Shiyue Weekly (Issue 15): Stop worrying and learn to love artificial intelligence

Original link: https://www.skyue.com/23032707.html

This is the 15th issue of Shiyue Weekly, and the title is imitated from the first article. Overall, this week I read a lot of content related to artificial intelligence and the history of science and technology, including several old articles, and the number after the title is the year of publication.

The last two issues have been read around a specific theme, and I have two feelings:

  1. It is necessary to actively search for information and be able to find some old content that has lasted for a long time.
  2. Able to find connections between different articles and connect different knowledge and viewpoints.

It’s a wonderful feeling.

Artificial Intelligence and History of Technology

There are a total of 5 articles in this part, except Muyao’s podcast is the most recent one, the other 4 articles are from before 2019, and the earliest one is from 1999.

How to Stop Worrying and Learn to Love the Internet – Douglas Adams / 1999

The article written by Douglas Adams, the author of “Guide to the Galaxy” in 1999 (died 2 years later in 2001), is a very insightful article. The recent shock of AI is estimated to be similar to that of the Internet in 1999, and the future looks back at history , The impact of the current AI is likely to be equal to or even greater than that of the Internet technology at the end of the 20th century.

The core point of the article is already reflected in the title: stop worrying and learn to love the Internet. Very suitable for the current AI moment to reread.

An excerpt is as follows:

  1. Things that were there when you were born, you take for granted.
  2. From birth to before the age of 30, new inventions will excite you, and if you are lucky, they may even become your career. [Think about Internet industry practitioners born in the 1980s and 1990s]
  3. And any new invention that appears after the age of 30 is against the laws of nature and is the terminator of civilization. Until the new invention lasts for more than 10 years, you will find that it is not so terrible.

From the perspective of human adaptability, it reminds me of Romain Rolland’s famous saying: “Most people die at the age of twenty or thirty. They become their own shadows, and the rest of their lives are just replicating themselves day by day. .”

In addition, the current price of ChatGPT is not cheap, the performance is average, and there are not many application scenarios. All of this is the same as the Internet at the beginning, and it will also be the same as the Internet later. In the future, AI services will become very powerful and easy to obtain. cheap price.

[Podcast] In the era of AI hurricanes, are people still valuable? – Mu Yao and others

The Mu Yao podcast has been updated, and it has been more than a year since the last time. This issue talks about the recent hot AI, and most of the content is involved in its Weibo. Some new knowledge points for me:

  • Societal value ranking. Also making a lot of money, the social recognition of beautiful anchors is not as high as that of highly educated people. The praise of “smartness” came after the work revolution (because after the industrial revolution, smart people have greater social value to a certain extent), it is a kind of idea. If AI can replace knowledge work, on the contrary, the care work mentioned by Graber is not easy to be replaced, and the value sequence will be rearranged.
  • It may take two generations to embrace a revolutionary new technology. Mu Yao mentioned that in 1840, Britain came to fight, and the civil servants must also know the power of the Industrial Revolution. But the imperial examination system was not abolished until 60 years later. This is because the grandchildren have grown up immersed in new technologies and have no historical burden, so they have the courage to innovate. [Same as the above article]

The History of Artificial Intelligence – Rockwell Anyoha / 2017

A brief article on the history of artificial intelligence published in 2017 on the Harvard website. The content is not long, but the key nodes of history are covered, and there are more high-quality external chain references, which is convenient for in-depth understanding.

for example:

etc.

The godfather of artificial intelligence Hinton’s battles – sayonly / 2018

Geoffrey Hinton (Geoffrey Hinton) is known as the “Father of Deep Learning”. This article introduces the transformation of deep learning from non-mainstream to mainstream in the field of artificial intelligence research and some of Hinton’s stories.

The article mentions three points of doubt about the neural network:

  1. Lack of theory: black box, uninterpretable
  2. Can’t reason: ChatGPT now has CoT (Chain of Though) capability
  3. Biology is not established: NN uses BP, and the change of brain neuron connection weights is based on STDP.

I have seen the first two articles in many articles, but this is the first time I have heard of the third article, which is very technical.

When discussing the suspicion of “lack of theory”, the author uses the relationship between “mathematics and physics” as a comparison. Historically, mathematical theories have been ahead of physics for a long time, but in modern times, physics has been ahead of mathematics, and some physics research has led to new mathematical progress.

For the study of neural networks, who said it would not be possible?

Regarding the success of neural networks, there is an extreme speculation: the underlying ability of human intelligence is not rationality (reasoning-related ability), but imitation.

The Bitter Lesson – Rich Sutton / 2019

An old article in 2019, the author Rich Sutton is a computer scientist who also works at DeepMind.

Regarding artificial intelligence, there used to be a line dispute: adding human knowledge rules to the system vs. pure statistics and calculations to achieve intelligence.

The core point of this article is to learn lessons from history. Although adding human rules to AI is effective in the short term, in the long run, as long as the computing power breaks through, the added human experience rules are insignificant.

GPT model technical dismantling

These two articles, which can also be considered as history, are the history of ChatGPT’s technical details.

Dismantling and tracing the origin of various abilities of GPT-3.5 – Fu Yao

GPT-3 was released in June 2020, and ChatGPT was released at the end of November 2022. There were more than 2 years in between. This article uses public information to sort out and speculate on the work of OpenAI in the past two years. How does GPT-3 work step by step? Evolved into amazing ChatGPT.

In-depth understanding of the emergent capabilities of language models – Yao Fu

This article describes emergent capabilities (another common translation is emergent capabilities).

Knowledge is mentioned in the article. ChatGPT has no external knowledge data, that is, all knowledge exists in the model, and the model contains 175 billion parameters. That is, world knowledge is implicitly stored in 175 billion parameters by the large model, which is very similar to human knowledge being implicitly stored in neurons (abstract level).

just looking around

Advice to young critics – Matt Zoller Seitz / 2014

If you write a movie review:

  • Watch more, including old movies, learn about film history
  • As you watch, keep a record of thoughts that come to you
  • Study history and psychology, two of the most covered topics in film
  • Write for two hours a day, whether published or not

Can I take a bite out of an apple? – neverland

A very interesting practice, Neverland exported all the data on Apple’s services and archived it, and at the same time studied what data Apple collected.

tool resources

An Attempt to Filter RSS Locally for Free: NetNewsWire – UNTAG

I am using NetNewsWire as the main RSS reader, and I just have the need to filter articles. I found this article, which uses AppleScript to realize the article filtering requirements, directly marks the articles that hit specific keywords as read, and indirectly realizes the filtering.

Direct use of the code in the article may report an error, and it needs to be modified appropriately in combination with your own subscription. I’m still researching and haven’t formed a perfect filtering ability yet.

This article is transferred from: https://www.skyue.com/23032707.html
This site is only for collection, and the copyright belongs to the original author.