Good morning. It’s Monday, January 12th.

On this day in tech history: In 1960, the ACM/GAMM committee convened to hammer out Algol 60, the pioneering block-structured language that revolutionized scientific computing with its portability and recursive capabilities. This laid the groundwork for modern langs like C and Python, enabling efficient AI algorithm implementations, from neural nets to search trees.

In today’s email:

• Anthropic clamps down on unauthorized Claude use

• Google rolls out a new protocol letting AI find, pay, and deliver for you

• Meta secures 6.6 GW of nuclear power to run AI supercluster 

• 5 New AI Tools

• Latest AI Research Papers

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

Anthropic clamps down on unauthorized Claude use

Anthropic is taking control. Third-party tools like OpenCode and competitors like xAI can no longer piggyback on Claude AI. OAuth hacks and IDE shortcuts are gone, leaving only the official Claude Code or metered API. The move closes a loophole that let flat-rate plans bypass usage limits, but it also secures stability and predictable scaling.

Pushing Claude into regulated markets, Claude for Healthcare now taps HealthEx, Apple Health, and clinical databases, letting patients and providers query records in natural language, with minimal data pulled per question and full consent. Powered by Claude Opus 4.5, the system delivers improved reasoning, factual accuracy, and extended-context performance.

Anthropic is also partnering with Allianz to deploy Claude in insurance workflows. Claude will handle claims, case prep, and document workflows under human-in-the-loop oversight, all fully auditable. Read more.

OpenAI wants your actual work, tests 'ChatGPT Jobs'

OpenAI is collecting authentic office work from contractors to train models that can handle complex professional tasks.

Through partner Handshake AI, OpenAI is asking third-party contractors to upload real deliverables from past or current jobs such as Word documents, Excel sheets, presentations, code repos. Workers must anonymize proprietary and personally identifiable information first, using a ChatGPT-based tool called “Superstar Scrubbing.” Legal experts say the approach relies heavily on contractors' judgment, creating potential IP risks if sensitive material slips through.

The company is also internally developing “ChatGPT Jobs,” a vertical AI agent for career guidance. It will help users optimize resumes, research opportunities, match roles to skills, and plan long-term paths, following the pattern of domain-specific tools like ChatGPT Health. The feature consolidates job-search workflows into one interface.

In a separate legal development, a federal judge has upheld an order requiring OpenAI to turn over 20 million anonymized ChatGPT conversation logs in a consolidated copyright lawsuit from major news outlets including The New York Times and Chicago Tribune. The court ruled that privacy protections are sufficient and the logs are relevant to evaluating infringement claims, setting an important precedent for discovery in AI-related cases. Read more.

Google rolls out a new protocol letting AI find, pay, and deliver for you

Google is expanding its AI-driven commerce ecosystem with the launch of the Universal Commerce Protocol (UCP), an open standard for end-to-end agentic shopping.

Debuted at NRF 2026, UCP enables AI agents to handle product discovery, checkout, and post-sale support, integrating with protocols like Agent Payments Protocol (AP2), Agent2Agent (A2A), and Model Context Protocol (MCP). Users in the U.S. can pay via Google Pay, with PayPal support coming soon, while shipping details flow through Google Wallet.

Branded Business Agents debut in Search to answer queries in retailers’ voices, with early adopters including Lowe’s, Michaels, Poshmark, and Reebok. Direct Offers let merchants deliver real-time discounts based on natural-language queries, while new Merchant Center attributes optimize listings for AI search. This sets Google against OpenAI, Amazon, and Perplexity in a sector McKinsey sees reaching $3 trillion by 2030.

Meanwhile, in the latest Search Off the Record podcast, John Mueller and Danny Sullivan advised against "content chunking." The chunking trend includes short paragraphs, question-form subheadings, and minimal texts intended to make content more consumable by LLMs, but Google says such tactics are largely ineffective and may backfire long-term. Read more.

Meta secures 6.6 GW of nuclear power to run AI supercluster 

Meta has locked in up to 6.6 gigawatts of nuclear power to fuel its next-generation AI infrastructure. The deals, spanning TerraPower, Oklo, and Vistra, secure 24/7 carbon-free electricity for the Prometheus AI supercluster in New Albany, Ohio.

TerraPower’s Natrium reactors will deliver 690 MW each, with options for six more units, plus 1.2 GW of integrated storage. Oklo’s advanced Aurora reactors, coming online by 2030, add 1.2 GW of clean baseload power using fast-reactor technology. Vistra’s 20-year agreements supply 2.1 GW from existing plants, including 433 MW in capacity upgrades.

Meta says the plan supports AI compute at scale, strengthens U.S. energy infrastructure, and creates thousands of construction and long-term operational jobs. All energy costs are covered by Meta, keeping the grid stable and clean without passing expenses to consumers. Read more.

• DeepSeek sets February for V4 rollout, claims it will outcode OpenAI and Anthropic

• China’s AI ambition is colliding with a hard limit—there simply isn’t enough compute to compete

• At CES 2026, Nvidia showed why system-level design beats raw hardware every time

• xAI is pouring $20 billion into Mississippi to build one of the world’s largest AI data centers

• Watch: Humanoid robots just stepped into the fight ring, MMA-style, at CES

• AI is intensifying a 'collapse' of trust online, experts say

• A little-known Korean chip startup is quietly testing its AI silicon with Meta—and now Elon Musk’s xAI

• Research backed by LeCun moves AI closer to predicting actions from raw video

• Alphabet's Wing drone delivery expands to 270+ Walmart stores, reaching 10% of U.S. population

• Vision and language models are converging on a shared “Platonic” representation of reality

• Gates cautions that open-source AI tools could enable the next global catastrophe

• Researchers build “Web World Models” to give AI agents infinite, consistent universes without ever saving a byte

• Global AI compute just crossed 15 million H100 equivalents

• AI is coming for collectibles next

• Quadrupling nuclear power for AI by 2050 could be America’s boldest and riskiest energy gamble yet

• Qatar and UAE set to join U.S.-led Pax Silica initiative to secure AI and semiconductor supply chains

• New Python tool Orchestral promises reproducible, cost-conscious LLM research

• Niji V7 by MidJourney promises next-level anime and illustration generation

• Stop overpaying for LLMs: semantic caching boosts cache hits from 18% to 67%

📄 Atlas: Orchestrating Heterogeneous Models and Tools for Multi-Domain Complex Reasoning

📄 Evolving Programmatic Skill Networks  ​​

📄 Token-Level LLM Collaboration via FusionRoute 

📄 Learnable Multipliers: Freeing the Scale of Language Model Matrix Layers 

📄 GDPO: Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization 

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!