Post

Today I announce Cantrip: On summoning entities from language in circles. In this book I unify the paradigm behind base models, chatbots, coding agents, RLMs, and RL agents, through the metaphor of magic. Code is provided. deepfates.com/cantrip

deepfates.comCantrip: On summoning entities from language in circles | deepfatesWith Cantrip, deepfates reimagines the fundamentals of language model agents. Available as a ghost library with generative test specification.

13 reposts68 likes

It's chaos magic as a tool call. Stripped of all the pomp and ceremony, it's just a series of tools that are variants of change_cognitive_state. Each takes some structured input, formats it, and echoes it back via the tool response channel. Because it's via tool response, it feels real to the LLM.

One of these tools is `become`, and takes (name, lens, environment) as argument. The other tools are `name`, `feel`, `ritual`, and `drugs`. These are tools humans use to change their cognitive state, and as such are deeply embedded in the training data as such. Metacog leverages this.

I've been creating cantrips in one way or another for a decade now. From the Markov chains of pkd_head, to early transformer/ReAct agents like @berduck.deepfates.com , to the marginalia geists on my website, and much more. Cantrip is the grimoire I wish I had then. www.deepfates.com/deepfates-we...

deepfates.comDeepfates website | deepfatesAt my website http://deepfates.com you can talk with persistent characters like Berduck and Deeperfates in the margins

Cantrip recognizes that what we call artificial intelligence now is at all levels a loop of language: predict next token, next message, next command, next sub-agent, in a feedback loop with your own effects. The static intelligence of language come alive www.deepfates.com/recursive-la...

deepfates.comRecursive language models | deepfatesIn my quest to understand the true nature of an Agent I have been thinking a lot about the loop and the actions and the environment. And I think i see where we're headed next.

I like the terminology and overall framework here, nice work. Minor gripe: I wish the axioms were line-wrapped. I had to copy and paste to another view to read them.