Google Rolls Out Gemini 3 Deep Think
Today's curated insights on AI product management from 100+ sources across X, LinkedIn, and YouTube.
Google Rolls Out Gemini 3 Deep Think
From X
AI Product Launches & Updates
Anthropic Interviewer pilot launched: Anthropic AI @AnthropicAI announced Anthropic Interviewer, a new tool to help us understand people’s perspectives on AI, now available for a week-long pilot.
Codex model available in Cursor: Cursor AI @cursor_ai announced the new Codex model is available in Cursor, free to use until December 11th, optimized via collaboration with OpenAI.
Gemini 3 Deep Think rollout: Demis Hassabis @demishassabis shared that Gemini 3 Deep Think is now available for Google AI Ultra subscribers in the Gemini app, incorporating gold-medal IMO & ICPC technologies to tackle complex maths & science problems.
AI Tools & Applications
Customizable MCPs in v0: v0 @v0 announced you can now bring your own MCPs (Modular Component Packs) to v0, using presets or configuring custom MCPs in just a few clicks.
Bug triage with coding agent: LangChainAI @LangChainAI shared an AI agent built with LangSmith Agent Builder that creates Linear issues from Slack messages, prioritizes and assigns tasks, and updates existing issues to save engineering time.
NPM superpowers unlocked in Base44: Base 44 @base_44 announced support for NPM packages like GSAP, Chart.js, Radix UI and more, enabling advanced 3D, motion, richer dashboards and file uploads via simple prompts.
Product Management Insights & Strategies
Listening skill deep dive: Shreyas Doshi @shreyas posted a previously private piece on the skill of listening, revealing an unorthodox but highly effective approach, complete with a 15-minute AI-generated audio podcast.
LinkedIn’s Full Stack Builder program: Lenny Rachitsky @lennysan highlighted LinkedIn’s new Full Stack Builder program, which teaches employees to build, design and ship products end-to-end, replacing its APM program.
Rethinking competitor analysis: George Nurijanian @nurijanian noted that most competitor analysis is documentation theater and suggested focusing instead on competitors’ beliefs and how those beliefs constrain their response.
AI Industry Developments & News
Titans architecture revealed: Google Research @GoogleResearch introduced Titans, a new architecture combining RNN speed with Transformer performance and deep neural memory to scale to contexts larger than 2 million tokens.
Devin agent at scale in finance: Cognition @cognition reported that Brazil’s largest bank, Itaú, deployed the Devin AI agent across its SDLC for 17,000+ engineers, achieving significant efficiency across hundreds of thousands of repositories.
Training AI models on Hugging Face: Clement Delangue @ClementDelangue shared how they used Claude code, Codex and Gemini CLI with Hugging Face skills to train high-quality AI models—even if you’ve never trained a model before.
From LinkedIn • Deeper Insights
AI Tools & Applications
Paweł Huryn’s guide to multi-agent systems warns against over-engineering with AI agents by default and recommends using code or visual orchestration for deterministic tasks, calling large language models and agents only when necessary. He evaluates five leading platforms—no-code options like Google Workspace Studio, low-code tools such as n8n and Make, and code-first frameworks like LangChain and LangGraph AI—and highlights n8n as a free, versatile choice for PMs to automate workflows efficiently.
Product Management Insights & Strategies
Geoff Charles’s overview of Ramp’s daily feature release process shows how teams leverage an early-access tier, automated AI-powered checks on goals, customer sentiment, discoverability, and cross-functional readiness, followed by a 48-hour leadership review. This template streamlines decision documentation and ensures quality at high velocity.
Brian Balfour’s deep dive into Dreambase.ai’s AI-first workflow highlights four pillars: consolidating vision in an AI Requirements Doc (AIR), running multi-tool “bake-offs” for ideation, starting with a data-first schema to align product and engineering, and splitting roles for focused AI coding versus customer engagement. This structure exemplifies how PMs can orchestrate teams around rapid learning and AI effectiveness.
AI Industry Developments & News
Guillermo Rauch’s post on Vercel’s self-driving infrastructure introduces Fluid compute for AI workloads, an AI Gateway acting as a token CDN, and an AI SDK for deploying intelligent agents. He cites Thomson Reuters as an example of shipping agents at scale, demonstrating how serverless platforms are evolving to support end-to-end AI applications.
From YouTube
AI dev 25 x NYC | Kay Zhu: How Genspark Built a Super Agent That Scales
Deeplearning.ai • December 05, 2025
Kay Zhu, CTO of Mainfunc (Genspark), presents the design and scaling of Genspark’s Super Agent platform—showcasing how its autonomous, mixture-of-agents architecture with 80+ specialized tools, context management and autoprompt capabilities enables rapid end-to-end task automation.
Key Takeaways:
- After launching its Super Agent suite in April 2025, Genspark achieved over $50 million in ARR and 10 million global users within five months, earning BVP’s “AI supernova” tag and joining OpenAI’s trillion-token club.
- The platform follows a “less control, more tools” philosophy, using mixture-of-agents orchestration, prompt caching, context isolation and error recovery to let agents autonomously plan workflows without rigid, predefined flows.
