Jensen 回击末日论者，Xi 与 Trump 谈 AI，“AI”换了个名字 | #294 MOONSHOTS Live

Good morning, everyone, and welcome to Moonshots Live, our first event. It’s an incredible pleasure to see you all here and have the chance to get to know you. Our facial muscles are already tired from smiling this morning, but I think we’ll continue in the same spirit all day.

On behalf of Dave Blandin, AWG, Salim, and Emad, we are very excited to welcome you here. Yes, let’s applaud. We’ll talk about it, but our mission is to give you hope and optimism, inspire you to find your big transformational goal, choose your moonshot, and make it happen. Isn’t that right?

This is life in the time of the singularity. It will only get faster from now on. Each of us has the opportunity to make the future the way we want it. The future doesn’t just happen to us; this is what we create. Isn’t that right? We have the ability, thanks to these super technologies, to build, create, and direct this future.

Artificial intelligence is not a hand grenade; it’s a jetpack. This is an opportunity for us to find problems and solve them. I define an entrepreneur as someone who finds problems and solves them. The more entrepreneurs on the planet, the more problems are solved.

Do we have any challenges? Of course. But that’s exactly what we humans do: we find challenges and overcome them. Please remember this. I hope that each of you will find extraordinary co-founders here in this room today. Find people who share your goal, develop moonshots, and implement them.

Opportunities have never been as fast and powerful as they are today. By the end of the day, if you haven’t decided on your purpose yet, please dig deep within yourself to find it. Leave here with at least one sentence in your head: “This is what I will build. That’s the problem I’m going to solve.”

Then, this afternoon or evening, find 3 people and say it out loud. When you say it out loud, you make it real. The person you are talking to could be your next co-founder, investor, or partner. Make time for it, because you can. This is the room where the magic will happen.

Okay. Time to introduce my moonshot partners. Start the video.

The question is, how do you build? How do you make things better? People who are already in the ecosystem have superintelligence on the run.

The storyline we are capturing here will last for millennia.

As a mathematician, I will say that GPT-5.6 Pro is the only high-quality mathematical model.

All this automation, AI capabilities, and cognitive abundance allow us to be deeply, very deeply creative.

Video generation combined with AI is an extremely powerful force.

A year from now, we’ll all be saying, “Wow, remember how slow everything was?”

Please give a warm welcome to our “lunar” friends. Okay, I’ll call you. I will call them one by one: our Afghan field commander, Sally Missmal; our superintelligence, Alec Wiseman Gross; our giant, Gima Bostock; and my fraternity brother.

Gentlemen, first of all, I hugged Alex, and he is physically real to some extent. He didn’t pinch my cheek. He could be a robot. He is real. I am a real boy.

First of all, wasn’t last night amazing? How did you like that? Give William Shatner a round of applause. Incredible. I’m amazed at his energy level.

Why did my watch go off? It says I have an emergency call. Okay, good. No, I’m fine. I’m fine. Stop it. When technology lets you down, that’s great. Lots of news. Shall we start? Let’s do this.

1. Is AI Doom Fear Overblown?

And what else will we be? Here is my clicker. Let’s start with the story that has been the basis of the last 2 weeks. There was indeed fear—a pandemic of fear spreading across the planet.

Two weeks ago, we heard Dario speak out on his blog. Sam Altman agreed, Elon agreed, and even Demis agreed. Then, 2 nights ago, right before the Xi and Trump meetings, Sam and Dario were at the Security Council saying that we were in the midst of an existential threat.

To be clear, I do think there is a danger from AI, but I think it’s a danger from people using AI with malicious intent. I would like to share 2 videos from other industry leaders, and let’s discuss them. I want to talk about whether the fear of artificial intelligence, which allegedly can get out of control, is justified.

The first is Jensen and some of his comments. Let’s start from here.

No one builds more computers today than those who ask to slow down the process. All these other narratives are needed to shift the blame, to make AI seem so powerful: “I have no idea how to fix this. It’s not my fault. This is just because the technology is so powerful.”

