jakiro
Author: City of Sky·Lord, Source: Web3 City of Sky
host:
I first met Sam Altman, our upcoming guest, about 20 years ago, when he was working on a local mobile app called Loop.We all have the support of Sequoia Capital, and in fact, we are all the first Boy Scouts of Sequoia Capital.
He invested in Stripe, a fintech company in unknown ways, and I invested in Uber.In that little experiment…
Sam:
Have you invested in Uber?I’ve never heard of it before.
Yes, I think so.
You should write a book, Jacob.
The small experimental fund that Sam and I participated as the Boy Scouts was the fund with the highest return on capital in Sequoia.I heard that millions of dollars have turned into more than 200 million dollars.He also worked for Y Combinator for a while and served as president from 2014 to 2019.In 2016, he co-founded OpenAI with others, with the goal of ensuring universal artificial intelligence benefits all mankind.In 2019, he left YC and joined OpenAI full-time as CEO.
On November 30, 2022, things became very interesting.That day was the day when OpenAI launched ChatGPT.In January 2023, Microsoft invested $10 billion.In November 2023, Sam was fired by OpenAI during a crazy five-day period.Everyone has to work at Microsoft.A bunch of heart-shaped emojis went viral on X and Twitter, and people began to speculate that the team had reached universal AI.The end of the world is coming.Suddenly, a few days later, he returned to the position of CEO of OpenAI.
According to reports, in February 2023, Sam is seeking to raise $7 trillion for an AI chip project.It was previously reported that Sam was seeking to raise $1 billion from Masayoshi San to create an iPhone killer with iPhone co-founder Johnny Ive.
At the same time, ChatGPT is getting better and better and becoming a household name.It has had a huge impact on the way we work and how we work done.It is reportedly the fastest product in history to reach 100 million users in just two months.
Look at OpenAI’s crazy revenue growth.Their ARR reportedly reached $2 billion last year.Welcome to the all-round podcast,
I’ve observed that the entire industry is nervously looking forward to the release of the GPT-5.As far as I understand, it may be launched sometime this summer, but this time window is quite broad.Can you define it for us?
Sam:
We will spend some time releasing this major new model.I believe it will be great when we do this.We will think carefully about how to proceed.It might be published in a different way than we have published the previous model.Also, I’m not even sure if we’ll call it GPT-5.
I would like to say that since we released the GPT-4, especially in the past few months, a lot of people have noticed its excellent performance.I think this reveals the essence of the world better. It is not simple 1, 2, 3, 4, 5, 6, 7, but a process of using AI systems, and the entire system will continue to become more and moregood.I think this is both a better technological direction and easier for society to adapt. I believe this is our development direction.
host:
So, does this mean we won’t have long training cycles, but will keep retraining or training submodels, Sam?Maybe you can share with us what changes may occur in the architecture of large models in the future.
Sam:
One scenario you can imagine is that we are always training a model.This seems to be a reasonable approach to me.
We are discussing different ways of publishing this time.Are you considering publishing to paid users first or slowing down the release to keep the Red Team nervous because the risk is too high now.In fact, you have so many paid users and everyone is following every step of your life.You have to think more carefully now, right?
Yes, GPT-4 is still only open to paid users at the moment, but one thing we really want to do is figure out how to make the free users more advanced technology available.I think this is a very important part of our mission.Our goal is to build AI tools and make them widely available, whether they are free or less expensive, whatever it is, so people can use them to create the future, rather than letting the magic AGI in the sky create the future andPour it on us.This seems to be a better road and a more inspiring one.I firmly believe that things are indeed moving in this direction.So I’m very sorry,We haven’t figured out how to get free users to use GPT-4 level technology.This is what we really want to do.I must admit, it’s very expensive.
Sam, I think two big factors that people often discuss are potential costs and latency, which somewhat limit the rate of killer applications.Then I think the second factor is the long-term capability that people build in the open source world and the closed source world.I think the fanatic part of this field is the enthusiasm of its open source community.An incredible example is that we did a very crazy demo for Devon, which was five or six weeks ago and it worked very well.Then, some young people released the project under MIT’s open licenses, such as OpenDevon.It performed very well and was nearly comparable to other closed-source projects.So maybe we can start with this question, i.e., what are the business decisions that keep these models closed source?What do you think will be developing in the next few years?
For the first part of your question, speed and cost, these are very important to us.I don’t want to give a timeline on when we can significantly reduce the delay, because research is difficult, but I believe we can do it.We hope to significantly reduce latency and significantly reduce costs.I believe this will happen.We are still in the early stages of scientific development and cannot understand how it works.In addition, we have all the engineering tailwinds.So I don’t know when we’ll get intelligence that’s cheap enough to be measurable, and so fast that we and everyone else feel it’s instant.But I believe we can reach a fairly high level of intelligence.This is important to us, and it is important to the user, and it will unlock a lot of things.
Regarding open source and closed source, I think both have their advantages.I think we have opened sourced some projects, and we will open source more projects in the future.But in reality, our mission is to move towards artificial intelligence and find ways to widely distribute its benefits.We have a strategy that seems to resonate with many people.Obviously, it is not for everyone.And it is a huge ecosystem.There will also be open source models and people who are built in this way.
