jakiro
Former Google CEO Eric Schmidt, who is “nor a Google employee now”, shared it at Stanford not long ago.
The share was filmed as a video uploaded to the official account of Stanford online class YouTube, which included more than 40 minutes of Schmidt’s session with student Q&A.
Because the views are too direct and the words are too realistic, Schmidt’s sharing is on the news.
The Stanford official account hid all the videos.
Finally, Schmidt apologized for the “error remarks” in an email interview.
Well-known technology bloggerLanxiThe key content shared by Schmidt, TLDR, is summarized.Schmidt’s full-time Q&A is also attached at the end of the article.
Why is Google now being suppressed by OpenAI in the field of AI?Because Google feels that getting employees home early and balancing their work is more important than winning competition.If your employees only come to the company to work one day a week, how could you compare to OpenAI or Anthropic?
Look at Musk, look at TSMC. The reason why these companies are successful is because they can recruit employees. You have to force employees tight enough to win. TSMC will let PhD in physics work in the factory for the first year, you can imagineDo doctoral students in the United States go to the assembly line?
I have made many mistakes, such as I once thought that Nvidia’s CUDA was a stupid programming language, but now CUDA is Nvidia’s most awesome moat. All the big models must be run on CUDA, and only Nvidia’s GPU supports CUDA.It is a combination that cannot be shaken by other chips.
I also felt incredible when Microsoft cooperated with OpenAI. How could Microsoft outsource the most important AI business to that kind of small company? As a result, I took a look at it again. Then I saw how warm Apple is in AI. Big companies are reallyAll of them are bureaucratic, and all struggles are starting a business.
TikTok teaches Americans a lesson. If you start a business in the future, you can steal music or something, hurry up and do it – it seems that TikTok is instigating pirated BGM in the early days – if you do it, you will have the money to hire the mostTop lawyers help you wipe your butt, and if you don’t do it, no one will sue you.
OpenAI’s Stargate said it would cost 100 billion US dollars during its publicity, but in fact it might not be able to catch 300 billion US dollars. The energy gap is too big. I have suggested to the White House that the United States will either establish good relations with Canada in the future, have abundant water and electricity resources, and have labor force.Cheap and close enough, either get close to Arab countries and let them make sovereign investments.
Europe is gone, Brussels (the location of the EU headquarters) has been destroying opportunities for technological innovation. Maybe France still has some hope, Germany is not good, let alone other European countries. India is the most important swing state among the United States’ allies, andThe United States has lost China.
Open source is very good. Most of Google’s infrastructure in history have also benefited from open source, but to be honest, the cost of the AI industry is too high and open source cannot afford it. The French big model Mistral, which I invested in, will be converted to a closed source route., not all companies are willing and capable of being suckers like Meta.
AI will make the rich richer and the poorer poorer, and so will the country. This is a game between strong countries. Countries without technical resources need to get tickets to join the supply chain of strong countries, otherwise they will miss the feast.
AI chips belong to high-end manufacturing industry, with high output value, but they are unlikely to drive employment. There may not be many of you who have been to chip manufacturing factories, which are all mechanized and produced without people. People are stupid and dirty, so don’t expect manufacturing industry.In revival, Apple moved its MacBook production line back to Texas not because Texas has low wages, because there is no need to hire people on a large scale.
Historically, after the introduction of electricity into factories, no more productivity was created than steam engines. It was after about 30 years that distributed power supplies transformed the workshop layout, promoted the emergence of assembly systems, and then began a leap in productivity.Today’s AI is as valuable as the power of the original, but it still requires organizational innovation to truly achieve huge returns. At present, everyone is still picking “low-hanging fruits”.
Click to follow to discover more AI entrepreneurs
1.Three AI technologies that will change the future
Host: What do you think of the development of AI in the short term?The short-term definition for you should be the next one or two years, right?
Eric Schmidt: Things are going too fast, and I feel like every six months, I have to give a speech about the future again.Is there any computer science major here?Can anyone explain to you what is a million token context window?
audience: The basic meaning is that asking prompt can use one million tokens or one million words, or something similar.
Eric Schmidt: So a million token means you can ask a question of one million word length.
audience: Yes, I know this is a big direction for Gemini at the moment.
