On The Other Side ep 70 - Governing AI + web3 data unions w/ Jake Brukhman
Primer: The AI revolution is speeding up. Are we headed toward some serious consequences? How can we govern AI to ensure a beneficent outcome? Jake Brukhman suggests Web3 and data unions. Let’s find out in this episode of On The Other Side.
Background
Has a tech background
His parents were both engineers
Learned to code when he was 14
Learned about Bitcoin in 2011
Started CoinFund, a blockchain-focused investment firm, in 2015. Have worked with 150+ crypto projects
Thoughts On Artificial General Intelligence (AGI)
AGI is the ability of an AI algorithm to perform tasks at the level of a human
If you ask AI experts in the late 2000s how long it would take to build AGI, they would say they would not know
However, some of them think it would be much simpler because they understood neural networks as a model of the brain. One of these individuals was Ilya Sutskever, Chief Scientist of OpenAI
GPUs became more efficient and powerful
There was a huge breakthrough — Neural networks were able to do a fantastic job of classifying images and giving descriptions of who was in that image
We are starting to see the results of image generative models, large language models (LLMs), etc.
There are people who believe that AI will become so intelligent that it will become dangerous
On The AI Doomsday Scenario
The AI Doomsday scenario is not about AI murdering humans, but private corporations
It creates a game theory where to be competitive, people have to get into AI
If AI is provided by a single company or a few companies, there’s a danger that we would have to give them all of our data
Preventing AI From Sucking Up Our Data / Development Of AGI
Data being sucked up by AI is orthogonal to the development of AGI
It’s more of a distribution question (How do people use AI? What AIs do they have access to?)
It’s where we are with OpenAI today. Safety is defined by a small group of people
Have to ask whether the model has a political bias
A lot of traditional institutions are failing and we are seeing the rise of creators
Decreasing Time-To-Market Of AI Models
OpenAI launched DALL-E in Jan 2022
In August 2022, Stability AI launched Stable Diffusion
It took 7 months for Stable Diffusion to go from closed and proprietary to fully open source
LLMs are taking lesser time to hit the market
Wants to have an alternative where he can run a model on his desktop instead of proprietary software like OpenAI’s ChatGPT
AI Models And Their Feedback Cycle
If your models are open, you get more feedback and innovation
When Stable Diffusion launched, people began to integrate it everywhere and it became widely available
In contrast, DALL-E was gate-kept from people for many months
There is a project called GPT4ALL. They fine-tune it on data from GPT and it could fit on your computer
AI Models As Public Goods And Web3
It starts with smart people training models on large buckets of data
You need to have good and cleaned data that is correctly labeled
The infrastructure that is needed for computation costs millions of dollars
Once you have the model, you can make an API for people to use it
The last piece of the pipeline is commercialization
The entire pipeline is owned by big tech today
Believe that Web3 will provide the primitives that will make every single part of that pipeline open:
Making the datasets publicly accessible and governable
Decentralized computation
Crowdfunding models
Thinks that commercialization around AI is actually not that valuable
Created a prompt in ChatGPT to teach him Spanish. It worked exceptionally well
“I think a lot of products that we pay for today, once you throw AI into the mix, are going to be less valuable, and it will need things like Web3 to make sense.”
- Jake Brukhman
The Value Of Governance In A Public World
Think of them as facilitation technology
Some artists are very dissatisfied that some of their works have made it into a training set for a model like DALL-E or Stable Diffusion
This is a legal grey area
DAOs can create datasets where artists could choose either to opt in or opt out of their content
Artists could get a royalty when someone generates a work that is based on their content
The enforcement will be a social phenomenon
There’s nothing technological about smart contracts that allows them to conclusively enforce royalties
Do We Have The Capacity To Govern AI As Public Goods?
Corporate governance tends to be not that diverse
Web3 has a wide design space for governance systems. We can come up with systems that work extremely well for particular purposes
The reality of Web3 today is that it’s a one-token, one-vote system
These systems end up centralized around a couple of whales
When Oracle purchased Java, they could exercise more governance power around the features of the language
How Will AI As Public Goods Come To Fruition?
There’s such a thing called data unions
There’s a push for consumer data to be unionized
A year ago, Apple cut off advertiser IDs. Merchants who want to advertise to users need to get their zero-party data and had to pay them for it
Web2 and Web3 e-commerce companies are building zero-party data products
Believes that this trend of data unions will continue
AIs are natural consumers of data union data
Paying For Datasets
There’s a difference between self-sovereign data and open public data
In the future, there will be free, open, public MIT-licensed datasets
There will also be very rich and comprehensive datasets that people need to pay for
Privacy On The Blockchain
Currently, we are sharing our financial data on the blockchain
In the future, some of our data could be privacy shielded
All information presented above is for educational purposes only and should not be taken as investment advice. Summaries are prepared by The Reading Ape. While reasonable efforts are made to provide accurate content, any errors in interpreting and summarizing the source material are ours alone. We disclaim any liability associated with the use of our content.