Alibaba Introduces Qwen 3.5 Small Models for Edge
Today's top 24 insights for PM Builders, ranked by relevance from X, YouTube, and LinkedIn.
Alibaba Introduces Qwen 3.5 Small Models for Edge
#1 đť•Ź
Qwen launched the Qwen 3.5 Small Model Series—0.8B, 2B, 4B & 9B-parameter native multimodal models with improved architecture and scaled RL, optimized for fast edge deployment and lightweight agents while closing the gap to larger models.
#2 đť•Ź
Google DeepMind launched Nano Banana 2, an AI tool boasting upgraded vibrant lighting, richer textures, and sharper elements to take your visual ideas to the next level.
#3 đť•Ź
Logan Kilpatrick announced that Gemini 3 Pro will be shut off on March 9, and users are encouraged to upgrade to the 3.1 Pro Preview, which bundles numerous feedback-driven improvements.
#4 đť•Ź
Claude now offers Memory on the free plan, and you can easily import or export your saved memories anytime.
#5 ▶️
Super Nested Claude Code Is Vibecoding On STEROIDS
All About AI
A controller agent using T-Max and nested Claude Code spawned six parallel cloud code instances to generate a procedural 3JS space galaxy and four instances to create a real-time microGPT training dashboard.
- The controller ran on the Opus model and launched six parallel Claude Code instances in T-Max for modules galaxy, objects, render, spacecraft, UI, and index, each receiving tailored prompts.
- Hostinger’s VPS (KBMT2 plan, $9.99/month with coupon code ALLABOUTAI, Germany region) deployed OpenClaw in about five minutes via automated setup using an OpenAI key.
- Four nested Claude Code instances—backend, dashboard, charts, and samples—ran microGPT.py (a ~200-line dependency-free Python script) for over 220 steps, showing live cross-entropy loss (2.5–7) and outputting names Alan, Mol, Anna, Maron, Pandla, Anan, and Jana.
#6 in
Guillermo Rauch introduced Queues, Vercel’s simple (two-method) durable event streaming API that powers higher-level DX tools like Workflow (useworkflow.dev) and can even serve as a serverless Celery backend.
#7 ▶️
Cloudflare just slop forked Next.js…
Fireship
Using AI, Cloudflare rebuilt the Next.js API on V into V-Next in three days—reaching 94% API coverage and enabling Next.js apps on Cloudflare Workers with up to 4.4× faster builds and 57% smaller bundles.
- Cloudflare used AI tokens costing about $1,100 to implement basic SSR, middleware, server actions, and streaming in one day and achieve full client hydration on Workers by day three.
- After a week of edge-case fixes and test-suite expansion, V-Next attained 94% coverage of the Next.js API.
- In Cloudflare’s own Trust Me Bro benchmarks, V-Next production builds were up to 4.4× faster than Next.js with client bundles 57% smaller thanks to Vit and rolldown (the Rustbased bundler).
#8 ▶️
How Coinbase scaled AI to 1,000+ engineers | Chintan Turakhia
How I AI Podcast
Chintan Turakhia scaled AI across 1,000+ Coinbase engineers by embedding Cursor-based rules for routine tasks, staging speedruns that generated thousands of PRs in minutes, and building an in-house Cloudbot agent in Slack and Linear to automate feedback-to-PR workflows.
- Between January and April 2025, Cursor was used hourly by leadership to define Cursor rules for unit tests and linting, and a “cursor-wins” Slack channel documented wins such as generating 20 unit tests in one session.
- In a 15-minute “Cursor speedrun,” 100 engineers created 70 draft PRs with the “create draft PR” command, and in a company-wide speedrun, 800 engineers pushed 3,400 PRs in 30 minutes, breaking GitHub’s performance limits.
- AI-assisted tooling reduced average PR review cycle time from 150 hours to approximately 15 hours (10× improvement) and an in-house “Cloudbot” agent now converts live audio feedback into Linear tickets and GitHub PRs in seconds via Slack.
