You’ve changed.

Maybe I’m projecting but, how have you picked directions in the past? What caused you to dive into react or bsky or write longform (and greatly appreciated) posts recently? I have to imagine there’s an emotional force, a passion/joy that guides some of that. What about that energy?

I’m getting a ton of value from the posts but it’s a slog. Because they’re long I break it up into sessions but the structure and style requires maintaining some context and history of the changes. Sometimes I come back and I’m lost requiring rereading earlier sections.

We live in a world where technology is already helping us achieve tasks faster and easier than ever before. We can order groceries for the week from our laptop in minutes. No agent needed.

As an FX artist who used to spend days crafting fluid simulations (fire, water, smoke), the gen AI stuff often falls short. The coke ad always make me cringe. But, this… This, is really good.

I’d argue 90% of the time an ‘agentic’ framework isn’t needed. I had to learn the basics about orchestration and distributed systems to realize that.

The initial appeal was easily constructing/modifying a graph. Define your fn’s, define the flow/relationships.
It took time for me to understand agentic flows, thinking complex systems needed a special framework.

“According to EPRI, a single ChatGPT query requires around 2.9 watt-hours, compared to just 0.3 watt-hours for a Google search, driving a potential order of magnitude more power demand.” about.bnef.com/blog/liebrei...

From the article:
Several screenwriters who’ve worked for the streamer told me a common note from company executives is “have this character announce what they’re doing so that viewers who have this program on in the background can follow along.”

I think this is why some agentic workflows produce seemingly novel results - they’re designed to iteratively explore until a condition is met. We don’t necessarily do that when chatting with an llm.

My take on this is that we’re iteratively exploring the latent space. The initial prompt and response puts us in the ball park. By asking “write better code” we’re probing the neighbouring, high dimensional space.

I copied the files into Claude, laid out my steps and had a mostly working refactor in seconds. What would have been a half day or more was done. But that tension still lingers. Anyone else feel this?

I think htmx and alpinejs are good additions. Lightweight, inline, llm-friendly.

Is there any way we can see some of these experiments? I’d love to get a sense of the the problem statements, how they’re structured, and what makes for a good response. I’m not an academic but I feel like the methodology could be transferable.

At first I wasn’t sure what Command offered over conditional edges but after reading docs it makes sense. BUT, I then discovered interrupt() and that just feels like a minefield of footguns. I’ll likely put it in the wrong place in a node. Or worse, run into issues with tracking multiple interrupts.