This is I think critically important to understand for building AI applications:
an ai application is just the following things:
1. context
2. memory
3. prompts/agent loops
4. tool calls
5. surrounding human-facing software
Christina says, "look, I cant in good faith resell you a token. my tokens are just claude tokens, and you should just get them from claude"
For many many companies this is true because, mapping to the above:
1. they don't own the context, so claude can fetch the context (in slack or notion or whatever)
2. they dont own the memory (same as above)
3. their agents are just claude/oai
4. they may have proprietary tool calls, but tools are mostly softwre these can be rebuilt in many cases.
5. they have proprietary software, but again, this can be rebuilt or is less useful in agentic world
I think some companies that have staying power here are the following:
1. apollo - they *do* have proprietary tool calls that are hard to replicate
2. rippling/palantir - they *do* have proprietary softare thats been painstakingly built via research about the real world. in the worst case, these become proprietary tool calls, in the best case its a foundation to compete as a standalone agent.
3. i initially had mixpanel on this list. I think it sort of is for similar reasons
4. stripe
You effectively have to rigorously encode so much business logic that an agent would look at your tool call API and say, "well yeah i aint doing that myself", and then uses its agentic credit card to subscribe to your service instead of building in house
hard to do, but I do think its time to bet the farm on it. not much time left
With respect, this is terrible advice. You need to prove -- to yourself more so than to investors -- that intelligence is more valuable inside your product, than on top.
If tokens *aren't* worth 4X inside your product, what are you even doing? You're just reselling the labs. 1/