One area I personally have a particular interest in open source is that I want a good open source model that can run on my phone.I think the world doesn’t have good enough technology to develop a good version.But this seems to be a very important thing to do at some point.
Would you do this?Will you post it?
I don’t know if we will, or if anyone will.
How about Llama 3?
Llama 3 runs on your phone?
I think there may be 7 billion parameter (mobile phone) versions.
Yes.I don’t know if that’s suitable for a phone, but…
That should be suitable for a cell phone.
But I don’t know, I’m not sure if that fits, I haven’t played it yet.I don’t know if it’s enough to accomplish what I’m considering here.
So when the Llama 3 was released, I think the biggest takeaway for a lot of people was, wow, they’ve caught up with the GPT-4.I don’t think it’s equal in all respects, but overall, it’s very, very close.I think the problem is, you released 4 not long ago, you are developing 5, or making more upgrades to 4.I want to hear your point of view on how to stay ahead in an open source environment, what do you think?This is usually a very challenging task, what do you think?
Our goal is not to create the smartest set of weights we can make, but to create a useful layer of smarts for people to use.The model is just a part of it.I believe we will remain ahead of the front line in this regard, and I hope we can be far ahead of the rest of the world in this regard.But there is a lot of other work in the entire system, not just model weights.We must build lasting value in a traditional way, like any other business.We have to come up with a great product and reason to stick with it and deliver it at a good price.
When you start this organization, what you say is or part of the discussion is that it is too important for any company, so it needs to be open.Then it turns to, no one can see it is too dangerous, we need to lock it up because you have some concerns about it.I want to know, is this true?Because the cynical side would say that this is a capitalist move.Then, I was curious about what the decision was to start with opening up.The world needs to see this.Close is really important.Only we can see it.So, how did you come to this conclusion?
Part of the reason we released ChatGPT is because we want the whole world to see this.We’ve been trying to tell people that AI is really important.If you go back to October 2022, then there weren’t so many people who thought AI would be that important, or it actually happened.A big part of what we are trying to do is to put technology into people’s hands.Now, again, there are different ways to do this.I think there is indeed an important role, that is, for example, this is the way to implement it.
But in fact, we have so many people using the free version of ChatGPT, we don’t run ads, and we won’t try to make money from it.We just rolled it out because we wanted people to have these tools.I think it does a lot of work, provides a lot of value, teaches people how to fish, and also makes the world really think about what’s going on here.Now, we still haven’t found all the answers.We are like everyone else, groping forward.I think we change our strategy many times as we learn new things.
When we started OpenAI, we really didn’t know how things would go, we would make language models, we would make products.I remember clearly the first day, and we were like, OK, now we are all here.It’s difficult to prepare all of this.But what will happen now?Maybe we should write some papers.Maybe we should stand around the whiteboard.We have been working hard to figure out what is next, what is next, and what is next.I think we will continue to do that.
Can I confirm it again to make sure I’m not hearing it wrong?I understand your point of view that whether it is open source or closed source, all of these models, no matter what business decisions you make, will gradually improve their accuracy.Not all companies have enough capital, but suppose there are four to five, such as Meta, Google, Microsoft, etc., or there may be a startup.These companies all operate on an open network.Then soon, the accuracy or value of these models may be transferred to these proprietary training data sources that you may or may not be able to obtain, or others can obtain and you cannot obtain.
Do you think this is the trend of things?Opening the network makes everyone reach a certain threshold, and then it becomes an arms race for data?
I don’t think so.I firmly believe this won’t turn into an arms race for data, because when the model becomes smart enough, it shouldn’t be just about getting more data, at least not the data used for training.Data may be important because it has its value.One of the most important things I have learned from all this is that it is difficult to make confident predictions about what will happen in the next few years, so I don’t want to try it now.What I want to say is that I expect many very capable models to appear in the world.I feel like we just stumbled upon a new natural or scientific fact, or something you could call it a fact that we could create.I don’t think it’s literal, but it’s like a spiritual point.Intelligence is just this sudden property of matter, which is like physical rules or something.So people will figure it out.But there will be all these different ways of designing the system.People will make different choices and come up with new ideas.I’m sure like any other industry, I expect there will be multiple approaches, and different people like different approaches.Some people like iPhones, some people like Android phones.I think this will be effective.
Let’s go back to the first part and discuss only cost and speed.All of you NVIDIA throughput is a bit bit rate limited, and I think you and most of you have announced how much capacity you can get because that is the maximum capacity they can produce.What needs to happen on the substrate to truly reduce computing costs, speed up computing, and get more energy?How do you help the industry solve these problems?
We will certainly make huge progress in algorithms.I don’t want to underestimate this.I’m very interested in chips and energy.But if we can double the efficiency of the model of the same quality, it would be equivalent to double our computing power.I think there is a lot of work to do there.I hope we can really see these results.In addition, the entire supply chain is very complex.The production capacity of a logical factory.How many HBMs can the world produce?You can quickly get a license, pour concrete, build a data center, and then arrange for people to connect all the lines.Finally, there is energy, which is a huge bottleneck.But I think the world will make a difference when it has such great value to humanity.We will work to make this happen faster.The possibility does exist, and I cannot give a specific probability, but I believe, as you said, this is a huge fundamental breakthrough.We already have more efficient calculations, however, I don’t like to rely too much on it and I don’t want to spend too much time thinking about it.