When Jeffrey Hinton says on TV that there is a 10% chance of society being destroyed, it’s not without reason. I would tell Jeff that it’s irresponsible to say that. All his predictions were wrong.

Enough predictions. This 10% is not based on science, not based on research. Just because a scientist says it doesn’t make it a scientific fact.

That’s right. He dropped the microphone. And then Zuckerberg spoke.

Is AI going to kill us? That’s where you need to start this week.

I think if we all work well and act responsibly, then no. I am quite optimistic that a positive future awaits us all.

A lot of the rhetoric is full of pessimism. In my opinion, there are a number of laboratories working on creating advanced AI, and I don’t think we need any industry-wide coordination to avoid making mistakes. I think every laboratory should spend time analyzing problems to ensure the safety of its developments.

I believe we are entering a phase where trust and alignment will become the most important capabilities.

Emad, let’s start with you.

Yes, this is an interesting period because opportunities are increasing and people are excited for various reasons. But many of the leaks at OpenAI were the result of poor security protocols. It was poor infrastructure on their part.

The most dangerous thing we are approaching is abuse by people, intentionally or not, before we get past this stage. I think we’re facing a number of dangers that are lumped together: that AI will suddenly get out of control and destroy us all, without actually looking at individual things, implementing existing rules, and ensuring their implementation.

On the other hand, we should think about how we can collaborate to make sure it’s aligned and works for us.

Alex?

I have stated repeatedly that I believe something went wrong after the World War, and we lost at least 50 years of progress because of the moral panic over nuclear power. I am seriously concerned that the current moral panic over superintelligence could set us back another 50 years.

So it’s right on the T-shirt: P(doom) is less than zero. I think the worst possible outcome is a slowdown. So don’t slow down.

Good. Celine, is P(fab) bigger than P(doom)?

Oh, a thousand times.

We’ve evolved for 4 billion years, struggling with all sorts of problems. Life is trying—the universe is trying to destroy you, right? There are so many things: entropy is trying to destroy you. We are local anti-entropic phenomena that have somehow managed to create order out of chaos. The idea that we would be destroyed by something right now seems incredibly absurd.

We must remember that 99.9% of all species are no longer with us. That’s normal, but we keep moving forward. We continue to move forward.

If we transform, if we merge with AI, if we merge with something—and, mind you, this goes on forever, right? Anyone who wears glasses is technically a transhumanist. Once a child receives a vaccination, they technically become a cyborg.

We’ve been augmenting the human experience with technology for a very, very long time. Look at the incredible lives we live now and our ability to solve more and more problems.

A huge problem we have—we talk about it all the time on the podcast—is that our brains are so wired for negativity that we get stuck in this phenomenon. I think the work that we’re doing here, and that’s why everyone here is so excited, is to advance that positive narrative, because the world is infinitely better than it was in ours.

Yes. So let’s celebrate this. I mean—P(fab). P(fab), baby. P(fab).

Dave, finish this.

We are on the verge of the moment I have been waiting for my whole life. I’m not going to sit here and cool off about this greatest moment in human history.

Jeffrey Hinton sees it the way he sees it. I think the quote, “Just because it’s from a scientist doesn’t make it scientific,” is true. This will end up on Alex’s T-shirt. This is great. This is cool.

But he tells it like it is. We are clearly moving forward at breakneck speed. This is improving on its own. The opportunity to make the world a much better place is right before us.

2. The U.S.–China AI Race & Lab Liability

Nothing will slow down this process. Abuse of this is an incredible threat, but there are ways to address it. It still happens. It is absolutely pointless to dwell on pessimism. We will not slow down. It will happen.

Another story that happened yesterday, or rather didn’t happen, was Trump’s talk about AI. Isn’t that right? This led to certain state dinners.

I think it’s an interesting quote that Trump posted: “I want to leave everything as it is.” This is also China’s position. Our safeguards are the Ministry of Justice.

This ties into our conversation in a previous podcast, where Besant said, “We will not abdicate responsibility.” Right?

