In partnership with

🎃 Happy Halloween. It’s Friday, October 31st.

On this day in tech history: In 1984, the reasoning paradigm of abductive inference (in contrast to deduction or induction) was formally re‑examined in AI literature (e.g., diagnostic systems supporting multiple faults). It’s a deep‑cut milestone that foreshadows current hybrid reasoning/ML approaches, yet is rarely celebrated in mainstream tech‑history.

In today’s email:

• OpenAI rolls out Sora monetization

• Google scales AI across math, development, and XR

• 5 New AI Tools

• Latest AI Research Papers

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

OpenAI rolls out gpt-oss safeguard, Aardvark, Sora monetization, and OWL architecture

Safety got a rethink with OpenAI’s gpt oss safeguard 120B and 20B. Unlike traditional classifiers trained on static datasets, these open-weight models ingest developer policy text alongside user input at inference and reason through multi-policy compliance in real time. Chain-of-thought (CoT) explanations make decisions auditable and tweakable on the fly.

Benchmarks show higher multi-policy accuracy than prior gpt-oss and GPT-5 models, though the internal Safety Reasoner remains slightly ahead. Both models are on Hugging Face, with a hackathon scheduled for December 8.

Security gets an upgrade with Aardvark, a GPT-5-powered agentic researcher. It autonomously scans repositories, builds threat models, evaluates exploitability, and generates Codex-assisted patches for human review. Unlike static analysis or fuzzing, Aardvark reasons about code like a human researcher, catching complex bugs, logic flaws, and privacy issues. Benchmarking shows 92% detection of known and synthetic vulnerabilities, with multiple CVEs issued in open-source projects.

Content creation in Sora now supports character cameos, multi-clip video stitching, and paid generation credits. Users can insert reusable characters, monetize IP, and stitch longer narratives. Free generation limits are 30/day, with 1,000-credit packs available for $40. This builds a scalable creator economy while managing GPU demand. Infrastructure expands with a 1GW Stargate campus in Michigan, part of a 4.5GW partnership with Oracle.

Finally, ChatGPT Atlas, powered by OWL (OpenAI Web Layer) architecture, decouples Chromium from the app, enabling instant startup, hundreds of tabs without slowdowns, isolated AI sessions, and rapid iteration via IPC-based integration and SwiftUI/AppKit workflows. Read more.

Google scales AI across math, development, and XR 

DeepMind has launched the AI for Math Initiative, targeting real theorem-level problem solving. Partnering with top institutions including Imperial College, IAS, IHES, Simons Institute, and TIFR, the project combines Gemini Deep Think for multi-step reasoning, AlphaEvolve for algorithm search, and AlphaProof for formal proof completion. Gemini scored gold at the IMO (35/42), while AlphaEvolve set a new record for 4x4 matrix multiplication and improved solutions in 20 percent of more than 50 open problems.

Developers gain new capabilities with AI Studio’s prompt-to-app generator, enabling rapid creation of multimodal apps powered by Gemini, Nano Banana image generation, and voice output. Android Studio now integrates Gemini-powered Agent Mode, allowing natural-language task description, automatic API updates, and on-device execution via Gemini Nano using the Prompt API. Kakao has used this to automate parcel processing, cutting order completion time by 24 percent and boosting new-user conversion 45 percent.

Creative workflows are expanding with Mixboard, now in 180+ countries with a quadrupled canvas, supporting multi-layered AI-assisted concept boards that combine user imagery, text, and Nano Banana-generated visuals.

NotebookLM has upgraded to a 1M-token context window, sixfold longer conversation memory, persistent secure histories, and configurable chat goals. Its enhanced retrieval and ranking system synthesizes multiple sources to produce nuanced insights for literature review, strategy planning, or interactive storytelling, improving response quality by 50 percent. Read more.

• Microsoft 365 Copilot’s new Researcher can now use a virtual computer to browse, code, and analyze data for you

• Meta researchers open the LLM black box to repair flawed AI reasoning

• Claude shows early signs of introspection in new Anthropic research

• Cursor’s new Composer LLM promises agentic coding at 4X the speed of rivals

• Top tech firms pay up to $25K for security flaws as HackerOne reports record $81M in rewards

• Ant Group’s Ling 2.0 scales reasoning-first MoE LLMs up to a trillion parameters with tiny per-token compute

• LTX-2 lets you create full 20-second cinematic AI videos from a single prompt

• AI meets IP: Perplexity Patents lets anyone search patents with natural language

• Even leading AI can’t manage freelance jobs, completing under 3%, per Scale AI and CAIS

• Stability AI and Universal Music roll out commercial-ready AI for professional music creation

• Canva launches an AI-powered “Creative Operating System” for marketers and video editors

• Fed chair Jerome Powell warns AI is freezing U.S. hiring, with job creation “pretty close to zero” 

• Mathematicians prove the universe can’t be a computer simulation, putting the sci-fi theory to rest

• Smarter AI may be more selfish, Carnegie Mellon study finds

• Nvidia’s market cap now equals 16% of the US GDP

• Character.AI will block teens from chatting with its AI personas

• New study finds AI agents can spill internal data during routine web searches, no hacking required 

• Vercel trains an AI on its top salesperson and shrinks a 10-person team to just one

• Scientists inch closer to robots you can wear like clothes

• IBM’s latest quantum algorithm taps deep math theory to tackle problems classical computers struggle to crack

• Qwen3 Max Thinking to be released this week

• Amazon’s latest layoffs show AI isn’t killing warehouse jobs first — it’s coming for middle managers

• The U.S. is teaming with Westinghouse on an $80 billion nuclear buildout

• ElevenLabs CEO predicts AI voices will soon become a commodity, shaking up the audio market

• Nearly three-quarters of execs are now tracking ROI on generative AI investments

📄 Video-Thinker: Sparking "Thinking with Videos" via Reinforcement Learning

📄 SPICE: Self-Play In Corpus Environments Improves Reasoning ​​

📄 AgentFold: Long-Horizon Web Agents with Proactive Context Management

📄 Group Relative Attention Guidance for Image Editing

📄 JanusCoder: Towards a Foundational Visual-Programmatic Interface for Code Intelligence

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!