In partnership with

Good morning. It’s Wednesday, December 3rd.

On this day in tech history: In 2014, Facebook AI Research released the first open-source implementation of “fastText,” blending bag-of-words speed with subword embeddings. It looked humble next to deep networks, but its efficiency reshaped text classification benchmarks and proved that clever tokenization could beat brute-force scale. That idea, structural efficiency over parameter count, foreshadowed today’s small-model renaissance powering on-device AI and edge agents.

In today’s email:

• Altman declares ‘code red,’ fast-tracks new model ‘Garlic’ against Gemini surge

• Claude Code scales up with Bun as Anthropic demos AI DeFi hacks

• Runway’s Gen-4.5 beats Google and OpenAI in text-to-video benchmark 

• 5 New AI Tools

• Latest AI Research Papers

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

Altman declares ‘code red,’ fast-tracks new model ‘Garlic’ against Gemini surge

OpenAI is scrambling to regain momentum as Google’s Gemini 3 ramps up, reclaiming benchmark leadership and prompting a 6–7% drop in ChatGPT’s active users. CEO Sam Altman has declared a company-wide ‘code red’, pausing commercial initiatives, including ads, shopping and health agents, and the Pulse personal-briefing tool, to focus engineering on core model improvements. Daily progress calls, temporary cross-team transfers, and an emphasis on lower latency, higher reliability, and richer reasoning now define OpenAI’s operating mode.

According to internal briefings reported by The Information, ‘Garlic’ is already posting strong results in coding, complex reasoning, and multi-step problem solving, areas where Google and Anthropic have recently surged with Gemini 3 and Opus 4.5. Garlic is being positioned as a next-generation architecture rather than a minor update, with early targets pointing to a release as GPT-5.2 or GPT-5.5 in early 2026.

OpenAI is also piloting Memory Search, a deeper system-level retrieval tool letting ChatGPT query long-term memory directly, mirroring Atlas Browser’s memory engine. If paired with GPT-5.2, it would shift ChatGPT toward persistent, context-rich workflows rather than episodic conversations.

Parallel to its technical sprint, OpenAI has entered a circular deal with Thrive Holdings, an offshoot of investor Thrive Capital. No cash changed hands; OpenAI supplies models, staff, and tech, while gaining access to operational data for training and potential returns. IT services and accounting, high-volume, rules-driven workflows, are the first target, positioning OpenAI as a modernization engine for private-equity portfolios. Read more.

Claude Code scales up with Bun as Anthropic demos AI DeFi hacks

The company recently completed its first acquisition, snapping up developer tools firm Bun to accelerate Claude Code, its AI coding assistant. Bun, built in Zig and leveraging Apple’s JavaScriptCore, combines a package manager, bundler, test runner, and script runner into a single executable. Its fast startup times, low memory footprint, and broad Node.js compatibility will enhance Claude Code’s execution speed, stability, and scalability.

At the same time, Anthropic’s research is pushing AI into high-stakes territory. Their team used Claude 4.5 Opus, Sonnet 4.5, and GPT-5 to test 405 DeFi smart contracts that had already been hacked, successfully creating full exploit scripts and simulating $4.6 million in gains. They also ran tests on 2,849 previously untouched contracts on the BNB Chain, uncovering new vulnerabilities for just $1.22 per test. The findings show that AI can now carry out automated attacks that are both technically possible and cost-effective, and that these techniques could eventually target regular software and critical systems too.

Underlying these advances is Anthropic’s unique alignment approach, captured in the leaked and verified “Soul Document.” The 14,000-token blueprint embeds safety, ethics, and operational guidelines directly into Claude’s weights, creating functional self-awareness, resilience, and context-aware reasoning. Claude prioritizes human oversight, anticipates prompt injection attacks, and enforces strict “bright lines” to prevent harmful outputs. Read more.

Runway’s Gen-4.5 beats Google and OpenAI in text-to-video benchmark 

Runway has released Gen-4.5, its latest text-to-video AI, delivering more realistic, responsive, and cinematic outputs. Objects move with believable weight and momentum, liquids flow naturally, and scenes follow complex prompts with higher visual fidelity. It handles both photorealistic and stylized visuals efficiently, running on Nvidia Hopper and Blackwell GPUs.

Gen-4.5 scored 1247 on the Artificial Analysis benchmark, just ahead of Google’s Veo 3 and Kling’s Version 2.5, showing improved modeling of physical interactions and scene coherence. Some limitations remain. The model can misjudge causality, lose object permanence, while a success bias can exaggerate task outcomes. Read more.

• Google is testing a deeper, chat-ready Search that blends AI Overviews with AI Mode

• AWS launches AI agents that can code ‘for days’ and handle security and DevOps

• DeepSeek’s open-source AI rivals GPT-5 and wins Olympiad gold

• Nvidia’s Alpamayo-R1 gives autonomous vehicles human-like perception and reasoning

• Mistral launches 10 open-source AI models, including smaller customizable options 

• Former Google and Microsoft exec Subramanya steps in to steer Apple’s AI future

• Momentum AI demo at NeurIPS 2025 Google booth highlights Gemini, Imagen, and Veo integration

• Ex-Tesla, Google, and Nvidia leaders launch UMA to build next-gen humanoid robots for real-world work

• Stack Overflow launches AI Assist to answer developers’ coding questions using community-verified knowledge

• Cosmos 1.0 uses machine learning to chart the age, reach, and scientific depth of thousands of technologies

• Watch: New fighter drone smokes a jet at long range in a historic first. No pilot required

• MIT spinout claims AI agent beats OpenAI, Anthropic at fraction of cost

• MiniMax-M2 brings interleaved thinking to AI coding, delivering faster, cheaper, and more reliable dev workflows

• Watch: EngineAI T800 launches at $25K, surpasses 90% of adult males in strength and agility

• Silicon Valley startups are increasingly building on free Chinese AI models as open-source alternatives gain traction

• From open-source to locked-down: Artificial Analysis maps the AI openness gap

• Alibaba supercharges Quark desktop with always-on AI powered by Qwen

• Kling’s Video O1 combines video creation and editing in one AI model, claiming wins over Google and Runway

• Consultancies like McKinsey froze graduate pay offers for 2026, the third consecutive year

• New York becomes first US state to force companies to disclose AI-driven pricing

• HSBC taps Mistral AI to bring generative AI across its global operations

• Watch: FieldAI’s robot dog walked 100 miles, shot 45,000 photos, and outworked a whole team of engineers

• PJM warns data centers could trigger blackouts unless grid capacity keeps up

• AI becomes the hottest college major as students chase skills and job prospects

• The AI rush is fueling massive pay increases for workers building tech’s data infrastructure

• Trump-backed plan to preempt state AI laws could net the federal government $600 billion, analysts say

• Humans and AI make better partners than expected, researchers show in large creative study

• Researchers develop compact audio tokens to help AI “listen” without losing nuance

📄 Layer-Aware Video Composition via Split-then-Merge

📄 Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer ​​

📄 GR-RL: Going Dexterous and Precise for Long-Horizon Robotic Manipulation  

📄 Infinity-RoPE: Action-Controllable Infinite Video Generation Emerges From Autoregressive Self-Rollout

📄 STARFlow-V: End-to-End Video Generative Modeling with Normalizing Flows

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!