Good morning. It’s Friday, February 20th.

On this day in tech history: In 1991, Guido van Rossum released Python 0.9.0, a language with clean syntax and dynamic typing that revolutionized scripting. This deep-cut milestone featured early modules for math and strings, fostering rapid prototyping. Indirectly, it became AI's lingua franca, powering libraries like TensorFlow and PyTorch for neural network training, data pipelines, and gradient descent optimizations in machine learning workflows.

In today’s email:

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

Google has debuted Gemini 3.1 Pro in preview, shifting the model from a static endpoint to an adaptive reasoning engine. The update introduces three adjustable compute tiers: low, medium, and high, allowing users to scale reasoning power per query.

The high setting functions as a Deep Think Mini, enabling multi-minute analytical runs. Benchmarks show a massive leap: a 77.1% score on ARC-AGI-2, more than doubling the 31.1% of Gemini 3 Pro and surpassing Claude 4.6 and GPT-5.2.

The "point one" update targets enterprise workloads by reducing orchestration overhead. New gains in "thinking" tokens and long-horizon planning are reflected in improved scores on Terminal-Bench 2.0 (68.5%) and MCP Atlas (69.2%). Practical applications now include text-to-animated SVG generation and live aerospace dashboards tracking the ISS.

For developers, Google launched CI Fixer for Jules, an SWE agent that autonomously diagnoses and patches continuous integration failures. Jules also gains an authorship setting, allowing it to commit code under a user’s identity for better repository governance.

Beyond the core model, Google reported that AI-driven safeguards blocked 1.75 million malicious apps from the Play Store in 2025. While internal violations dropped, Google Play Protect identified 27 million malicious apps on third-party channels, suggesting a shift in threat actor tactics.

On the creative front, Google DeepMind’s Lyria 3 is now live in the Gemini app. The model generates 30-second tracks from text, images, or video, complete with lyrics and precise control over style and vocals. All Lyria outputs include Nano Banana cover art and a SynthID watermark for AI verification. Additionally, the Pomelli tool for SMBs added Photoshoot, which generates professional product images from simple uploads to replace traditional studio photography. Read more.

OpenAI is moving to lock down the three pillars of the next economy: digital assets, cultural IP, and the global talent pipeline.

The company just launched EVMbench, an adversarial security benchmark designed to test an agent's ability to exploit and patch smart contracts securing over $100 billion in crypto. Technical results show GPT-5.3-Codex hitting a 72.2% success rate in exploit mode, more than doubling its predecessor's performance. While the model is a potent offensive tool, OpenAI is committing $10 million in credits to defensive research and $7.5 million to The Alignment Project to curb safety risks from misaligned AI.

OpenAI is now finalizing a massive funding round at a $730 billion valuation. In a pivot to immediate liquidity, Nvidia swapped its non-binding $100 billion multiyear framework for a $30 billion direct equity stake. Total capital could exceed $100 billion, with SoftBank contributing $30 billion and Amazon weighing up to $50 billion. The cash fuels a staggering roadmap: OpenAI expects to spend $600 billion on compute by 2030, betting that raw scale is the only viable defense.

The company recently hired Instagram veteran Charles Porch to lead global creative partnerships, tasked with licensing celebrity likenesses for Sora, its video-generation engine. This "cultural land grab" follows a billion-dollar deal with Disney, aimed at bridging the trust gap with creators before current labor protections expire.

Meanwhile, OpenAI is upgrading its developer experience with interactive Code Blocks, allowing users to write, edit, and preview code or mini-apps directly within ChatGPT.

The company's most aggressive infrastructure push is now in India. OpenAI partnered with Tata Group to secure 100MW of AI data center capacity, scaling to 1GW, under the Stargate initiative. This local compute supports a massive educational rollout, putting ChatGPT Edu in the hands of 100,000 students at elite schools like IIT Delhi.

However, the lead isn't guaranteed; Epoch AI reports that Anthropic is growing 10x per year and could surpass OpenAI in revenue by mid-2026. The rivalry was on full display at the India AI Impact Summit, where CEOs Sam Altman and Dario Amodei shared a visible moment of awkwardness onstage. Read more.

📄 GLM-5: from Vibe Coding to Agentic Engineering

📄 Multimodal Fact-Level Attribution for Verifiable Reasoning ​​

📄 EditCtrl: Disentangled Local and Global Control for Real-Time Generative Video Editing

📄 SLA2: Sparse-Linear Attention with Learnable Routing and QAT

📄 RynnBrain: Open Embodied Foundation Models

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!