Don't Really Trust and Verify

I’ve been using Fable 5.1 and Astra 6 for some of the most important work at Hypercubic.

As I was working on a new feature related to writing documentation, I was surprised at how good the quality of the documentation was by certain measures. At the same time, I was surprised at how abysmal it was by other measures.

For example, I’d given it instructions to write prose in a specific way, and it did that really, really well, making the documentation almost put you in a state of flow by how well-written it was.

At the same time, it failed to use basic React components in the right way, leading to a large number of UI/UX ugliness. All of this is, in spite of the fact that I had dedicated agents to do the QA of this whole documentation site, making sure everything looked good and was written properly.

Even as these models are approaching general intelligence, there’s something still quite off about them.

A core hypothesis at Hypercubic is that even if these models reach AGI, we still need external verification methods to check and prove their work. The importance of this verification is even more necessary pre-AGI.