Good morning. It’s Friday, January 30th.

On this day in tech history: In 2017, Carnegie Mellon University's Libratus AI system wrapped up its dominance in a grueling Heads-Up No-Limit Texas Hold'em tournament against top pros at Rivers Casino. Leveraging abstraction and equilibrium-finding algorithms, it handled imperfect information with subgame solving, marking a leap in strategic AI beyond chess or Go.

In today’s email:

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

Google DeepMind just opened Project Genie to US subscribers paying $250/month for AI Ultra. It generates interactive 3D worlds in real time from text or images. Type a prompt, get a navigable environment at 1280×720 and 24fps. Sessions cap at 60 seconds. You can sketch, explore, and remix what it creates. DeepMind's play here is to train AI agents in simulated worlds instead of static datasets.

Gemini 3 Flash added Agentic Vision. Instead of one-shot image recognition, the AI investigates: zooms, measures, annotates objects step by step using Python. Accuracy jumps 5–10% on vision benchmarks. Use cases include analyzing building plans, labeling objects with bounding boxes, generating charts from images. Available through AI Studio and Vertex AI.

Google Maps lets walkers and cyclists ask voice questions hands-free -no leaving navigation. Chrome's Auto Browse handles multi-step web tasks: trip planning, form filling, appointment booking, subscription management. It uses the Universal Commerce Protocol (built with Shopify, Target, Etsy, Wayfair) to search products and apply discounts. It stops before payments or social posts and asks permission. AI Pro and Ultra subscribers only.

For builders, Stitch gets design systems, Figma export, and a "Deep Design" mode. Jules, the coding assistant, now auto-fixes CI errors and connects to Linear, Supabase, Neon, and Stitch. Future versions run 24/7, handling backlog and bug fixes without prompting. AI Pro/Ultra plans now include Developer Program access and Cloud credits for AI Studio, Gemini CLI, and Vertex AI. Read more.

OpenAI is preparing for a Q4 IPO and is moving quickly to clean up its product lineup while sharpening its long-term strategy. The company is in talks with Wall Street banks and has strengthened its finance team with two key hires: Ajmere Dale as Chief Accounting Officer and Cynthia Gaylor leading corporate finance & investor relations. The IPO is expected to raise significant capital for compute infrastructure and AGI research while establishing a public valuation before Anthropic or others beat them to market.

OpenAI is retiring several older models next month: GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini. Even though GPT-4o still has fans for its warm, conversational style, only 0.1% of users access it daily. The vast majority have already switched to newer models (especially GPT-5.2). OpenAI is focusing resources on high-engagement models that offer stronger personality, creativity, and reasoning. Developer API access to the retired models will not be affected.

After six months, ChatGPT Agent is being discontinued. It launched with 4 million weekly active paying users but lost 75% of them very quickly. Most people never clearly understood what it was for. The virtual browser it offered overlapped heavily with features already available in other ChatGPT modes (coding, web research, image analysis). OpenAI is now shifting toward simpler, clearly defined specialized agents, starting with tools like Shopping Research, that solve one job well and are easier to grasp.

Sora debuted with 100,000 installs on day one but has quickly lost momentum. Downloads dropped 32% in December and 45% in January. Stronger competition from Google Gemini video generation and Meta’s Vibes, combined with tighter copyright restrictions, has hurt growth. A limited Disney IP-licensing deal has not meaningfully lifted engagement.

Elon Musk’s companies are increasingly linked. xAI launched the Grok Imagine API, which can turn text or single images into 15-second 720p videos with realistic motion, object interaction, continuity, and built-in audio and dialogue.

Users can edit scenes, add or remove objects, animate characters, and change lighting, seasons, or aspect ratio.

Grok ranks first on benchmark tests for text-to-video and image-to-video, outperforming competitors like Runway, Kling, and Veo 3.1, and is faster and cheaper than Sora or Veo.

𝕏 is testing “manipulated media” labels to flag edited or AI-generated images, though detection methods and appeals are not yet fully defined. Tesla is preparing Optimus Gen 3 for production in Q1 2026, with improved hands, a dedicated Fremont line, and a goal of 1 million units per year for industrial, home, and medical applications.

Tesla has invested $2 billion in xAI, integrating Grok into vehicles and sharing energy infrastructure. Early merger discussions are exploring a combination of SpaceX, Tesla, and xAI, potentially before a SpaceX IPO. Musk has said AI data centers in orbit could lower costs using continuous solar power and reduced cooling needs, supporting large-scale AI operations both on Earth and in space. Read more.

📄 Harder Is Better: Boosting Mathematical Reasoning via Difficulty-Aware GRPO and Multi-Aspect Question Reformulation

📄 Visual Generation Unlocks Human-Like Reasoning through Multimodal World Models ​​

📄 Youtu-VL: Unleashing Visual Potential via Unified Vision-Language Supervision

📄 A Pragmatic VLA Foundation Model

📄 Advancing Open-source World Models

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!