Good morning. It’s Wednesday, November 19th.

On this day in tech history: In 1992, Sepp Hochreiter’s Munich thesis exposed the vanishing/exploding gradient curse that doomed deep RNNs on long sequences. Ignored until Schmidhuber’s lab revived it, this insight birthed LSTM in 199, and ultimately forced the invention of attention that powers today’s Transformers and LLMs.

In today’s email:

• Gemini 3

• 5 New AI Tools

• Latest AI Research Papers

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

_{In partnership with Public}

🚀 The investing platform for people who take it seriously

Public puts stocks, bonds, treasuries, options, crypto, ETFs, direct indexing, IRAs, and high-yield cash in one place.

Top benefits:
• 3.6% APY on cash
• 5.6% bond yields locked in at purchase
• 1% uncapped match on IRA transfers & rollovers
• Direct Indexing from $1,000 (tax-optimized)
• Options trading with rebates and no per-contract fees

Plus elite tools: AI-powered earnings recaps, Key Moments, Income Hub, Investment Plans, and Queue.

^{Thank you for supporting our sponsors!}

Google stakes AI leadership with Gemini 3

Google has taken a clear lead in AI with Gemini 3 and Gemini 3 Pro, dominating benchmarks in math, science, multimodal reasoning, and agentic workflows. Gemini 3 Pro now tops the Artificial Analysis Intelligence Index, beating GPT-5.1 and Claude 4.5 across GPQA Diamond, MMLU-Pro, LiveCodeBench, SciCode, and the new AA-Omniscience knowledge test. Its sparse mixture-of-experts architecture, trained on licensed, synthetic, and user data, handles one-million-token contexts, while Deep Think mode boosts multi-step hypothesis generation.

The model powers Search, AI Studio, Vertex AI, Chrome, and the Gemini app. In Search, it produces interactive visual layouts, real-time simulations, and custom tools tailored to queries. Developers also gain advanced agentic and vibe coding, automating multi-step workflows, updating legacy code, and prototyping apps faster. The new Antigravity platform introduces an agent-first architecture for asynchronous, verifiable coding workflows, enabling autonomous agents to plan, act, and verify tasks across IDEs and enterprise systems. Early users report measurable gains in speed, accuracy, and workflow efficiency.

Gemini 3 is enterprise-ready via Vertex AI and Gemini Enterprise. Gemini app gets a new upgrade with new generative interfaces, immersive visual layouts, and Gemini Agent for multi-step task execution. Scholar Labs accelerates research by analyzing questions, mapping relationships, and retrieving relevant papers.

Education sees the launch of three Gemini AI certifications for K–12 students, university students, and educators, offering verifiable credentials to validate AI proficiency and promote responsible use.

DeepMind’s WeatherNext 2 with Gemini 3 outperforms a supercomputer in both accuracy and efficiency, proving its capabilities extend to complex scientific forecasting.

Following the launch of Google’s newest AI model, the DeepMind CEO Demis Hassabis and Google Gemini VP Josh Woodward joined the Hard Fork podcast to break down the capabilities of what they call their “most intelligent” AI yet. Read more.

xAI launches Grok 4.1 with stronger reasoning and emotional intelligence

xAI has officially rolled out Grok 4.1, moving the chatbot from its earlier, unpredictable persona into a more reliable and emotionally aware digital assistant. The update sharpens real-world reasoning, creative problem solving, and collaborative interactions while reducing hallucinations and improving response speed.

During a silent rollout from November 1–14, users preferred Grok 4.1 over the previous model in roughly 65% of blind tests, showing clear gains in interaction quality and accuracy. Grok 4.1 (Thinking) briefly held the top position on LMArena with a normalized Elo score of 1,483 before being overtaken when Google released Gemini 3, which posted an even higher 1,501 Elo.

The update also pushed Grok 4.1 to a leading 1,586 on EQ-Bench3 for emotional intelligence and kept it among the strongest performers in Creative Writing v3 and LLMArena’s tests for versatility, cultural context, and linguistic precision. With refined learning pipelines and context-aware algorithms, Grok 4.1 positions xAI to compete more effectively with ChatGPT and Claude. Read more.