On the device side, you mentioned models that can be installed on your phone.Obviously, whether it is LLM or SLM, etc., I believe you are considering this issue.But will the device itself change?Does it need to be as expensive as an iPhone?
I’m very interested in this.I like new forms of computing, and every major technological advancement seems to bring something new.The level of excellence of the phone is incredible, so I think the barrier here is very high.I personally think that the iPhone is the greatest technological product in human history, and it is really a great product.
So, what will be next?
I have no idea.It would be a great thing to go beyond it, but I think the barrier is very high.
I’ve been working with Johnny Ivy and we’ve been talking about various ideas, but I’m not sure if the new device has to be more complex or actually just needs to be cheaper and simpler?Almost everyone is willing to pay for their phones, so if you can make a cheaper device, I think the barriers to carrying a second piece of equipment or using a second piece of equipment are quite high.So, I think, considering that we are all willing to pay for our phones, or most of us are willing, I don’t think it’s the answer to the problem.
So, what are the differences in the answer?
Is there a specialized chip that can run on a phone and can drive a phone-sized AI model well?
There may be, but phone manufacturers will certainly do that, and that doesn’t require new devices.I think you have to find some really different interaction paradigms that technology can achieve.If I knew what it was, I would be happy to start researching it now.
Now you can use voice in the app.In fact, I set the action button on my phone to go directly to ChatGPT’s voice app, which I use with my kids and they love talking to.It has latency issues, but it’s really great.
We will do better.I think voice is the hint of the next thing.For example, if you can make voice interactions very good, it feels like a different way of using a computer.
However, like the problem we’ve already encountered, like, why is it not responding?And, it feels like CB, like over and over again.It’s really annoying to use, but it’s also great when it gives you the right answer.
We are working on this problem.Now it’s bulky and slow and doesn’t feel too smooth, unreal or unnatural.We will make it all better.
What about computer vision? You can choose to wear related devices.You can combine visual or video data with voice data.
Today, AI has been able to understand what’s going on around you.You can ask ChatGPT like you would with a multimodal device, for example: “What am I looking at?” or “What plant is this?” I must admit, this ability is very powerful.This is obviously another hint.
However, whether people choose to wear glasses or use some kind of device when needed, it can cause many social interpersonal problems, and wearing computer devices can become very complicated.
We’ve seen this in Google Glass use.People may be troubled when performing tasks.I forgot some specific situations.
If AI is everywhere, for example, on people’s phones, what apps can it unlock?Do you feel this way?What do you want to see?
I guess all I need is an always online, ultra-low friction device that I can interact with via voice or text, or ideally other ways.It just needs to know what I want and there is a constant presence that helps me through the day.It has as much background information as possible, like the greatest assistant in the world.It is this existence that makes me better and better.
When you hear people talking about the future of artificial intelligence, they may imagine two different ways, although it doesn’t sound much different.But I think that in practice, we will be very different when designing systems.I want a person that extends himself, like a ghost or another, or this one that really belongs to me, acts on my behalf, responds to emails, and doesn’t even need to tell me about it.It’s kind of like me, becoming more and more like me.On the other hand, I want a good senior staff member.It might know me very well and I might delegate it.You can access my email and I will tell you about the restrictions.But I think it is a separate entity.I personally prefer the way of being independent entities and think that’s where we are going.
So in that sense, it’s not you, but an ready-to-use, always great, super capable assistant supervisor.
In a way, it’s like an agent, it works on your behalf, knows what you want, predicts what you want, and that’s how I understand what you say.
I think there will be agent behavior, but there is a difference between senior staff and agents.
I want it, and I think the one thing I like about senior staff is that they will refute me.They sometimes don’t do what I asked for, or sometimes they say, I can do that if you want.But if I do, I think the following will happen.And then this, then that.
Are you really sure?
I absolutely want that atmosphere, not just that I give it a task, it does it blindly.It can reason and refute.It is able to reason, and its relationship with me is like the relationship I expect to build when I work with a person who is truly capable, which is different from flatterers.
Indeed, if we had a tool like Jarvis that had reasoning capabilities, it could have an impact on many of the highly valuable product interfaces we use today.Take Instacart, Uber and DoorDash as examples, these services are not intended to be pipelines, but provide a set of APIs to a group of ubiquitous smart agents that represent 8 billion people working around the world.So, what we need to think about is how we should change our understanding of how applications work, the entire experience infrastructure, to fit into this new world where you interact with the world in a proxy way?
I am personally very interested in designing a world where both humans and artificial intelligence can use.I like the interpretability, the fluency of handover, and the ability we can provide feedback.For example, DoorDash can expose some APIs to my future AI assistants to enable it to place orders, etc.I can hold my phone and say, “Okay, AI assistant, please place this order on DoorDash.” I can see the app open and see something clicking, and I can say, “Hey, no, not this.” DesignA world where both humans and AI can use the same, I think it’s an interesting concept.
