Build Teammates on Perplexity with any frontier model

Today's top 20 insights for PM Builders, ranked by relevance from X, YouTube, GitHub, and LinkedIn.

Build Teammates on Perplexity with any frontier model

#1 š•

Aravind Srinivas said users can build Teammates on Perplexity with any frontier model they choose, claiming no other product offers this capability.

#2 š•

OpenAI shared physicist Nathaniel Craig’s account of working with GPT-6 Astra to develop a candidate lattice definition of the Standard Model. The article is the first post on OpenAI’s new science blog.

#3 š•

Sundar Pichai retweeted @_mohansolo’s statement that the context window for Argon, Carbon, and Hydrogen in Antigravity is being doubled.

Also covered by: @Demis Hassabis

#4 ā–¶ļø

Meta’s Muse: How we launched the most successful consumer product this year | Mona Sarantakos

Lennys Podcast

Mona Sarantakos recaps how Meta released Muse seven months after Nat Friedman recommended OpenClaw. Its three-phase launch process spanned product definition, reliability/safety/security work, and internal dogfooding plus external alpha testing, with agent swarms analyzing conversations, product data, and user feedback to generate synthetic data and feed evaluations.

Also covered by: @Lenny Rachitsky, @Alexandr Wang, @There's An AI For That

#5 š•

Qwen announced that Qwen4-Max is available to build with through Vercel’s AI Gateway.

#6 š•

Julien Chaumond announced that llama.cpp has reached version 1.0.0, stating that it works everywhere and asking where users will run it.

Also covered by: @clem šŸ¤—

#7 šŸ™ NVIDIA/SkillEvaluator

NVIDIA/SkillEvaluator - **NVIDIA/SkillEvaluator**, shared by NVIDIA, is an Apache-2.0 CLI framework with source code, reusable skills, examples, and documentation for validating agent skills, detecting overlap, and comparing runs with and without a skill. It has 2,265 GitHub stars and a 953 daily trending gain. Worth trying if you need structured skill evaluation and can supply the provider access, supported agent, and sandbox required for live testing.

#8 ā–¶ļø

How I fight token fraud as an AI founder

How I AI Podcast

Claire Vo shared how ChatPRD combats token fraud after abuse—including custom MCP-based 24/7 compute workers—cost the company tens of thousands of dollars over the past year. ChatPRD eliminated free trials and promo codes, added Stripe, Vercel, and Clerk protections, and uses a ZDR Jev loop to identify suspected abuse and ban or downgrade accounts, alongside hourly reviews of recent signups and anomalous chats.

#9 š•

Guillermo Rauch announced xeet, a hermetic, pixel-perfect X post renderer in safe Rust/wasm that requires zero browser, network, filesystem, or system fonts and renders PNGs in ~1ms—0.4 ms for @jack’s ā€œsetting up my twttr.ā€ Rauch says it is 100x-1000x faster than browser screenshots while just as accurate, using an ā€œautorepoā€ workflow to verify tens of thousands of text-layout rules.

#10 šŸ™ kajisho5/ffmpeg-skill

kajisho5/ffmpeg-skill - **kajisho5/ffmpeg-skill** is a local video-editing skill for Claude Code, Cursor, Codex, and MCP clients, with 42 typed FFmpeg tools, a machine-readable contract, an MCP server, examples, and verification workflows. The MIT-licensed repo has 6,187 stars and 331 daily trending gains. Worth trying if you need agents to edit and verify video offline, with FFmpeg and Python 3 on PATH.

#11 š•

Harrison Chase shared a post from @vitalflux recommending that builders write the evaluation runner before the AI agent. The original poster said they are building an agent using LangChain deep agents.

#12 šŸ™ google-labs-code/stitch-skills

google-labs-code/stitch-skills - **google-labs-code/stitch-skills** is a library of reusable Agent Skills and plugins for turning frontend code into Google Stitch designs and generating React and React Native components and Remotion walkthrough videos. It was observed with 9,595 GitHub stars and a daily trending gain of 540. Worth trying if you need design-to-code workflows and can configure the required Stitch MCP server.

#13 š•

Qwen said Qwen4-Max made significant improvements, particularly on coding and finance tasks, referencing a Vals AI evaluation.

#14 in

Udi Menkes recaps OpenAI’s announcement on the Friday preceding his post of its ChatGPT Wallet, saying ChatGPT’s personal agent, dot, can complete purchases after approval without seeing card details and with security checks, in a limited U.S. rollout for Pro subscribers through ā€œdots.ā€ In a demo—not a real order—dot suggested a $59.97 herb box for an approximately $100 budget, highlighting the need for accurate product data, agent access, and agent-ready risk systems; Menkes also says more than 5,000 product leaders follow GenAI PM.

#15 š•

Sebastian Raschka shared a link to the full article about understanding positional embeddings.

#16 š•

Harrison Chase shared a post from @vitalflux discussing whether to use a LangGraph workflow or a deep agent for a procure-to-pay (P2P) agent on LangChain. The source is truncated, so the final choice and implementation details are unavailable.

#17 š•

Kevin Yien commented that dismissing AI products as ā€œjust chatā€ is like calling SaaS ā€œjust forms.ā€ He loosely believes many distinct chat products will emerge because they’re easier to understand—not necessarily five different AI chat products for one user, but five different people using five distinct chat products for their respective jobs.

#18 š•

Yann LeCun argued that open foundation models, like roads and the Internet, require government, industry, and academia to build and operate the infrastructure while allowing anyone to build services on top. He said diverse AI assistants are needed for the same reason as a diverse press.

#19 in

Ben Erez commented that OpenAI and Anthropic are not the whole AI product job market: AI-centric roles also span big-tech frontier-lab organizations, AI application companies, and companies applying AI to improve core products. He said strong career opportunities exist outside frontier labs and recommended asking hiring managers how their requirements have changed since last year.

#20 š•

DeepLearning.AI announced that Dylan Couzon is scheduled to lead a hands-on workshop on building agent memory with retrieval at AI Dev NYC, featuring a simulated fly brain that remembers bad experiences and is tested on avoiding new hazards. The event runs November 30–December 1, 2026, at Chelsea Piers in New York.

Get tomorrow's brief first

Join AI product managers receiving the latest brief before it reaches the public archive.

Subscribe free