57) Adoption

Behind-the-scenes building Vambrace AI, a company on a mission to figure out its mission. Please pardon the stream-of-consciousness style. Subscribe to follow along or visit the site here:

(typos are to make sure you’re paying attention)

Introductory Remarks

Dear Vambracers —

In last week’s post, Open Source, we looked at Palantir CEO Alex Karp’s recent comments on CNBC against frontier models like OpenAI and Anthropic. His criticism was obviously in pretty transparent self interest, but it caused a stir in the industry nonetheless. And I do think there were some valid nuggets nestled within the diatribe. I suppose only time will tell. Moving on!

Adoption

I wanted to write briefly today about the most important piece of the AI revolution: Adoption. First, an overview of the AI value chain as we understand it today (I’ll take every opportunity to create a purple picture):

• Raw Materials & Inputs: the raw stuff coming from the earth used to build the components and equipment that generate, store, and animate the intelligence

• Chips & Hardware: the components that go into the equipment that generates, stores, and animates the intelligence

• Data Centers & Compute: the equipment that generates, stores, and animates the intelligence

• Foundation & Frontier Models: the companies that create the intelligence and work to make it better (faster, smarter, cheaper)

• AI Applications: the user- (or agent-) facing digital (and, eventually, physical) interfaces that leverage artificial intelligence to do something (typically, to transform some input or set of inputs, x, into some desired output, y)

• End-Use & Adoption: the actual end-user that interacts with the AI Applications to transform x into y; they most likely have historically been responsible for the transformation of x into y or are now newly-responsible for the transformation of x into y

This post claims that, within this value chain as it exists today, the End-Use & Adoption piece presents the most challenges. Specifically, the models have become incredibly powerful, but to actually harness that power for real, quantifiable organizational impact is really really difficult—and there are always many more hurdles and obstacles to a reliable enterprise-grade AI-enabled production system than people realize.

Adoption of AI tools, ironically, requires a deeply human approach to: (a) understanding the nuances of some workflow that transforms x into y, (b) applying significant organic intelligence and compute to understand x and y, and some transformation f(x) that gets us from x to y, and (c) working closely with the organization to help promote actual use of some AI-enabled product, properly structuring incentives, building trust within the end-user cohorts, crafting a compelling business case to executive stakeholders, and monitoring and quantifying performance.

The general sense is that there is some meaningful gap between the promise of AI and the actual results on the ground. We all have probably felt the power of AI, but I don’t think we’ve truly cracked the code on reliable, reproducible, and cost-effective enterprise solutions. I think this is because we have to build backward from adoption, beginning with the workflow and the actual business challenge, and the humans currently part of that process.

In other words, there’s too much model-push and not enough adoption-pull for enterprise AI solutions—and that’s the basis for the opportunity that really excites me.

Looking Forward

Simple post today on the value chain as it currently exists. But I do think it’s vital that we nail the actual adoption and use of AI if we want to defend the massive capital investments that have gone towards training and data centers. I’m optimistic that we’ll get there, but I think it’s going to be harder than people realize. And the hard challenges are the ones worth vigorously attacking.

Have a great week!

Sincerely,

Luke