Teaching Machines the Frameworks Humans Take for Granted

Lex Fridman asks David Ferrucci how hard it is to encode shared human knowledge—centuries of history, wars, and social experience—so that a computer can use it. Ferrucci answers that it is possible to program a machine to acquire knowledge on a similar interpretive foundation, not merely to store facts. Humans bring a frame of reference to the world, he says, and they use multiple frameworks to interpret it.

For social and political situations, Ferrucci describes a substrate through which people interpret events: who the people are, what groups exist, what goals they have, how survival and quality of life shape those goals, how scarce resources create economics, and how power and influence operate. Humans apply this framework because they grew up with historical events and similar experiences. A computer could be imbued with this taken-for-granted substrate and then acquire the details of a situation within it.

Once a machine has that foundation, Ferrucci argues, it can predict not only outcomes but explanations in terms humans share. He notes that people can interpret events differently when they use different frameworks, citing The Matrix's premise that humans were really batteries. Still, most social interpretation relies on a shared framework rooted in similar goals and emotions.

Fridman asks how much knowledge there is. Ferrucci says detailed knowledge is tremendous and that one can imagine an effectively infinite number of unique situations and configurations. But the frameworks needed to interpret them, he believes, are finite. Fridman summarizes this as frameworks being more important than the bulk of specifics.

Ferrucci says frameworks give a system the ability to interpret and reason over specifics in ways other humans understand. The specifics are acquired by reading and talking to other people. He acknowledges that common-sense knowledge—manipulating objects, drinking water—is like an onion in robotics: even basic tasks reveal how much knowledge is required.

Fridman asks whether the details are written down somewhere. Ferrucci says they must be learned through experience. For basic physics, he sees a combination of mechanisms: fundamental pattern matching where enough examples lead to predictions from similar inputs, even if the system cannot explain why. A machine can learn quickly that a released object falls without being able to explain gravity.

Fridman notes that people relied on objects falling long before they understood gravity. Ferrucci agrees: the experiential generalization is valuable, but it must connect to theoretical knowledge. The challenge is integrating something as basic as falling objects with high-level socio-political reasoning and decision-making.

Fridman asks how hard it is to be intelligent about both gravity and politics. Ferrucci says the problem is solvable and that learning is fundamental. Current learning is primitive, he argues, because machines have not been taught to learn the frameworks humans use. People usually do not annotate data with the frameworks underlying their understanding; they just operate with the data.

To reason over data in common frameworks, a computer must have access to, acquire, and learn frameworks, then connect them to data. Ferrucci says machine learning with enough examples can learn basic dynamics, but it will not relate those dynamics to gravity unless it also acquires theories and connects experiential knowledge to theoretical knowledge. He points to architectures designed to learn specifics and find patterns while also acquiring frameworks and connecting data to them.

Fridman asks whether such systems will look like neural networks or like older expert systems based on logic and graphs, where automated knowledge acquisition was the challenge. Ferrucci says asking whether they look like neural networks is a red herring. He has experimented with architectures that combine machine learning and neural networks to learn classes of knowledge and find repeated patterns for inductive guesses, then connect those learnings to frameworks for human-style reasoning.

At Elemental Cognition, Ferrucci says, they build architectures that do both. There is a fundamental knowledge representation—graphs of logic, if you will—and neural networks acquire certain classes of information and align them with frameworks. There is also a mechanism to acquire the frameworks themselves.

Fridman observes that frameworks require collaboration with humans. Ferrucci agrees, with a clarification: he is interested in designing intelligence that can communicate with humans in frameworks humans understand. One could independently create a machine-learning system—an alien intelligence—that does some things better than humans but cannot explain its framework; its framework might be incomprehensible to humans. Ferrucci says that is a reality, but his goal is the communicable case.

His approach is to have machines acquire knowledge from humans and communicate knowledge to humans. Inductive machine learning is good at observing patterns in language, images, or video and inducing generalizations; then the machine should work with humans to connect those patterns to human-comprehensible frameworks. The machine may have a richer or longer memory and more rigorous, deeper reasoning, making the relationship complementary.

Fridman brings up Tesla Autopilot as an example of human-machine collaboration: a human monitors a system that fails every few miles, producing hundreds of millions of failures a day, and each failure is a moment of interaction. Ferrucci says this is a primitive example of collaboration: the machine has learned something, fails, communicates the failure, and the human fills in or corrects it; the machine can take that learning.

In a richer example, the machine says it can predict the next word but does not understand why, and asks the human to explain the framework. It can then reason with that framework and reuse it the next time it reads. Ferrucci compares this to his daughter's first-grade reading assignment about electricity produced by water flowing over turbines: she could match “created” and “produced” and copy the answer, but had no framework for what water flowing over turbines or electricity meant. Ferrucci describes a framework as a set of not-too-mathematical axioms of ideas brought to interpretation, built up with experts and shared understanding.

Fridman asks whether humans on Earth share a set of frameworks and how many there are. Ferrucci says it depends on how you bound them; there are many, new ones appear, and they are layered. There is a small set of primitives that provide a foundation for building frameworks, and humans can acquire and reuse many frameworks. He cites reasoning by analogy: someone who has never heard of soccer can interpret it through similarities to basketball—goals, getting the ball in the hoop, guards—and note the differences.

Ferrucci says political and social frameworks are still largely similar but can differ in fundamental assumptions and values. The reasoning process for processing a framework may not differ much. Different fundamental values or assumptions can still lead to very different conclusions.

Fridman notes that a Democrat and a Republican could read the same statement and derive different conclusions. Ferrucci says what we want from AI is to show that one set of assumptions leads here and another leads there, helping people see where their differences lie. He rejects the NLP idea that there is one way to understand a statement; there are many interpretations, and if people are committed to understanding each other they decompose interpretations toward more primitive components until they see why they disagree.

Ferrucci sees AI as complementing human intelligence by helping overcome biases and the predisposition to be persuaded by shallow reasoning—for example, assuming one is right because a point of view is labeled Democratic or Republican or contains certain keywords. A machine can help break down an argument and ask what someone really thinks, holding them accountable to critical thinking. Fridman calls this an empowering use of AI for public discourse, which he says is disintegrating.