In partnership with

Good morning. It’s Wednesday, April 9th.

On this day in tech history: 1979: The first fully functional Motorola DynaTAC prototype, a precursor to the modern mobile phone, was demonstrated.

In today’s email:

• Gemini 2.5 Pro

• Deep Analysis of Llama 4

• MS Copilot Turns Vision into Context

• New AI Tools

• Latest AI Research Papers

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

Google Rolls Out Deep Research and New AI Tools with Gemini 2.5 Pro

Google has launched Deep Research for Gemini Advanced users on the experimental Gemini 2.5 Pro model, now accessible via web, Android, and iOS. Designed as a personal research assistant, it excels in reasoning and structured synthesis, surpassing competitors with a 2-to-1 user preference margin. Features like Audio Overviews provide podcast-style narration, making insights more portable. Positioned not as a chatbot but a full-stack tool, it integrates deeply into real-world knowledge workflows.

In tandem, Google Research introduced a Geospatial Reasoning framework, enhancing crisis response and urban planning by leveraging remote sensing models trained on satellite imagery. On a separate front, Google Cloud has deepened its collaboration with the Allen Institute for AI (Ai2), unlocking open-source models like OLMo 2 for sectors demanding transparency. Google also previewed its AI-powered 3D remastering of The Wizard of Oz for Sphere’s 16K immersive screen, pushing cinematic boundaries.

On the AI tools front, Google has upgraded both AI Mode and NotebookLM. AI Mode now includes visual recognition, enabling users to upload images for contextual insights and suggestions. Meanwhile, NotebookLM allows users to pull live web content into their notes, integrating external sources for enhanced generative capabilities. These updates mark a significant leap in AI-driven, real-world applications across Google’s platforms. Read more.

Meta's Llama 4 Models Impress in Some Areas, But Face Criticism Over Long-Context Tasks

Meta's submission of the customized Llama 4 models to LM Arena has stirred up a storm of questions, particularly around transparency. The "Llama-4-Maverick-03-26-Experimental" version was fine-tuned for human preference, but this wasn’t made clear initially. Meta's VP of generative AI, Ahmad Al-Dahle, has denied rumors that the company artificially boosted its Llama 4 models' benchmark scores.

In response, LM Arena dropped over 2,000 battle results, highlighting how style and tone swayed evaluations. They’ve also revamped their leaderboard policies, reinforcing their commitment to fair and reproducible tests. Artificial Analysis has also updated its Llama 4 Intelligence Index scores for Scout and Maverick, following adjustments to account for discrepancies in Meta’s claimed MMLU Pro and GPQA Diamond results.

On the performance front, Meta’s Llama 4 models—Maverick and Scout—impressed with strong scores in reasoning, coding, and mathematics, outpacing rivals like Claude 3.7 and GPT-4o-mini. Maverick clocked 49 points, while Scout followed with 36. However, when it came to long-context tasks, both models hit a wall. Maverick managed just 28.1%, and Scout trailed even further at 15.6%. Meta points to ongoing tweaks and optimizations as the models are gradually rolled out.

Meanwhile, NVIDIA has turbocharged Llama 4’s inference on its Blackwell B200 GPUs, pushing the models to over 40,000 tokens per second. With a multimodal, multilingual architecture and TensorRT-LLM optimization, these models now handle tasks like document summarization and image-text comprehension with impressive speed. Read more.

Microsoft’s Copilot Turns Vision into Context

Microsoft has extended Copilot Vision from web to mobile. Now accessible via the Copilot app for iPhone users subscribed to Copilot Pro, the feature turns a phone’s camera into a real-time visual search tool. Point your camera, and Copilot deciphers what it sees—plant health, product specs, interior tweaks—all processed in real time.

The tech runs inside Voice mode of the Copilot app and only activates with permission. Powered by OpenAI’s latest models, it’s a cognitive overlay for the physical world. Microsoft’s broader push is clear: collapse the distance between seeing and knowing. $20/month buys you early access and premium speeds. Read more.

• Stanford's 2025 AI Index Maps a Costly, Crowded Race

• ElevenLabs Launches MCP Server for Voice Agent Access via Text Prompts

• Runway Launches Gen-4 Turbo: 10-Second Videos in Just 30 Seconds

• Amazon expands Bedrock with speech, video, and multilingual coding AI

• Together AI Launches Open-Source DeepCoder-14B, Challenging Code-Crunching Giants

• New open source AI company Deep Cogito releases first models and they're already topping the charts

• Nvidia's new Llama-3.1 Nemotron Ultra outperforms DeepSeek R1 at half the size

• OpenAI Introduces the Evals API: Streamlined Model Evaluation for Developers

• Google is allegedly paying some AI staff to do nothing for a year rather than join rivals

• AI heads for $4.8T by 2033, cementing its place as the next great industrial terrain

• Anthropic announces 100 roles in Europe, new EMEA head

• China’s quantum computer cracks billion-parameter AI, nudging frontier physics into machine learning

• Nvidia and Supermicro lead chip and AI stock rebound after tariff hit

• Amazon says its AI video model can now generate minutes-long clips

• IBM releases a new mainframe built for the age of AI

• $115 million just poured into this startup that makes engineering 1,000x faster — and Bezos, Altman, and Nvidia are all betting on its success

• Sakana AI Introduces The AI Scientist-v2

• ArXiv Launches Deep Research for Instant Lit Reviews

• China: First-ever mega humanoid robot training hub opens with plans to teach 100+ models

• KPMG's AI bot slashes interview scheduling time by 60%, saving over 1,000 hours

• Andreessen Horowitz seeks to raise $20 billion megafund amid global interest in US AI startups

• EU moves to trim GDPR, eyeing leaner data rules in coming weeks

• Tesla and Warner Bros. Win Part of Lawsuit Over AI Images from 'Blade Runner 2049'

• White House orders agencies to develop AI strategies and name leaders

• Hyundai expands robot fleet with thousands of Atlas units from Boston Dynamics

• Microsoft AI chief Suleyman sees advantage in building models ‘3 or 6 months behind’

• China’s DeepRoute.ai to team up with Qualcomm to develop advanced driver assistance solutions

• Sam Altman defends AI art after Studio Ghibli backlash, calling it a 'net win' for society

📄 One-Minute Video Generation with Test-Time Training 

📄 T1: Tool-integrated Self-verification for Test-time Compute Scaling in Small Language Models

📄 MATH-Perturb: Benchmarking LLMs' Math Reasoning Abilities against Hard Perturbations

📄 OmniSVG: A Unified Scalable Vector Graphics Generation Model

📄 An Empirical Study of GPT-4o Image Generation Capabilities

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!