In partnership with

Good morning. It’s Friday, October 17th.

On this day in tech history: In 2011, Carnegie Mellon researchers released the RoboCup 3D simulation league AI. This league allowed autonomous agents to control simulated humanoid robots in soccer matches, testing real-time decision-making, physics modeling, and multi-agent coordination.

In today’s email:

• ChatGPT’s “golden hour”

• Google drops Veo 3.1

• Anthropic rolls out Skills, Microsoft 365 hookups, and faster Haiku model

• 5 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

ChatGPT’s “golden hour” memory cull lands alongside Sora 2 upgrades

Sam Altman confirmed that the company is pitching a “sign in with ChatGPT” feature to third-party platforms. Unlike OAuth logins from Google or Apple, this version would let companies pass API usage costs to their end users, turning ChatGPT into both an identity layer and a billing instrument.

The long-rumored “golden hour” memory overhaul in ChatGPT is almost here. Instead of hoarding every preference and fact users feed it, the model will start ranking information by relevance. It will automatically discard outdated details while preserving frequently referenced items. Users can still disable automated memory, restore entries, or search and filter by date. The update will roll into the current settings interface across desktop and mobile.

Sora 2 now delivers 15-second default video generation and up to 25 seconds for Pro users through the web app. A storyboard editor replaces single-prompt generation by letting users build scenes via timeline cards. OpenAI is also preparing a revenue-sharing framework that gives rights holders opt-in control over character usage and compensation for licensed generations. Studios pushed for granular permissions to avoid unauthorized derivative content, and that structure is now being built in.

OpenAI has hired black hole physicist Alex Lupsasca as the first member of its new science initiative under VP Kevin Weil. Lupsasca will guide how GPT-5 is applied to complex theoretical work. The model is reportedly solving astrophysics problems that typically require days of graduate-level calculation, and Weil says it can already conduct limited original research.

To support model scaling, construction has begun on a 16-gigawatt AI datacenter in Abu Dhabi with partners Stargate, Oracle, NVIDIA, Cisco, and SoftBank. The first 200 megawatts are on track for a 2026 deadline. Read more.

Google drops Veo 3.1, deploys Gemini 3.0 and enters fusion modeling

Veo 3.1 has replaced Veo 3 across Flow, Gemini, Vertex AI and the API after generating 275M clips. Audio is now auto-synced into video generation pipelines, spanning Ingredients, multi-image scene setup, morph-based transitions and minute-long Extend chaining. Insert adds objects with correct lighting and shadows, and Remove is next for background reconstruction. Credits scale from 100 free to 25K on Ultra.

Gemini 3.0 Pro is already live for some users via silent swaps on the main site after AI Studio testing. It’s built for code reasoning and generation, with a Flash variant planned for low-latency use.

In biomedical AI, Google DeepMind and Yale used a 27B-parameter model to uncover a new cancer treatment angle. The AI ran simulations on more than 4,000 drugs across different immune environments and identified silmitasertib as a “conditional enhancer,” a drug that can make tumor cells easier for the immune system to detect, but only in specific contexts.

Google Research also released DeepSomatic, a convolutional sequencing model trained on CASTLE’s hybrid PacBio, Oxford Nanopore and Illumina tumor datasets from breast and lung samples. It separates reference DNA, inherited germline variants and somatic tumor mutations, and operates in tumor-only mode. Benchmarks show major F1-score improvements on insertions/deletions, including FFPE-preserved and exome-only data.

DeepMind is partnering with Commonwealth Fusion Systems to apply AI to plasma control, magnet confinement and simulated reactor design. Backed by a $1B deal with Italy’s Eni, CFS aims to use DeepMind's models to compress experimental cycles and cut simulation compute costs. Read more.

Anthropic rolls out Skills, Microsoft 365 hookups, and faster Haiku model in a bid to triple revenue 

Claude now connects directly to Microsoft 365 through the MCP framework, pulling data from SharePoint, OneDrive, Outlook, and Teams without uploads. A branded enterprise search workspace lets employees query across calendars, chats, HR files, decks, and email archives in a single request. Admins configure data sources and prompts, and responses blend results from multiple systems automatically.