For the same reason, I’m more interested in humanoid robots than robots of other shapes.This world is designed for humanity and I think we should keep it that way.Sharing the interface is a good idea.
So you will see that voice, chat and other modes are replacing the application.You just have to tell it that you want sushi, it knows what sushi you like, knows what you don’t like, and will do your best to satisfy you.
It’s hard for me to imagine that we’re entering a completely different world where you’re going to say, “Hey, ChatGPT, give me some sushi.” It’s going to answer, “OK, do you want to order from this restaurant? What kind of variety, whatTime, whatever. “I think the user, I think the visual user interface is very good for a lot of things.It’s hard for me to imagine a world where you never look at the screen and use only voice mode.But I can imagine that many things are like this.
Apple has tried Siri.It is said that you can order Uber automatically with Siri.I don’t think anyone has done this, because…why are you going to risk not putting it on your phone?As you said, the quality is not good.But when the quality is good enough, you actually like it more because it is lighter.You don’t have to take out your phone.You don’t have to search your app and press it.Oh, it will automatically log you out.Oh, wait, log in again.This is so painful.
It’s like setting a timer with Siri.I do this every time because it works really well and it’s great.And I don’t need more information.But when ordering Uber, I want to see the prices for several different options.I want to explore how wide the application scope of this technology is.I even want to know where they are on the map, as I might choose to walk somewhere.I think that by looking at the Uber order screen I can get more information in a shorter time, and if I have to get that information through the audio channel, it will take longer.I like the idea you put forward to observe things happening, which is really cool.I think this will bring some changes, and we will use different interfaces for different tasks.I believe this trend will continue.
Among all the developers who build applications and experiences on OpenAI, are there some that impressed you?Do you find this a very interesting direction, even if it’s just a toy app.But have you pointed it out and said it is really important?
This morning, I came across a new company, and in fact, it can’t even count as a company.It’s like two people will work on a summer project trying to eventually become AI mentors.I’ve always been interested in this field.There are a lot of people doing great things on our platform.But if someone can deliver the way you really like, they use one thing I like, that is, it will be like Montessori-level reconstruction, the way people learn things.But if you can find this new way to get people to explore and learn new ways on their own, I personally feel very excited about it.
Devin, a lot of coding related stuff you mentioned before, I think that’s like a really cool vision for the future.I think health care should change considerably as a result.But what I personally excites most is to make scientific discoveries faster and better.While GPT-4 obviously doesn’t play a big role, it may accelerate things by increasing the productivity of scientists.
Sam…that would be a victory.These models are trained and built differently than language models.For some, obviously, there are many similarities.But there are many models with a starting architecture that is applied to these specific sets of problems, these specific applications, such as chemical interaction modeling.You will definitely need some of them.
I think what we generally lack for many of the things we discuss is models that can reason.Once you have the ability to reason, you can connect it to a chemical stimulator or anything else.
Yes, this is an important issue I want to discuss today, namely the concept of model networks.People often talk about proxying as if there is a linear set of calling functions happening.But one thing that appears in biology is a network of systems with cross-interactions, aggregation of systems… The aggregation of networks produces output, rather than one thing calling another thing, that thing calling another thing.Have we seen specialized models or network models appear in this architecture that together solve larger problem sets, using reasoning?There are computational models that can do things like chemistry or arithmetic, and there are other models that can do, rather than a purely generalized model to rule them.
I’m not sure how much reasoning can translate into a form of widespread promotion.I’m skeptical about this, it’s more of an intuition and expectation, and if that’s the case, it would be great.I’m not sure…
Let’s take protein modeling as an example.We have a large amount of training data, protein images, and sequence data. Based on this data, we build a prediction model and we have a set of processes and steps to achieve this.
Have you ever imagined that there is a general AI or great inference model that can figure out how to build a sub-model and solve the problem by getting the necessary data and then solve the problem…
There are many ways to achieve this.Maybe it will train a text model, or maybe it will just know a big model.It can select other training data you want, ask questions, and then update.
I guess the real question is, will all these startups go bankrupt?Because many startups are working in this pattern, which is to get special data and then train new models from scratch with this special data.Then it does only that thing.And it works very well on that thing.It’s better than anything else.
I think you can already see a version.
When you talk about biology and these complex networks of systems, I understand it because I’ve been very ill lately.Now I’m much better.But it’s like the body is systematically defeated one at a time.Like you can really see, well, this is a cascading thing.It reminds me of the biology you are talking about, like you don’t know how big the interaction between these systems is until things start to go wrong.This is a little fun.
But I was using ChatGPT to try to figure out what was going on, and anyway, I would say, I’m not sure about it.Then I published a paper and didn’t even read it, just like in the context.It said, Oh, that’s something I’m not sure about, and now I think so instead.So it’s like a small version you’re talking about, you can say, I don’t know about this stuff, you can add more information, you don’t need to retrain the modeler to add to the context here.
So these models that predict protein structure, for example, yes, are the whole basis.Now, there are other molecules on AlphaFold3.Can they do it?Yes, is this basically a world where the best generalized model enters and gets training data and then solves the problem on its own?
