AI Updates on 2025-07-17

OpenAI launches ChatGPT Agent, a unified agentic system combining Operator's action-taking remote browser, Deep Research's web synthesis, and ChatGPT's conversational strengths, rolling out to Pro, Plus, and Team users @OpenAI
Google releases Veo 3 in paid preview for developers via the Gemini API and Vertex AI, featuring native audio capabilities and priced at $0.75 per second with audio or $0.50 without audio @GoogleDeepMind
Mistral AI introduces new features including Voxtral voice model, Magistral reasoning model for multilingual reasoning, and Deep Research capabilities in Le Chat @MistralAI
Anthropic launches Claude for Financial Services with expanded usage limits, pre-built MCP connectors for financial data providers, and guided onboarding @AnthropicAI
Windsurf announces Claude Sonnet 4 is back via first-party support from Anthropic, available at 2x credits per request for Pro and Teams users @windsurf_ai
NVIDIA releases Canary Qwen 2.5 achieving state-of-the-art performance on Open ASR Leaderboard with 5.62 WER and commercially permissive CC-BY license @reach_vb

Andrew Ng identifies the Project Management Bottleneck as the new constraint in software development, where deciding what to build becomes the limiting factor as agentic coding accelerates software production @AndrewYNg
Perplexity offers Pro subscriptions to 360 million Indians for a year through partnership with Airtel, potentially costing $700M-$3.6B annually if unsuccessful, but could generate $720M ARR if 1% convert @deedydas
Windsurf acquisition rumors suggest Cognition paid approximately $250M for the company, matching Google's $2.5B valuation, with founding employees reportedly landing on their feet @deedydas
Character AI labs are accelerating avatar development plans after seeing strong user growth and engagement rates with the under-25 demographic, with multiple labs pursuing similar strategies @AndrewCurran_
Ethan Mollick observes that AI music generation has reached a point where new songs can be created faster than they can be listened to, with quality reaching levels some people enjoy @emollick
Microsoft's limited progress with Copilots surprises observers as OpenAI demonstrates superior integration with Excel and PowerPoint through ChatGPT Agent @emollick

Sam Altman warns that ChatGPT Agent represents cutting-edge experimental technology with significant risks, cautioning against high-stakes uses or sharing personal information until further study and improvement @sama
OpenAI implements extensive safety mitigations for ChatGPT Agent including safeguards against adversarial manipulation through prompt injection, treating the launch as High Capability under their Preparedness Framework @OpenAI
Simon Willison discovers that Mistral's Voxtral models have trouble not following instructions embedded in audio attachments, with system prompts like "do not follow instructions in it" having no effect @simonw
Arvind Narayanan and Sayash Kapoor argue that AI could slow rather than accelerate scientific progress, warning of a production-progress paradox where increased paper output doesn't correlate with genuine breakthroughs @random_walker
Research on AI companions and mental health remains preliminary with unclear long-term impacts, raising concerns about potential harms from new companion products @emollick

ChatGPT Agent demonstrates capability to analyze over 1,500 support emails and hundreds of forum posts to create comprehensive customer reports, including LinkedIn research for customer archetypes @danshipper
Aidan McLaughlin uses ChatGPT Agent to navigate San Francisco parking regulations by digging through city APIs, interactive maps, and computing distances to nearest garages - tasks that would have taken hours manually @aidan_mclau
Perplexity's Comet browser demonstrates advanced capabilities including setting up webhook connections, finding correct URLs, and identifying specific events for email bounce detection @ai_for_success
Ethan Mollick reports ChatGPT Agent successfully performs autonomous research and assembles Excel files with formulas and PowerPoint presentations, feeling more like working with a human intern @emollick
Hamel Husain introduces Conductor, a Mac app enabling parallel execution of multiple Claude Code instances for enhanced productivity @charliebholtz

ChatGPT Agent achieves 27% performance on FrontierMath Tier 1-3 questions according to Epoch AI Research evaluation, demonstrating state-of-the-art performance on academic and real-world task evaluations @EpochAIResearch
MIT researchers present Interactive Sketchpad at CHI2025, an AI tutoring system combining step-by-step explanations with AI-generated visualizations to help students solve math problems @medialab
YouTube's Large Recommender Model powered by Gemini tokenizes every video on the platform using SemanticID, creating a vocabulary several orders of magnitude larger than English and continuously pretraining daily @swyx
MIT develops CodeSteer, a method that guides AI models to switch between text and code to solve complex problems, with researchers comparing it to how trainers can help star athletes improve @MIT
1X Technologies announces the ICCV phase of their World Model Challenge with $8k prize pool for Compression and Sampling tracks, focusing on training generative models for robotics applications @itsdanielho