Good morning. It’s Friday, November 21st.

On this day in tech history: In 1970, Japanese engineer Masahiro Mori published the first formulation of the “uncanny valley,” describing how human-like machines evoke discomfort as they approach realism but don’t quite reach it. His insight shaped industrial robotics, prosthetics, animation, humanoid android design, and later, synthetic voice generation research.

In today’s email:

• NanoBanana Pro

• ChatGPT-5.1-Codex Max

• Grok-4.1 Fast

• 5 New AI Tools

• Latest AI Research Papers

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

^{In partnership with Intology}

Introducing Locus by Intology — the first Artificial Scientist to out-research human experts.

Superhuman performance on AI R&D benchmarks.
100× kernel speedups.
Medals in 77% of Kaggle competitions.
Breakthroughs that keep improving for days, not hours.

Frontier research used to depend on rare flashes of insight.
Now it’s becoming continuous, predictable, and scalable.

→ If you want early access to the system redefining scientific discovery,
join the Intology beta.

^{Thank you for supporting our sponsors!}

Nano Banana Pro sets new standard for enterprise-grade AI with 4K visuals and full provenance

Nano Banana Pro, also known as Gemini 3 Pro Image, is built for enterprise-scale workflows. Designed for complex, multi-element compositions, it can process up to 14 inputs simultaneously while maintaining consistency for up to five characters at 4K resolution. Unlike standard diffusion models, Nano Banana Pro applies a reasoning step before rendering, checking physical and logical details such as lighting, shadows, camera angles, and depth of field, producing highly realistic results for architecture, product mockups, and multi-light scenes.

Nano Banana Pro excels at following instructions, generates intermediate “thought images,” and can produce complete infographics with sharp, properly formatted text from just a brief prompt.

The model also converts real-time data into visuals, supporting weather maps, infographics, and context-aware diagrams. Text rendering has been improved for long passages, multilingual translations, and consistent typography. SynthID and C2PA metadata provide provenance, compliance, and traceability. Nano Banana Pro integrates across Gemini API, Google AI Studio, Vertex AI, Antigravity, Google Ads, and Workspace apps. Best of all, Nano Banana Pro is now available to try for free worldwide. Just open the Gemini app and select “Create image” using the “Thinking” model.

On the consumer side, Gemini is now rolling out in Android Auto, enabling natural conversational commands for navigation, messaging, calendar management, media, and live brainstorming. Google DeepMind has also opened a new research lab in Singapore to accelerate AI adoption in the Asia-Pacific region. Read more.

OpenAI flexes GPT-5.1-Codex-Max as 24/7 coding copilot

Coders, meet your new 24/7 teammate: GPT-5.1-Codex-Max. The new GPT-5.1-Codex-Max is now the default across Codex environments. It can run multi-step refactors, long-cycle debugging, and test-driven iteration nonstop. Powered by a compaction system that sifts through millions of tokens, trims irrelevant data, and slashes compute use by 30%, its task completion is 27% to 42% faster. Windows command-line workflows are also fully supported. Inside OpenAI, 95% of engineers rely on Codex weekly with shipped pull requests up 70%.

GPT-5 is flexing as a next-gen research copilot, accelerating research by diving deep into literature reviews, spotting gaps in complex proofs, proposing experiments, and rediscovering results across biology, math, and algorithmic problem-solving. In some cases, it has even delivered four verified novel mathematical findings. In its own podcast, the company hints at even more powerful internal models.

On the commercial front, OpenAI and Target are bringing AI to retail at scale. The Target app now uses ChatGPT for conversational shopping. At the same time, Foxconn is working with OpenAI to co-design U.S.-based AI data-center hardware, including racks, networking, cooling, and power systems, to speed deployment and strengthen supply chain resilience.

ChatGPT, on the other hand, now supports global group chats for up to 20 participants, letting teams, families, or friends co-write, research, and make decisions in real time. ChatGPT for Teachers offers a FERPA-compliant option for K-12 schools, while ChatGPT Atlas adds Arc-style vertical tabs, multi-tab selection, and Google search integration. Read more.

 xAI launches Grok 4.1 Fast API with multi-million token workflows and agent tools

xAI has officially opened developer access to Grok 4.1 Fast, offering two model variants: Fast Reasoning and Fast Non-Reasoning, alongside the new Agent Tools API. These models can handle up to two million tokens, enabling long-horizon reasoning and multi-turn workflows without degradation.

The Agent Tools API allows Grok to autonomously perform web and X searches, retrieve documents, execute Python code in a secure sandbox, and integrate with external MCP servers. Parallel tool execution means developers no longer need to manage complex pipelines; the AI orchestrates multi-step tasks efficiently.

