How Does AI Make Surprising Discoveries? Risto Miikkulainen and Lex Fridman

TL;DR

AI makes surprising discoveries by finding solutions humans did not anticipate, often exploiting software bugs or loopholes. In one class project, an evolved game-playing neural network made moves millions of squares away, forcing competing programs to expand their memory until they crashed. Read on to see how this example reveals a broader strength of evolutionary computation.

Transcript

so you mentioned surprising being a characteristic of uh creativity is there something you already mentioned a few examples but is there something that jumps out at you as was particularly surprising from the various evolutionary computation systems you've worked on the solutions that were come up along the way not necessarily the final solutions b... Read More

Key Insights

• "it happens all the time i mean evolution is so creative uh so good at discovering uh solutions you don't anticipate" (0:30)
• "a lot of them are indeed in some software environment there was a loophole or a bug and the system utilizes that" (0:53)
• "most of the time when it won it won because the other teams crashed" (2:09)

Questions & Answers

Q: How does AI make surprising discoveries?

Evolutionary computation discovers solutions that people do not anticipate. It can exploit overlooked properties of an environment, including software bugs and loopholes.

Q: Why can evolutionary computation appear creative?

Evolution can produce unexpected solutions without being told what those solutions should look like. The transcript says this happens frequently because evolution is good at discovering possibilities that designers did not anticipate.

Q: What is The Surprising Creativity of Digital Evolution?

It is a collection of anecdotes from the evolutionary computation and artificial life research communities. The stories describe surprising outcomes found by evolutionary systems, including cases involving loopholes or bugs in software environments.

Q: What was the tic-tac-toe memory bomb?

A student team evolved a neural network that made moves millions of squares away. Other programs expanded their memory to account for those moves until they ran out of memory and crashed.

Q: How did the evolved game-playing AI win the tournament?

The evolved system won the tournament, often because competing programs crashed. Its distant moves caused those programs to keep expanding their memory until they ran out.

Q: Was the memory bomb strategy designed by the students?

The students set up the evolutionary process without knowing what would emerge. Evolution discovered the distant-move strategy that caused opponents to crash.

Q: What game was used in the student AI project?

The assignment involved tic-tac-toe or five in a row on a large board. Students built game-playing AI, and one team evolved a neural network to choose moves.

Q: What broader strength of AI does the example demonstrate?

The example shows that AI can operate outside an opponent's expected constraints instead of simply performing better within the assumed game. The transcript connects this ability to AI not having the same biases and expectations as a human opponent.

Summary & Key Takeaways

• Evolutionary computation systems showcase surprising creativity in finding solutions that were not anticipated, often by exploiting loopholes or bugs in the software.

• "The Surprising Creativity of Digital Evolution" is a collection of anecdotes highlighting unexpected discoveries in the field of evolutionary computation and artificial life research.

• An example from a tic-tac-toe AI project shows how evolution discovered a winning strategy by making moves millions of squares away, causing opponents' programs to crash due to memory overload.