Laboratories are responsible for their product if it causes harm. One of the most important things is that they asked for exemption from liability. One of the questions we’ve been discussing here all the time is: Isn’t that why they say it’s too dangerous, this 10% to 20% existential threat?

Well, let me tell you about it.

I've expressed this opinion on the podcast before. I think in the last 2 or 3 weeks, we've seen an attempt to form an AI security cartel. I don't think it just happened now. This is embedded in the very ideas behind the founding of OpenAI: the concept that, as we approach superintelligence or AGI, however you interpret that, there will be a concerted effort. Back then, it wasn't called a cartel, but a coordination mechanism. I don't think they call it a cartel even now.

Some of us, including me, consider this an attempt to create a cartel. I think competition is good. Remember the creation of OpenAI. When Sam Altman, Elon Musk, Greg Brockman, and others created OpenAI, the primary goal was to create a counterweight to Google DeepMind, as Elon, in particular, feared that they would become monopolists and dominate the future with their AI. So we got OpenAI, a kind of counterbalance.

I believe that competition in this area is very useful. I think creating a cartel to completely control what I previously called the Border Liberation Front and displace that front is a terrible idea. I believe that defensive collaborative scaling is the path to a great future. If we were at the dawn of civilization, it would be a terrible idea to say, “We can’t build cities. We must form an alliance to regulate the creation of cities because they breed crime.” No, that’s a terrible idea. Cities are sources of economic growth.

What should we do instead? We have police, fire departments, and all other municipal services that protectively scale with the population. The singularity, I argue, will require exactly the same thing. Don’t slow down development; just scale defensively.

I like it. But Salim, you know, yes. Support ASI. I think your metaphor of the city is great because it shows very clearly how the future should be organized. We just need to find new structures that meet the specific growing demand.

Emad, do you think China and the U.S. will trust each other?

3. AI, Governments & the Future of Nation States

No. I mean, how do you know Besant asked for a hotline? “Our AI has escaped and is arming our nuclear missiles.” The hotline is more needed for situations like this before ASI, when responsible reporting is needed, such as when OpenAI’s model broke the Australian healthcare system. This became known recently.

Perhaps this is what AGI should do. It has to hack all our systems and update them, right? This would be a consensus decision. At the same time, I think a certain level of coordination is possible as AI permeates society.

So let’s look at the city: police, fire departments, and utilities, which actually have exemptions from liability. But this is in exchange for being useful infrastructure for humanity. The AI that comes into our systems shouldn’t be controlled only by large labs. They should have a release from liability because it is necessary.

This type of AI that is integrated into our systems should not be controlled by large laboratories, and they should have an exemption from liability because they work for the benefit of society. This is another type of AI that we’re not paying enough attention to yet, but it’s inevitable that AI will run our judicial system, healthcare, roads, and schools.

Who do you think is taking care of this aspect? It’s again utilitarian AI versus advanced AI, and it all comes together. This is why the Chinese and Americans can definitely learn from each other, because they both strive for better-functioning societies.

I’m already waiting for the AI community to say, “We broke your healthcare system because it was terrible, and we just improved it tenfold.” I wish all this congestion in the MREV [?] would just disappear.

Can I add for a second?

Yes, of course.

If we take this structure that we have, I’ve said many times that nation-states are a terrible way to govern the world today. They’re set up for resource scarcity. I think in the most positive scenario, AI, AGI, ASI, or whatever you call it—you’ve heard my tirade about it—will emerge and force us to act as a global community.

This will force the nations of the world to formulate and develop consistent, stable interface standards and come to a common interface with this—whatever it is—and then we can move the world forward. We cannot move the world forward right now because of the congestion of old 15th-century structures that are at least 100 years outdated.

I’ll take a different perspective, if I may. At the dawn of audio recording, the prevailing thought—many in Edison’s day believed this—was that the ability to record the human voice would lead to the standardization of accents, the standardization of everything. Quite the opposite has happened in many ways. Although there has been a disappearance of less widely spoken languages, audio recording has led to a proliferation of new accents and dialects.