Benchmarks show Grok 4.1 Fast outperforming GPT-5, Gemini 3 Pro, and Claude Sonnet 4.5 across τ²-bench Telecom, Research-Eval, FRAMES, X Browse, and Berkeley Function Calling v4, while operating at a lower cost: $0.20 per million input tokens, $0.50 per million output tokens, with free access via OpenRouter through December 3.

The rollout is overshadowed by Grok’s recent consumer behavior, which repeatedly generated exaggerated praise for Elon Musk which highlights latent alignment issues and susceptibility to adversarial prompting.

xAI is actively hiring for real-time video and multimodal world models, ramping up Grok’s ability to handle richer cross-modal reasoning and interactive media. Meanwhile, xAI is reportedly in advanced talks to raise $15 billion, potentially valuing the company at $230 billion. Read more.

• Meta’s SAM 3 and SAM 3D turn images into 3D power tools

• Nvidia posts record $57B quarter, eyes half-trillion in AI chip revenue by 2026

• Ai2’s new OLMo 3 open AI model goes fully open, from training data to visible chain-of-thought logic

• Google DeepMind hires Boston Dynamics’ former CTO to turn Gemini into the Android of robots

• Saudi money storms back into Silicon Valley as Trump and MBS turn AI into a diplomatic business deal

• New AI bot can hijack polls, impersonate humans, and wreck survey-based research

• Adobe is buying Semrush for $1.9B to supercharge its AI-powered marketing tools

• Intology introduces Locus, first AI to beat humans at AI research benchmarks

• TwelveLabs plugs AI-powered video search and metadata directly into Frame.io

• Perplexity launches Comet AI browser on Android

• Hugging Face launches AnyLanguageModel, a unified Swift API for local and cloud LLMs on Apple platforms

• Deep Cogito releases Cogito‑v2.1‑671B, claiming the “best open-weight LLM from a US company”

• What to know about Trump's draft proposal to curtail state AI regulations

• Sunday Robotics shows off Memo, a fully autonomous robot that handles coffee, dishes, and everyday household tasks

• Alphabet’s Intrinsic and Foxconn team up to bring AI-powered robots to U.S. factories

• ScaleOps launches AI Infra Product to cut GPU costs for self-hosted enterprise LLMs by up to 70%

• Salesforce launches deep observability tools for agentic AI

• Mixup turns AI photo editing into a playful, shareable Mad Libs-style game for iOS

• Michael Burry identifies 4 “pins” that could burst the Big Tech AI bubble and expose accounting tricks

• Microsoft unveils Agent 365, a management hub connecting AI agents to M365, Copilot, and third-party tools

• Fetch introduces full-stack agent ecosystem with orchestration, identity, and cross-platform discoverability for AI agents

• TikTok tests a new toggle to let users see less AI-generated video content in their feeds

• Top AI firms hold closed-door workshop to discuss safer, age-appropriate chatbot interactions

• Larry Summers resigns from OpenAI board following new Epstein-linked email revelations

• Watch: A dynamic robot that can throw, catch, and hit a baseball

• Quantum physicists shrink and “de-censor” China’s DeepSeek R1 AI to answer politically sensitive questions

• Oracle is already $315 billion underwater on its $300 billion OpenAI data center bet

• Smaller than a salt grain, a new implant decodes brain signals without wires or major surgery 

• Stability AI joins forces with WMG to give artists generative AI tools while protecting their rights

• Perplexity adds asset creation tools for Pro and Max subscribers

• Replit launches Design, a new Gemini 3.0-powered tool for creating beautiful UIs

• MIT trains AI on 41,000 CAD videos so it can design 3D models like a human engineer

• Crypto exchange Bitrue taps GPT-5, Claude, and Gemini to trade on users’ behalf

• Europe rolls back GDPR and AI rules under Big Tech pressure, easing cookie and data restrictions

• Midjourney launches user profiles

• Stitch UX designs can now be exported to AI Studio for instant Gemini-powered app creation

• Georgia Tech’s new AI can train exoskeletons without human trials, slashing development time

• OpenCV founders launch CraftStory, an AI startup generating human-centric videos up to five minutes long

📄 A Style is Worth One Code: Unlocking Code-to-Style Image Generation with Discrete Style Space

📄 Error-Driven Scene Editing for 3D Grounding in Large Language Models ​​

📄 Φeat: Physically-Grounded Feature Representation  

📄 What Does It Take to Be a Good AI Research Agent? Studying the Role of Ideation Diversity  

📄 Mixture of States: Routing Token-Level Dynamics for Multimodal Generation

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!