Forging Light with AI
AI is the most powerful tool to come out in an era. It will undoubtedly change the world. Let’s work... View more
Organizer:
- Organized by
-
-
3 basic AI stack layers and Semantics
-
3 basic AI stack layers and Semantics
As with any technological revolution, the semantics, lingo, and branding is being hashed out day by day.
One way to simplify things and improve the signal to noise ratio is to think of the AI technology in 3 layers:
1) Models : the base layer is typically a Large Language Model (LLM).
There are also LMM’s (large multi modal models (which can see, hear, speak, and draw, as well as interact via text chat; and there are LAM’s large action models which know how to talk with)
2) Model Runner / UX : next above that is some sort of environment software that makes it easy to load a chat with models.
Examples include: LocalAI, Ollama, TextGen Webui, LM Studio, and many others.
3) Orchestration : above the model runner there is a layer of software that may or may not be used. This layer allows users to create teams of AI’s with different roles and which interact with each other, even in an organizational structure similar to a business. A “team” of AI’s is synonymously referred to as a “team”, “swarm”, “crew”, “ai agency”, and other terms which are presently jockeying for lingo dominance. It can be experimentally demonstrated that a team of 3 AI’s (1 manager, 1 coder, and 1 tester) will outperform a single coder AI working on it’s own.
Examples of orchestration software include: AutoGen, ChatDev, CrewAI, and many others
Semantics…
On-Grid refers to cloud based, closed source AI’s which are owned by big corporate interests and are censored, for instance: ChatGPT, ClaudeAI, Google Bard/Gemini, CoPilot, and others. These models will do things like send you to the CDC for ivermectin questions, or lock out your account if you ‘violate’ their usage policies. They also form an interior model of your behavior and preferences.
Off-Grid refers to open source models which are free to use. They are created typically by large organizations with powerful computers (Ex: Stabillity AI and Meta), then thousands of tinkerers work on these models to prune them down, “fine tune” them, “embed them” with specific data, “quantize” them to shrink their size to run locally on consumer-grade CPU’s not just expensive computers, and “uncensor” them by removing guardrails so that, for instance, the AI will not just send you to the CDC if you mention the word “Ivermectin”.
Prompt Engineering refers to the art and science of asking questions to an AI. Anyone can chat colloquially with an AI, however, learning to craft context rich questions, and queue the AI to pay attention to a particular topic, can powerfully focus the results. Prompts are also used to define specific roles to different AI’s in a team of AI’s that are working on a problem together, bringing different skills and abilities to the task at hand.
Context …
Tokens …
Long Term Memory …
Embedding …
Fine-Tuning …
CPU vs GPU vs TPU …
Edge AI …
👆More on these later!
-
2 Replies
-
Thank you for that summary … very helpful!
-
Thanks Steve, this is helpful
Log in to reply.
