But why would 10% to 20% go to Anthropic rather than being mostly diffused across all providers with the majority actually going to cheap open model providers who provide 90% of the intelligence for the lowest cost?
Here's the thing though: like labor today, employers will constantly be pushing for the lowest possible price. Especially in a world where open models are 90% as capable, the cost of AI labor will quickly drop. Thats not a good thing for Anthropic or any other frontier AI lab. There's a reason most companies don't just hire the top 5% of intelligent individuals. Its really just not needed for most business.
Also another aspect: think of the most highly paid workers right now in the US and the companies they work for. Assume that normal people no longer have jobs due to AI. How are these companies going to continue making revenue to pay Anthropic? Anthropic might get a quick buck but a long term growth business is not macro-economically possible for frontier AI businesses.
Its already commoditized. Thats why things like OpenRouter are useful and successful because it lets you route inference at the click of a button between different providers after balancing cost and availability. Exactly how commodities behave.
Hey Frenck, longtime HA user here. I know this feature feels pretty safe but I hope you can give http://cablej.io/blog/k-anonymity/ a read on how this ultimately still puts trust in HIBP, a third party cloud service. In my opinion this might help with understanding why people are rightfully worried, especially when they thought they are running a very local and private HA instance.
Thats the thing though, it's not safe [1] and puts trust in a third party - HIBP.
To do that automatically without any notice to the user breaks the fundamental value proposition of a local home assistant which over and over again preaches keeping things local and private.
As someone who relies on home assistant every day, I'm quite disappointed to hear Frenck's response to this.
From the issue author: "I can give you a list as long as my arm about how I could abuse this to loot credentials from every HA install on the planet.".
From the linked article I can count 2 with following conditions:
1. relies on a malicious or compromised lookup server, linking lookups to the same user, the user checking two predictably similar passwords, both passwords being breached, the attacker having their plaintext candidates, and knowing the account identifier for account takeover
2. relies on a malicious or compromised lookup server, linking lookups to the same user, the user checking multiple predictably related passwords, sufficient computing resources to search the original password space, and knowing the account identifier for account takeover, the attack demoed seems to work if you have a weak password to starts with (10 chars alphanum).
Yeah, I think the situation is that if your password is already weak/known to a level that it should already be considered compromised, a malicious HIPB could surface that compromise a lot more reliably/quickly.
Can someone explain to me the economics of AI and how it intersects with billing by the hour?
I think there would be a strong incentive not to use tools that speed up your work because you'd effectively be able to bill less time?
I'm sure there are some firms out there with more work than people, but still wouldn't it be more effective to hire another human who can then bill at a high rate for many hours?
Lawyers are client-facing. Client wants lawyers to use AI. There are justifiable reasons to use AI (in measured doses). Hence, firms will use AI. They already have all kinds of strategems to increase their billing which AI slots into.
we can only estimate token <> cost in our cloud agent, but we also OSS'd the skills and in those cases the cost is ~$0 (part of one's LLM subscription): https://github.com/usetrmnl/trmnl-agent-skills
the agent also has a model switcher, and the estimate you quoted is based on Haiku when we launched.
reply