Introduction to Gemini’s official website (Chinese translation is plug-in effect, thanks to immersive translation)
Eric Schmidt: No, their goal is to reach ten million.Anthropic has reached 200,000 and is still growing.The goal is one million or more, and it is conceivable that OpenAI has similar goals.Can anyone give us a technical definition and explain what an AI Agent is?
audience: AI agent is to perform tasks online, buy things on your behalf, and various similar operations.
Eric Schmidt: So agent is something that performs a certain task, and another definition is a large language model with memory functions.Let me ask another question, a student in computer science, can anyone explain what Text-to-Action is?
audience: It means expanding text to more text, inputting text, and then the AI triggers operations based on the text.
Eric Schmidt: Another definition is to convert the language into Python – a programming language that I never thought would continue to survive.But now everything in AI is done in Python.Recently, there is a new language just released called Mojo, which seems to finally solve the problem of AI programming, but we also need to see if it can survive under the situation of Python.
Let me ask another technical question, why is Nvidia worth $2 trillion while other companies are in trouble?
audience:For technical reasons.I think this is mainly due to the optimization of the code running.Most of the code currently needs to run in an optimized environment, and currently only Nvidia’s GPU can do this.In fact, other companies have the ability to develop various technologies, and may have up to ten years of software development experience, but they do not have teams that optimize specifically for machine learning.
Eric Schmidt: I like to regard CUDA as the C language of GPU.This is my favorite way of understanding.It was born in 2008 and I always thought it was a bad language, but it became mainstream.There is now a complete set of open source libraries, all highly optimized for CUDA.This is ignored by everyone who builds these tech stacks.We call it vlm technology, plus other similar open source libraries, they are all optimized for CUDA.This is difficult for competitors to replicate.
What do the above mean?
Over the next year, you will see the capabilities of larger context windows, Agents, and Text-to-Actions.When they are applied at scale, the impact will be even greater than the huge impact of social media we see now, at least in my opinion.In the context window, you can use it as short-term memory, and it is shocking to be so large, and technically service and computing are very complex.
The interesting thing about short-term memory is to let it read 20 books, enter the text of these books as a query, and let it tell you what the book is.The human brain forgets the middle part.There are some people now building basic LLM Agents.The way they work is, for example, read the content of chemistry, discover the chemistry principles, then test it, and then add the results to their understanding.This is very powerful.
The third point is the text to action I mentioned.For example, the government is now considering banning TikTok.We don’t know if it will happen.If TikTok is banned, I suggest you say to your LLM: Copy a TikTok, get all users, get all music, join my preferences, generate and publish within 30 seconds.If there is no fire within an hour, then change to a similar approach, this is the command.Bang bang bang, it will be done immediately.
Do you understand?If you can directly generate any digital instructions from any language, this is basically what Python does in this scenario.Imagine that everyone has a programmer who can work as you ask, and no longer those who work for me but are disobedient.(Laughs) Programmers all know what I’m talking about.Imagine a programmer who is not arrogant and does what you ask for without paying that much money.And these programmers are unlimitedly available.And these…
host: It will be realized within the next one or two years.
Eric Schmidt: It will be realized soon.I am very confident that they will happen in the next wave of technology.
Listener: You mentioned that the combination of extended context windows, proxy, and Text-to-Action will have unimaginable effects.First, why are these combinations important?Second, I know you can’t predict the future, but why do you think this will be beyond our current imagination?
Eric Schmidt: I think it is mainly because extended context windows can solve the problem of timeliness.Current AI models take about a year to train, including 6 months of preparation, 6 months of training and 6 months of fine-tuning, so they are always a bit lagging.But the extended context window allows you to enter the latest information, which is very powerful and can be updated in real time just like Google.
Regarding the Agents model, I give an example.I built a foundation, funded a nonprofit, and they started a project with a tool called Chemcrow, a system based on a large language model to learn chemistry.They use this system to generate chemical hypotheses on proteins, and then the laboratory will do tests at night, and the system will continue to learn.This has greatly accelerated research progress in the fields of chemistry and materials science.
