Good morning. It’s Monday, January 5th.

On this day in tech history: In 2001, Leo Breiman published “Statistical Modeling: The Two Cultures,” drawing a sharp line between data modeling and algorithmic modeling. Though not framed as AI, the paper foreshadowed modern ML debates: interpretability vs. performance, theory vs. scale. Breiman’s argument legitimized black-box learning just as compute and data began to explode

In today’s email:

• OpenAI's ‘Gumdrop’ could be an AI that lives in a pen-sized device

• Google engineer says AI built in one hour what took her team a year

• 2026 belongs to recursive language models scaling to 10M tokens

• 5 New AI Tools

• Latest AI Research Papers

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

OpenAI's ‘Gumdrop’ could be an AI that lives in a pen-sized device

OpenAI is gearing up to launch its first consumer-edge AI device, Project Gumdrop, with Foxconn now handling production in Vietnam or the US. The device, still in design, could be a smart pen or portable audio tool, equipped with a microphone and camera to capture handwritten notes and upload them straight to ChatGPT.

Roughly the size of an iPod Shuffle, Gumdrop is designed to be lightweight and portable. Launch is expected in 2026–2027. Technical challenges remain, including software bugs, privacy issues, and incomplete cloud infrastructure, but Foxconn will manage the entire supply chain, from cloud support to consumer delivery.

Meanwhile, OpenAI’s UAE Stargate facility is scaling high-performance AI compute with four 340MW Ansaldo Energia AE94.3 gas turbines. Extreme desert heat limits output to 1GW instead of 1.3GW, but Phase 1’s 200MW target by the end of 2026 is on track. GPUs are deploying smoothly, in contrast to slower rollout at Texas’ Abilene site.

Co-founder Greg Brockman donated $25 million to Trump’s MAGA Inc., part of $102 million raised in the second half of 2025. Analysts suggest big donors may be trying to influence federal AI policy, which the Trump administration plans to centralize, while Brockman’s role in “Leading the Future” shows AI leadership increasingly intersecting with politics. Read more.

Google engineer says AI built in one hour what took her team a year

Google engineer Jaana Dogan revealed that Anthropic’s Claude Code generated a distributed agent orchestration system in one hour, a problem her team had spent over a year building. Minimal prompts produced a working prototype, though not production-ready. Claude Code’s creator Boris Cherny recommends self-checking loops, parallel agents, and integrations with tools like Slack, BigQuery, and Sentry to boost output. This milestone shows AI-assisted coding has jumped from single-line completions in 2022 to full codebase orchestration by 2025.

Google’s Gemini 3.0 Pro also flexed serious multimodal muscle by decoding handwritten notes scribbled in a 1493 Nuremberg Chronicle. It pulled in paleography, historical context, and biblical timelines to explain the marginalia as calculations reconciling Septuagint and Hebrew Bible chronologies.

A few minor math slips aside, the analysis was spot-on and deeply grounded, proving large models can unlock centuries of archived knowledge. Read more.

2026 belongs to recursive language models scaling to 10M tokens

AI is about to break the context ceiling for good. MIT researchers just made the full paper on Recursive Language Models (RLMs) available, with much more expansive experiments compared to their initial blogpost from last year.

Recursive Language Models (RLMs), the paradigm that will own 2026. Instead of cramming everything into one giant prompt, RLMs turn the input into code they can manipulate. A root model like GPT-5 runs a Python REPL, slices massive contexts, spins up sub-models on chunks, caches results, and stitches everything back together.

The result is accurate reasoning over 10 million tokens, 100x beyond today’s limits, without losing the plot. Benchmarks like S-NIAH, BrowseComp-Plus, and OOLONG show RLMs crush retrieval agents on dense, quadratic tasks. Prime Intellect’s RLMEnv isolates tool use in sub-models while the root orchestrates clean reasoning.

Alex Zhang calls it the real shift from probabilistic guessing to structured, programmable reasoning. LLMs now manipulate their own context for structured, deep outputs. Read more.

• Anthropic moves to own and operate one million TPUv7 units

• Yann LeCun slams Alexandr Wang as ‘inexperienced’ and warns of Meta AI exodus

• Grok use for illegal content faces full enforcement

• Stanford’s new AI lets robots imagine how objects move before they act

• Humans and their AI lovers survived a chaotic couples retreat in the woods

• Nokia pivots from phones to data centers, lands a $1 billion Nvidia AI deal

• Watch: Unitree’s $30K humanoid execute jump side and moon kicks

• AI agents can now orchestrate multi-institution experiments with the Science Context Protocol

• ByteDance’s StoryMem gives AI videos memory, keeping characters from morphing across scenes

• Watch: Flying cars take off as China pushes low-altitude airspace strategy

• US Army launches dedicated AI officer track to build in-house machine learning expertise

• NASA praises 17-year-old for teaching old telescope data new tricks with AI

• Watch: Quadruped robot tackles fires, floods, and rugged terrain at rescue drill

• Japan bets on AI-powered robots to save shipyards from a crippling labor shortage

• An NYU professor fights AI cheating with voice-powered oral exams that cost 42 cents per student

• Epoch AI finds Chinese AI models average seven months behind US

• Watch: Mainland China arms drones with machine guns as Taiwan tensions rise

• US towns push back as AI data center boom collides with power and water limits

• UniX AI positions humanoid robots for deployment, not display, at CES 2026

• Bill Gates-backed wind tech to showcase new way to power AI data centers at CES

• Why Sweden produces more AI unicorns per capita than anywhere outside Silicon Valley

• CLOiD, a humanoid butler that wants to do everything from dishwashing to breakfast prep

• ‘Intelition’ redefines AI as a continuous, co-creative system where humans and machines act as one.

• Meet the $5,599 WALL‑E cousin that rolls around your home doing light work

• Plaud launches a new AI pin and a desktop meeting notetaker

• Someone made a huge profit predicting Maduro's capture. Here's what happened

• Big Pharma bets on AI as first AI-discovered drug nears Phase 3 trials

📄 DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models

📄 SpaceTimePilot: Generative Rendering of Dynamic Scenes Across Space and Time ​​

📄 FlowBlending: Stage-Aware Multi-Model Sampling for Fast and High-Fidelity Video Generation

📄 Dynamic Large Concept Models: Latent Reasoning in an Adaptive Semantic Space

📄 JavisGPT: A Unified Multi-modal LLM for Sounding-Video Comprehension and 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!