yes - it will definitely have a bias.

Language models are a mirror of the data they're trained on. ALL data has an inherent Bias.

A bias is simply a "model of the world", or an "opinion" or a "worldview". Everything that's ever been written, has a worldview (or bias).

Therefore, every LLM ever built will have....you guessed it...a Bias!

The key is not to try and remove bias (impossible) but to have MNAY models, each with their own biases.

As Bitcoiners, our bias, and model of the world is different to the mainstream, fiat model of the world.

As a result, as the model is trained on more and more Bitcoin data, it will represent, more and more, some "aggregation" of the bitcoin model of the world.

Does that make sense? Let me know if I can clarify further :)

ooh.. Good question.. haha

The Unfettered Mind.

Mmm.

I actually have a contrarian viewpoint on this. I think it will take a while for this to really happen at scale, because all of the current accounting system (especially with the big companies) is denominated in USD.

Doing machine to machine payments in sats will create an accounting nightmare for them - especially in the near term.

But...that doesn't mean we're not going there. We are for sure.

I just think it's going to take alot longer than what some of us Bitcoiners think..

ok..

So a GPT wrapper is simply taking a bunch of data (say, your docs, a few books, whatever), chunking them up and then storing them in what's called a "Vector Database". Once you've done that, you have ChatGPT (or other LLM) reference the most relevant "chunk" from your vector database when someone asks a question.

This is kind of like inserting an example when you're using chatGPT. What you're essentially doing is inserting relevant context so when the "model" responds, it's more relevant.

It's a great, cheap, quick, "hacky" way to get a model to give more bitcoin-like responses, but it's not really a "Bitcoin model", nor has the model actually been "trained".

Where it falls over is when you ask anything that's not specifically in the vector store. Then you're basically back to normal chatGPT.

On the other hands, a model that's been trained (like Spirit of Satoshi), has actually had a TON of data transformed and formatted for training, then a bunch of GPU cycles spent on actually CHANGING the weights and biases of either a pre-existing open source model, or (the harder alternative) using all that data to do a ground-up "pre-training".

These are both very different processes to a "ChatGPT wrapper". The latter actually involve training a model, whereas the former uses another model with some reference material.

Training from scratch ultimately delivers the best possible result, but it takes WAY longer and is orders magnitude more expensive.

Hope that makes sense man!

Just basic physics. The more general you make something, the less specialised it can be.

Google is a good example, but so are reddit and forums. Google might give you a good high level overview on something, but when you want to go deep into a topic or subject, you go to reddit or a forum.

Same thing will happen with models. You'll have mainstream general ones that are "good enough" with general stuff.

Then when you want to go deep on a topic (especially those outside of mainstream, ie; Bitcoin, self defence, home schooling, alternative health, etc) you will want to seek out a domain-specific language model.

We will look to make the method we've used to train Satoshi and the training app (train.spiritofsatoshi.ai) available for other domains.

Honestly - it's like having a computer. Like any other technology, "AI" is just a tool.

In fact, it should probably called "IA" (Intelligence Amplification), because when used intelligently, that's what it does (once again, like any other tool, or technology).

There's no real magic here. There's no "sentient being" that will come out of the circuits. All that "AGI" talk is just fairy tales from wierdos who wish they could create their own god. It's silly.

At the end of the day, it's a tool. Use it where necessary and use it wisely, and you can do more work, in less time.

You can also use it stupidly (like how some people might use a hammer - hitting themselves in the head - or social media - doom scrolling all day) and you will make yourself more dumb (ie; never think for yourself, and just ask chatGPT everything)

haha - yeah this is a tough one. How the model answers it will be a little emergent from what it's been feed. We will see.

🤣🤣🤣

One day...when there's enough data about my dumb ass, and I have nothing better to do with my life, I'll build a Ghost of Svetski.

Or maybe my kids will. Until then, I'll try do more useful things..

yeah - that's a RAG model, kind of like what I described to @kr below. It's just a general base model that references his data, maybe with a slight LoRa fine-tune.

I reached out to do something similar with my data, so I can see what they're doing in greater detail.