I think “Text-to-Action” can be understood as the effect brought by a large number of cheap programmers.But I don’t think we really understand what happens when everyone has their own programmer. What they do is your expertise, not just a simple thing to turn on and off the lights.
You can imagine a scenario, such as if you don’t like Google.Just say, help me create a Google competitor, search for web pages, build interfaces, add generative AI, and do it in 30 seconds. Let’s see the effect.These established companies, such as Google, are likely to be threatened by such attacks, and we’ll see.
2.“I’m no longer a Google employee.”
host: You worked at Google for many years, and they invented the Transformer architecture, and Peter (Peter Norvig, former director of engineering at Google Research) is one of the leading ones.Thanks to smart people like Peter and Jeff Dean.But now, Google seems to have lost its advantage in initiative, and OpenAI has caught up.Anthropic’s Claude comes in the latest rankings I’ve seen.I asked Sundar (Sandar Pichai) and he didn’t give me a definite answer.Maybe you have a clearer or objective explanation about what exactly happened there.
Eric Schmidt: I am no longer a Google employee.Frankly speaking, Google pays more attention to work-life balance, getting off work early and working from home seems to be more important than winning a battle.The secret to success of a startup is that employees work hard.I’m sorry, it’s so straightforward, but that’s the truth.If you start a company after graduation, you won’t let employees come to the company only one day a week and work from home most of the time.This won’t work if you want to compete with other startups.
host: The early situation of Google is very similar to Microsoft at that time…
Eric Schmidt:Yes.
In our industry, there is a common phenomenon:Some companies have won the market in a very innovative way, completely dominating one field, but failing to transition to the next stage smoothly.
There are many such situations.I think the founders are important, and it is a very important issue, they are at the helm of the company.While founders are often difficult to get along with and demanding on their employees, they also push the company forward.
While we may not like some of Elon’s personal behaviors, look at what he does at work.The day I had dinner with him, he kept flying back and forth.I was in Montana and he was flying to the early morning meeting with xAI at 10 o’clock that night.
When I went to Taiwan, I felt that different places had different cultures. What impressed me was that TSMC had a regulation that a newly-employed PhD in physics must first work in the basement of the factory.Can you imagine letting a PhD in the United States do this kind of work?Nearly impossible.
The results of the work are different.The reason why I am so harsh about work is because of the network effect of these systems.Time is very critical, and in most industries, time is not that important, they have enough time.Coca-Cola and Pepsi will always exist, and the competition between the two will continue, changing slowly like glaciers.
When I work with a telecom company, it takes 18 months for a general telecom contract to be signed.I don’t think it’s necessary for this long, things should be done as soon as possible.We are now at the peak of growth and earnings, and we still need some crazy ideas.
For example, when Microsoft decided to work with OpenAI, I thought that was one of the stupidest ideas.It was incredible that Microsoft handed over AI leadership to the teams of OpenAI and Sam.Today, however, they are gradually becoming one of the most valuable companies, comparable to Apple’s competition.Apple has no good solution for AI, and it seems that Microsoft’s strategy has worked.
3.The gap in the model is widening
Eric Schmidt: You just asked, what will happen next, every six months, my thoughts will fluctuate.We are now in a periodic fluctuation of odd and even oscillation.As of now, the gap between cutting-edge models—only three models now—and other models seem to be widening.Six months ago, I thought the gap was narrowing, so I invested a lot of money to some small companies, but now I’m not so sure.
I started talking to big companies, and big companies told meThey need 10 billion, 20 billion, 50 billion, or even 100 billion funds.
Host: The goal is 100 billion, right?
Eric Schmidt: Yes, it’s very difficult.Sam Altman and I are good friends, and he thinks it might take 300 billion, or more.I told him that I had calculated the power needed.I went to the White House last Friday and told him openly that we need to have a good relationship with Canada because Canada is not only good, but also helped invent AI and has a lot of water and electricity resources.And our country does not have enough electricity to support this development.
Another option is to have the Arab countries contribute.I personally like Arabia and have been there for a long time.But they won’t comply with our national security rules, and Canada and the United States can work together.
host: That’s right.So these data centers worth 100 billion and 300 billion,Electricity will become a scarce resource.
