Blurbs

Diogo Almeida has been hinting that Typesafe.ai is going beyond models that produce the primitives that JEV provides. I've been pondering what those could be.

Typesafe's primary thesis is that intelligence you can drop right inline with your code is what unlocks your ability to generate real value. JEV provides probabilities your deterministic code can use to decide a next action.

But what other primitives will typesafe give us that we could inline into code?

What about estimation? Given these priors can you estimate a quantity or a distribution? That's useful in code.

What about simulation? Given this state, and this change, where are the vectors most likely to go? That's useful in code.

These sorts of ML tools already exist today - but they're not general.

But imagine if they were. Imagine if the 'intelligence' returned by some general tool like that is fast and cheap. Imagine what you could do.

I can't wait to see what's next.

I'm getting close to the end of a project and I'm finally publishing design system documentation for future-humans that may one day work on the project. I'll say that again: I'm shipping the design system docs towards the end of the project.

We never created drawings in Figma or other design tools for this project. The design system has existed exclusively in code from the start.

The living visual documentation artifact is actually just more code that renders how things fit together in the system.

To date our agents have been able follow the six ADRs and two skills that describe token architecture, component taxonomy, fluid typography, third party foundations, tenant level customizations, and state escalation.

Our humans have been able to validate the components and their composition in the application itself, not in some vector drawing.

Why bother making docs at all? Future humans and their agents might appreciate something that describes how everything builds up from tokens and composes all the way to full page layouts. Then again, the most likely future human to benefit from it is me.

AI smells. (I've replaced the nouns with nonsense words to protect the original 'author')

The lorren-over-vessa framing at 18 mavels old is the tellan. Thats the narell of someone building an avora, not chasing talven share.

The conventional rellan of early Velora is wrong. The real sorren was never about the ferren — it was about who wasn't thinking of it as a ferren.

I couldn't agree more:

OpenAI Dots is just one example of us starting to talk directly to a cloud harness.

But even for basic chat, I wouldn't be surprised if we're not prompting directly into the model, but going through a cloud harness that's sprinkling secret sauce on top of what the model does on it's own.

The shift is happening and new 'invisible' harnesses are going to sprout up everywhere.

I really like this take. AI is just like electricity. Not just because electricity itself must be handled with care, but we can look back in horror at the wildly unsafe electric powered products from the early days.

We were basically putting glowing resistance wires inside of an ordinary household objects and connecting them to a wall.

Sending ALL my email to an LLM never seemed like the safest way to get intelligence for my inbox.

So, I built a private email manager where I control of which emails are handled by basic deterministic rules, which are evaluated by JEV for routing, and which items can safely get processed by my agents.

Lessons learned:

• There's an art to creating a set of question that produce high confidence score with JEV.
• You can use JEV and organic evolutionary algorithmic loop to come up with better questions.
• It feels natural use JEV as the logical routing gate in your code.
• Himalaya CLI mail client with Sirup are invaluable.

8,000,000 input tokens on my first serious JEV experiment cost $0.36. The experiment's approach was naive and inefficient, but yielded accurate results (compared to ground truth).

With a little experimentation and parallel calls, speed and token counts both went down by 50% with the same accuracy.

The most interesting insight, is that you have to put on your thinking cap when constructing your specific approach to a problem. LLMs aren't used to outputting this kind of fuzzy code yet.