- In live demos, power users turned raw Excel rent rolls into investor-ready presentations in 10 minutes (vs. 10 hours manually), mined SEC filings to build real-estate prospect decks, and converted 90-minute podcast transcripts into 4-minute audio summaries and a custom web page.
AI Dev 25 x NYC | Hatice Ozen: Build a Deep Research Agent with One API Call
Deeplearning.ai • December 04, 2025
Hatice Ozen demonstrates how to use Groq’s compound AI system to create a fast, server-side deep research agent that performs real-time tool integrations like web search, code execution, and browser automation with a single API call.
Key Takeaways:
- Groq’s custom LPU hardware delivers AI inference up to 10–40× faster than traditional GPU-based APIs, drastically reducing latency in agent workflows.
- The Groq compound model bundles the best third-party tools server-side—such as web search, code execution, and browser automation—so developers don’t need to manually orchestrate state or API calls.
- Agents built with Groq compound support include/exclude domain parameters for data-source filtering and will soon allow custom tool integration via MCP servers for greater flexibility.
AI Dev 25 x NYC | Jacky Liang: Why Agents Can't Find the Right Docs (And How Postgres Fixes It)
Deeplearning.ai • December 04, 2025
Jacky Liang demonstrates how pure vector search struggles with exact matches—like mixing Postgres 16 and 17 docs—and walks through implementing hybrid search in PostgreSQL using PG Text Search and PGVectorScale with reciprocal rank fusion to deliver precise documentation retrieval without external search stacks.
Key Takeaways:
- Vector-only search often returns semantically related but incorrect documents—e.g., suggesting Postgres 16 connection guides for a Postgres 17 query—because embeddings ignore exact terms and version differences.
- Hybrid search fuses semantic vectors and BM25-based keyword matching via the reciprocal rank fusion (RRF) algorithm (1/(K+rank)) to rank results that score highly in both similarity and exact keyword presence.
- TigerData’s native PostgreSQL extensions—PG Text Search for modern BM25 ranking and PGVectorScale for high-performance embeddings—enable single-database hybrid search, removing the need for separate vector databases, keyword engines, and ETL pipelines.
AI Dev 25 x NYC | JoĂŁo Moura: Design, Develop, and Deploy Multi Agent Systems with CrewAI
Deeplearning.ai • December 04, 2025
In this talk, JoĂŁo Moura, CEO of CrewAI, reveals how CrewAI runs 450 million AI agents per month and 1.1 billion agentic executions quarterly, outlines the emerging 'agentic systems' design pattern for blending deterministic flows with autonomous agent crews, and shares case studies that cut CPG reimbursement from three days to ten minutes and KYC verification from one week to under 30 minutes.
Key Takeaways:
- CrewAI operates 450 million AI agents per month and executed 1.1 billion agentic executions in the last quarter, illustrating rapid scale.
- JoĂŁo Moura introduces 'agentic systems', which combine deterministic workflows with optional multi-agent crews to balance simplicity and dynamic decision-making.
- Real-world deployments include a CPG company reducing reimbursement validation from three days to ten minutes and a global financial institution automating KYC to be more accurate than humans while cutting process time from one week to 15–30 minutes.
AI Dev 25 x NYC | Gary Qi: TRAE: Redefining Coding Agents
Deeplearning.ai • December 04, 2025
Gary Qi introduces TRAE Solo, ByteDance’s AI-powered IDE that transforms coding agents into a context-aware “context engineer” platform unifying development tools and multi-agent workflows to automate end-to-end software creation.
Key Takeaways:
- ByteDance’s TRAE IDE merged its VS Code extension and standalone product into one platform in April 2025, achieving over 1 million monthly active users, 6 billion lines of code processed, and 1.5 million daily queries.
- TRAE Solo integrates IDE, terminal, doc view, browser, and Figma into a context-rich environment with Solo Builder for zero-to-one prototyping and Solo Coder—a responsive coding agent featuring plan mode, diff view, and sub-agent workflows.
- Solo’s multi-agent architecture connects to internal tools like the Links front-end framework and CI/CD pipelines, automates debugging via direct browser error logs, and supports voice control for fully automated development.
The end of product managers? Why LinkedIn is turning PMs into AI-powered “full stack builders”
Lennys Podcast • December 04, 2025
Tomer Cohen, LinkedIn’s CPO, details the Full Stack Builder program that replaces traditional PM roles with AI-assisted end-to-end builders by rearchitecting LinkedIn’s platform for AI, deploying custom agents, and introducing new career tracks and performance incentives.
Key Takeaways:
- LinkedIn data shows that 70% of skills required for current roles will change by 2030 and 70% of today’s fastest-growing jobs didn’t exist a year ago, driving the push for AI-powered builders.
- The Full Stack Builder model collapses 15+ research sources and multiple review steps into a fluid human+AI workflow, enabling small cross-functional pods to research, prototype, design, code, and launch products end-to-end.
- LinkedIn has created custom AI agents—Trust, Growth, Research, Analyst, and QA maintenance—rearchitected its codebase into AI-friendly composable UI components, and replaced its APM track with an Associate Full Stack Builder program emphasizing AI fluency.