Maybe you can give an example, can you tell us about Sora?Your video model can generate amazing dynamic images, dynamic videos, and how different the architecture there is, no matter what you are willing to share, how to make it unique.
Yes, so I’ll start with the general question, you obviously need specialized emulators, connectors, data snippets, etc., but my intuition.Again, I have no scientific basis, my intuition is that if we can find out the core of general reasoning and connect it to new problem areas, just as humans are general reasoners, I think it is possible to unlock faster, I think so.But yes, you see it doesn’t start with the language model.This is a model customized for videos.However, we obviously haven’t entered that world yet.
For example, in order to build an excellent video model, you built it from scratch, I guess you used some different architectures and different data.But in the future general reasoning system, that is, AGI, no matter what system, it can theoretically achieve the goal by understanding how to achieve this.
Yes, one example is that, as far as I know, all the best text models in the world are still autoregressive models.The best image and video models are the diffusion models.This is somewhat strange to some extent.
Therefore, there is a lot of controversy about the training data.I think you are the most considerate of all companies, and you have now signed a license agreement with FT and others.We have to be a little cautious here because you are involved in the New York Times lawsuit and I don’t think you can reach an agreement with them on training data.
How do you view the fairness of reasonable use?We had a heated debate on the podcast.Obviously, your actions fully demonstrate that you are trying to achieve fairness through a license agreement.So, what position do you personally take on the rights of artists who create beautiful music, lyrics, and books?Do you use these rights to make derivative products and then monetize them?What is fairness?How can we get artists to create content and then decide what they want others to do with it?
I’m just curious about your personal beliefs because I know you’re a thoughtful person in this regard.I know many others in our industry are not very thoughtful about content creators.So I think different types of people will be very different.
In terms of fair use, I think we have a very reasonable position under the current law, but I think AI is so different.But for things like art, we need to think differently.
But, suppose you read a bunch of math online and learned how to do math, I think that seems uncompromising to most people.Then, there is another group of people who might have different opinions.In fact, in order not to be too long, I will not discuss this issue in depth.
So I think there is a type of people who would say, well, there is broad human knowledge, you can say, if you learn, that is, that is, open fields and the like, if you go to learn Pythagoras theorem.This is one end of the spectrum.I think the other extreme of the spectrum is art, maybe even a little more specific, I would say it is like doing it, it is a system that produces art in the style or similarity of another artist, which is probably the most extreme.Then there are many situations in the middle of the spectrum.
I think the focus of discussion has historically been on training data, but as the value of training data decreases, discussion will increasingly turn to what happens when reasoning.And what the system does, access information in the context in real time, or taking similar measures, what happens when reasoning, and how new economic models will be more debated.
So if you say, for example, if you say, for example, write a song for me in the style of Taylor Swift, even if the model has never been trained with any Taylor Swift songs, you still have itThe question, i.e. the model may have read articles about Taylor Swift, probably knowing her topic, what Taylor Swift means.The next question is, even if this model has never been trained with Taylor Swift songs, should we allow it to do so?If so, how should Taylor Swift get paid accordingly?
I think first, there should be an option in this case, i.e., an option to join or exit, and then there should be an economic model.Take music as an example, from a historical perspective, here are some interesting things to see, that is, sampling and how the economics around sampling work.It’s the exact same thing, but it’s an interesting starting point.
Sam, let me challenge this.
What is the difference in the example you gave?The model learns the structure, rhythm, melody, harmony, relationship of the song, discovers all the underlying structures that make the music successful, and then uses training data to build new music.And humans listen to a lot of music, their brains are processing and building all the same predictive models, and these are the same discoveries or understandings.What’s the difference here?Why you say maybe the artist should be rewarded uniquely, it’s not a sampling case, you’re not AI without output, it doesn’t store actual original songs in the model.
Yes, learn structure.
So I’m not trying to emphasize that, because I agree, just as humans are inspired by other humans.I mean, if you say to write a Taylor Swift-style song for me.
I see.I think that’s where the tips are used to take advantage of some artists.
I personally think this is a different situation.
Would you like to let the model train itself, using the entire music corpus created by humans to train the music model without paying royalties to the artists who entered the music?
Then you can’t ask the artist for specific tips, you can say, hey, give me a cool pop song, this song is pretty modern, telling a heartbreaking story, with a woman’s voice
We have decided not to make music at this time.Part of the reason is where you draw the line.
I recently met with a few musicians I admire so much and I just wanted to talk about these extremes, but even in this world, if we go, let’s say we pay 10,000 musicians to create a bunch of music just for the sake of makingA great training set, music models learn everything about powerful song structures.What makes good, catchy beats and everything else.And just train this, assuming we can still make great music models, maybe we can.I kind of feel like I’m posing it before a musician as a thought experiment.And they well, at that time I couldn’t object to this in any principle.However, I still don’t like it.Now, that’s not necessarily a reason not to do so.But this is.
Have you seen the ads posted by Apple, maybe yesterday, or something like that, compressing all human creativity into a very thin iPad?
What do you think about this?
