Q&A: what should middle school students learn now if they want to work in artificial intelligence in the future?
Q&A: what should middle schoolers study now if they want to work in AI in the future?
From Day Lesson: Great Changes Not Seen Since the Enlightenment
Reader Natural Jungle: My child is in middle school and plans to work in AI in the future, what subjects should I focus on now?
Reply from Wan Wan Gang -
College majors are important in selecting people for the AI industry. Generally need talents in computer science, natural language processing, statistics, data processing, computer graphics and so on, some universities directly have AI majors. For college students, it is best to take one of the above as the main major, and then minor in a major like cognitive science, brain science, psychology, philosophy, and so on, that is simply a customized AI talent.
For middle schoolers, it’s partly a matter of making sure you get into a good college that offers these majors, and partly a matter of doing some advance preparation. The most important thing is math. Having strong math muscles allows you to quickly understand and grasp various abstract concepts and understand the logical structure of, say, a program. The next step is to read widely and have a competent understanding of what the world is all about.
With ChatGPT, English and programming are now in a very delicate position. On the one hand AI has pretty much eliminated foreign language barriers, and it can also help people program. I’m sure people will be using natural language programming in general when your kids grow up. But on the other hand, learning foreign languages and programming isn’t just about those skills themselves, it’s also about opening up the brain. But the good news is that AI has made it easier to learn both foreign languages and programming, and now it’s half the battle, so why not?
On the flip side, popular extracurricular programs like calligraphy and music are time-consuming, expensive, and becoming less and less cost-effective relative to other programs. If your child is not really interested, you should give up.
Reader Charles: The rise of the mobile web has made google and baidu’s market space smaller. Will the rise of ChatGPT change the source of information, and thus the means of promotion for commercial organizations?
Reply from Vanguard Steel -
ChatGPT and Bing Chat are already a strong threat to Google’s search. after Bing Chat came out, Google’s stock price fell and search traffic is dropping every day. Of course, the decline has not exceeded 10%, but the trend is very scary.
Google’s search advertising revenue is $162 billion per year, which is its lifeblood. It’s relatively natural to add ads to a search page; your eyes have to go through the entire page of results anyway, and there’s nothing inconvenient about seeing a few ads along the way. But if you want to insert ads into a chat, it’s just too much of a distraction. The former is like watching an ad before a TV show, while the latter is like adapting a TV show with ads.
After Bing Chat came out, Google quickly launched its own chatbot, Bard, which caused a great deal of dissatisfaction even before the ads were inserted, just because Bard replied with a wrong fact about the Webb Space Telescope. It’s clear that people have high expectations of the chatting experience.
I’ve been using Bing Chat for a while now, and I’ve had a few occasions where it’s inserted ads into the outer layer of the dialog box, which may be the way to go, but the effectiveness of that remains to be seen.
So far, neither Google nor Microsoft has found a good solution for ads.
There are already several new search engines on the market, such as Neeva, which are sold as ad-free. Perhaps the next step for search engines is to pay for a subscription search service without ads.
It is worth mentioning that GPT-based search requires more than ten times the computing power of traditional search. We usually always love to say that the marginal cost of software services is zero - in fact, it is not absolutely zero, it requires a lot of computing power, a lot of chips and servers, especially now that AI chips are so expensive. This cost is also a consideration.
We can imagine that now Google feels very difficult.
From Day Lesson: Dangerous Fabulousness
Reader User 73119051: Since an AGI learns like a human brain, is it possible for an AGI to train a human brain in turn? A child grows up learning in a human intellectual environment, and grows up with cognition that struggles to comprehend things outside the realm of reason. So if a person grows up in an AGI environment will he or she also have a super brain? Is there a limit to the human brain?
Wei Wei Gang replied -
The storage capacity of the brain is massive, far from touching the limit. The main bottleneck of the brain as a device is that the speed of input and output and logical operations are too slow. A computer can read a book in less than a second, and the human brain can’t do that with any amount of effort. But we have two consolations, and I don’t think people need to get too hung up on the fact that they don’t have enough brains.
One is that while knowledge is infinite, ideas are finite. * Our columns have talked about “full-coverage reading,” “cultural self-awareness,” and “adjusting one’s strength.” As long as a person has some sense of control over what the world is like and what the logic of his field is, he can do things well.
The other is that AI is our friend, the second brain, so to speak. * Why do you have to carry the answer with you if you can always find the right answer?
AGI reverse-training the human brain is a great idea, and the fact is that we’re already using new technology to train our brains. People today have access to knowledge and can participate in training that was simply unimaginable in the past and should take advantage of these conditions.
AI Topic 3: The Enlightened Moment of Language Modeling
Reader SpongeBob: Two questions for Wan Sir:
(1) Like those paid programs online like paid courses, are the paid courses obtained not in the search scope of ChatGPT? If these paid programs are of higher quality compared to other information online, is it fair to say that ChatGPT’s INPUT is not actually of high quality?
(2) I’m a bit incredulous about this sentence in that talk “People have to be meaner than AI”: “As a language model, GPT is limited by the training material, like GPT-3’s knowledge is cut off at 2021.” Why can’t something as powerful as ChatGPT update its training material in real time? Especially now that the world is changing so much, I can’t believe ChatGPT’s knowledge is still 2021.
