Model Routing For Young Ladies

Model routing is not just about boring saving money, but about building a multi model AI that uses many models to create something better.

The model is the modern AI, and many companies produce them, countless enthusiasts customize them.

If you happen to have a fancy computer, you download a 10 - 20 gguf gigabyte file.

Not having a fancy computer is not a bad thing, because small models, are super-cheap in the cloud, and they are bigger and faster.

For most, model routing is about using smaller models for simpler things, and more expensive models for planning and creativity.

But there is more to it, a new AI has come up on the radar, it is an old idea but it is a thinking thing, called a Decision Model.

And decision models, are millisecond fast, except you ask a question, but must also provide choices it can pick the answer from.

Different models are good at different things, some write well, other code well, switching between the gives you an advantage.

And on top of this, consider asking your AI to create a router, that also supports symbolic AI, which is more predictable and precise.

Your agents - which are models configured to wear different hats, so that they can examine things from different perspectives.

Also benefit from having a choose of models, agent’s planning ability, should use expensive models, whereas programming is more relaxed.

But this also means that your multi-agent framework must be aware, of the router and integrate well with it.

Models, Agents, Routers, Multi Agent Frameworks, Conventional AI, is all very confusing at first, but there is a nifty trick you can use.

Ask your AI to create a Bootstrap 6 based Web Desktop, with a virtual file system.

Present Agents with human avatar photos, and models with robot faces.

And your SKILL library, the text files that define your agent’s abilities, can now associate SKILLS with different Models (or robots).

The skills that make up an agent, will now prefer different models.

Most people route based on some easy to check properties, complexity and file size, and may use a fast model or a decision model.

Aside from saving money, and giving you an easy peasy ability to switch providers.

Wise model routing, can also give you a massive speed boost,

Once you have powerful AI at home, you are likely to only route the most complex request out to SOTA models.

Finally, please use the Web Desktop pattern often, don’t just create one, each time you are learning something big ask your AI to make a desktop for it.

It is something that you are already familiar with, and it covers a lot of complexity from login down to a useful data tree of files.

Consider creating a Web Desktop, for 3D games written in three.js. And route player behavior to different tiny WebLLM models.