REACH
How far a system gets to put its hands: a coding agent through a thousand tool calls, an open-weight model into somebody's sandbox, one policy across a whole robot body.
One curl, and Meta's coding agent lives in your terminal
Muse Code is a terminal agent that plans changes, writes them, and validates the result across large repositories. The part worth stealing is the runtime: it keeps a local event log, which makes a run replay-exact and restart-safe, so an agent that dies at tool call 600 resumes instead of starting over. Background agents run async against the same log. It ships with bundled skills — /plan, /grill, /goal — and is driven by Muse Spark 1.2, co-trained with the harness rather than dropped into it.
curl -fsSL https://dev.meta.ai/install.sh | bash
# Muse Spark 1.2 gains reported on:
Terminal-Bench 2.1 · DeepSWE 1.1 · Meta Internal Coding Bench
The model mapped the attack surface, then wrote the fuzzer
Malte Ubl points an open-weight model, Kimi K3, at production infrastructure: it enumerated the attack surface and built fuzzers on its own. It did not get out. The window between "an open model can do reconnaissance" and "an open model can finish the job" is the thing Ubl is measuring, and his argument is that you should spend it buying defensive review now — frontier models running DeepSec turned up IDORs, XSS and SSRF, and Vercel pays tens of thousands of dollars a quarter for full passes.
Offense · open weights
Kimi K3 maps the surface and authors fuzzers against a sandbox. No escape yet. The capability is downloadable and the price only falls.
Defense · frontier
Sol 5.6 under DeepSec finds IDORs, XSS, SSRF. Quarterly full reviews at tens of thousands of dollars — cheaper than the alternative.
Fifteen seconds at 2K, and the sound is already in the file
MiniMax H3 generates video up to fifteen seconds at 2K with native stereo audio — the track is produced with the picture instead of dubbed over it afterwards, which removes the sync pass that eats most of a short-form pipeline. Per second, it lists at under a third of mainstream model pricing at 2K and under half at 768p. MiniMax says weights are coming "in the coming days, subject to applicable laws and regulations." That clause is the whole question: whether this lands as an open model or an API with an open-model press cycle.
Legs and arms on one policy, and <200 examples to fit a new body
Gemini Robotics 2 is a vision-language-action model that drives whole-body motion, so a humanoid's legs and arms come off the same policy rather than being treated as an arm bolted to a cart. The on-device variant adapts to a different embodiment in a few hours of work, typically on fewer than 200 examples — which is the number that decides whether a lab can afford to switch hardware. Multi-robot collaboration lets unlike robots split a workflow neither finishes alone.
On-device adaptation: hours, not weeks · typically < 200 demonstrations per new embodiment
He mailed a potato with his face on it, and the customer replied
Jake Sciotto's job at PostHog is getting users to talk to him, and he wrote down what actually works. Trade thirty minutes for merch. Build somebody a dashboard before asking for anything. Tell a power user they are in the top ten of literally all PostHog users, then send a merch code. When none of that lands, print your face on a potato and post it. His own read is the useful part: the merch is a pretext, and what earns the reply is a message that sounds like a person wrote it.
The ladder
- 30 minutes of your time for some merch.
- A dashboard built for them, unasked.
- "You are in the top 10 of all PostHog users" + a code.
- A potato. With his face on it.
DEPT · Customer success
RULE · No corporate voice
That's today.
CC · Five stories, five sources, none repeated
RULE · Nothing older than 30 days, nothing shipped before
Today's picks came from Meta AI Research, Vercel, MiniMax, Google DeepMind and PostHog, surfaced through the daily sweep of newsletters, model-lab feeds, engineering blogs and Telegram channels. The rubric, in order: tools you could adopt this week, creative and generative software, dev tools and agentic coding, privacy and security when the signal is high, research with a practical kernel, and anything a senior engineer or founder could act on tomorrow.
OpenAI · Anthropic · Meta AI Research · Google DeepMind · MiniMax
GitHub · Sentry · Vercel · PostHog · Cloudflare · Fly.io · Next.js
@seeallochnaya · @TochkiNadAI · @denissexy · @rvnikita_blog · @ProductsAndStartups