🧵 How to think about building AI models (LLM Edition)

How do LLMs work? 🏭
How do you scope a custom LLM? 📃
What are the tradeoffs between popular Foundational Models? 💱
Is Context enough? 📓
When is tuning worth it? ✍

[Instructions] + [Context] + [System Prompt] -> [Tuning filter] ->[Foundation Model]

Instruction: summarize the main points of this article 
Context: the article
System Prompt: behavior, censorship, capabilities (e.g., web browsing)
Tuning Filter: specialize a model
Foundational Model: GPT 3.5; Claude; Gemini; Llama-2

How do you scope a custom LLM? 📃

Have a specific goal: search? summarize? generate?
Understand the resource constraints: proof of concept vs horizontal model (mixture of experts) vs vertical model (master many domains).

The more ambitious the goal, the more likely you need tuning.

What are the tradeoffs between popular Foundational Models? 💱

In general, with enough context and tuning, models of similar size/complexity converge in quality.

version control (deterministic testing), cost control, enterprise support, hosting, and avoiding getting rugged!

LLMs = Chaotic = easy to break behavior.

hey gang,
If you have chatGPT Plus, I trained a custom GPT on the Flipside 101 dashboard. It's just a proof of concept, but for Ethereum specifically it's helpful for starting off a query
 
https://chat.openai.com/g/g-O5l1MMvHb-flipsidecrypto-pro

Context is the easiest way to format a model's behavior.

Simply *append* relevant information to every single user's request (i.e., exploit large context windows)

Be careful: these can be *misleadingly* good in proof of concepts, but struggle to be aMaZinG in the wild

Context is the easiest way to format a model's behavior.

Simply *append* relevant information to every single user's request (i.e., exploit large context windows)

Be careful: these can be *misleadingly* good in proof of concepts, but struggle to be aMaZinG in the wild

https://warpcast.com/charliemktplace/0xbdce4dd1

When is tuning worth it? ✍

Tuning is *not* reducing model complexity. It's more like biasing some weights over others.

You provide 50 -100+ pairs of custom input -> perfect dream output

"squeeze" a foundation model to your preferred type of outputs for similar inputs.

HUGE quality boosts for often < $100!

To summarize: 

customizing LLMs is becoming easier every day! 

The outstanding problems from my view are: 
- stochastic responses make testing hard
- the best models are liable to rug you by fiddling with system prompts you can't control 
- open source models are still weak
- Making great tuning data is a GRIND!!