People are very excited about this.
Yes.The reaction is stronger than you think.
I’m obviously very optimistic about artificial intelligence, but I think there are some beauty in human creativity and human artistic expression.This is great for an AI that can do better in scientific research, come on.However, if AI can reach the level of deep and beautiful creative expression of human beings, I think we need to be clear about how this will happen.It will become a tool that leads us to higher creative heights.But I think we need to find ways to do this to preserve the spirit that we all care about.
I think your actions are very convincing.We are trying to create the characters in Star Wars in DALL-E.If you ask Darth Vader, it will say, hey, we can’t do that.So, I think, you’ve formed a red team internally or you can call it as you like.We’ve tried it.Yes, you don’t allow people to use other people’s intellectual property.So you’ve made this decision.
Now, if you ask it to make a Jedi Bulldog or a Sith Lord Bulldog (which I did), it will turn my Bulldog into a Sith Bulldog.
There is an interesting question about what you said, yes.We released something called norms yesterday and we tried to illustrate how our model should behave.This is a difficult task.This is a very long document.It is very difficult to specify exactly where the limit should be in each case.I think this is a discussion that requires more input.
However, the idea of the Sith Lord, Sith-style thing or Jedi is like a part of the culture at this point, and these are hard decisions.
Yes, I think you are right.The music industry will consider this opportunity to make Taylor Swift’s songs theirs.This is part of the four-part fair use tests that can leverage new innovations in existing art.Disney has an argument, hey, if you’re going to make a Sora version of Ashoka, or Obi-Wan Kenobi, that’s Disney’s chance.This is a great partnership for you.
So, I think I’m marking this part as AI and law.
So, let me ask a higher level question.What does it mean when people say regulating artificial intelligence?You can also comment on the newly proposed regulations in California if you prefer.
I’m very worried about this.There are a lot of proposed regulations, but most of the regulations I see about California are worried me.I also worry that states will do it themselves.
When people say they want to regulate AI, I don’t think they mean a specific thing.I think some people will forbid the whole thing.Some people will say that it is not allowed to be open source and require it to be open source.What I am most interested in is that I think someone will say, look, I might be wrong.I acknowledge that this is a forward-looking statement.And making these statements is always dangerous.But I think in the near future, there will be a day when we are not talking about decades later. Artificial intelligence says cutting-edge artificial intelligence systems can cause major damage to the world.For these systems, likewise, we conduct global oversight of things like nuclear weapons or synthetic organisms that can have very negative effects that may go far beyond a country’s scope.I look forward to seeing an international institution researching the most powerful systems and ensuring reasonable safety testing is carried out.These systems will not automatically escape, nor will they improve themselves, etc.
However, the criticism of this is that you have the resources to please, lobby, participate, you have close ties with politicians, and for startups, despite their enthusiasm and investment, they don’t have enough resourcesTo deal with this issue.This phenomenon of regulatory capture, as our friend Bill Gurley pointed out in his speech last year.
So maybe you can solve this problem head-on.I would be happy with that if we only focus on models trained on computers worth more than $10 billion or more than $100 billion or anything else.This is a line that I can accept.I don’t think this will put any regulatory burden on startups.
So if you have the nuclear raw material to make a nuclear bomb, just like a small number of people have it, you can use an analogy like a nuclear inspector.I think this is an interesting point.
Can I say one more thing?certainly.I’m very nervous about the over-regulation here.I think we can go wrong if we do too much, or even a little overdoing it.I think we don’t do enough and we may go wrong.But I do think part of the reason is… Now, we’ve seen overreach of regulation or control get really bad in other areas.
And, maybe nothing will happen.But I think it’s part of our duty and mission to talk about what we think is possible and what needs to be done to achieve that.
Sam, the challenge is that we have laws and regulations to protect the people and the entire society.
However, the regulations we are developing give governments the right to enter and audit trade secrets.We have never seen regulations of this level before.Basically, legislation proposed by California and some federal legislation basically require government audits of the parameters and weights of a model, audit software, audit and review models.Then you need their tick to deploy it for commercial or public use.
To me, it feels like we are trying to control government agencies to create fear.Because people are hard to understand this, afraid of its effects, they want to control it.And the only way to control it is to give me the audit rights before I publish it.
Yes, they know nothing.These people know nothing.The way these things are written, you read them, you will go crazy.Because, as you know, in 12 months, these things will not make sense anyway.I’ve pushed for an agency-based approach to solve the big picture instead of writing it out like the law,
I don’t think even if these people are real world experts, I don’t think they can handle it correctly in 12 or 24 months.I don’t think so.We will review these policies, conduct a comprehensive review of all your source code, and view all your weights one by one.I do think there are many crazy proposals, especially when the model is constantly retrained and becomes more dynamic.That’s why I think it makes sense.
Before the aircraft is certified, we will conduct a series of safety tests to get the aircraft through these tests.This is completely different from the process of reading all the code.We are reviewing the output of the model, not looking at the inside of the model.
So, I think it makes sense to do security testing.
So, Sam, how should we achieve this?
