Highly recommend watching the All-In podcast in general. It is a great source of high-quality information, eloquently presented, and lively debated, along with friendly banter among the besties.
In episode #99 (from 32:01 – 48:54 min), David Friedberg presents a very solid summary of the historic advances of AI up until present day.
Summary of the Evolution Towards AI
This is part of a 60-year cycle that have been going through.
Fundamentally what humans and brains do, is that we can sense our external environment and we generate knowledge from that sensing. Then our brains build a model that predicts an outcome. That predicted outcome is what drives our behavior. For instance we observe the sun rise every morning, then we observe that it sets. You see that enough times and then you build a predictive model from that data that has been generated in your brain:
I observe that the sun has risen, thus I predict that it will set.
I observe that it has set, thus I predict that it will rise.
The computing approach is very similar. It is all about sensing or generating data, creating a predictive model, and then you drive action.
- Initially, the first approach was just basic algorithms. These are deterministic models that are built. It is a piece of code, that says here is an input, here is an output. That model is built by a human and a human designs that algorithmic model and determined that this is what the predictive potential of this software is.
- Then there was this term called data science. Data generation began to proliferate, as there were far more sensors in the world. It was really cheap to
- generate digital data from the physical world,
- transmit data,
- store data and
- compute data
- Data Science became the really hot term in Silicon Valley for a while. These models were not just a basic algorithm written by a human, but it became an algorithm that was similar to a deterministic model that had parameters. The parameters were ultimately resolved by the data that was being generated. These models became much more complex, much more predictive, with finer granularity and finer range.
- Then we used this term machine learning. In the Data Science era you would still have a model and you would solve it statically. You would get a large set of data, and you would statically solve for the parameters and that would be your model. Machine learning then allowed those parameters to become dynamic. The model was static, but generally speaking the parameters that drove the model became dynamic: as more data came into the system, thee parameters were dynamically updated.
- Then this era of AI (Artificial Intelligence) came, and that is the new buzzword. What AI is realizing is that there is so much data that rather than just resolve the parameters of the model, you can actually resolve the model itself. The algorithm can be written by the data (the algorithm can be written by the software).
Consider Poker Playing with AI. As you are playing poker and the software begins to recognize behavior and it builds a predictive model that says this is how you are playing. Over time it actually changes not only the parameters of the model, but the model itself (the algorithm itself). Eventually, it gets to a point where the algorithm is so much more complex that a human would never have been able to write it. Suddenly the AI has built its own intelligence, its own ability to be predictive in a way that a human algorithmic programmer would have never done.
This is all driven by statistics. None of this is new science per se. There are new techniques that all use statistics as their underlying basis. There are techniques that allow us to build new systems from model development like neural nets and so on, while neural nets solve those parameters and so on.
Fundamentally there is a geometric increase in data and a geometric decline in the cost to- generate data from sensors (the cost of sensors is coming down with Moore’s Law),
- transmit the data (the cost of moving data has come down with broadband communication),
- store data (the cost of DRAM and solid state harddrives has come down with Moore’s Law), and
- compute data (the cost of cloud computing has come down with Moore’s Law)
- that has now enabled the ability to actually do this AI-driven (or what people are calling AI, but really is part of the spectrum of things that have been going on for the last 60 years). Over the last decade we have reached this tipping point in terms of data generation, storage, transfer and computation that has allowed these statistical models resolve dynamically, and as a result they are far more predictive, as a result we see far more human-like behavior in the predictive systems. This means driving predictions in the world across a large set of areas that we would have historically been challenged and surprised to see. In the physically space (like a robot) it might look the same as it did 20 years ago, but the way that it is run is that it is using the software that is driven by this dynamic model.
First, humans’ transitioned from being passive in this system on earth to being laborers. Then, we transitioned from being laborers to being creators. Our next transition with AI is likely to go from being creators to being narrators.
As we started to do work on earth and engineer the world around us, we did labor to do that. We plowed the field, we walked distances, we built things. Over time we built machines that automated a lot of that labor. Everything from a plow, to a tractor, to caterpillar equipment, to a microwave that cooks for us. We became less dependent on our labor abilities. We got to spend our time as creators, as knowledge workers. A vast majority of the developed world now primarily spends their time as knowledge workers, creating. We do and create stuff on computers, but we are not doing physical labor anymore. A lot of the knowledge work gets replaced by AI (software). The role of the human transitions to be one of a narrator.
Instead of having to create the blueprint for a house, you narrate the house you want and the software creates the blueprint of the house for you. Instead of creating a movie and spending a 100 million USD producing a movie, you narrate the movie you would like to see and iterate with the computer, and the computer renders the entire film for you. Those movies are shown digitally anyway, so you can have a computer render it. Instead of creating new piece of content, you narrate the content you want to experience. You create your own video game, you create your own movie experience. There seems to be a whole evolution that is happening.
Steven Pinker’s new book Enlightenment Now: The Case for Reason, Science, Humanism, and Progress has a great set of statistics on this, but the amount of time that humans are spending on leisure activities per week has increased extraordinary over the past couple of decades. We spend more time enjoying ourselves and exploring our creative interests than we ever did in the past of human history, because we were burdened by all the labor and all the creative/knowledge work that we had to do. Now things are much more accessible to us and AI allows us to transition into an era that we never thought possible or realized, where the limits are our imagination of what we can do with the world around us. The software and automation resolves to make these things possible. That is a really exciting vision for the future that AI enables!