56) Open Source

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, Comparative Advantage, we explored a framework (sort of) for assessing optimal token usage within an organization. My claim is that different people with different skillsets are better able to drive some desired outcome through less token usage than others. That ultimately minimizes wall-time spent on a task and should minimize token usage (and AI spend) and accelerate outcomes across the organization. So, even though anybody can theoretically do anything now, that doesn’t mean that they should.

Current Situation

There have been some changing winds lately within the broader AI landscape. Over this past week, Palantir’s CEO Alex Karp went on CNBC to discuss the fears he claims to hear from the heads of government and enterprise customers around their relationships with the frontier labs, specifically OpenAI and Anthropic. The All In podcast also discussed this ad nauseam.

Essentially, the claim is that the frontier labs, through their commercial partnerships with enterprise companies, are extracting trade secrets and knowhow through their work with those clients, and then launching competitive offerings that effectively transfer the alpha (or unique, moat-driven, durable competitive advantages) from the clients to the labs. Claude Code and Claude Design are well-known examples.

To me, it shares a lot of characteristics with anti-competitive claims made against Amazon around them creating Amazon Brand products that benefit from all the interest and pricing data they have from their massive distribution platform. This is most prominent across fungible commodity-type products. If Amazon sees that pencils are ripping, they can undercut the market with low-cost pencils my possible by their vertically integrated model.

This is similar to what’s happening with the AI labs and their customers. Right now the frontier labs “own” the means of production, or at least some of the most popular, reliable, trusted, and robust production apparatuses right now—which drives distribution and trust, and lets them compete on end-user experience and in end-user product categories. If the labs want to help enterprise clients deploy AI solutions, how can the clients get comfort that the labs won’t drink their milkshake (s/o There Will Be Blood)?

On a related but less important note, cost has also come to the fore lately, with a lot of stories being shared across X and other platforms of engineering (and other) teams ripping through their AI quotas and stuff. So this, in tandem with security and competitive concerns, have conspired to put some heat on the labs.

Open Source & Model Routing

In response to these concerns, people are more intensely exploring intelligent model routing (and other cost reduction measures) and leveraging open source models, which at this point are pretty close to performance-parity with the frontier labs. Companies can leverage open source models, fine-tuned to their proprietary data and specific needs, to achieve whatever workflow enhancement that they would otherwise build with frontier labs. This lets them own the means of production, maintain data sovereignty, and retain their business’ alpha (or so the argument goes).

I don’t really know where I land on all of this, truthfully. But I also do understand concerns that frontier labs can and will launch competitive products that can almost 1:1 drive equity value compression for their commercial partners. To me, the most fascinating piece of the equation is how quickly broader business sentiment can shift in favor of and against the players within the AI space. Anthropic goes from beloved to hated on a multi-week cycle, and the same is true of OpenAI. Chinese open source models are amazing, but also scary, and they’re maybe the answer to high-cost and competitive overreach from the frontier labs, but also do we want to support a geopolitical competitor?

All I know for sure is that nobody really knows. And that’s what’s simultaneously so fun yet disorienting about being in this space. But I wouldn’t have it any other way.

Looking Forward

We’re still so early in the broader diffusion and commercial development of AI, and there are still so many challenges across cultural, corporate, competitive, and technological domains. We will see where things go, but it’s never been a more exciting time as a technology lover and business enthusiast.

Have a great week!

Sincerely,

Luke