|
Welcome to The Weekly Fizz!
The Drip
Anthropic is shipping fast to close the gap OpenClaw opened up. Routines. Managed agents. The kind of scheduled, always-available features that make AI actually do things instead of just answer questions.
We set one up for Justin's daily planning workflow — it fired at 8am, kicked off the session, waited for him to pick it up. It works. Parts of our content machine could probably run this way too.
But here's the pull: these tools will never tell you there's nothing to do. They'll always find something to build, some issue to "fix." Justin's been writing about what he's calling the art of redaction — the time-tested skill, across cuisine, art, music, writing, of knowing when to stop. Of taking things away to create clarity.
The "always-on autonomous agent constantly working for you" story misses that. Kellan calls it a Venn diagram problem: the overlap between what AI can automate and what's actually worth automating is smaller than people want it to be. The skill worth building isn't tool selection — it's discernment.
Inside The Bottle
We've been tinkering with open source models running locally. Why? A few forces are pulling the same direction.
Providers want you in their ecosystem. Anthropic already blocks subscription use for third-party tools. The days of popping open Gemini with unlimited pro compute are probably numbered. Cost reconciliation is coming: either prices go up, costs come down, or subsidies enter. Subsidies aren't happening. So something's gotta give.
Open weight models are catching up. Sonnet 3.5-level capability now fits on medium hardware — call it a $2,000 desktop. If you're paying for Haiku-level work at the API, that's a real ROI conversation. Opus-frontier size needs a heavier GPU, but that's a Mac Studio, not a server farm.
Model selection is a discipline, not a default. Justin was talking with a CIO recently who hadn't cemented the idea yet — that within a single application, you might use Haiku for retrieval, Opus for orchestration, and an open-weight model for high-throughput tasks where cost-to-serve matters more than frontier capability.
The broader point: the value isn't just in the model anymore. It's in the ecosystem around it. There's a model. There's a system layer. There's an application layer. There's the handling of messages between them. That's why using Claude in Claude.ai feels different from using a Claude model inside Microsoft Copilot. Solution architecture is never just a model.
Open source is a route where you take more control of those layers. If the ROI makes sense.
Lab Notes
| ■ | Justin's note: Posting your team's $117K monthly token bill as a badge of honor on LinkedIn? That's a signal about your architecture, not your intelligence. |
| ■ | Kellan's note: If you're looking at the technology and working backwards into places where you could use it, you end up going down rabbit holes. Start with the painful process. Then see if the tool fits. |
Latest Article
Remembering the art of redaction
What Stopped Our Scroll
|
