Google Updates Gemini 3 Deep Think

#1 𝕏

Google AI unveiled Gemini 3 Deep Think for tackling complex science and engineering challenges and rolled out the Genie-3-powered Waymo World Model for advanced autonomous scenario mastery.

Also covered by: @Jeff Dean

#2 𝕏

Philipp Schmid announced the Gemini Interactions API now supports multimodal function calling, letting agents natively see, process, and return real images (not just text) with Gemini 3’s image processing and mixed text/image outputs.

Also covered by: @Jeff Dean

#3 𝕏

DeepLearning.AI: Mistral released the open-weights Ministral 3 family (14B, 8B, 3B params) via its new cascade distillation method, yielding vision-language models that rival or beat similarly sized competitors while using far less training data and compute.

#4 𝕏

Qwen launched Qwen AI Slides, powered by Qwen3 Agent and Qwen-Image 2.0. It auto-researches, structures, and generates polished slide decks from simple text or document inputs.

#5 𝕏

Google DeepMind launched Project Genie on Google Labs, letting U.S. Google AI Ultra subscribers start creating AI-powered projects immediately.

#6 𝕏

Fei-Fei Li rolled out Marble’s new Advance model, giving users editable controls to build true 3D worlds—add scenes outside windows, tweak room vibes, plant decor, and generate multiple variations via an image→pano→edit→create workflow.

#7 𝕏

Logan Kilpatrick just made paying for the Gemini API 10x easier by letting you upgrade to a paid account, track usage, and filter spend by model directly within AI Studio.

#8 𝕏

LlamaIndex 🦙 PostHog has integrated LlamaIndex into its LLM Analytics, demoing an agent workflow that uses LlamaIndex, LlamaParse, and OpenAI to match product specs to user needs.

#9 𝕏

Santiago unveiled an open-source agentic coding toolkit featuring parallel isolated agents, headless CI/CD mode, and ACP integrations for Neovim, Zed, Emacs, making it barrier-free to try.

#10 in

Chun Jiang launched Synthetic Users, enabling hundreds of parallel agent sessions so PMs can A/B test onboarding flows and value props during product discovery.

#11 𝕏

Guillermo Rauch highlights Vercel’s new content-negotiation approach for agent-friendly pages, letting a single endpoint auto-deliver JSON to AI agents and HTML to browsers.

#12 in

Kevin Weil: GPT-5.2 Pro discovered a general closed‐form for “single-minus” gluon amplitudes at arbitrary n, and an internal scaffolded model then proved it—overturning decades-old assumptions in theoretical physics.

#13 𝕏

Boris Cherny: Spotify’s top developers haven’t written a single line of code since December, using Claude Code to ship 50+ features via Slack and even patch bugs from their phones.

#14 📝 PromptLayer Blog

The Emergence of Agent-First Software Design - The post argues a shift from traditional explicit-programming toward agent-first design, where engineers orchestrate agents rather than hard-code decision trees. It frames this as an emerging paradigm for building software with LLMs and agentic systems.

#15 𝕏

Jason Zhou praises tinyClaw for its streamlined setup—he spun it up in just two minutes with two commands (curl -fsSL https://raw.githubusercontent.com/jlia0/tinyclaw/main/scripts/remote-install.sh | bash tinyclaw start) after finding ClawBot more cumbersome.

#16 ▶️

My 24/7 Claude Code AI Agent’s Biggest Win...that won’t happen again

All About AI

#17 ▶️

Stop Shipping AI Slop. Design with Weavy AI, Claude etc.

Greg Isenberg

#18 𝕏

Guillermo Rauch warns that splashy, hype-driven launches are increasingly uncorrelated with success and vulnerable to AI-driven Sybil attacks, so founders should start with one user for validation, then expand through close design partners, deepening their moat before going p...

#19 📝 Simon Willison

Thoughtworks - Key takeaways from a Thoughtworks retreat: junior developers remain highly valuable because AI helps them overcome early productivity deficits, while the bigger problem is retraining mid-level engineers who lack fundamentals; apprenticeship and rotation models were discussed but no solution was identified.

#20 in

Peter Yang notes that AI-native startups like Factory, Ramp, and Linear delegate tasks to AI agents across engineering, PM, design, and sales, letting humans focus on context, systems, and feedback loops.

#21 𝕏

DeepLearning.AI: The Wikimedia Foundation teamed up with Amazon, Meta, Microsoft, Mistral AI, and Perplexity to launch high-speed API access to Wikipedia’s datasets, accelerating AI model training while funding its infrastructure.

#22 𝕏

NVIDIA AI TogetherCompute’s optimized inference stack on NVIDIA Blackwell powers DecagonAI’s real-time voice concierge, handling hundreds of queries per second with sub-second latency and 24/7 reliability.

#23 📝 Simon Willison

The evolution of OpenAI’s mission statement - Explains that OpenAI's nonprofit must file annual IRS returns and that the mission description on those returns has legal significance; the post traces how OpenAI's stated mission has evolved over time and links to a longer article for full details.

#24 📝 Armin Ronacher

The Final Bottleneck - AI speeds up writing code, but human accountability and the capacity to review still impose hard limits.

#25 𝕏

Peter Yang praises Linear’s new homepage as a live demo: every left-nav item is interactive, and you can Slack-message or assign tasks to AI agents directly on the page.