Eric Schmidt:Yes.Following this idea, if 300 billion is invested in Nvidia, do you know what stocks to buy, right?(Laughs) Of course, I’m not recommending stocks.
host: That’s right.We will need more chips, Intel is getting a lot of money from the U.S. government, and AMD, they are all working to build chip factories.
Eric Schmidt: If there is a device using Intel chips on site, please raise your hand (the listener raises your hand).Its monopoly seems to have ended here.
host:Intel was indeed a monopoly.And now it is Nvidia’s monopoly.So, are there any other companies that can do technical barriers like CUDA?I talked to another entrepreneur a few days ago that he would switch between TPU and Nvidia chips based on the resources he could obtain.
Eric Schmidt: Because he has no other choice.If he had unlimited funds, he would definitely choose Nvidia’s B200 architecture today because that was faster.I’m not suggesting anything, competition is of course a good thing.I discussed this matter in detail with AMD’s Lisa Sue. They developed a system that can convert CUDA architecture into their own, called Rocm.They are still improving.
4.We will experience a huge bubble,Then the market will adjust itself
Listener: You are very optimistic about the prospects for AI.What do you think drives this progress?Is it more money?Or more data?Or a technological breakthrough?
Eric Schmidt: I basically look at all projects, because I can’t say which one can succeed.Moreover, now there is a lot of funds coming in with me.I think part of the reason is that early investment has made money. Nowadays, investors with large funds do not understand AI very much, but they think that every project needs to be added with some AI elements, so almost all investments have becomeAI investment.They can’t tell the difference between good and bad.The AI I understand is the kind of system that can really learn, and I think that counts.
In addition, there are now some very advanced new algorithms that are no longer limited to the Transformer architecture.I have a friend, who is also my long-term partner, who has made a brand new non-Transformer architecture. A team I sponsored in Paris also said that they have similar innovations, and there are many new trends at Stanford.
Finally, it is widely believed in the market that developing smart technologies will bring huge returns.For example, if you invest $50 billion in a company, you definitely want to make a lot of money back through smart technology.So we may experience a huge investment bubble and the market will adjust itself.This has always been like this in the past, and it may be no exception now.
host: You mentioned before that the leading companies are getting closer and closer.
Eric Schmidt: Yes, it is indeed the case now.There is a company called Mistral in France, and they do a good job and I invested in them.They launched a second edition of the model, but the third edition may be closed because the cost is too high.They need income and can no longer provide models for free.
The debate between open source and closed source is very fierce in our industry.My entire personal career has been based on people’s willingness to share open source software.The technical work I do is open source, and many of Google’s core technologies are also open source.But now it mayBecause the capital cost is too high, the way software development may undergo fundamental changes.
I personally think that the productivity of software programmers will at least double.There are now three or four software companies working hard to achieve this goal, and I have invested in these companies.Their goal is to improve the efficiency of software programmers.A very interesting company I met recently called Augment.I always think of individual programmers, but their goal is actually those large software teams that may have millions of lines of code, but no one can figure out the details of running all the code.This problem is very suitable for solving with AI.Can they make money?I hope it can.
host: So, there are still many issues to discuss.
Audience: I don’t think there are much discussions about non-Transformer architectures, but now they have made more progress. What new progress have you seen in this field?
Eric Schmidt: I don’t have a deep understanding of mathematics, and the mathematics here is very complicated.But basically, they use different methods to do gradient descent and matrix multiplication, which is faster and better.Transformers is a systematic way of performing multiplication operations at the same time, so I understand it.It’s similar to this, but the mathematical principles are different.
Listener: You are an engineer, and considering the capabilities these models may have in the future, do we still need to spend time learning programming?
Eric Schmidt: It’s like you already know how to speak English, why do you still need to continue learning English?Learning can always make people take a step forward.You have to understand how these systems work.
5.Distributed computing cannot solve itComputing power issues in AI
Listener: Two simple questions: First, is the economic impact of large language models slower than the market impact you expected at the beginning?Second, do you think the academic community should receive AI subsidies?Or should we cooperate with large companies?