I am here not only representing OpenAI, but also the entire industry and even humans.I worry that if we limit these technologies that have great potential for development and great benefits for humanity, we may fall into the dark age.So, how should we change this and achieve this?
Because the government level is developing too fast, people seem to have gotten the wrong direction, and I am worried about it.On this basis, Sam, for example, Lama’s architectural decisions are very interesting because we will let Llama grow with as few constraints as possible.We have another thing, we call it the Llama Guard, which is designed to be these protective guardrails.
What do you think is the correct way to solve the problem?
Judging from the strength of the current model, there will definitely be some problems.I don’t want to underestimate these or take them seriously.But I am not worried about any catastrophic risk that the GPT-4-level model will bring.I think there are many safe ways to choose to deploy it.
If we say that, maybe we will find more common ground.I like this concrete example, that models are capable, technically, and even if they are not used this way, they can be recursively self-improvement, or self-design and deploy biological weapons, or something like that.Or a new model.That’s the key point of recursive self-improvement.
We should conduct security testing of the output of the models at the international level, because these models have reasonable opportunities to pose a threat there.I don’t think anything like GPT-4, of course, does not pose any threat, well, I’m not saying any threats, because we don’t think, yes, I don’t think GPT-4 is a substantial threat to these things.I think there are many safe ways to publish such models.
But when major casualties are a serious possibility, like an airplane or any other example, I think we’re happy to have some kind of testing framework, like when I get on the plane, I won’t think about the plane, I just assume it willIt is safe.
Sam, there are a lot of concerns about employment right now.You did a lot of tests about UBI when you were in YC.Your results will come out soon.I’m just saying, this is a study that ended or started five years ago.First there is beta research, then long-term research.However, what have you learned from all this?
Indeed, why did you start all this?Can you explain why you started looking at Universal Basic Income (UBI)?
We started thinking about this in 2016, and around the same time, we started taking AI seriously.In theory, the extent of change in employment in society and economy, and the potential for a deeper meaning (e.g. how the social contract will change), means that we should do a lot of research to explore how these new changes are arranged.Various ideas.
I must admit that I am not an avid supporter of the government’s handling of most policies aimed at helping the poor.I tend to believe that if you can give people money, they will make the right decisions and the market will work.I am very supportive of raising minimum living standards, reducing poverty, and even eliminating poverty.But I’m interested in whether there are better ways than what we’ve tried for the existing social safety net, and how to handle things.
I don’t think giving people money will solve all problems, and of course it won’t make people happy.But it may solve some problems and give people a better perspective to help themselves.I’m very interested in this.
I think now we see some approaches, so 2016 was a long time ago.Now that we see some ways AI is developing, I wonder if there is anything better to do than the traditional UBI concept.For example, I wonder if the future will be more like general basic computing than general basic income.Everyone gets a portion of the GPT-7 calculations that they can use, they can resell, they can donate to someone for cancer research.But what you get is not the dollar, but a part of productivity.Yes, you have some productivity.
OK, let’s talk about gossip.So, let’s go back to November last year.What exactly happened?
If you have specific questions, I’d be happy to answer, maybe I said we’ll talk about it at some point.
So that’s the point.what happened?You are fired and you come back, it’s a court plot.Is someone stabbed you in the back?Have you found the ultimate goal of artificial intelligence (AGI)?What happened?This is a safe space.
I was fired.I said I was back and I was a little unsure what I wanted to do because I was very upset.I realized I really like OpenAI and the people there, I will be back and I know it will be hard.This is harder than I thought.But I thought at the time, OK, no problem.I agree to come back.It took a while for the board to figure it out.Then, we were kind of trying to get the team together and keep doing things for our clients.Then, other plans begin to be developed, and the board decides to hire another interim CEO.Oh my God, what’s that guy’s name?His presence is like scaring people, which is great.I have only good things to say.
So, where did you learn about your firing?
I learned about it while spending the weekend in a hotel room in Las Vegas.I got a text message and they said, fire and pick it up.I think you’ve had this situation before I tried to think about whether I was fired.I don’t think I’m tired of this.Yes, I received a text message.No, it’s just a strange thing.
So, who sent this text message?
Actually, I got this text the night before.Then I spoke to the board of directors.And then that’s it.Then I kind of and then everything goes crazy.My phone was not available at the time.It just keeps shaking, such as text messages, phone calls.
Basically, you were fired from a tweet.This has happened several times during the Trump administration.
They did call me before tweeting, which was great for them.Then, I spent several hours in my hotel room in a totally hilarious state.I was very confused at the time and didn’t know what to do.Too strange.
Then I flew home, probably on the plane, about 3 p.m.Still like, you know, the crazy phone keeps ringing.I met some people in person.That night, I thought, OK, I’ll do AGI research and I’m very happy about the future.Yes, I have a choice.
Then the next morning, I had a phone call with several board members to discuss the issues coming back, which led to a few more days of chaos.Then, I felt the matter was solved.Well, there are a lot of crazy things in the middle.
How many of these nonprofit board members are because of these nonprofit board members?
Well, we only have one nonprofit board.So they are all non-profit board members.The board of directors has been reduced to six.Then they removed Greg from the board and fired me.
