Open-source builder Jason argues vertical integration enables in-distribution training that outperforms prompting for tool calling

@HarryStebbings 1) models cna be trained in distribution to application, which will make it better than just prompting. 2) incentivized to get the most value for the customer. Which is a longer term game than token spend. 3) being efficient in fool calling requires training an in distribution

@HarryStebbings Will be irrelevant when on device/local LLMs match Fable (which inevitbaly they will soon enough). Until then, third party tools with the flexbility to use your own API keys will prevent single company workflow tie in and its inherent risks of a switch off or monopoly pricing.

@HarryStebbings its funny how the talking points of the harness companies (including our own) converge

@HarryStebbings There's a constant trade-off between competition and vertical integration, Apple and SpaceX being examples of the latter.

@HarryStebbings But we’re here:

Goodhart's Law: When a measure becomes a target, it ceases to be a good measure =Tokenmaxxing.

Now we’re: Valuemaxxing!!!

@HarryStebbings Separation creates competition on speed and cost. It also creates the seam where confidentiality can live, since the routing layer no longer needs to see the data.

@HarryStebbings This is wrong because the the model providers themselves are also competing with each other to lower cost and provide good service. Now if model providers are oligopolies that have cartel behavior, this changes.

@HarryStebbings The weak point is assuming the app layer has cleaner incentives by default. If it gets paid on usage, it can token-max too. The stronger protection is transparent evals, portability, and easy switching costs staying low.

@HarryStebbings * inevitably

@HarryStebbings I've just cleared 1.6B of tokens on Codex this week ($8,000+ on API), on a $200 monthly subscription

So objectively he is wrong

This is exactly the opposite of what happens. A standalone application layer benefits from consuming more tokens. A vertically integrated provider pays the inference bill itself, so every unnecessary token directly hurts margins. Its incentive is to reduce token usage, improve efficiency, and deliver more value per dollar. The current market distortion comes from subsidized pricing, not vertical integration.

@HarryStebbings This doesn't make sense, application providers can't compete on pricing

@HarryStebbings It optimizes the wrong unit. Customers don't buy tokens, they buy a finished task. Compete the model down to cheap tokens and you've won the wrong race. Real lock-in lives one layer up, in the context users can't export. What actually holds your users?