One approach, many problems

This module widened two parts of the story.

In the late 1960s, Shakey joined planning to physical action in a small world.
In 2020, retrieval-augmented generation joined a language model to fetched documents.
In 2022, ReAct showed a language model alternating between reasoning and external actions.
Diffusion research moved from a 2015 generative method to much stronger image systems in 2020 and practical latent designs soon after.

Language-model agents, image generators, speech systems, recommenders, robots, and scientific models are built differently. They share a move away from writing every rule by hand. Examples, feedback, and evaluation shape learned numbers instead.

The next module does not add one more mechanism. It asks what this history leaves unresolved. Who chooses the goals and data? Who gains from the systems? Who carries the cost when they fail?

The whole chapter, simply

No single kind of machine is best at words, pictures, sounds, movement, and science.

So people build different kinds that learn the patterns needed for different jobs.

This module widened two parts of the story.

In the late 1960s, Shakey joined planning to physical action in a small world.
In 2020, retrieval-augmented generation joined a language model to fetched documents.
In 2022, ReAct showed a language model alternating between reasoning and external actions.
Diffusion research moved from a 2015 generative method to much stronger image systems in 2020 and practical latent designs soon after.

The next module does not add one more mechanism. It asks what this history leaves unresolved. Who chooses the goals and data? Who gains from the systems? Who carries the cost when they fail?

The whole chapter, simply

No single kind of machine is best at words, pictures, sounds, movement, and science.

So people build different kinds that learn the patterns needed for different jobs.