Eric Schmidt: I have been working hard to build a data center for the university.If I were a professor in the computer science department here, I would be very dissatisfied because I couldn’t develop those algorithms with graduate students and were forced to cooperate with those big companies.In my opinion, these companies are not doing enough in this regard.I’ve talked to some professors and many of them have to spend a lot of time waiting for Google Cloud usage quotas.This is a thriving field, and the right way to do it is to provide resources to the university, and I am working hard to promote this.
As for the impact of the labor market you mentioned, I basically believe that high-skilled college education and related work should be fine because people will work with these systems.I think these systems are no different from the previous wave of technology, and those dangerous jobs and jobs that do not require human judgment will eventually be replaced.
Listener: Have you ever studied distributed environments?I asked this because it is difficult to build a large cluster, but the MacBook is still very powerful.There are many small machines all over the world.Do you think ideas like Folding@home can be used for training?
Note: “Folding@home” is a project that utilizes global distributed computing resources, using idle resources from global participants’ computers to perform protein folding calculations.
Eric Schmidt: Distributed environments are indeed a challenge.It is indeed not easy to build a large cluster, but every MacBook has its own computing power.There are so many small machines around the world, and the idea of uniting them does have potential.This can be used for training, but there are still many technical details to be solved.
We have studied this problem in depth, and the working principle of these algorithms is as follows: You have a very large matrix, which is basically multiplication.You can imagine that the process is repeated.The performance of these systems depends entirely on how fast data is transferred from memory to the CPU or GPU.In fact, Nvidia’s next-generation chips have integrated these functions into one chip. Now these chips are very large and their functions are integrated together.Moreover, the packaging process is very fine, and the chip and packaging are completed in a clean room.So at present, supercomputers and light transmission, especially the interconnection between memory, are the key factors.Therefore, I don’t think it’s possible to achieve what you said in the short term.
Host: Is it possible to separate the large language model?
Eric Schmidt: To do this, you have to have millions of such models.And the way you ask questions will get very slow.
6.In the future, we may not understand AI.But they need to be known
host: I want to change the topic and talk about something philosophical.Last year you and Henry Kissinger and Daniel HuttenlocherWrite an article together, explores the nature of knowledge and its evolution.I have also talked to others about this topic recently. Most historical periods, human understanding of the universe was mysterious until the arrival of the scientific revolution and the Enlightenment.Your article says that models are becoming more and more complex and difficult to understand, so that we are no longer as clear about their internal mechanisms.
Feynman once said,“I can’t understand what I can’t create.“I have mentioned this sentence recently, but at present, people seem to be creating something that they don’t even understand.Does this mean that our understanding of knowledge is changing?Do we need to start accepting the conclusions of these models, even if they cannot give clear explanations?
Eric Schmidt: Let me give an example, this is a bit like a young man.If you have teenagers in your family, you know they are humans, but you don’t always know what they think.However, our society has learned how to adapt to their existence and knows that they will eventually mature.So, we may have some knowledge system,We cannot fully understand, but we can understand their boundaries.We know what they can do and what they can’t do.This may be the best result we can expect.
Host: Do you think we can grasp these restrictions?
Eric Schmidt: I think we can handle it.The small team we discuss every week feels that we may use that kind of confrontational AI technology in the future.Imagine that in the future, there will be companies that will specialize in this. If you give them money, they will help you test the AI system and find vulnerabilities, just like the “red team” today, but this time they use AI.The entire industry will engage in such AI against AI, especially those parts that we don’t understand very well.I think this is quite reliable.Stanford can also consider this direction.If a graduate student is interested in how to crack these big models and study how they work, it is a good skill point for them.So I think these two things will improve together.
Audience: You mentioned the comments related to confrontational AI. In addition to the obvious improvement of the AI performance model, what other problems do we need to solve?What are the main challenges in order for AI to really do what we want?
Eric Schmidt: It is indeed necessary to improve higher performance models.You have to assume that as technology advances, the hallucination of AI will decrease, although I’m not saying it will disappear completely.You also have to assume that there is a way to verify the effect, so we need to know whether the results have met expectations.
For example, the example of TikTok competitors I just mentioned.By the way, I’m not suggesting you illegally steal everyone’s music.If you are an entrepreneur in Silicon Valley – I hope you all will become such entrepreneurs –If your product becomes popular, you will hire a large number of lawyers to help you solve the problem; but if no one uses your product, it doesn’t matter even if you steal all the content, it doesn’t matter.But don’t take my words seriously.