#9 ▶️
Claude Code & MCPs built my $145K marketing machine
Greg Isenberg
Cody Schneider live-builds a GTM engineering workflow with Claude Code agents, orchestrating Phantom Buster, Instantly AI, Refonic, Railway.com and the Facebook Ads API to automate LinkedIn outreach, bulk-generate and publish 100 Facebook ads, and dynamically optimize ad performance with real-time dashboards in under 30 minutes.
- Creates a “Graph Growth Agents” folder with a .env file holding API keys for Intercom, SendGrid, HubSpot, Cow.com, Perplexity, Facebook Ads API, MillionVerifier and Instantly AI to standardize Cloud Code integrations.
- Runs a Claude Code LinkedIn responder agent that, over 15 minutes, auto-comments on all posts containing the keyword “triage” by ingesting a Notion document and the post URL without manual intervention.
- Uses Claude Code to generate 1080×1080 React-based Facebook ad creatives—sourcing pain points from Reddit via Perplexity API—exports them via HTML-to-canvas as a ZIP, bulk-uploads 100 drafts into a specified ad set via the Facebook Ads API, then employs a Graph MCP to pause high-CPM ads and boost top performers into separate ad sets.
#10 in
🥞 Carl Vellotti demos four ways to connect Claude Code (Direct APIs, MCP servers, CLI tools, Chrome extension) to tools like Jira, Figma, Slack and Google Docs—highlighting CLI tools as the zero-token, zero-context “hidden gem.”
Also covered by: @Peter Yang
#11 in
Peter Yang shares Carl’s top five Claude Code integrations—Google Workspace for meeting prep, Linear for ticket creation, Slack for status updates, and Reddit for monitoring.
Also covered by: @Peter Yang
#12 đť•Ź
Santiago shared six core Claude coding rules—add them to your CLAUDE .
#13 đť•Ź
Philipp Schmid shares five practical tips to evaluate AI agents: define outcome, process and style goals; run 20–50 real-case tests; apply deterministic graders and LLM style judges; then grade outputs to measure improvement.
#14 đť•Ź
LlamaIndex 🦙 LlamaParse’s layout image saving feature now returns cropped screenshots of figures, charts and other layout elements directly in the parsing response.
#15 đť•Ź
NVIDIA AI Kuo Zhang, President of Alibaba, explains how AI agents like Accio slash product-sourcing timelines from weeks to hours, empowering entrepreneurs to compete globally.
#16 đť•Ź
Boris Cherny shares that running `/setup-terminal` in Apple Terminal enables native paste support—see code.claude.com/docs/en/terminal-config for setup details.
#17 đť•Ź
Santiago warns that Cowork’s multi-step file workflow—granting folder access, copying files, then generating a plan—slows you down when dealing with many or changing files, and suggests native chat-based file support for a smoother experience.
#18 đť•Ź
Claude rolled out a new Memory feature—head to Settings to get started and import your data at https://claude.com/import-memory.
#19 in
Udi Menkes turned four All About AI videos into a step-by-step GenAI PM guide and used it to build a 30-second, 60-word AI avatar clip for just $1.38. He leveraged 11 Labs for voice, Kling AI Avatar for lip-sync consistency, and Remotion for editing and branding.
#20 in
Claire Vo and Coinbase’s Chintan Turakhia explain how they scaled AI to 1,000s of engineers—building a mini Cursor app, a video/voice→Linear→Claude feedback pipeline, and adopting a code-first leadership model.
#21 in
Marc Baselga notes product leaders now see lack of Claude Code access to repos as a red flag when choosing a company. Connecting Claude Code lets PMs get instant, structured answers to deep code queries instead of lengthy engineer discussions.
#22 in
Greg Isenberg urges PMs to rebuild every SaaS tool—Notion, Slack, Stripe, etc.—as agent-native (payments, communication, memory) because the coming machine-to-machine economy will feature billions of software agents as customers.
#23 đť•Ź
Google DeepMind launched a model that lets you modify creations to exact specs, offering outputs in any aspect ratio. It also upscales content from 521px to 2K and 4K, giving full creative control.
#24 đť•Ź
Logan Kilpatrick removed the chat-removal feature after overhauling conversation history—its absence was hurting agent performance—and the team is now working to reinstate it.