_{In partnership with}

Good morning. It’s Friday, January 16th.

On this day in tech history: In 2018, Alibaba's iDST team dropped an AI that smashed Stanford's SQuAD benchmark with 82.44, beating the human ceiling of 82.30. Powered by deep bidirectional transformers, this extractive QA breakthrough proved attention mechanisms were ready to dominate. It directly paved the way for BERT and the modern LLM era.

In today’s email:

• Google builds first personal AI ecosystem

• OpenAI ‘walks away’ from Apple, and now building its own U.S. robotics supply chain

• Real-world Claude data forces Anthropic to halve productivity estimates

• 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 Scroll}

Turn any knowledge base into enterprise-grade chatbots and agents

Build AI experiences that people actually trust.

Scroll.ai's platform delivers accuracy, nuance, and depth that generic models can't touch. Your users will feel the difference from the first message.

⚡ Accelerate your team with on-demand knowledge 📚 Centralize fragmented information 🔧 Automate repetitive knowledge workflows 🧘 Free yourself from reactive work

Use the AI-BREAKFAST-2026 coupon to get two free months of the Starter plan ($158 value).

^{Thank you for supporting our sponsors!}

Google builds first personal AI ecosystem

Google is building the only AI ecosystem that actually integrates your digital life. The company launched "Personal Intelligence," an opt-in system connecting Gmail, Photos, YouTube, and Search through Gemini.

Users control which apps link in, can disable personalization per chat, and see exactly when personal data gets referenced. The data produces context-aware responses but never trains the model. Guardrails protect sensitive information like health records.

The infrastructure upgrades extend across Google's product stack. Trends now automatically identifies, compares, and contextualizes queries with a revamped Explore page that doubles rising topics and suggests related searches.

NotebookLM gained Data Tables, converting unstructured notes into interactive, structured formats. Meeting notes, clinical trial results, and historical readings become analyzable datasets instead of static text.

Google DeepMind has deployed a math-specialized Gemini internally to prove a novel theorem in algebraic geometry, demonstrating AI’s growing role as a research collaborator rather than just a verification tool.

The company also released TranslateGemma, an open-weight model supporting 55 languages across mobile, consumer, and cloud deployments. Three versions handle different use cases: 4B for phones, 12B for laptops, 27B for servers. TranslateGemma retains multimodal capabilities including image text translation and functions as a chatbot through integrated general instruction data. Read more.

OpenAI ‘walks away’ from Apple, and now building its own U.S. robotics supply chain

OpenAI is done playing the model game. Bloomberg reports the company is building a U.S. supply chain for robotics, AI hardware, and data centers, including domestic silicon, motors, actuators, and cooling systems. A separate Financial Times report adds that OpenAI has ‘strategically’ walked away from being Apple’s custom model provider, choosing instead to build its own AI devices that could compete directly with major consumer platforms.

The hardware play runs deeper than chips. OpenAI now backs Merge Labs, Sam Altman's brain-computer interface startup that raised $252M seed. Merge uses ultrasound-based BCIs with molecular reporters for high-bandwidth, non-invasive neural interfaces.

On the product side, ChatGPT Translate launched quietly: standalone, no account, 25 languages. Testing verticalized consumer tools beyond core chat interface.

At the same time, OpenAI is scaling its backend: a partnership with Cerebras will add 750MW of ultra low-latency inference capacity through 2028, leveraging wafer-scale chips optimized for real-time workloads.

Complementing that infrastructure, OpenAI has released GPT-5.2 Codex to developers, expanding multimodal coding, reasoning, and security analysis capabilities via the Responses API. Read more.

Real-world Claude data forces Anthropic to halve productivity estimates

Anthropic measured what actually happens when people use Claude and the results forced them to cut productivity forecasts in half.

The fourth Economic Index tracked one million conversations from November 2025 using five metrics: task complexity, skill level, purpose, autonomy, and success rates. The findings: Claude accelerates high-skill, college-level tasks 12× faster than humans, but success rates dip slightly for complex work. Adjusted for reliability, AI could lift U.S. labor productivity by 1–1.2 percentage points annually which is significant, but well below early hype.

Adoption remains uneven. Wealthy countries use Claude for work and personal tasks; lower-income regions lean on it for education. Within jobs, Claude handles nearly half of sampled tasks, concentrating on computer and math work, and favors higher-education tasks, raising potential deskilling concerns. Users increasingly collaborate with AI (52% of interactions) rather than fully delegate.

Separately, Claude Code got a major infrastructure upgrade and now supports MCP Tool Search using ‘lazy loading’ to eliminate context bloat. Developers can now connect hundreds of APIs without memory penalties. Read more.

• xAI's unreleased Grok 4.20 discovered a new Bellman function in five minutes

• GLM-Image shows open-source AI can beat Google in complex visuals 

• Thinking Machines Lab faces a major shakeup as CTO and co-founders head back to OpenAI

• Apple is losing its grip on TSMC as Nvidia gobbles up chip capacity

• Microsoft unveils OptiMind, a new AI research model for optimization on Hugging Face

• Nvidia joins $1.45B startup bet on AI that proves math problems autonomously

• Users spend $99 each month to access a Tony Robbins chatbot

• 𝕏 revokes API access for attention-farming crypto projects, walks away from millions: "We don't want it," head of product says

• Raspberry Pi’s new AI HAT+ 2 brings 8GB of RAM to run generative models locally

• Microsoft, Meta, and Amazon pay for enterprise Wikipedia access to power AI

• Watch: Two humanoid robots hold an unscripted conversation for the first time

• Former OpenAI policy chief launches nonprofit to push independent audits of AI labs

• AI startup Replit launches feature to vibe code mobile apps

• Pebble returns with open-source wearables, ultra-long battery life, and a playful AI twist

• Comet gets a new unified attachment selector in Perplexity’s latest test

• A New York judge warns lawyers: ignoring AI could soon be malpractice

• Breaking through AI’s memory wall with token warehousing

• Nvidia’s latest app update brings next-gen image clarity and a unified Control Panel experience

• Bandcamp bans fully AI-generated music to protect human artists

• Cursor shows what happens when hundreds of GPT-5.2 agents code together

• NBC Sports brings AI to the Olympics, letting fans follow any player in real time

• McConaughey draws a line in the sand for AI: his voice, his rules

• Microsoft doubles down on AI with $500M Anthropic spend despite OpenAI ties

• Robot learns to lip sync by watching YouTube

• Deutsche Telekom puts AI voices on the phone with Elevenlabs

• Rice University uses AI to design millions of DNA circuits in record time

• MIT’s MechStyle AI lets you style 3D objects without sacrificing structure

• AI and deepfakes fuel a record $17 billion in crypto scams, making fraud faster and far more convincing

• Liberal arts majors get a shot at McKinsey as AI expertise becomes the new hiring currency

• X says Grok will no longer edit images of real people into bikinis

• Robinhood CEO predicts AI could trigger a “job singularity”

• Your voice gives away valuable personal information. Expert raises privacy concerns

• Sakana AI proves its ALE-Agent can solve hours-long human problems

📄 Controlled Self-Evolution for Algorithmic Code Optimization

📄 Solar Open Technical Report ​​

📄 MAXS: Meta-Adaptive Exploration with LLM Agents  

📄 ShowUI-π: Flow-based Generative Models as GUI Dexterous Hands

📄 ArenaRL: Scaling RL for Open-Ended Agents via Tournament-based Relative Ranking 

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!