Silicon Valley will conduct these tests and solve these problems.This is how we usually deal with it.So I believe that in the future we will see more and more high-performance systems, and the tests will become more and more refined, and eventually there will be adversarial testing to ensure that AI is within a controllable range.Technically, we call it “chain thinking reasoning”.It is expected that in the next few years, you will be able to generate 1000 steps of chain reasoning, just like cooking by recipes.You can follow the recipe step by step and then verify that the final result is correct.That’s how the system works.Unless you are playing the game, of course.
7.Fake information seems unsolvable in the short term
Listeners: How to prevent AI from creating false information in public opinion, especially in the upcoming elections?Is there any solution in the short and long term?
Eric Schmidt: In the upcoming elections, and even globally, most false information will be spread through social media, and social media companies do not have enough power to manage this information at the moment.If you look at TikTok, some people criticize TikTok for being biased towards some kind of false information rather than another.I feel like we are in a mess in this regard and we need to learn how to think critically.It can be a tough challenge, but just someone tells you something doesn’t mean it’s true.
Listener: Will you go to the other extreme?No one believes the truth?Some people summarize this phenomenon as an “epistemological crisis.”
Eric Schmidt: I think we are facing a trust crisis now.I think the biggest threat to society is false information, because we will become more and more powerful in creating false information.The biggest problem I encountered when I was managing YouTube was that people would upload fake videos and even let someone die. We had a “no death policy” at that time, which sounded shocking.
Note: YouTube does not allow any content that encourages dangerous or illegal activities that can result in serious bodily injury or death.
It was really painful to try to solve these problems, and there was no generative AI at that time.So to be honest, I don’t have a particularly good solution.
host: Technical means are not a universal solution, but there is a way that seems to be able to alleviate this problem, which is public key authentication.For example, when Biden came on stage to give a speech, why can’t he add digital signatures to what he said like SSL did?Or when celebrities or public figures speak, can they have their own public key?Just like when I gave my credit card information to Amazon, I knew that the recipient was indeed Amazon.
Eric Schmidt: This is indeed a way of public key authentication, coupled with other verification methods, to ensure that we know the source of the information.
I have written a paper with others, and what I support is your argument just now, but unfortunately, this paper has no effect at all.So maybe the system is not organized to solve this problem as you said.
Overall, CEOs are aimed at maximizing revenue, and in order to do this, they must pursue maximum user engagement.Maximizing participation means ignite more anger.The algorithm will give priority to pushing outrageous content because it will bring more revenue.Therefore, there is a tendency to tend toward extreme content as a whole, and this is not divided into camps.This is a problem that must be solved in our society.
We have talked about TikTok’s solution in private before.When I was a child, there was a rule called the “Equal Time Rule”.Because TikTok is actually not social media, it is more like TV, and there are programmers controlling content.Data shows that TikTok users in the United States spend an average of 90 minutes watching 200 videos a day, which is quite large.The government may not set rules for equal time, but some form of balance is necessary.
8.Big model is a competition that only a few countries are eligible to participate in
Listener: In terms of national security or interests, what role do you think AI will play in the competition with China?
Eric Schmidt: I served as Chairman of the AI Committee, which examined this issue in detail.The report has 752 pages, you can check it out.Let me briefly summarize: we are leading now, we need to continue to lead, and we need a lot of money to achieve this.
The general situation is that if cutting-edge AI models continue to develop and a few open source models are involved, then only a few countries are eligible to participate.Those countries with a lot of money, a strong education system, and are determined to win.The United States is one of them, and so is China.Maybe there are other countries.But it is certain that in your lifetime, the competition in the knowledge field between the United States and China will be the greatest confrontation.
The U.S. government has basically banned the export of Nvidia chips to China, and while they don’t allow this, they do it.We are about 10 years ahead of China in chip technology.We are also leading the way in lithography technology by about 10 years.I guess we can lead a few more years in the future.The chip bill was a decision of the Trump administration and was approved by the Biden administration.
