Today · agents that see, remember, and get counted
Give your agent the ability to find out for itself.
— WebKit, on handing Safari to an agent · story 02
Generative Video · NotebookLM
Your notes, now a sixty-second reel.
NotebookLM's new Short Video Overviews turn whatever you upload — a PDF, a pile of notes, a dense paper — into a 60-second vertical clip that walks through one idea, phone-shaped and grounded in your own sources rather than the open web. Gemini Omni Flash drives the motion; Nano Banana 2 Lite paints each frame in about four seconds. Live now for AI Ultra and Pro on web and mobile, English first, with free access said to follow.
Dev Tools · Agents
Hand the agent a real browser.
WebKit shipped an MCP server that plugs an agent straight into a live Safari window: read the DOM, take screenshots, run JavaScript, watch network and console traffic, click and type, and flag missing ARIA labels or weak contrast. One line wires it into Claude Code. Stop writing the perfect bug report for your agent — let it look for itself.
Research · Agent Memory
Memory that keeps the point, drops the transcript.
Microsoft Research's Memora splits what an agent stores from how it recalls it — full content on one side; six-to-eight-word index phrases and flexible "cue anchors" on the other — so a long-running agent stops re-reading its whole history to stay on task. It clears Mem0, RAG, and full-context on the standard long-memory suites while spending a fraction of the tokens. Code is on GitHub; the paper lands at ICML 2026.
98% fewer tokens
- 86.3%
- LoCoMo accuracy
- 87.4%
- LongMemEval accuracy
- ~½
- memory entries stored
Privacy · Crawler Economics
Fifty thousand crawls. One visitor back.
Cloudflare's Attribution Business Insights dashboard shows publishers the number that matters: how many times each AI company crawled you against how many visitors it sent back. Ratios run from 118:1 to nearly 50,000:1. It sorts bots into Training, Search, and Agent, with volume, country, and bandwidth per operator — so a licensing conversation can start from data instead of vibes.
50,000:1crawls per referral, worst case
TrainingSearchAgent
Creative · Agents
An agent that chases the algorithm.
Viral Bench is an open agent that runs the whole social-growth loop on its own, then grades itself on the views it actually earns — turning "can a model go viral" into something you can measure. The repo, from the doublespeed / lightreel crew, is small and readable: a working template for anyone wiring an agent to a publishing API. Whether that thrills or unsettles you, the harness is on the table.
01Read the trendScans what's climbing on the feed that day.
02Make the postGenerates the carousel and copy to match.
03Ship to TikTokPublishes autonomously through the API.
04Score the viewsReal engagement becomes the benchmark.
Creative · Ambient AI
The diary writes back.
Maxime Rivest wired a reMarkable e-ink tablet to Claude Fable 5 and rebuilt Tom Riddle's diary: you write by hand, the ink fades from the page, and a reply appears in the same looping script. It is a toy, but a pointed one — the e-ink latency and the handwriting round-trip make the model feel less like a chat box and more like an object that answers. The cheapest way to feel where ambient AI is heading.
“who are you?”
…the ink sinks into the page…
I am the diary. Write, and I will answer.
That's today.
Six picks from a picked-over week: Google's NotebookLM, WebKit, Microsoft Research, and Cloudflare, plus two builds surfaced via @denissexy and @seeallochnaya. Chosen against one test — could a senior engineer or founder act on it this week — from roughly eighteen feeds of model labs, engineering blogs, and Telegram. Issue 077 · Monday 06 July 2026 · Zürich.