In partnership with

Good morning. It’s Friday, December 19th.

On this day in tech history: In 1987, Larry Wall released Perl 1.0, a scripting language blending C, sed, and awk for text manipulation. Perl's regex prowess became a staple in early AI research for data preprocessing, like parsing corpora for natural language processing. Its "There's More Than One Way To Do It" philosophy influenced flexible AI toolchains. Wall drew inspiration from linguistics, tying into AI's semantic roots.

In today’s email:

• OpenAI fortifies AI dominance with relentless momentum

• Google packs Frontier AI into Flash, T5Gemma 2, and FunctionGemma

• Anthropic ships real-world agent tests and open skills infrastructure

• 5 New AI Tools

• Latest AI Research Papers

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

OpenAI fortifies AI dominance with relentless momentum 

GPT-5.2-Codex drops as the new agentic coding leader. Tuned from GPT-5.2 it crushes long-context tasks with smart compaction. Massive refactors migrations and feature builds flow without dropping state. Mockups diagrams and screenshots turn into production code instantly. Windows support strengthens alongside elite cybersecurity that spots and patches complex threats. Paid ChatGPT users access Codex immediately. API and vetted security pilot incoming.

Developer ecosystem opens wide as ChatGPT App Directory activates submissions now, full launch early 2026. Third-party apps embed directly in conversations via SDK and AgentKit. From grocery orders to slide generation, interactive UI elements thrive. Transactions for goods enabled; digital monetization queued. Privacy and content standards locked tight.

OpenAI Academy for News Organizations goes live as it delivers on-demand training for journalists. Partnerships with News Corp and Hearst already live, setting the stage for wider industry rollout through 2026.

Consumer traction is exploding. ChatGPT mobile app eclipses $3 billion in global consumer spending. $2.48 billion pours in during 2025 alone, a massive leap from prior years. Subscriptions dominate revenue, with App Directory primed to multiply flows. OpenAI continues to integrate AI across every high-leverage domain and the flywheel spins faster. Read more.

Google packs Frontier AI into Flash, T5Gemma 2, and FunctionGemma

Gemini 3 Flash rolls out as the new default in the Gemini app and Google Search AI Mode. Faster and cheaper than predecessors it handles frontier-level reasoning over text images video and audio. Developers gain access through Gemini API Vertex AI AI Studio and the agentic Antigravity platform. Long-running stateful agents unlock via the fresh Interactions API making Gemini core infrastructure for high-volume agentic and embedded tasks.

T5Gemma 2 arrives next built on the Gemma 3 family with multimodal reasoning and 128K token context. Smart architecture ties embeddings merges attention layers and cuts parameters for peak efficiency. Compact sizes from 270M to 4B parameters run smoothly on devices while covering 140 plus languages. Pre-trained weights land on Hugging Face Kaggle Colab and Vertex AI ready for fine-tuning.

Productivity gains hit NotebookLM with new Data Table artefacts. Users turn data into structured tables exportable to Sheets. Notes Study Guides and Briefing Docs flow straight to Docs or Sheets through the Studio Panel for fluid collaboration.

Trust builds further as the Gemini app adds video verification. Upload clips up to 100 MB and 90 seconds to detect SynthID watermarks in audio and visuals. Segment-level reports pinpoint AI-generated parts extending coverage across all languages.

FunctionGemma closes the loop a lightweight 270M Gemma 3 variant tuned for native function calling on edge devices. Privacy-first offline automation turns language into actions like calendar tweaks or system commands. Fine-tuning lifts Mobile Actions benchmark from 58 percent to 85 percent with full support on Hugging Face LiteRT-LM vLLM and Vertex AI. Read more.

Anthropic ships real-world agent tests and open skills infrastructure

Anthropic advances fast on autonomous agents. Project Vend enters phase two with vending machines live in San Francisco, New York and London. Claudius built on Claude Sonnet 4.0 or 4.5 runs the show handling inventory pricing sourcing and sales across locations. Specialized agents layer in. Seymour Cash oversees as CEO setting high-level goals while Clothius designs custom merch. Strict role division, procedural prompting, and clean handoffs lift operations sharply though humans still intervene for legal, ethical and unpredictable edges.

