Stephen

[Stephen @ashortstephen]

Related thought on incentives. I’ve been a little taken back recently at just how often Claude Code arrives at “you know what would make this project even better, give me an anthropic api key”.

[Henry Dowling @henrytdowling]

Since the standard is to hide thinking traces now, the frontier lab pricing model from the perspective of a user, is basically "just trust me bro here's what it cost".

Like, the user is charged for a bunch of thinking tokens that they never got to see on a usage basis.

This creates misaligned incentives between model providers and consumers. If you're Anthropic, and you find a way to reduce thinking tokens by 10% without compromising response quality, are you really gonna press that button to reduce your top line revenue by 10% right now, as you gear up for a potential IPO?

[Stephen @ashortstephen]

I had a weird aha moment with Claude last week that recouped me €1.7k. I set Claude loose on bank statements from all my accounts, mainly just wanting to get a lay of the land on where my money was actually going.

Claude picked up that I’d been charged twice for mortgage fees, once through the notary and once by the bank directly.

It had access to my ‘second brain’ directory structure, including the notary statement, but the transaction itself wasn’t broken down beyond a generic bucket of fees. It managed to marry part of that up with a random €1.7k charge from the bank, on a different account, a month later.

Slightly embarrassing to admit I wasn’t on top of this. When the second charge came in, while unpleasant, I recognised what it was, but being a month after the notary transaction, I just didn’t compute that I’d already paid it.

Such a trivial example, but one that I doubt would never have been caught by me or the bank. Claude casually found this buried across my accounts and documents when I didn't even ask It to look. Makes me think of what's happening in the world of fraud detection with more powerful models and hardware and targeted application.

[Stephen @ashortstephen]

@rawC7Z @joshperrone98 @_brinxinx Most people when they lose their phone send a message using find my “lost please contact #” I received a phishing Apple ID login when it happened to me. Also some old sims even include phone number on the card itself.

[Stephen @ashortstephen]

A lot of people are jumping on the ISO 24495 / ASD-STE100 Simplified English bandwagon, and I can see the value for getting Claude to respond with less fluff and more directly.

But for anyone using it for copywriting, sense-checking emails, etc., or anyone already worried about their content looking like AI slop and using light-touch prompts like “make this sound more coherent, but grammatically correct, while sticking to my original writing style”, I’d be wary. Adopting these standards just strips all the personality and humour from your writing.

People weren’t meant to write to each other like the instruction manual for a Philips toothbrush.

[Stephen @ashortstephen]

Something I keep coming back to is how much better positioned orgs would be to build their company brain and institutional intelligence if shared network drives were still a thing.

With O355/GSuite everything just became a link to a specific file in someone’s personal space. Few took the time to set up proper team/department structures in the cloud like they would network drives.

It means most orgs have little open information and no permission model for AI to inherit.

[Cerebras @cerebras]

https://t.co/FvKy1StDpQ