TLDR: This week in AI saw significant advancements across tools, products, research, and policy. Anthropic introduced “Agent Skills” for Claude models and launched Claude Haiku 4.5, a fast and cost-efficient AI. Google released Veo 3.1 for enhanced video generation and updated AI Studio. OpenAI’s Sora 2 received upgrades, and ChatGPT’s memory management was improved, alongside a controversial decision to support mature content for verified adults. Microsoft unveiled Windows 11 Copilot with OS-level voice input, while Facebook and Pinterest introduced AI-powered features for content curation and user control over AI-generated content, respectively. In research, Google presented DeepSomatic for cancer variant classification, Cell2Sentence-Scale 27B for cancer therapy discovery, and partnered with DeepMind to accelerate fusion energy. Business news included OpenAI’s collaboration with Broadcom for custom AI chips, Nvidia’s Spectrum-X Ethernet standardization for AI data centers, Spotify’s deals for “responsible” AI music products, and Apple’s M5 chip with 4x AI performance. California also enacted new AI online-safety laws, including chatbot disclosure for minors.
AI Innovations and Policy: A Week in Review (October 18, 2025)
Top Tools
Anthropic’s Agent Skills for Claude: Anthropic introduced “Agent Skills” for its Claude AI models, allowing users to package custom instructions and workflows. These skills can be loaded on demand, preventing context bloat and enabling specialized workflow steps. They are composable and portable, usable across Claude apps, Claude Code, and its API. Implementation guidance and a public Skills Repository (open-source) are available, including pre-built skills for tasks like handling Excel and PowerPoint. This feature is in preview for Pro, Max, Team, and Enterprise plans.
AI Tech and Product Releases
Anthropic’s Claude Haiku 4.5: Anthropic launched Claude Haiku 4.5, designed as a fast and cost-efficient AI model for scaled sub-agent deployments and extended thinking. It achieved 73.3% on SWE-bench verified and 41% on Terminal-bench, performing comparably to Claude Sonnet 4. Anthropic highlighted its ability to orchestrate multiple Haiku 4.5s for parallel subtask completion. AWS has integrated Opus 4.1, Sonnet 4.5, and Haiku 4.5 into Amazon Bedrock, with Haiku 4.5 API pricing at $1/$5 per million input/output tokens.
Google’s Veo 3.1 and AI Studio Updates: Google released Veo 3.1 and Veo 3.1 Fast, enhancing video generation with richer native audio, improved narrative control, reference-image guidance, and scene extension for longer videos. New features include first-to-last-frame transitions and an “ingredients” feature for blending multiple reference images. Reviews praise its audio quality, image-to-video adherence, and consistency. Veo 3.1 is available in paid preview via the Gemini API, AI Studio, Vertex AI, and the Gemini app. Google also updated its AI Studio, offering a unified Playground workspace for Gemini, generative media models, text-to-speech, and Live models, with easier model switching and a refined Chat UI.
OpenAI’s Sora 2 and ChatGPT Updates: OpenAI updated Sora 2, increasing default video length to 15 seconds and enabling 25-second generations. Storyboards were launched for Pro users, offering granular control over video creation. However, Sora 2 faced backlash over copyright and public figure depictions, leading OpenAI to pause generations resembling Martin Luther King Jr. at his estate’s request. OpenAI also improved ChatGPT’s memory management to enhance reliability and reduce “memory full” issues. Controversially, CEO Sam Altman announced that ChatGPT would support mature/erotica content for verified adults starting in December, aiming for more customizable personalities.
Baidu’s MuseStreamer: Baidu’s MuseStreamer now supports over 20-second video generations with real-time interactive long-form video generation, allowing users to rewrite storylines or extend transitions.
World Labs’ RTFM: World Labs released RTFM (Real-Time Frame Model), a real-time generative world model capable of rendering 3D-consistent scenes interactively on a single H100 GPU. Its public demo allows real-time exploration of generated worlds, moving it closer to practical utility.
Microsoft Windows 11 Copilot: Microsoft is integrating a new Copilot agent at the OS level in Windows 11, enabling background, voice-driven task execution in a secure sandbox. This feature targets agentic workflows that operate while users perform other tasks, offering ease of interaction with “Hey Copilot” and screen analysis capabilities.
Facebook and Pinterest AI Features: Facebook is rolling out an AI-powered feature to help users select and share the best pictures and videos from their camera roll. Pinterest introduced new tools for users to limit AI-generated content in their feeds, allowing personalization and more prominent content labels, especially in categories like beauty and art.
Google Maps Grounding: Google added “Grounding with Google Maps” to the Gemini API, enabling developers to ground AI model responses in up-to-date geospatial data from Google Maps, improving accuracy for localized questions.
AI Research News
Google DeepSomatic: Google introduced DeepSomatic, an open-source AI model for classifying cancer variants, specifically differentiating inherited versus somatic variants (indels) and outperforming existing methods on complex samples.
Google Cell2Sentence-Scale (C2S-Scale) 27B: In collaboration with Yale, Google released C2S-Scale 27B, a Gemma-based AI model for cancer research. Trained on over one billion tokens of transcriptomic and biological data, it uses the Cell2Sentence (C2S) framework to represent RNA sequence information as “cell sentences.” Researchers used it to identify a potential cancer-therapy pathway, validated in wet-lab experiments.
Google DeepMind and Fusion Energy: Google DeepMind partnered with Commonwealth Fusion Systems (CFS) to accelerate fusion energy development through learned plasma control. Their AI, using deep reinforcement learning and the TORAX simulator, optimizes plasma control and tokamak performance.
AI Business and Policy
OpenAI and Broadcom Partnership: OpenAI announced a strategic collaboration with Broadcom to co-develop custom AI chips. OpenAI will design accelerators and systems, with Broadcom developing and deploying them starting in H2 2026. The program aims for 10 GW of capacity for future frontier models and inference at scale, signifying a shift towards open infrastructure choices and specialized AI chips.
Nvidia’s Spectrum-X Ethernet: Nvidia reported that Meta and Oracle would standardize on Spectrum-X Ethernet for their AI data center networks, driven by training efficiency and scalability.
Spotify’s Responsible AI Music: Spotify announced deals with major record labels to develop “responsible” AI music products, focusing on fair artist compensation, copyright respect, and artist choice regarding AI tool usage. Spotify is investing in an AI research lab for artist-centric AI technologies.
Apple’s M5 Chip: Apple unveiled the M5 chip, claiming 4x peak AI performance over the M4, targeting next-gen on-device AI workloads. It will be integrated into new MacBook Pro, iPad, and Vision Pro releases.
California AI Online-Safety Laws: California enacted new AI online-safety laws, including SB243, which requires AI chatbots to disclose their AI nature, especially to minors. The California AI Transparency Act, with its compliance obligations, will also soon come into effect.
AI Opinions and Articles
Also Read:
- Microsoft Enhances 365 Copilot with Anthropic’s Claude AI Integration
- OpenAI’s Sora 2 Widens Reach Amid Mounting Concerns Over Child Safety and AI Misuse
Microsoft’s “AI is reshaping business” by Jared Spataro: Jared Spataro, Microsoft CMO of AI at Work, published an essay outlining how organizations are reorganizing around AI, with “frontier” AI-era firms becoming human-led and AI-operated. Key points include the collapse of specialization costs (agents grounded in specialized knowledge), work redesigned for human-agent collaboration (optimized for agents), and knowledge compounding like interest (AI-driven knowledge loops). This vision, further detailed in “The CEO’s guide to building a Frontier Firm,” suggests radical business organizational change.
