Intro

I have been following talks on latest developments in AI and I want to share one that caught my attention very much.

In this talk, Tom Bilyeu talks with Emad Mostaque, who is founder of Stability AI. Here I am dumping my raw notes on the talk.

PS: I will be using “AI” without adding “models”, “apps” or “services” afterwards. You know what I mean.

PS #2: I am using the emoji (👀) for marking my observations.

Keywords

• AGI → artificial general intelligence
• ensemble of models
• talented young people
• open models and private data
• augmentation vs generalization

Notes

Humans vs AI

AI won’t replace humans. Humans with AI will replace the humans that don’t use AI

Bilyeu says we will have better friends who will be AI apps. But if you have introvert personality and have problems of drug abuse, many will completely cut off from the world. Since feeling “no need”

Emad is using gpt4 as therapist. Also he shares his own story: his autistic child, searched for causes (there can be multiple) Used AI to curate latest research in the literature

Curating knowledge

Wouldn’t it be great if someone can find the latest knowledge easily?

He states that we (humans) can never scale intelligence & expertise. So, his vision is that everything will be personalized AI. Like a personal trainer.

💡 Talented young individuals

Another example from him on scaling intelligence & expertise:

Imagine that you are not alone anymore. You have a team of talented young people. And this team can also build other teams !

👀 This part has a huge impact on me! Since models can do a lot of things (both correctly and incorrectly), they are like new graduates. So with AI, finding and working with new graduates will take less time and resources.

Attention is all you need & inner workings of GPT

The working principle of GPTs: not all data is important → paying attention to most important parts of speech.

From Shannon’s info theory ❝ Information is valuable only if it changes ❞

A Gpt model is just a single file, like a codec. Gets input and maps it to some of its content It is just compression of intelligence or a filter Similarly, a book is compressed knowledge

Emad suggests viewers to watch alphago documentary, it plays with itself → hallucinations.

Emad’s aim in Stability AI: Layer-1 for AI

• open models and private data
  • for addressing privacy concerns.
  • models will be shared publicly, anyone can find and use them
  • apps will be built using these open models but not sharing the data they generate with external third parties (like OpenAI).
• models will be small & specialized
  • gpt-4 has all the knowledge that it don’t need
  • like if you won’t ask calculus, there is no need for storing those data
• multiple models
  • as if they are teachers who specialized different areas”
  • not just one model but an ensemble of models, a model combining other models with a right flow, chaining one’s output to another’s input
    • just like a talented youngster organizing other talented youngster in a right flow
• He is not building big enough models that can be used as AGI on purpose !
  • because AGI may benefit or destroy the humanrace. Two very polarized results
    • “embed guardrails, standards, make it predictable & boring”
  • His objective is augmentation not generalization !

Effects of AI on Education

It will transform many things! Like personalized tailored content & learning methods according to your skills & tendencies !

Current education needs 1-on-1 attention. With AI, this can be done 1-to-many, much more efficiently.

Since every person is unique and different needs, multiple AI models can guide you by adapting your needs (such as audio/visual learner etc). So AI can generate content according to your level. This is also called adaptive learning.

As well as single child can have multiple “AI teachers”, an AI model can guide multiple children, which makes teaching efficient.

on Medicine

Doctors will be the most affected since AI will have more empathy

Alignment

Emad thinks that:

There is no way to align outputs, we can align inputs, in other words, raising a child right

This is done by controlling the AI, making it predictable, “boring”. Tom says that this is not alignment, this is enslavement.

👀 I can related to arguments of both sides, controlled freedom is required. But wait, isn’t this concept already employed on humans? 🧐

Web3

Emad says web3 has no AI in it, it is just solving problems of identity and value transfer.

But AI will play well with the ledger idea, since it is automating the value transfer.

Autonomous AI agents won’t be using banks but will use crypto.

Tom mentions “protocol 6551” ~russian nesting doll, wallet + storing other data But Emad says that storage value of blockchain may not be that important, but the identy concept is important.

I can relate to that because (currently) storage is not so efficient

Future of Culture & Society

How would your life/society/community/business change if you have infinite talented young people ?

Emad thinks AI as:

• talented young people
• army of analysts & teachers

• How will we adapt to potential widescale job loss ?
• What are the cults, religions and political movement over 5 to 10 years?

This question is important because those will be hyper-organized/persuasive utilizing AI. Tom says these may even be started by AI !

On the one hand, we will have youths with AI girlfriends. On the other hand, youths that want to believe something bigger to fill the void. Here who is gonna step in?

Emad emphasizes that we build many things around stories and whole AI discussions are built around “power of the controller who tells the story”

Emad also mentions positive & negative liberty:

• Positive → freedom to believe in any “ism”s
  • fascism, communism, islam(ism)..
• Negative → freedom for being told what to do
  • capitalism, consumerism

👀 Are we really free in negative liberty? Or are we being manipulated in some other way (without our consent) ?

👀 I think that main difference between positive & negative libery is:

• in positive → we “surrender” our free will with our consent
• in negative → we choose(?) what to do, but may be manipulated (i.o.w directed without our consent)

Conclusion 👀

• There shall be smaller AI models that specializes more specific areas, instead of a AGI. He calls these “talented young individuals”.
  • I remember that Daniel Miessler mentioned similar things in one of his newsteller article.
• Businesses will transform rapidly by those who have many “talented young individuals”.
• Our culture is not ready for so rapid changes, we need to ask questions and think more about that.