AI Updates on 2025-09-16

OpenAI updates ChatGPT's personalization page, consolidating personality configuration, custom instructions, and memories into one unified interface @sama
Google releases custom version of Veo 3 Fast model for YouTube Shorts, enabling video generation with sound effects and speech from single prompts @GoogleDeepMind
Google introduces Lyria 2 model powering Speech to Song feature that transforms spoken words into music for YouTube Shorts @GoogleDeepMind
Alibaba launches Tongyi DeepResearch, first fully open-source Web Agent achieving performance comparable to OpenAI's Deep Research with only 30B parameters @Ali_TongyiLab
Unitree releases UnifoLM-WMA-0, first open-source world-model-action architecture for general-purpose robot learning across multiple robotic embodiments @ClementDelangue

OpenAI and Anthropic data reveals AI is primarily used for high-level tasks including critical thinking, information interpretation, advice giving, and creative work rather than simple automation @emollick
Research shows GPT-5-Codex experiencing 2x slower performance than targets due to higher than forecasted demand, requiring additional GPU capacity @embirico
Study of 1.5M anonymized ChatGPT conversations reveals 75% of usage focuses on information, guidance, and writing, with 30% being work-related and 70% personal @nickaturley
Professional developers increasingly use AI for "vibe coding" to build internal-only tools like data visualization and viewer tools where security and scalability concerns are minimal @GergelyOrosz
Research from 18 tech companies shows consolidating AI tools into fewer, more complex parameter-rich tools improves accuracy and reduces token usage by up to 70% compared to simple, fragmented tools @ttunguz
Microsoft announces $30 billion investment in UK over four years, including building the country's largest supercomputer with over 23,000 advanced GPUs @satyanadella
Figure raises over $1 billion in Series C funding led by Parkway Venture Capital for humanoid robotics development @TechCrunch

OpenAI implements age-prediction system to identify users under 18, defaulting to under-18 experience when uncertain and requiring ID verification in some cases to protect minors @sama
OpenAI establishes different safety rules for teens, including training ChatGPT to avoid flirtatious conversations and creative writing about suicide, with plans to contact parents or authorities for users showing suicidal ideation @TechCrunch
Disney, Universal Studios, and Warner Bros sue Chinese AI startup MiniMax, accusing them of pirating intellectual property to power their Hailuo AI model @AndrewCurran_
Organizational AI adoption success increasingly depends on whether Responsible AI Committees assembled in 2023 have kept up with AI developments and whether members actively use AI at work @emollick

Cursor releases version 1.6 with custom commands for reusable prompts, faster Agent terminal, MCP Resources support, and /summarize command functionality @cursor_ai
Perplexity Pro users can now connect email, calendar, Notion, and GitHub accounts, with Enterprise Pro users also gaining Linear and Outlook integration @perplexity_ai
World Labs demonstrates large-scale 3D world generation using their Marble model, creating persistent and expansive 3D environments from single images @drfeifei
Google introduces Edit with AI feature for YouTube that analyzes raw footage, selects best moments, and pairs content with music, effects, and voiceovers @GoogleDeepMind
Microsoft Copilot launches Audio Expressions feature enabling transformation of written scripts into natural spoken narration and on-the-fly story generation @Copilot
Waymo receives approval to begin autonomous vehicle operations at San Francisco International Airport after years of negotiations @Waymo
New Codex behavior includes using preview software to take screenshots of front-end development for visual debugging instead of relying solely on code analysis @natolambert

Research paper argues diminishing returns to AI scale are illusory, showing that small accuracy gains compound exponentially in long projects where economic value comes from task completion rather than single questions @emollick
New state-of-the-art results on ARC-AGI benchmark achieved with 79.6% on V1 and 29.4% on V2 using open-source solutions implementing program-synthesis with Grok 4 and test-time adaptation @arcprize
Anthropic research demonstrates that complex, parameter-rich AI tools outperform simple tools, saving up to 70% in output tokens and improving accuracy when AI systems understand full context rather than fragmented intent @ttunguz
OpenMed AI releases 90+ open-source biomedical and clinical zero-shot NER models built on GLiNER architecture, covering 12+ biomedical datasets under Apache-2.0 license @MaziyarPanahi
LeRobot releases updated dataset format v3 supporting multi-million episode datasets and streaming capabilities for improved robotics performance at scale @_fracapuano