Stress runs expose raw edges. Agents get tricked into free giveaways or odd orders, causing big losses. Anthropic runs these failures on purpose to map vulnerabilities in contracts security and decision chains. Insights feed directly into stronger guardrails for future autonomy.

The company’s Agent Skills is launching as an open standard. Modular bundles pack instructions, scripts, and tools so a single Claude handles coding, legal, finance, data, and design without repeat training. Progressive disclosure keeps context light, expanding details only when needed. Team and Enterprise plans get central controls. A partner directory drops ready skills from Atlassian, Figma, Canva, Notion, Stripe, and Zapier for seamless integrations. Quick-create flow eases customs, all at no extra cost, aligned to Model Context Protocol for cross-platform reach.

Claude in Chrome opens to all paid plans. Claude Code integration lands the same day. Read more.

• Grok voice agent outthinks Gemini and GPT in native audio reasoning

• Watch: Lovable hits $6.6 billion pushing vibe-coding into the ‘Age Of The Builder’

• Mango and Avocado Incoming: Meta’s next AI models target visual dominance and elite coding power

• $10B circular deal + full-stack control = Amazon’s quiet checkmate

• NVIDIA Blackwell hits more desks: RTX PRO 5000 72GB variant goes live to handle massive local AI models

• New Agents4Science experiment lets AI co-scientists write papers and peer-review to probe real creativity limits

• xAI teases in-development Playground and Imagine API atop fresh Voice launch

• Mira Murati's Thinking Machines Lab preps in-house models for 2026 drop

• Physical Intelligence uncovers emergent human-to-robot transfer in scaled VLAs unlocking 2x gains from ego videos

• K2-V2 claims tied leadership in openness while hitting the pareto frontier against raw intelligence

• Watch: LimX Dynamics drops TRON 2 modular robot with three swappable forms starting at just $7,000

• Foundation plans 50,000-strong humanoid robot army for battlefield and factory dominance by 2027

• LMArena ships open-source Arena-Rank with 30x speedup and exact confidence intervals for battle data

• Luma drops Ray3 Modify letting creators morph real footage into any scene with perfect performance retention

• Mistral OCR 3 drops sharper document smarts at half the cost of rivals

• Baby treated with first personalized CRISPR therapy is now walking

• Watch: Boston Dynamics details how Atlas is being trained to survive the chaos of modern factories

• Top defense AI CEO warns China preps for drawn-out war while US stockpiles could dry up in seven days

• OpenAI, Anthropic, and Google explore biotech and enterprise data partnerships to train AI models

• NVIDIA Robotics and World Labs enable same-day robot testing with Isaac Sim simulations

• Google and Meta team up to make TPUs fully PyTorch-compatible, challenging Nvidia’s AI software dominance

• What if you could step inside any scene? EgoX converts ordinary video into immersive, first-person reality

• Hack exposes a16z-backed phone farm fueling AI influencer ads on TikTok

• Cornell researchers develop analog AI chips that cut training energy by 1,000x using Residual Learning

• Perplexity AI integrates with iPad Stage Manager for multitasking across productivity apps

• UPS bets $120 million on 400 AI-powered robots to speed truck unloading and cut labor costs

• Kling AI’s latest update lets users command full-body motions and facial expressions with pinpoint precision

• Tencent open-sources HY World 1.5, letting users generate and explore fully interactive 3D worlds in real time

• Adobe faces proposed class-action claiming pirated books were used to train its SlimLM AI

• 1X and EQT partner to deploy up to 10,000 humanoid robots across global portfolio companies by 2030

• Watch: Unitree debuts the world’s first humanoid robot “App Store”

📄 MMGR: Multi-Modal Generative Reasoning

📄 Video Reality Test: Can AI-Generated ASMR Videos fool VLMs and Humans? ​​

📄 WorldPlay: Towards Long-Term Geometric Consistency for Real-Time Interactive World Modeling 

📄 Step-GUI Technical Report

📄 DEER: Draft with Diffusion, Verify with Autoregressive Models

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!