Seeds

We’re in the era where large models that run in the cloud can handle most straightforward tasks fairly quickly and accurately. The downside being, they are large, expensive models that run in the cloud. There’s many tasks that could be done quickly and effectively with smaller, specifically trained local models on my own machine with my own compute, without a dependency on Anthropic, OpenAI, OpenRouter, or whatever the cloud provider du jour is. What if we made making these smaller distilled models a core part of our workflows?

I’ve not sussed out a great interface for it yet, but the general workflow is that a larger model (or its harness) can identify when you are repeating a similar task multiple times. If it’s a task that it deems “simple”, it creates its own dataset of examples to train a much smaller model on. Some good cases for this could be categorization (eg, organize these downloads), simple data extraction (eg, get the first names out of this json), summarization, image conversion (eg, does this image that the user just downloaded need to be converted to a jpeg?), and similar tasks.

The advantage here being that, when tuned for one specific task, a smaller model can be much more efficient at tasks that it is capable of doing. It can be loaded into memory faster, so we can essentially run them as “on-demand” models. Their cost-effectiveness and speed would allow us to have a handful (or more) of very specific models that are able to do one thing, and do one thing well. Unlocking cost-effective automations that work online and offline, and are speedy, private, and essentially free because they run on our own hardware.

The dream would be some level of OS (or browser?) integration so that it can proactively identify and automate tasks that the user does frequently, so that the user doesn’t have to think about and identify tasks that could be effectively automated or accelerated by a model.

Challenges#

• How does a model determine repetition, and “simple tasks”? That’s a hard problem to solve. Maybe v1 has the user manually specify these models, then publish them to some kind of marketplace for? Then you have the discovery issue, and the fact that the user (or agent) still needs to figure out which tasks are worth pulling a smaller model to automate.
• Models struggle with correctness. How do you verify that this small, well-trained model is actually working properly?
• As we create more and more models for specialized tasks, how do we keep track of them? Or update them? It could quickly turn into a model wrangling problem.

There’s an interesting phenomena where, with AI agents, I feel like if I don’t have my agents running on tasks, then I’m missing out or wasting time. Part of this is because orchestrating agents is fun, but I think the “not being wasteful” portion of this is also real. Other people have written about feeling like they are missing out if they don’t have agents running, too.

We’re seemingly moving in a direction where token costs are no longer subsidized by the inference providers with their subscriptions (see Github’s recent changes to the Copilot plans). I wonder if, when I have to pay per token, I’ll be less incentivized to work with agents on the evenings and weekends? We know that agents require careful tending, and low-effort prompts or leaving them to make their own decisions often doesn’t turn out well. When I have to pay for their mistakes (instead of my subscription’s 5hr quota absorbing it), will I feel less like I’m squandering my time if I’m not using agents when I’m not at my best?

For example, I use Obsidian on desktop with markdown files, use Vim to edit it, and it’s great. I can type quickly, dictate, edit properties, etc.

On mobile, on the other hand (ie, iPad, phone, etc), plain text is painful. Special characters are buried, and it’s much easier to do a few taps to random places on the screen than it is to type out a handful of characters; precisely the opposite of when I have a keyboard.

Another interface is a pencil. Great for writing text, not so much for complicated symbols. This is an in-between of keyboard and touch; you have high precision with text, but for more complicated operations, it’s still easier to tap a button than doing a series of keystrokes.