Vanguard Steel replied -
Both questions deserve a dedicated answer. Paid courses are not in the search engine’s scope of search because they are not on the public web. In fact, not only paid courses, including Taobao products and other information, are now blocked to search engines, so the value of search engines was already decreasing.
But language modeling is another story. According to the analysis [1], there is no law or case law in the U.S. on whether or not you can use a copyrighted corpus to train a language model. The rule of copyright law is to directly copy someone’s content and output it in large chunks, and that’s definitely not okay. But training a model is not copying, it’s digesting and transforming, and OpenAI has been invited to send a document to the U.S. Patent and Copyright Office [2] explaining its understanding that it is legal to use copyrighted content to train models. However, it is true that there is no official word on this yet.
The most recent high-profile case is a class-action lawsuit [3] that has been launched against Microsoft and its Github website, saying that Github’s AI-assisted programming product, Copilot, infringes on the copyrights of some open-source code. The program code itself is open source, but it’s also copyrighted, and copyright requires that you share and use it, but only if you retain the original author’s attribution - and Copilot gives some of the code directly to other programmers to use without retaining the original author’s attribution.
Artists have also recently launched lawsuits against AI image generation sites [4], saying you used our work to train AI but didn’t compensate us. There are also news organizations suing OpenAI, saying that their articles should not be used for training.
All of these cases haven’t been definitively decided yet, so let’s wait and see. I hope all of them will be allowed, so that AI’s knowledge can be maximized for the benefit of human progress.
The second issue is that we must understand that training a large language model is very difficult and requires a lot of arithmetic power and feeding a lot of corpus. So you can’t train it every week. It’s hard to train once, hundreds of billions of parameters in the model are fixed, the magic weapon is refined, and the rest is reasoning.The model behind ChatGPT is GPT-3.5, which should have been refined last year, and the corpus used is as of 2021, which is very reasonable.
There are two ways to make GPT handle new knowledge. One way is to feed the model with more material and continue to train it, just as we did with the philosopher Dennett’s robot in the previous talk. This is called fine tuning and is more time consuming.
Another way is to let the model “learn” new knowledge on an ad hoc basis, which is what Bing Chat and many of the gadgets on the web today that call APIs to read books do. This is essentially not training, but small sample learning. The disadvantage of this approach is that the total amount of information that can be input is limited.
AI Topic 4: Replacing Rules with Predictions
Reader Dish: When AI replaces us in the future to make the best decisions, our lives are more orderly and certain. But isn’t a life of never making mistakes also less exciting and fun? After all, sometimes mistakes bring us bad results, and sometimes they bring unexpected surprises. When we no longer have these surprises, isn’t it time for humanity’s inane ideas, mistakes, entertainment, and humor …… to be unleashed in the virtual world? Won’t the meta-universe flourish as well?
WV Steel replies -
You’re right, life needs mistakes and surprises. But AI doesn’t cancel out our mistakes and surprises. First of all there are random variables in the AI output, you can have it specialize in ideas that are not reliable but interesting. Further, and more importantly, AI is only providing suggestions, the decision is still in the hands of the human.
As we’ll talk about next week, one of the key ideas in the book Power and Prediction is to split decision-making into ‘prediction’ and ‘judgment’ parts, with the AI only being responsible for predictions and the human being being responsible for making judgments based on the AI’s predictions. Simply put, it means that the AI can accurately tell you whether it will rain tomorrow or not - but it’s still your own judgment as to whether to bring an umbrella or not. I think this has the potential to become a behavioral norm for AI in the future.
One can always be a maverick and not listen to the AI’s advice, just as the main character in the “I, Robot” movie turned off the car’s autopilot and drove his own car to continuously overtake the car.
However, insurance companies will have something to say about this. Not listening to AI advice could lead to higher premiums. But, many will argue that’s worth it. Moreover, the AI will be able to work out how much higher the premium should be for this - not as a punishment, but just to make the system fairer and more reasonable: after all, people driving honestly shouldn’t be made to pay for your capriciousness.
- So, people would still be free, they would just feel the responsibility and cost behind each of their choices more often. *
Q&A Zhao Erlong: Wan Sir, a question suddenly came to mind, if AI can predict everything, then can we predict everything in sports lottery and so on, then anyone can win the first prize, and then it will be meaningless? Will there be restrictions on AI to maintain market order?
Wei Wei Gang replied -
No! AI can’t predict everything, especially not lottery numbers. as good as AI is, it’s a product of math, it won’t cancel out chaotic phenomena, and it can’t do anything about chaotic epochs. Even for weather, AI can, at best, provide more accurate probabilities - not tell you 100 percent whether it will rain five days from now. For areas as chaotic as the stock market, AI can already only do a little bit of work in very short time zones, and it doesn’t always work yet. With the lottery, because it’s designed with a mechanism to be as random as possible, AI is inherently unpredictable.
Annotation
[1] https://www.thefashionlaw.com/ai-trained-on-copyrighted-works-when-is-it-fair-use/
[2] https://www.uspto.gov/sites/default/files/documents/OpenAI_RFC-84-FR-58141.pdf
[3] https://www.theverge.com/2022/11/8/23446821/microsoft-openai-github-copilot-class-action-lawsuit-ai-copyright-violation- training-data