[Reading Record -154] AI Lectures as Liberal Arts

To understand AI correctly, one must know its history.

In November 2022, OpenAI released ChatGPT to the public.
With that, AI transformed from a laboratory technology into a tool that everyone uses in their daily lives.

However, this book, "AI Lectures as Liberal Arts", does not start the story there.

The author of this book, Melanie Mitchell, is an AI researcher who earned her PhD under cognitive scientist Douglas Hofstadter. This book is neither a work that praises AI nor one that rejects it. It is a book in which the author, from the perspective of someone who has been involved in AI research for many years, calmly re-examines its progress and limitations.

The history of AI is a repetition of "expectations and disappointments".

The beginning of AI research dates back to the Dartmouth Conference of 1956.

Researchers at the time thought that the day when human intelligence could be reproduced by machines was not far off. They predicted that within about 10 years, computers would defeat the world chess champion and take over most of the work performed by humans. Such predictions were discussed seriously.

However, reality does not work so easily.
The "symbolic AI" that described rules could not handle complex reality, while on the other hand, neural networks that imitated the human brain could not produce sufficient results with the computing power of the time. Expectations would swell and then lead to disappointment, and research funding would dry up... such things were repeated.

So-called "AI winters" have visited many times.

What did the deep learning revolution change?

The trigger for AI to return to the spotlight was the development of deep learning. However, the deep learning revolution did not happen due to a single technology.

I have listed the related technologies below, so if you are interested, please purchase the book and check them out.

• Advances in algorithms

  • Practical application of backpropagation

  • Development of CNN (Convolutional Neural Networks)

  • Network designs such as AlexNet

• Advances in hardware

  • Higher performance of GPUs

  • Parallel computing became possible, allowing large-scale neural networks to be trained in a realistic amount of time.

• Explosive growth of data

  • Spread of the internet

  • Spread of digital cameras and smartphones

  • Development of large-scale datasets like ImageNet

• Development of internet and cloud environments

  • It became possible to collect and share massive amounts of data.

  • The cloud made it easier to utilize enormous computing resources.

Yes, it was not a success because algorithms suddenly became geniuses, but rather it blossomed all at once as peripheral technologies such as computer performance, networks, and data storage matured.

Mechanisms to collect vast amounts of data and the computing power to process it.
Without these two things, AI research might never have reached the practical stage.

Does AI understand "meaning"?

AI identifies objects using clues that are different from those used by humans.

Humans see objects as a "collection of meanings".
For example, when a human sees a "dog," they understand "this is a dog" by connecting various pieces of information such as it having four legs, fur, the shape of its face, how it moves, and past experiences interacting with dogs.

On the other hand, image recognitionAI looks at "pixel values".
AI searches (by learning from massive amounts of images) for what numerical features frequently appear in images labeled as "dog." This is because what is important for AI is finding clues that allow it to correctly predict the label "dog."

While humans connect shapes, movements, past experiences, and relationships with the surroundings to understand an object as a meaningful whole,what AI learns is the statistical correlations that appear repeatedly within massive amounts of data.

Yes, just because AI provides the correct answer does not necessarily mean it is doing so for the same reasons as a human.

Thinking precisely because it is the AI era

There is something I feel when using AI.
That is,AI is good at expanding conversations.
If you input a few lines, it will write out related information in a few seconds that is many times the length of your input.

However, as you continue the dialogue, the information keeps expanding, and sometimes you cannot return to the initial state. I think that is where AI struggles.

Taking scattered information andstructuring it into a single story or theory.
Among various interpretations,defining what is essential.
Out of 465 pages and 350,000 characters,deciding what to use as the axis for writing a book review.

AI presents a massive amount of statistical correlations.
However, from among them,what to keep, what to discard, where to draw the boundary, and how to assemble it into a single theory.

That editing work might be the essence of human thinking.