SUBJECT Signal Brief: Multimodal search goes on-device, release notes become videos, and everyone ships PRs
Knocket
Knocket — launched October 12 — is a website-visitor chat that works the way a maker does: paste one script tag and a chat bubble appears on any site (or publish a contact page for your bio), visitors start typing with no account, messages land in your Telegram where a quote-reply appears on the visitor's screen, and when you are away an AI agent answers from your docs and hands off to you when it matters — free, no credit card, live in minutes.
WHY IT MATTERS Every PM's site has silent visitors who bounce without a word — a zero-friction chat that meets them with no signup and an AI first responder trained on your docs turns the marketing site itself into a discovery channel.
knocket.trtc.io →Google EmbeddingGemma 2
Google EmbeddingGemma 2 — launched October 12 — is Google DeepMind's open, lightweight multimodal embedding model: it maps text, code, images, video, and audio into one shared space so a single query searches across all of them, and at 740M parameters it runs fully on-device (about 191MB of RAM text-only, 567MB for the full multimodal model on a Pixel 11 Pro) — Apache 2.0 licensed, weights on Hugging Face and Kaggle, with the MTEB Code score jumping from 68.76 to 78.68.
WHY IT MATTERS On-device embeddings this small make private, offline-first semantic search and RAG shippable inside a product instead of around it — one model for text, image, audio, and video collapses a whole retrieval pipeline into a component.
blog.google →ngram 2.0
ngram 2.0 — launched October 12 — is an AI video agent for product and marketing teams that turns release notes, specs, screenshots, docs, URLs, and product recordings into polished, on-brand videos: it plans the message, writes the script, builds the storyboard, and creates visual variants, and 2.0 adds animated explainers, screencasts with key moments pulled from your recordings, an AI avatar presenter, and chat-based refinement — for launches, demos, onboarding, support, sales, and social.
WHY IT MATTERS Product updates deserve better than a text-only changelog — an agent that turns the artifacts a PM already writes into launch and onboarding videos removes the agency bottleneck between shipping the feature and telling the story.
ngram.com →Base Code
Base Code — launched October 12 — lets a whole team build directly on your existing product from the browser: connect the GitHub repo, describe the change in plain language, preview it live on an isolated branch, and send the finished work through a standard pull request — no new workflow for engineering, no code required for everyone else.
WHY IT MATTERS PMs feel every small change queue behind engineering — a workspace that turns plain-language product tweaks into reviewable PRs on the real codebase moves the PM from writing tickets to shipping the change.
base44.com →Manifold MCP
Manifold MCP — launched October 12 — is one hosted MCP server that gives your AI agent live data from every marketing channel: Google rankings and backlinks, what ChatGPT and Gemini say about you, the TikTok, YouTube, and Instagram posts winning your niche, Reddit threads asking for what you make, competitors' live ads, and verified emails for buyers looking right now — ask in plain words from Claude, ChatGPT, Cursor, or Claude Code, free to start.
WHY IT MATTERS Ask an agent for growth help today and it guesses, because it cannot see rankings, competitors' ads, or live buyer signals — wiring those channels into the agent over MCP turns generic marketing advice into decisions on your actual data.
manifoldmcp.com →