• Jeff Bezos joins Project Prometheus as co-CEO to lead ‘AI for the physical economy’

• Microsoft forms new AI alliance with Anthropic, expands autonomous AI

• LeJEPA could be LeCun’s final Meta project, using theory-driven design to outpace massive pretrained models

• Meta’s DreamGym lets AI agents train entirely in text, cutting RL costs and complexity

• NVIDIA launches Apollo, a family of open AI physics models aimed at accelerating industrial and scientific simulation

• Meet AI-Newton, the AI that’s figuring out physics from scratch while giant LLMs struggle

• Call of Duty’s AI images spark congressional backlash as Ro Khanna demands worker protections and anti-automation taxes 

• Behind CoreWeave’s AI gold rush hype lies a company teetering on debt and insider sell-offs

• Weibo’s open-source AI flips the script on LLM scale, besting DeepSeek-R1 at a fraction of the cost

• Bitcoin miner Bitfarm is ditching crypto for AI, betting its data centers are more valuable than coin

• China’s Unitree Robotics clears pre-IPO hurdles as it eyes $7 billion Star Market debut

• ElevenLabs launches all-in-one Image & Video platform to turn ideas into polished AI-generated content

• New wearable lets users command robots on the move with AI that cancels motion noise

• Watch: Disney’s smart robots train to fall, roll, and land safely without damage

• Watch: AI takes the checkered flag. Germany’s TUM claims victory in first-ever autonomous racing championship

• Meet Volodya, the first humanoid flight attendant to greet passengers mid-air

• Companies are experimenting with AI, but experts say human-centered upskilling is the real key to adoption 

• Cloudflare acquires Replicate, giving developers access to 50,000+ AI models via a single API

• Perplexity named the AI giant most likely to implode as investors favor Anthropic over OpenAI

• GlobalFoundries acquires Singapore’s Advanced Micro Foundry to scale AI-ready silicon photonics

• AI advocacy super PAC goes after New York lawmaker leading state-level regulation

• AI pay boost may be peaking sooner than expected, Penn professor warns

• Bank of America ramps up AI to turn 15-client bankers into 50-client powerhouses

• Boaz Weinstein’s Saba sells credit protection on Big Tech as AI-fueled debt surges

• AI security startup Runlayer raises $11M, lands eight unicorn clients including Gusto and Instacart

• Laptops are finally catching up to AI, with NPUs, unified memory, and smarter chips bringing local LLMs within reach

• Poe adds group chats for 200 users, pushing collaborative AI to the next level

• Scientists develop a way to stop AI models from being mathematically stolen

• New AI model flags short videos likely to trigger suicidal thoughts before they go viral

• Phi-4 shows “data-first” fine-tuning is the key to outperforming larger models

• Meta beats FTC in antitrust case, retains Instagram and WhatsApp

• Nvidia powers two new Japanese supercomputers to accelerate AI and quantum research

• The Milky Way, fully mapped: AI makes 100-billion-star simulations possible

• Arm and Nvidia team up to link CPUs and AI GPUs with NVLink Fusion for faster cloud workloads

• The rise of AI in recruiting is creating chaos in hiring, with trust at an all-time low

• Cloudflare confirms global disruption, attributing outage to bot mitigation service failure

• Physical Intelligence’s π*0.6 brings self-learning robots to real-world tasks

• Intuit is bringing TurboTax, QuickBooks, and Credit Karma into ChatGPT with a $100M+ OpenAI deal

• Instagram CEO says AI engineers don’t need Ivy League degrees, just scrappiness and fast learning

• GOP lawmakers plan a new push to override state AI laws before year-end

• Hugging Face CEO warns the AI hype may be an “LLM bubble” set to burst next year

• Gen Z students are turning to double majors as an AI-proof strategy in a volatile job market

• Stack Overflow is remaking itself into an AI data provider

• Two-thirds of companies will slow entry-level hiring due to AI

• Alibaba launches upgraded Qwen chatbot as it makes a serious push into the consumer AI market

• Manus AI launches Browser Operator extension

• Autonomy Association and NASA test AI-powered data fabric that lets flying vehicles make real-time decisions in cities

📄 MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling

📄 P1: Mastering Physics Olympiads with Reinforcement Learning ​​

📄 DoPE: Denoising Rotary Position Embedding 

📄 Souper-Model: How Simple Arithmetic Unlocks State-of-the-Art LLM Performance  

📄 MMaDA-Parallel: Multimodal Large Diffusion Language Models for Thinking-Aware Editing 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!