In partnership with

Good morning. It’s Friday, November 28th.

On this day in tech history: In 1999, MIT researchers released the RoboCup Simulation League platform, a 2D virtual soccer environment where autonomous agents learned strategy through multi-agent cooperation and competition. It looks quaint today, but it helped establish emergent behaviors in swarm robotics and game-theoretic AI. Concepts tested in simulated “soccer” later shaped reinforcement learning for logistics, traffic optimization, and even StarCraft-playing bots.

In today’s email:

• Claude’s new framework bridges sessions, trims errors, and powers ahead in benchmarks

• xAI charts the frontier of autonomous, pixel-driven AI

• Karpathy’s weekend “vibe code” hack reveals the hidden layer of enterprise AI orchestration

• 5 New AI Tools

• Latest AI Research Papers

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

Claude’s new framework bridges sessions, trims errors, and powers ahead in benchmarks

Anthropic is leveling up Claude agents to work more like human software engineers. Multi-session projects used to trip them up, forgetting previous work, overbuilding features, or marking tasks completed too early. The new two-agent system changes the game. An initializer agent sets up the project, seeds a git repository, and logs progress, while a coding agent tackles features step by step, produces clean, documented outputs, and runs end-to-end tests via Puppeteer. Early trials show agents can now sustain continuity across days-long projects, though Anthropic is still testing whether a multi-agent setup could push performance even further.

The results are showing up in benchmarks. Claude Opus 4.5 snagged top-two spots in web development challenges and top-three in text arenas, just behind Gemini 3 Pro and Grok 4.1.

Meanwhile, CEO Dario Amodei has been called to testify before the House Homeland Security Committee on December 17 about a Chinese state-linked cyber-espionage campaign that leaned heavily on Claude Code. This is the first documented case of an AI running a near-complete cyberattack. Google Cloud CEO Thomas Kurian and Quantum Xchange CEO Eddy Zervigon have also been asked to weigh in on how commercial AI can both fuel and defend against attacks that operate at machine speed. Read more.

xAI charts the frontier of autonomous, pixel-driven AI

Shen Zhuoran, a technical member at xAI, outlined a breakthrough AI system capable of reading and understanding computer interfaces from raw video, reasoning under tight time constraints, and executing actions with precision - all without APIs and within 150 milliseconds.

OpenAI Five and DeepMind’s AlphaStar mastered Dota 2 and StarCraft II using direct API access, giving them perfect game-state knowledge and superhuman precision, while Grok 5, the new xAI system Shen Zhuoran describes, operates purely from raw video input, reading the screen, reasoning under tight time limits, and executing clicks and commands like a human, aiming to generalize across any computer interface without specialized APIs.

xAI is addressing the energy demands of its massive Colossus data center in Memphis by planning an 88-acre solar farm capable of generating roughly 30 megawatts, around 10% of the center’s power needs. Local authorities have temporarily permitted 15 turbines through January 2027. xAI previously announced a 100-megawatt solar-plus-battery project funded with a $414 million interest-free USDA loan. Read more.

Karpathy’s weekend “vibe code” hack reveals the hidden layer of enterprise AI orchestration

Andrej Karpathy, former Tesla AI lead and OpenAI researcher, released LLM Council, a simple orchestration framework where multiple large language models debate, critique, and synthesize answers under a “Chairman.” Built with FastAPI, React, JSON storage, and OpenRouter for API integration, it runs GPT-5.1, Google Gemini 3, Claude Opus 4.5, and Grok 4 as swappable components.

The prototype proves multi-model orchestration is technically possible but highlights the missing operational essentials that keep commercial platforms in play. Frontier models are increasingly swappable, but orchestration, governance, and observability remain the differentiators that determine safe, scalable deployment.

Karpathy also turned to education, arguing that policing AI-generated homework is a losing battle. He recommends a “flipped classroom” approach with the goal of having dual competency. Students must use AI effectively while retaining the ability to reason independently. Through his startup Eureka Labs, Karpathy is exploring AI-native classrooms. Read more.

• OpenAI’s growth model faces soaring costs and rising liability

• Amazon employees raise alarm over aggressive, “all-costs-justified” AI expansion

• Luma scales 10B-parameter models with 25× faster inference

• DeepSeek Math V2 goes head-to-head with OpenAI and DeepMind, now open-weight and IMO-ready

• Hackers can now rhyme their way past AI safety filters as poetry proves a surprisingly effective jailbreak

• Prime Intellect’s INTELLECT-3 shows how 100B+ MoE models can master math, code, and reasoning on end-to-end stacks

• UBTECH kicks off the world’s first large-scale industrial rollout of humanoid robots with Walker S2 units

• Slop Evader browser extension lets users block all AI-generated content by freezing searches before 2022

• Watch: Flexion’s humanoid robot tackling ‘mountain-grade’ terrain while autonomously picking up litter

• MIT’s “digital twin” analysis shows AI can already handle millions of American jobs

• China warns of a looming humanoid robot bubble

• Fed warns of a K-shaped U.S. economy

• Wuhan researchers develop RGMP, a new framework that lets humanoid robots manipulate objects with 87% success in unseen environments

• Michael Burry versus Nvidia could be the real Thanksgiving drama this year

• Jeff Bezos’ Project Prometheus quietly snaps up AI startup General Agents to power next-gen agentic systems

• Meta kicks ChatGPT and Copilot off WhatsApp, leaving only its own Llama-powered AI

• Uber poaches PhDs for Project Sandbox, then ends AI training contracts after just a month

• AI models skew left in global elections, rarely matching actual voter outcomes

• Bipartisan coalition of state attorneys general urges Congress to let AI rules proceed

• For the first time, Chinese developers lead global adoption of open AI models

• Humanoid robots are gearing up to be the next smartphones

• LTX’s Retake lets creators tweak dialogue, emotion, and framing after a video is already rendered

• Tencent shrinks OCR power into a 1B parameter VLM that tackles documents, receipts, and video subtitles in one pass

• McKinsey cuts 200 tech jobs as AI reshapes the consulting giant’s workforce

• New 360-degree robotic head combines radar, depth sensors, and AI to sense its surroundings

• AIOZ Network launches decentralized AI platform to give developers control and token rewards

• Federal judge flags AI-written use-of-force reports by immigration agents as a risk to accuracy and privacy

• AI can now read pianists’ hidden muscle movements from video, replacing invasive EMG sensors

• Deloitte faces new scrutiny over suspected AI-generated mistakes

• New AI explainability method maps internal “thought circuits” behind image recognition

• Neuroscientists are now the secret weapon of Big Tech’s AI arms race

• Alibaba’s AgentEvolver lets AI agents teach themselves, cutting data costs while boosting performance by 30 percent

📄 CLaRa: Bridging Retrieval and Generation with Continuous Latent Reasoning

📄 Latent Collaboration in Multi-Agent Systems ​​

📄 PRInTS: Reward Modeling for Long-Horizon Information Seeking  

📄 GigaEvo: An Open Source Optimization Framework Powered By LLMs And Evolution Algorithms

📄 Budget-Aware Tool-Use Enables Effective Agent Scaling

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!