Alibaba Introduces Efficient Qwen 3.5 Series
Today's top 25 insights for PM Builders, ranked by relevance from X, LinkedIn, Blogs, and YouTube.
Alibaba Introduces Efficient Qwen 3.5 Series
#1 đť•Ź
Qwen released its 3.5 Medium Model Series (Qwen3.5-Flash, 35B-A3B, 122B-A10B, 27B), with the 35B-A3B outperforming larger 235B predecessors. Qwen3.
#2 đť•Ź
NVIDIA AI launched the Red Hat AI Factory with NVIDIA, combining Red Hat AI Enterprise and NVIDIA AI Enterprise software to streamline how organizations develop, deploy, and scale AI workloads on NVIDIA-accelerated infrastructure.
#3 in
Udi Menkes announces a multi-year Intuit–Anthropic partnership to bring Claude Code’s AI to 100 million customers and let mid-market businesses build custom AI agents via the Claude Agent SDK.
#4 in
Guillermo Rauch announced Vercel AI Gateway now supports video generation and released an open-source xAI Grok Creative Studio (v0-grokstudio.vercel.app) offering free Grok Imagine Video & Image until tomorrow.
#5 đť•Ź
Google AI partnered with Google DeepMind and YouTube to launch Music AI Sandbox, a suite of AI-driven tools for pro musicians, and Grammy-winner Wyclef Jean shows in a behind-the-scenes video how he used it to co-create his track “Back from Abu Dhabi.”
#6 đť•Ź
Harrison Chase made evals a Day 0 requirement for monday.com’s AI service agents using LangSmith, cutting feedback loops from 162 s to 18 s (8.7× faster), running hundreds of tests in minutes, and enabling real-time, end-to-end production monitoring.
#7 đť•Ź
Santiago unveiled Claude Code, a Nimble-powered Claude skill that scrapes live structured data from any website in real time. It outputs normalized tables (e.g., 2-bedroom rental prices, square footage, URLs) ready for immediate spreadsheet use.
#8 đź“ť Simon Willison
First run the tests - Argues that automated tests are essential when working with coding agents because agents can quickly modify code; tests ensure AI-generated code actually works and guard against deploying unexecuted code.
#9 ▶️
Notion’s AI-powered prototype playground: How designers are building better products faster
How I AI Podcast
Notion AI’s design team uses a Next.js monorepo called Prototype Playground, powered by Claude Code (Opus-4.5) and Cursor, to rapidly build production-ready prototypes with AI-assisted tooling.
- Prototype Playground is a Next.js app deployed on Vercel, with an app directory namespaced by designer under app/[username], standalone pages without a global layout, and shared Notion UI templates importing colors, typography, and a library of over 5,000 icons.
- Designers run Claude Code in plan mode within Cursor’s terminal UI, enter voice prompts via Monologue, and rely on a global cloud.md and an uncommitted cloud.local.md file to configure project tools (Bun, Tailwind), workspace paths, and MCP server settings for Figma and Chrome Dev Tools.
- Custom slash commands and Claude skills include /create-prototype (auto-generates page.tsx and metadata), /figma (imports a Figma frame via Figma MCP, generates code, then loops with Chrome Dev Tools MCP for up to three verification iterations), a find-icon skill (writes a TypeScript script to scan 5,000+ icon files for correct names), and /deploy (uses GitHub CLI to create a branch, commit, open a PR in the browser, and poll CI and Vercel deployment statuses every 60 seconds until all checks pass).
#10 ▶️
How I Use Obsidian + Claude Code to Run My Life
Greg Isenberg
Vin demonstrates integrating Obsidian’s Markdown vault with Claude Code via Obsidian CLI to feed interlinked notes as context and run custom LLM commands for workflows like daily planning, idea generation, and tracing idea evolution.
- The /today command pulls calendar entries, tasks, iMessages and the past week’s daily notes into Claude Code, outputting a prioritized plan for the day.
- The /trace demo command scanned all interrelated vault files and traced Vin’s Obsidian usage over a 13-month period—first appearing January 11, 2025—identifying phases like “Discovery and skepticism” (Jan–May 2025) and “Explosive building” (Jan 2026).
- The /ideas demo command took over five minutes to gather vault structure and context from sources labeled “Obsidian orphans”, daily notes, “new context” and “personal agent infrastructure” before producing an actionable idea report divided into structural highlights, tools to build and systems to implement.
#11 đť•Ź
Cursor built secret redaction for model tool calls and validated it end-to-end by having its AI agent record a three-chapter walkthrough video of the local build.
#12 đť•Ź
Andrej Karpathy highlights N8python’s GitHub gist that handcrafts neural weights for a mini addition model. It provides a clear, small-scale view of mechanistic interpretability in LLMs.
#13 in
Udi Menkes built a Claude Code skill called /one-step-better that pulls today’s GenAI PM briefs, analyzes your current project, and tells you the single insight to actually apply (like Nat Eliason’s $3,596 OpenClaw side-project hack).
#14 đť•Ź
LlamaIndex 🦙 finds that OmniDocBench’s 1,355-page OCR benchmark is maxing out—models like GLM-OCR hit 94.6% accuracy yet still misread complex financial, legal, and domain-specific docs.
#15 đź“ť PromptLayer Blog
How do you observe LLM systems in production? - LLM observability is essential once models are live because they can hallucinate, generate unexpected costs, or slow down in ways traditional monitoring misses. The article outlines connecting inputs, outputs, latency, cost, and quality to get a single picture of model health.
#16 đź“ť Ampcode Chronicle
Mainframe Magic - A practical guide to migrating COBOL/mainframe code using an agent, covering the workflow and benefits of automating migration tasks. It shows how an agent can assist with analysis, transformation, and verification during migration.
#17 đť•Ź
Dharmesh Shah advises packaging your GTM expertise into “knowledge agents” on agent.ai to automate common use cases and potentially offer them as a paid perk to your community—email him for a direct intro to the team.
#18 đť•Ź
Teresa Torres and Petra Wille reveal how PMs stretching into bug fixes, tech debt, and architecture breeds burnout and poor quality, challenging legacy IT mindsets and the “CEO of the product” myth to redefine healthy product–engineering boundaries.
#19 đť•Ź
Lenny Rachitsky found that 30 current and recent job seekers use AI not just to polish resumes but to build custom feedback tools, question predictors, and story-surfacing workflows for interviews.
#20 đť•Ź
Lenny Rachitsky has launched a free live workshop series with MavenHQ on “The AI-Native PM,” featuring top product leaders across three themes: AI workflows, technical skills, and product sense & influence.
#21 đť•Ź
Aravind Srinivas rolled out an upgraded voice mode on Comet, enabling full hands-free browser control. The Comet iOS app with this feature arrives in a few days and is available for pre-order.
#22 đť•Ź
Kevin Yien disables Stripe’s successful payments notification—while most alerts stay on by default—to avoid spam when business is booming.
#23 in
🥞 Carl Vellotti calls out Opus 4.6 for needlessly loading eight files to answer a two-sentence question and rarely spawning context-saving agents. He shares a “Context Management” snippet to drop into your CLAUDE.md to fix it.
#24 đť•Ź
Peter Yang demystifies APIs, Skills, and MCPs with a professional kitchen analogy—APIs are the kitchen itself.
#25 đť•Ź
Cursor swaps out code diffs for demo videos, enabling agents to record and share working software walkthroughs directly.