Good morning. It’s Wednesday, December 24th.

On this day in tech history: In 1965, Daniel Bobrow's STUDENT program solved high-school algebra word problems in natural language, parsing ambiguous English into Lisp equations. This early NLP feat handled units and context, influencing symbolic reasoning systems. An underappreciated gem in bridging human language to formal computation, predating modern semantic parsers.

In today’s email:

• 50+ customizable Claude skills just dropped

• GPT-5 cracks elegant math as OpenAI wrestles with unpatchable prompt exploits

• Google is steadily building a deeper AI moat

• 5 New AI Tools

• Latest AI Research Papers

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

50+ customizable Claude skills just dropped

A subtle but powerful resource just landed for anyone serious about leveraging Claude at scale. A new GitHub repository has dropped with over 50 ready-to-customize "Claude Skills," essentially structured workflows that train Anthropic's AI to execute specific tasks consistently and on demand.

These skills span nine practical categories: Document Processing (Word, PDF, PowerPoint), Development Tools (Playwright, AWS, Git), Data Analysis, Business & Marketing, Communication, Creative Media, Productivity, Project Management, and Security. Each one lives in its own folder with a clear SKILL.md file laying out the prompts and logic.

You can load them directly in Claude.ai via settings, drop them into Claude Code's config, or hook them up through the API. The whole repo is Apache 2.0 licensed and openly welcomes contributions. As Claude Skills become a de facto standard for structured AI workflows, this repository offers a practical toolkit for anyone looking to operationalize AI beyond basic prompts. Read more.

GPT-5 cracks elegant math as OpenAI wrestles with unpatchable prompt exploits

A paper by Swiss mathematician Johannes Schmitt reports that GPT-5 independently solved a long-standing open problem in enumerative geometry, producing a novel inequality proof by drawing on non-obvious techniques from algebraic geometry.

The work was conducted in a multi-model workflow: Gemini 3 Pro handled boundary cases, Claude assisted with exposition, and formal verification was completed in Lean using Claude Code and ChatGPT 5.2. Schmitt published full prompts, transcripts, and authorship labels, a new norm for transparency in AI-assisted mathematics.

Meanwhile, OpenAI is staring down a colder truth about agentic systems. With tools like ChatGPT Atlas that browse and act on the web, prompt injection attacks may never be fully eliminated. Admitting there may never be a perfect fix, the company’s response includes: stack defenses deep, patch fast, and train an RL-powered automated attacker to red-team exploits before release. Security researchers continue to advise constrained permissions and breach-first assumptions.

OpenAI also just launched "Your Year with ChatGPT." This Spotify-Wrapped-style recap delivers personalized themes, stats, awards, poetry, and images for users with memory enabled. SoftBank meanwhile is reportedly racing to raise $22.5 billion by year's end to meet its commitment to OpenAI's Stargate. Masayoshi Son has already cashed out the full Nvidia stake, trimmed others, and locked down spending. Read more.

Google is steadily building a deeper AI moat

Gemini 3 has now outperformed rivals including ChatGPT on speed, reasoning, and benchmark coverage, leading 19 of 20 major tests and setting a new high on Humanity’s Last Exam, a benchmark designed to approximate near-human reasoning. The jump has drawn rare public praise from figures such as Geoffrey Hinton and Marc Benioff, fueling the view that Google may be regaining a dominant position in frontier AI.

The advantage runs deeper than models. Cloudflare data shows Google’s AI crawlers see more than three times as much of the web as OpenAI’s, a byproduct of Google Search acting as the internet’s default gateway. That imbalance is now attracting regulatory scrutiny, with the UK’s CMA considering whether Google should be forced to split its search and AI crawlers to prevent data dominance from spilling into model training.

Infrastructure is the other lever. Alphabet’s $4.75 billion acquisition of Intersect Power, plus assumed debt, secures co-located renewable energy and data center projects projected to deliver 10.8 gigawatts by 2028. The strategy is designed to bypass grid bottlenecks and lock in power at a scale few competitors can match.

Distribution is moving just as fast. Gemini 3 is being embedded directly into YouTube through Playables Builder, letting creators generate games from simple prompts. At the productivity layer, Google is testing a new “Lecture” mode in NotebookLM, capable of generating 30-minute, single-voice audio lectures that connect and explain dense source material.

Language selection is built in, and new narration options, including British English voices previewed in partnership notebooks, are slated for 2026. Google isn’t just chasing the AI race anymore. It’s trying to own the terrain. Read more.

• Meta AI open-sources PE-AV, a unified audio, video, and text encoder for cross-modal understanding at scale

• Yann LeCun calls general intelligence “complete bs” and Demis Hassabis pushes back

• GLM-4.7 tops GPT-5.2 and Claude 4.5 in open-source AI benchmarks

• Four AI coding agents try to rebuild Minesweeper and the results are a mix of brilliance and disaster

• China deploys humanoid robots at Vietnam border to manage personnel and logistics 24/7

• WIth the RAISE Act, New York joins California in pushing AI transparency while Trump pushes back

• Watch: Physical Intelligence fine-tunes robot model to wash pans, clean windows, and make peanut butter sandwiches

• Watch: Kinsi Robotics' KR-1 humanoid masters towel folding autonomously through kinesthetic teaching demos

• Operational misstep, not software bug, behind Waymo’s robotaxis gridlocking streets

• Sigma’s new AI-native browser runs LLM locally, no cloud needed

• TikTok owner ByteDance to spend nearly $23B building AI infrastructure next year

• Salesforce’s Agentforce hits 18,500 customers, generating $540M ARR

• Mozilla confirms AI kill switch for Firefox as community questions the future of the browser

• Micro robots the size of microorganisms gain autonomy, computation, and communication in liquid environments

• MIT demonstrates short-term alignment can unlock the potential of stubborn neural networks

• Alibaba’s Qwen models can clone voices from just three seconds of audio and generate expressive speech from text

• ARC-AGI sees Poetiq break human baseline using multi-LLM meta-system

• xAI secures Pentagon deal to deploy Grok AI for 3 million military and civilian users

• Light-powered LightGen achieves 100x Nvidia speeds for image and video generation while slashing energy use

• Asia markets rise on AI momentum as Wall Street tech rally spills over 

• ICE invests hundreds of thousands in AI bounty hunters to track persons of interest

• Next-level quantum computers using neutral atoms aim for error correction by 2026

• Gen AI use costs Clair Obscur its ‘indie game of the year’ crown

• Ex-Tesla AI Chief Andrej Karpathy maps four ways startups can compete with OpenAI through LLM apps

• World’s largest nuclear plant may host data centers, Japan plans

• Inside the data center built to withstand a nuclear blast

• MiniMax drops M2.1, a multilingual AI coder built for real-world office and agentic tasks

📄 DataFlow: An LLM-Driven Framework for Unified Data Preparation and Workflow Automation in the Era of Data-Centric AI

📄 When Reasoning Meets Its Laws ​​

📄 The World is Your Canvas: Painting Promptable Events with Reference Images, Trajectories, and Text

📄 Probing Scientific General Intelligence of LLMs with Scientist-Aligned Workflows

📄 PhysBrain: Human Egocentric Data as a Bridge from Vision Language Models to Physical Intelligence

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!