Aramadai.sh - is an infinite canvas for all of your agents and development tools. There is a master agent that can control everything. For example, tell the master agent to start a new workspace with claude code and codex and build some project. It will do it and orchestrate the agents. You don't need to jump between agents, terminals or workspaces, you can just talk to the master agent and it will handle it.

resetting in 3..2..1...

I love people who post $30k mrr on my app, but then never actually share what their app is!? Why gatekeep the app but share the mrr?

I have usage fomo - as long as I have usage I keep adding features 🤣😂

Very close to 2k followers - hoping to get there this weekend! I have had my x account since 2012 but didn't start posting on it until just a few months ago. I had 200 followers when I started posting again. One of my biggest regrets is not posting for the last 14 years. Imagine where I would be if I had!!

Anthropic says an internal model traced the NIST curve seeds to a 1990s random generator seeded with the clock, in under half an hour. Filippo Valsorda confirmed it. Nothing changes for P-256. The lesson is about provenance.

Mercor mapped 29 AI benchmarks onto 923 US jobs: only 208 have a task tested by at least five prompts, and about 90% of work hours have little or no coverage. Coding is well tested. Most other work isn't. Build your own evals.

Reflection's Beam is out: 501B params, 23B active, 1M context, Apache 2.0, and trained to use about a third of the tokens on some agent tasks. Its own product lead says fewer tokens may not mean a lower bill. How I'd test it.

Noah Hawley gave Earth to five corporations in 2120. I think he was off by about ninety years, and too generous by one company. Four companies now own the four floors of the AI economy, and they're wired into each other.

Postgres 19 slipped from Oct 29 to Nov 19 and pulled 8 to 10 features after beta. A Register headline blames AI reviewers. The project's reason is narrower, but a committer's take on AI review and the long tail is worth reading.

Vercel Drives is GA: a persistent folder your sandboxed agent mounts and keeps across runs, with snapshots for readers and forks for parallel agents. About $6 a month for 100 GB in Vercel's example. One writer, one region. How I'd use it.

Microsoft finished pretraining MAI-Base-2: 2.8T params, 16,000 GB200s, 38 days, and a claimed 4 to 12x efficiency edge over DeepSeek's best open base. No API, no weights, no price yet. What's checkable and what I'd do.

I have released a ton of new features to Armadai.sh recently. If you are using it, you can update it by clicking the download button in the bottom left.

Claude Code Review is out of preview: agents review each PR in 6 to 10 minutes, verify findings and never block a merge. The catch: repos added from the admin page review on every push, billed by tokens. How I'd set it up.

OpenRouter's new Armor screens every agent tool call against plain-English rules before it runs, adding about 279 ms. Useful. But a trolley-problem test this week shows decision models shift with wording. How I'd roll it out.

Meta's new business terms, live Oct 30, let it crawl sites tied to business accounts and train on the content. No opt-out except leaving. Meta says robots.txt still works. What the clause says and what I'd lock down.