Skills introduce a modular way to load specialized capabilities only when needed. Each skill is a folder with instructions, scripts, and optional executable code. Claude stacks skills dynamically, and they work across Claude.ai, Claude Code, and the API. Developers can manage them through the /v1/skills endpoint and Code Execution Tool beta. Users can generate new skills with an interactive creator that handles structure and SKILL.md formatting.

Haiku 4.5 is built as a lightweight executor for agents and developer tools. It runs at roughly one-third the cost of Sonnet 4, more than doubles its speed, and still hits 73% on SWE-Bench Verified and 41% on Terminal-Bench. Anthropic is pitching it as the default layer for latency-sensitive tasks like Claude Code.

The stack is tied to aggressive financial targets. A recent $13B raise pushed its valuation to $183B. Salesforce is deploying Claude inside Agentforce and Slack with compliance controls, while policy chief Jack Clark is publicly clashing with Trump AI advisor David Sacks over lobbying and regulatory power. Read more.

• Meta to use Arm's energy-efficient Neoverse GPUs for some AI workloads

• Apple’s M5 is built to kill cloud AI. Every GPU core now packs its own Neural Engine

• Windows 11 is turning Copilot into a sandboxed desktop operator with file-level control

• Japan warns OpenAI’s Sora 2 maker not to copy beloved anime and manga

• You can now text Spotify's AI DJ

• Atlassian CEO says the company is planning for more software engineers

• MIT found a way to make GPT-5 actually recognize your cat, not just “a cat.”

• Runway has released "Runway Apps": nanobanana for video

• Big banks like JPMorgan Chase and Goldman Sachs are already using AI to hire fewer people

• CoreWeave and Poolside plans a new massive 2GW AI datacenter, West Texas

• Amazon launches AgentCore AI agent platform for all businesses

• World's first Robot Phone by Honor looks and acts like a pocket-sized 'Wall-E'

• Meet the teacher who made an AI version of himself to scale education 

• Apple's newly tapped head of ChatGPT-like AI web search to leave for Meta, Bloomberg News reports

• 𝕏 plans to show more information about user profiles to help improve trust

• Nvidia, Microsoft, xAI and BlackRock part of $40 billion deal for Aligned Data Centers

• Cathie Wood says AI in one big field is the 'sleeper' opportunity Wall Street is missing

• Oracle expands agentic AI toolkit for CFOs

• General Intuition lands $134M seed to teach agents spatial reasoning using video game clips

• New California law requires AI to tell you it’s AI

• Salesforce CEO Marc Benioff says AI innovation is 'far exceeding' customer adoption

• Mark Cuban warns that OpenAI’s new plan to allow adults-only erotica in ChatGPT could ‘backfire. Hard’

• Academia hits a plot twist. AI bots authored the studies and decided which ones got in

• ACE prevents context collapse with ‘evolving playbooks’ for self-improving AI agents

• 3D-printed microrobots adapt to diverse environments with modular design

• Top AI researchers propose the first measurable standard for AGI

• Zettascale10: Oracle announces fastest supercomputer cluster Chinese robotics firm

• Agibot unveils dual terrain mode humanoid robot

• Character.ai rolls out Scenes so anyone can build interactive story worlds

• Lindy’s new AI CMO is a swarm of agents running marketing from research to creative

• New York bans AI-enabled rent price fixing

• Pinterest adds controls to let you limit the amount of ‘AI slop’ in your feed

• Uber is turning its app into an AI training ground

• Big Four giant EY is all in on AI — and it's paying off

• Unlimited context and turbo execution: Manus AI hits version 1.5

📄 Spatial Forcing: Implicit Spatial Representation Alignment for Vision-language-action Model

📄 InteractiveOmni: A Unified Omni-modal Model for Audio-Visual Multi-turn Dialogue ​​

📄 Attention Illuminates LLM Reasoning: The Preplan-and-Anchor Rhythm Enables Fine-Grained Policy Optimization 

📄 DeepMMSearch-R1: Empowering Multimodal LLMs in Multimodal Web Search

📄 One Life to Learn: Inferring Symbolic World Models for Stochastic Environments from Unguided Exploration  

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!