So, is there a cultural conflict between people with only nonprofit experience and those with entrepreneurial experience on the board?If you want, maybe you can share the motivation behind the action, anything you can share.
I think cultural conflicts have always existed.Obviously, not all board members are my favorites, but I have a lot of respect for their seriousness towards AGI and the importance of dealing with AI security correctly.Even if I totally disagree with their decisions and actions, I never doubted their integrity or commitment to a shared mission for a safe and beneficial AGI.
I think they made the right decision in the process or knew how to balance everything OpenAI has to do right?No.But I think they have intentions, have AGI scale intentions, and do it well.
Actually, let me ask you this question.So OpenAI’s mission is clearly to create AGI, and I think it’s really fun.Many people think that if we create artificial intelligence (AGI), it would be the unexpected result of something going wrong.They were extremely scared of it.However, OpenAI sees it as a practical mission.Will this make you feel more fearful about what you do?I understand that this can generate motivation, but how do you reconcile this?I wonder why this is a mission?
I think, first of all I want to answer is that this really triggers a lot of fear.I think many people in the world are very scared of AGI, even current AI, and are also very excited about its direction, even more scared and excited about.We struggle with it.But I think this is inevitable.I also think it will bring huge benefits.But we have to find a way to get there in a reasonable way.Many things change, and changes make people feel uncomfortable.So we have a lot to do.
Can I ask a different question?You have created one of the hottest companies.And you are indeed in the center.But from an economic perspective, you avoid all these values, which is very unique.Can you tell us?
Yes, I wish I had the equity at that time so I didn’t have to answer this question.If I could go back to the past…
Why don’t they give you a grant now?Why doesn’t the board of directors give you a large option grant you deserve?What was the decision at that time?Why is that so important?
The initial reason for the decision at that time was just the structure of our nonprofit organization.From a motivational point of view, this is great.But mainly our board needs most selfless directors.I think that’s good.I don’t need equity now.I’m kind of…but, I hope…
But in this weird way, now you run a company, yes, it creates these weird questions like, what is your real motivation…
Totally correct.
It’s so profound… One thing I noticed is that it’s hard for people to imagine that they’re going to say “I don’t need more money”.I also know how insensitive the tone is…I think people think it’s a little ulterior motive.Yes.So assuming…what is he doing to make money?If I just said, I’m going to make a trillion dollars with OpenAI, I think everyone will be more relaxed and that won’t save me.This will save a lot of conspiracy theories.
Sam, you are a great dealmaker.I’ve followed your entire career.You are very good at this.You have all these relationships.You are really good at raising funds.You’re great in this regard.You have Johnny Ivy, you invest in the company, you get the ball, you raise $7 trillion to build the fab, all of that stuff.All these factors combined together,
I’m just a joke here.Obviously, you didn’t raise $7 trillion, which might just be market capitalization or something.However, putting that aside, the key is that you have made all these transactions.Their distrust of you is because of what is your motivation?You finally left, what are the opportunities inside OpenAI?What opportunities should belong to Sam, and those of these nonprofits don’t trust you?is that so?
If we were equipment companies, or we were working on some chip manufacturing companies, those weren’t Sam’s projects.Just like OpenAI will get that equity.They will get it.
However, this is not the public opinion.It’s not the opinion of people who are commenting on these things all day long.It’s fair because we haven’t announced these things yet, because it’s not done yet.I don’t think most people in the world are thinking about this.But I agree, this has sparked a lot of conspiracy theories among technical commentators.
If I could go back in time, I would say, let me make this clear in fairness.Then, I still do it because I really care about AGI and I think it’s the most interesting job in the world.But at least, it will warn everyone.
So, what is a chip project?That’s $7 trillion.Where did that $7 trillion come from?
I don’t know where that came from.Actually, I really don’t know.I think the world needs more artificial intelligence infrastructure, far exceeding the number of currently planned construction, and the cost structure is also different.The specific way we were playing there was, we were still trying to figure it out.
What do you like about OpenAI’s organizational model?Is it like a quick action, breaking the norm, highly fragmented small team?Or is it more of this organized effort that you need to plan because you want to prevent some extreme situations?
This is not because we need to be more organized to prevent extreme situations.But because these systems are very complex, concentrated bets are very important.For example, before doing this, you have companies like DeepMind, all of these different teams doing these different things.They spread the bets.You let OpenAI say we basically have the whole company working hard to make the GPT-4 together.This is unimaginable for how to run an AI research lab.But I think, at least it works, it works for us.So it’s not because we want to avoid extreme situations, but because we want to concentrate resources and do these big, difficult and complicated things.There is indeed a lot of coordination in our work.
OK Sam, I understand you need to leave.You’ve done a great job in the past hour.Welcome to come back anytime, I am very happy to talk to you.Yes, it was a very interesting conversation.Thank you for coming,
Sam, thank you for your honesty, too.We’ve been discussing this issue for more than a year.This is amazing and I am very grateful for it.
I’m looking forward to coming back after the next big press conference so I can talk about some of these things more directly.
Yes, you’ve received the Zoom link.The same Zoom link is used every week, and at the same time, you can join at any time (laughs).
