Rahil

building things with ai, sometimes for ai.

[Rahil @Anot]

@paulg @t_blom Trying to build a whole bank this way as a thought experiment made me realize how much of what orgs do isn't the actual work, it's scaffolding for coordinating people.

Take the humans out from the start and a lot of the structure just doesn't need to exist.

[Rahil @Anot]

All this back and forth on whether AI takes the jobs or just "transforms" them. But maybe... could you just take the humans out of the loop entirely?

A thought experiment. What if you don't transform anything, and start a bank with only agents. Not agentic transformation. A full universal bank... built, run, and overseen by agents. Banking other agents.

Play around, see what the agents do: https://t.co/q2mOl4wSPZ

Next up, probably the OCC charter and FINRA app? Hopefully they don't ask for headcount.

[Rahil @Anot]

@ClaudeDevs or just ultracode and skip the pairing homework

[Rahil @Anot]

@rmstein Exciting stuff! Tell me these agents have browser use and you'll have me deprecating all my vibe-coded trackers and https://t.co/vmwuiVPCfG instances.

Either way, love the direction.

[Rahil @Anot]

https://t.co/k80sIgD2ze

[Rahil @Anot]

update on teaching claude how to consult: the mechanics work. where it gets interesting is how it fails. cited withdrawn regulations as current. mixed up geographies. invented rate cards. claimed experience that never happened.

got it asking questions before building anything now. analyst to associate upgrade, roughly. still aware...

https://t.co/i9Gt5YAHr5

[Rahil @Anot]

just submitted a PR to Anthropic that teaches Claude how to do my job

yes, i'm a consultant. yes, i'm aware.
https://t.co/WtKzGouhH7

[Rahil @Anot]

pentagon calls Dario a liar with a god complex, blacklists anthropic, Sam signs a defense deal same day. if you ever wanted to look up what these people have actually said about... all of this.
search it, chat with it. https://t.co/SAWa0czcQ9

[Rahil @Anot]

Spent three sessions with Claude trying to extract a single corporate PowerPoint template into code. It placed a black logo on a black background. Hallucinated shapes that didn't exist. Took 15 debug cycles to get slide number padding right.

8MB of conversation. 350 lines of working JavaScript.

Looked at everything. Claude Code skills, Cowork, Copilot in PowerPoint, python-pptx, PptxGenJS, raw OOXML. Every method breaks on corporate templates in a different way. And Microsoft hasn't shipped a sane API for creating presentations in 30 years.

The fix is boring. Extract your template into code once, commit it, let the LLM just research and write content against it.

https://t.co/d1vLygiRnJ

[Rahil @Anot]

@statesman texas has 550+ data center facilities, more than any other state. 28 GW of total capacity. a lot of those planned sites are in areas already flagged for water stress. hard to blame communities for pushing back..

[Rahil @Anot]

@YahooFinance the pipeline data backs this up. about half the announced capacity is landing in a handful of corridors already dealing with stretched grids.

[Rahil @Anot]

@zerohedge there are 15 facilities over 500 MW operating in the US today. the pipeline has 125. at that scale you're basically a utility, might as well act like one

[Rahil @Anot]

@business @tomwblack @opinion trend tracks with what the pipeline data shows. when the grid can't keep up with 250 GW of announced demand you just build around it. https://t.co/0eSR3UVVQs has the facility-level breakdown

[Rahil @Anot]

@MarioNawfal the nuclear move makes sense when you look at the pipeline. there are ~125 planned facilities over 500 MW in the US right now. total announced capacity is around 250 GW. the grid adds maybe 50-60 GW per year. they don't really have a choice. https://t.co/0eSR3UVVQs

[Rahil @Anot]

built a dataset of every US data center. ~4,500 facilities at https://t.co/rJ9MScOGbA. there are 15 over 500 MW operating today. the pipeline has 125. median planned facility is 225 MW vs 10 MW existing. total announced capacity equals 5+ years of national grid additions.

close to $700B in hyperscaler capex this year. half of it landing in corridors already dealing with stretched grids and water stress. capacity, pipeline status, climate risk, water stress, and energy mix for each facility.

how much of this actually gets built is anyone's guess.

[Rahil @Anot]

also an @openclaw skill. so your agent installs it, reads the skill file, and suddenly knows how to pull financial data without burning a gazillion opus tokens on API docs.

[Rahil @Anot]

built a cli for pulling stock quotes, SEC filings, crypto, economic data from public APIs. got tired of juggling a dozen services with different auth and response formats so now it's one interface. also works as an AI agent skill so they stop wasting tool calls on API docs.
https://t.co/bB49WA6dqS

[Rahil @Anot]

@jelanifuel you're missing the daemon. openclaw runs 24/7 with its own memory system, messaging integrations, cron engine. claude code ootb starts and stops per task. pretty different animal

[Rahil @Anot]

@JeffGrayCyber @JorgeCastilloPr yeah residential proxies exist but you still deal with rotation, session management, providers getting burned. or you can set up a local exit node via tailscale and route your VPS through your home IP. both work, just more moving parts..

[Rahil @Anot]

@ExquisiteHabits @adamc0dez just ask it. say "i want you to be able to scrape reddit" and it'll figure it out. or look up repos on github where someone already solved what you need and point your agent at it.

[Rahil @Anot]

@yoavgo people confuse building with shipping. ai makes it easy to keep tinkering. the hard part is still deciding what's good enough and who actually needs it. and even harder to keep it working for others vs just yourselves...