The article do mentions why they don't use multimodal retrieval. Also I think this approach is cheaper (compute wise) than multimodal retrieval. From the article:
Multimodal retrieval does not suit this domain. CLIP-style embeddings wash out exactly the fine detail that matters in charts, tables, and annotated screenshots, and short technical queries ("how do I configure X") give too little signal to match against image vectors
What do you think they are, what do you want them to be, and how much revenue and growth do they need before low margins make for a monster of a company?
From all their desperation in making sure api keys are not used in contexts where they are not supposed to, I would say that they actually appear to have services where their profit is negative, if a customer is actually using their api to the limits they set, they lose money. They wouldn't have been this desperate in trying to shut off OpenClaw if it wasn't this way. Most companies that provide api infrastructure love when a killer app using their api is made by outsiders.
And while you can beat low margin with scale, there is the famous joke "we lose money on every sale, but make it up in volume".
If you scale a low margin operation, you can become giant. If you scale a loss making operation, you go bankrupt.
Yup - the subscriptions are a VC subsidy. They've been phasing their enterprise customers directly onto API-pay-per-usage pricing (hence the recent reports from Uber and Microsoft about phasing out Claude code). Rest assured, many of their customers are happy paying for the value they get from Claude code.
The subscription is the loss leader to show you how good it is. And people think it's good and worth paying for.
There is some reason to think their margins will improve, also: they couldn't really plan for the capacity they've needed so far this year, so they're paying through the nose for it. That's fine because they can pass the cost onto customers and give a more reliable service at cost. But in a few years, they should be able to get those costs under control (presuming some ops excellence. Something Google has in spades)
reply