Dario Amodei’s, CEO of anthropic, essay “Machines of Loving grace” is a must read of anyone interested in AI. One concept that stands out to me is the idea of “return on intelligence”.
That is: how much value can we gain in a certain area or task if we are able to deploy above human expert intelligence? How much will powerful AI accelerate our understanding and ability to solve problems.
Such powerful AI is not what we have today – and irrespectively of whether his predictions holds true and we will have powerful AI in a few years – the idea of return on intelligence is a powerful concept to contemplate even without it.
Instead of thinking about a potential future, I want to discuss what we have today — pervasive knowledge systems. Or at least the ability to design such systems.
Is knowledge a requirement for intelligence?
Let us for a moment contemplate how much of our daily “intelligence” is a matter of knowledge recall? Intelligence certainly exists without knowledge, but knowledge is the framework that provide practical application of intelligence – it is the canvas on which intelligence paints.
However, gaining knowledge and learning how to apply it takes time. We also have a limited amount of bandwidth in terms of recall hampering our ability to draw connections between diverse areas of knowledge. Here classical machine learning, current AI and just pure knowledge bases becomes really interesting.
Pervasive knowledge
So how can we put knowledge into all our interactions? What if you had perfect recall and context in every interaction? We often look up stuff online during conversations – is that true? When was that song released, how many Celcius is 103 Fahrenheit?
But this type of knowledge consumption is active and disjoint from your experience. It leads to abrupt disconnects with the conversation or action at hand. What would designing knowledge systems that infuse our actions look like?
The immediate problem would be information overload, something we already have plenty of. Moreover, our capacity and interest in knowledge is also different. Consider your local barrista coffee being infused with knowledge: some might be pleased to get complete knowledge on where their coffee they just got was grown, harvested, processed, and what flavours to expect relative to other types of coffee. Some just want a bit of warmth and a kick of caffeine.
Our interactions with knowledge are highly personalized, and will ebb and flow with our general mental bandwidth.
So to have pervasive knowledge we need to figure out how to design it. In that design we need to take into account that
– our capacity is different
– our thirst for information is different
– we can experience decision fatigue with too many options
– our interests are constantly evolving
Knowledge system design
There are likely many good solutions to this, but I want to focus on a few that I think are key:
1. Aid curiosity
2. Choose your own adventure
3. Prevent easy mode
A good knowledge system aids your curiosity – it allows for your own discoveries, your own linking of fragments of insights to form your own mental model of the subject. It is like a well written comic book, you get the essential frames and fill out everything in between.
This keeps your mind engaged and allows for pivots in your exploration and understanding.
You need to have agency – your ability to choose your own adventure. We operate on very different levels in terms of thinking based on our previous training, schooling, context etc. Something that might be pedestrian to one person, could easily be a challenge for another. A good pervasive knowledge infusion sets to level just right, both in terms of wording, level of depth and abstraction.
This leads to the final and perhaps most important thing to get right — it should avoid easy mode. It should not let us get lazy. A perfect recall system might give you an answer up front, but that is not what we are interested in. A knowledge system should allow us to form our own opinion, align the subject into our own value set and world view. Be a canvas for our intelligence.
We already have too many echo chambers, and a pervasive knowledge system would easily make that way worse. In a sense it should be slightly adversarial. Either by presenting contrasting viewpoints, make certain information harder to gain leading to a deeper engagement, or ask questions based on previously presented knowledge. It is all about striking the balance between being helpful and challenging at the same time.