Good morning. It’s Wednesday, December 17th.

On this day in tech history: In 1976, AI researcher Jen Golbeck was born on December 16. A pioneer in explainable AI and social network analysis, her work on trust metrics in recommender systems (like predicting user behavior via graph theory) is a deep cut for semantic web nerds. Her "Computing with Social Trust" framework influenced hybrid AI models blending human data with ML, still relevant in bias detection today.

In today’s email:

• OpenAI's Triple Strike

• Gemini's total system dominance

• Anthropic tests agentic tasks mode for Claude

• 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 Triple Strike: GPT Image 1.5 launch, GPT-5 79x bio gains, and updated voice models

OpenAI’s new flagship image generator, GPT Image 1.5, is live, integrated into ChatGPT and available via API. This update cuts through it with massive gains in speed, razor-sharp instruction-following, and pixel-level precision. Now you iterate directly on images, shift styles, and refine details without ever regenerating the whole thing. The new Images workspace drops preset filters, trending prompts, and curated starters. Early tests put its quality on par with Google’s Nano Banana Pro.

Meanwhile, GPT-5 is now an active scientist. Partnering with Red Queen Bio, the model entered the wet-lab, optimizing protocols and adapting based on live experimental feedback. The payoff was massive: a 79x efficiency gain in molecular cloning and proving AI can combine biological concepts innovatively.

OpenAI also strengthened its voice layer with three new Realtime API models, reducing transcription errors, cutting text-to-speech word error rates by 35%, and boosting instruction-following by 22%, alongside expanded multilingual support.

OpenAI has also reversed the automated model router, removing the unpredictable system that frustrated users with inconsistent results. This rollback signals a return to a predictable default experience, prioritizing user control and stability at scale following a critical failure of optimization.

Former UK Chancellor George Osborne has been appointed to lead OpenAI’s global Stargate data center expansion, focusing on international infrastructure and government adoption. At the same time, Chief Communications Officer Hannah Wong will step down in January, leaving a team she guided through high-profile challenges. Read more.

Gemini's total system dominance: agent clarity and workflow acquisition

Google is no longer just iterating; it is executing total system upgrades across its AI infrastructure.

Eliminating friction in your daily output, Google’s new CC assistant is designed to kill the time-wasting morning scroll. Integrated deeply with your Gmail, Calendar, and Drive, CC delivers a personalized "Your Day Ahead" summary and executes tasks immediately (drafting, scheduling). This is contextual intelligence leveraged for maximum productivity.

Concurrently, the upgraded Gemini 2.5 Flash Native Audio model tightens the voice control loop, increasing developer instruction compliance to 90% and demonstrating a superior 71.5% accuracy on complex function calls, making voice the most precise way to control your workflows.

Google is collapsing the distance between raw research and executed action. NotebookLM, the research powerhouse, is now directly integrated into Gemini. Users can select multiple notebooks as a Retrieval-Augmented Generation (RAG) context, instantly turning deep documents into flexible knowledge bases.

This integration fuels the overhauled Gems Manager, which now features Super Gems (Opal apps) and an auto-generating Workflow Builder. The system is building a unified environment where research insights, personalized context, and precise action can be deployed immediately and at will. Control and context are now centralized. Read more.

Anthropic tests agentic tasks mode for Claude

Anthropic is testing a new Agentic Tasks Mode for Claude that reframes the assistant as a system for getting work done, not just answering prompts. The update introduces a switch between traditional chat and agent mode, steering users toward structured task delegation instead of open-ended conversation.

In agent mode, Claude presents five core work paths: Research, Analyze, Write, Build, and Do More, each tuned for a different stage of knowledge work. Research emphasizes source control and effort tuning, while Analyze focuses on validation, comparisons, and forecasting. Write generates structured outputs across documents, slides, and spreadsheets, and Build supports visual layouts, interactive assets, or code. A progress tracker and context manager show task steps and active inputs in real time. Read more.

• Meta updates AI glasses with conversation amplification and Spotify visual cue playback

• Scale Is Dead. Nvidia's Efficiency Revolution Begins

• Zenflow turns AI coding from “vibe” to structured workflows, letting Claude and OpenAI models cross-check each other

• Zoom tops AI’s toughest benchmark with a federated system that mixes OpenAI, Google, and Anthropic models

• AI is flooding podcasting with synthetic hosts

• Mozilla plans an AI mode for Firefox that mixes multiple models without selling out user trust

• DeepMind lets AI design its own reinforcement learning algorithm

• Terence Tao breaks down the rise of “artificial general cleverness” versus real intelligence in modern AI

• Meet MiMo-V2-Flash, Xiaomi’s open-source powerhouse LLM

• Firefly’s new AI video feditor brings frame-level control, camera angle replication, and FLUX.2 integration to Adobe users

• DoorDash rolls out Zesty, an AI social app for discovering new restaurants

• AI racks are so heavy that legacy data centers can no longer handle them

• AI meets supply chains: Nauta’s new tool gives importers real-time visibility down to the SKU level

• This low-power AI chip can predict your next move in rock-paper-scissors in real time

• Gallup data shows nearly half of U.S. workers now use AI a few times a year or more, led by knowledge-sector roles

• Trump administration launches 1,000-person Tech Force to accelerate federal AI and tech projects

• Stanford experts predict 2026 will be the year AI faces real-world evaluation over hype

• Alibaba updates Qwen Code to v0.5.0 with full VSCode and TypeScript support

• BBVA doubles down on AI with OpenAI partnership, bringing ChatGPT to 120,000 staff worldwide

• Watch: Stanford’s RAI Institute builds a bike robot that jumps, flips, and rides like a world-class athlete

• Mocha launches no-code app builder for non-technical users

• Codex Mortis pushes AI-generated games forward with 100 percent AI-designed code, art, and music

• China opens its first humanoid robot-friendly city zone to train robots in real-world urban environments

• Ai2 launches Bolmo, the first open byte-level language models designed for multilingual and noisy-text applications

• Manus rolls out 1.6 and 1.6 Max, with Max hitting 19% higher benchmark performance

• Motif-2-12.7B proves disciplined training beats scale for enterprise reasoning tasks

📄 LongVie 2: Multimodal Controllable Ultra-Long Video World Model

📄 Memory in the Age of AI Agents ​​

📄 PersonaLive! Expressive Portrait Image Animation for Live Streaming

📄 NL2Repo-Bench: Towards Long-Horizon Repository Generation Evaluation of Coding Agents  

📄 Towards Scalable Pre-training of Visual Tokenizers for 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!