Host: Do you think the current government and Congress listen to your suggestions?Do you think they will make such a large investment?In addition to the Chip Act, will large-scale AI systems continue to be built?
Eric Schmidt: As you know, I led an informal group, which is not an official group, and this group includes all the common AI sector participants.The recommendations made by these participants have become the basis for decisions in the Biden administration’s AI field, a bill that may be the longest presidential directive in history.
Note: President Biden of the United States issued the Executive Order on Addressing United States Investments in Certain National Security Technologies andProducts in Countries of Concern
host: You are advancing a special competitive research project.
Eric Schmidt: This is the actual implementation bill of the Executive Office.They are busy putting into the details and have done a good job so far.For example, last year we discussed a question: how to detect potential dangers in the system.This system may have learned some dangerous things, but you don’t know what to ask.In other words, this is a core issue.The system has learned something bad, but it won’t tell you what you learned, and you don’t know how to ask questions.There are many threats here, such as it may have learned how to mix chemicals that you don’t understand.So now many people are working hard to solve this problem.
Finally, we set a threshold in the memorandum called floating point operation at the power of 10^26, which is a measure of computing power.When this threshold is exceeded, you must report your behavior to the government.This is part of the rule, the EU sets the threshold to the 25th power of 10, but the difference is not big.I think these technical differences will eventually disappear. The current technology can be “federal training”, that is, different parts can be combined for training.So we may not be able to completely avoid the threats brought by these new technologies.
host: I heard that OpenAI has had to do this, partly because the power consumption is too high and no place can bear all the calculations alone.
9.AI is a game for rich people.The richer are richer
Listener: The New York Times sued OpenAI for training models with their works.What do you think this means for data usage?
Eric Schmidt: I have a lot of experience in music copyright.In the 1960s, there was a series of lawsuits that eventually led to an agreement that every time your song was played, no matter whether the listener knew who you were, you would get a certain royalty and the money would be deposited into your bank.Account.I guess the situation will be similar in the future, with many lawsuits and eventually some kind of agreement is reached that requires a certain percentage of income to be paid when using these works.You can look at the examples of ASCAP (Association of Composers, Writers and Distributors) and BMI (Broadcast Music, Inc., an American Performance Rights Organization), and while it seems a bit outdated, I think that will be the case in the end.
Listener: It looks like there are several companies that dominate and will continue to be in the AI field, and these companies seem to be the focus of all antitrust laws.What do you think of these two trends?Do you think regulators will split these companies?What impact will this have on the industry?
Eric Schmidt: In my career, I have pushed for splitting Microsoft, but it has not been split.I also tried to keep Google from being split, and it wasn’t split either.So in my opinion, as long as these companies avoid becoming monopoly giants like John D. Rockefeller (founder of Standard Oil), the trend is not a split.This is the origin of the Antitrust Law.
I don’t think the government will take action.The reason you see these big companies dominate the market is that they only have the funds to build these data centers.So my friends Reed Hastings (Netflix and CEO) and Elon Musk are doing it.
Therefore, the rich get richer, and the poor can only do their best.This is the fact, this is a game for the rich country, requiring huge capital, a large amount of technical talents and strong government support.There are many other countries with various problems, they do not have these resources, so they have to work with other countries.
Listener: You have spent a lot of time helping young people create wealth and are very enthusiastic about this.Do you have any suggestions for this stage of your career and future for the students here?
Eric Schmidt: I am very impressed by your ability to quickly show new ideas.In one of the hackathons I was involved in, the winning team was tasked with getting the drone to fly between the two towers.They completed this mission in a virtual drone space, allowing the drone to understand the meaning of “between…”, wrote code in Python, and successfully allowed the drone to pass through the tower in the simulator.If you are a professional programmer doing this, it may take a week or two.
What I want to say is,The ability to quickly make prototypes is really important.As an entrepreneur, one of the problems is that everything happens very quickly.Now, if you can’t prototype with various tools in a day, you have to think about it because your competitors can do it.
So my advice is that when you start thinking about starting a business, it is good to write a business plan, and you should have the computer help you write a business plan. It is very important to use these tools to quickly turn your ideas into prototypes.Because for sure, someone is doing the same thing in another company, another university, or a place you’ve never been.
