Good morning. It’s Monday, February 24th.

On this day in Tech History: In 1955 Steve Jobs, the visionary co-founder of Apple Inc., was born in San Francisco, California.

In today’s email:

• Google’s VEO 2 Price Drop

• AI Learns To “Shut up”

• Apple Plans Texas Factory

• 5 New AI Tools

• Latest AI Research Papers

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

Google DeepMind Prices Veo 2 AI Video at $0.50 Per Second

Google DeepMind has set the pricing for Veo 2, its latest AI video generation model, now available through Google Cloud’s Vertex AI platform. The cost is $0.50 per second, or $1,800 per hour, significantly lower than the $32,000 per second spent on Hollywood productions like Avengers: Endgame. However, this places it at a higher price point than OpenAI’s Sora, which offers an unlimited-use subscription for $200 per month.

While both Veo 2 and Sora represent major advancements in AI-generated video, neither eliminates the need for human refinement, iterative testing, and post-production work. Achieving professional-quality results still requires manual adjustments to scene composition, motion fluidity, and visual consistency. Read more.

AI Gets Smarter by Knowing When to “Shut Up”

A Johns Hopkins University study finds that AI isn’t just getting better at answering questions—it’s getting better at knowing when it shouldn’t. By extending processing time, researchers discovered that models sharpen both accuracy and self-awareness, challenging the long-held belief that AI should always provide an answer, even when uncertain.

Testing DeepSeek R1-32B and s1-32B on math problems from the AIME24 dataset, researchers adjusted reasoning depth to observe changes in decision-making. While both models performed similarly under normal conditions, DeepSeek R1-32B stood out when stricter confidence thresholds were applied, showing that longer computation time helps models make smarter calls on when to step back.

The study introduced three risk levels—from low-stakes decisions to critical, high-penalty environments—revealing how AI adapts under different pressures. Researchers argue that future AI assessments should prioritize not just accuracy, but the ability to recognize when silence is the smartest move. Read more.

Apple plans Texas factory for AI servers, 20,000 research jobs

Apple is making a decisive move in AI and research, announcing a 250,000-square-foot AI server factory in Texas, built with Foxconn and set to launch by 2026. This initiative is part of Apple’s $500 billion US investment wave, which also includes 20,000 new research and development jobs, solidifying its domestic AI dominance.

The Texas facility will manufacture AI servers fueling Apple Intelligence, the company’s machine-learning suite for automation, content generation, and predictive computing. Alongside this, Apple is doubling its Advanced Manufacturing Fund to $10 billion, with major backing for TSMC’s Arizona silicon production, ensuring AI hardware stays under tight strategic control.

Apple is also launching a Michigan-based manufacturing academy, offering free training in process optimization and industrial strategy. Read more.

• Microsoft Cancels Data Center Leases as OpenAI Turns to SoftBank for Compute

• Apple may be preparing Gemini integration in Apple Intelligence

• Meta AI Unveils MLGym to Benchmark and Train AI Research Agents

• Freepik Partners with Google to Launch Veo 2 AI Video Model

• EngineAI: The world's first humanoid robot to perform a front flip

• Humanoid robots building more robots 

• Google Colab lowers the barrier to AI and data science with free cloud computing

• Why Convergence’s Proxy is outpacing OpenAI’s Operator in the AI agent race

• OpenAI bans Chinese accounts using ChatGPT to edit code for social media surveillance

• Alibaba plans to spend $53 billion on AI in a major pivot

• Nvidia CEO Jensen Huang says market got it wrong about DeepSeek's impact

• Researchers discover how AI models process time-specific knowledge

📄 SurveyX: Academic Survey Automation via Large Language Models

📄 MaskGWM: A Generalizable Driving World Model with Video Mask Reconstruction ​​

📄 Logic-RL: Unleashing LLM Reasoning with Rule-Based Reinforcement Learning

📄 S*: Test Time Scaling for Code Generation

📄 Intuitive physics understanding emerges from self-supervised pre-training on natural videos 

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!