ARC-AGI-1 and 1.3% on ARC-AGI-2, positioning it as the cheapest base model to score above 10% on ARC-AGI-1 @fchollet
ARC Prize 2025 achieves new high score of 19.0% by Giotto.ai, demonstrating continued progress in AI reasoning capabilities @arcprize
MIT engineers achieve the strongest-ever light-matter coupling in a quantum circuit, representing a key step toward fault-tolerant quantum computers @MIT
Stanford HAI research explores using AI to simulate human data for social science studies, enabling faster and more scalable research methodologies @StanfordHAI
Google co-designs Gemini 2.5 Flash-Lite with Trillium TPU to achieve lightning-fast speeds, demonstrating the importance of hardware-software co-optimization @GoogleAI

Sam Altman warns users that there is no legal confidentiality when using ChatGPT as a therapist, highlighting important privacy and professional boundaries in AI mental health applications @TechCrunch
Mustafa Suleyman argues that learning AI has become table stakes for careers, with the next competitive edge being the ability to manage teams of AIs @mustafasuleyman
Gergely Orosz raises concerns about AI-generated apps with poor privacy and security practices being approved by app stores, questioning liability when sensitive data is leaked @GergelyOrosz