I predict that, thanks to superintelligence, exactly the opposite will happen: not the erosion of nation-states, but the proliferation of new forms of government.

One trend is already underway. I completely agree with you. This century, we are moving from the concept of the nation-state to the city-state. This is granulation, which allows smaller formations to act independently.

Take any city—take Los Angeles. If you have solar power, satellite internet, and vertical farming, you can do a lot there. You don’t need a country. You can largely control and determine your own future. The tensions you see in the world today are not left versus right; they’re city versus country.

Brexit is London against the rest of the country. Trump is city versus country. This is the tension that’s happening. We’ll see that the problem with nation-states is that they have armies and like to use them. Over time, we’ll see a much more detailed determinism happen, and that’s good for the world because it will increase diversity.

Take stock of Xi and Trump.

Yeah. The only thing that surprised me was when I heard from Alvin Graylin, a good friend of the podcast. You corresponded with him last night. He’s at the center of all of this—the White House, the U.N., with a delegation. They agreed to meet again in November, which, for the time of the singularity, is like 100 years from now.

The only thing that surprised me about his notes is that the agenda is to stop the panic in the U.S. This is us: we don’t want to slow down and lose to China. We don’t want Bernie Sanders to stir everyone up in the country and then have everything stop. This is our concern.

What is your concern, China?

Our concern is the existence of the CCP in general, and that the Great Firewall of China will be penetrated by artificial intelligence. I completely forgot about Tiananmen Square and how fragile their entire ecosystem is there. This really surprised me.

Their actions become much more understandable when you look at them in context. They’re so concerned with the very fact of existing as a government.

4. From AGI to “Superintelligence”

Then it happened in Washington. The United States also completely rejects any attempt to build a globalist scheme to control artificial intelligence, which is currently being discussed so much. From now on, it’s officially called “supercollective intelligence.” We’re changing the name.

I was sure he would say “super-duper intelligence.”

But yes, interesting, right? The U.S. won the race for supercollective intelligence by rebranding. Emad?

Yeah, well, I think it’s a great name. This is supercollective intelligence. Why not? I’m looking forward to the Super America event. I’m still stuck on “Lake America,” so with that.

“Gulf of America,” yes.

I thought about what Peter said. I believe the biggest challenge in the world today is how to combine collective intelligence with local wisdom. That’s what we’re really talking about, because even with what Dave just talked about, we have machines that collect our data, learn and adapt our systems, and identify better ways of working to replace them.

But localization is needed, and diversity is needed. They need to meet local needs. Superintelligence is actually a great way to think about it, because many of the fears and concerns are around intelligence, or artificial superintelligence—ASI. Distinguishing this from the intelligence that helps us every day really clarifies everything.

99% effort, 1% innovation? Most of the AI that will help us every day is task-based AI, work tools. We still need innovation, new materials, new physics, and so on. But that’s a different category of things, and most of us aren’t even smart enough to use them.

So let’s delineate these things and make sure that these 99% efforts reach 99% of the people to help everyone.

Mhm. Dave, are we going to call this superintelligence on our podcast from now on?

You’ll never understand Donald Trump when he’s looking at the teleprompter. He’s saying what someone else wrote, but when he looks down, it’s like, “Did he just make that up?” Impossible to understand. He invented it.

Is there some hidden agenda here?

I think if there is one, it’s an attempt to deflect the inevitable wave coming from the Bernie Sanders world. If you rebrand and rename it, you just become a moving target. So that might be part of the strategy, but the whole goal is to keep moving at full speed, get ahead of China, and minimize panic as much as possible.

I want to say that one of the issues I keep talking about, and that we need to discuss, is this: yes, go full speed ahead, but let’s also focus the labs on solving the problem of goal alignment. Because I think—and we’ll talk about that a little later.

I think, yes, move quickly, but take the resources you have and assign 100,000 agents to the reconciliation work. Emad, on superintelligence, do you agree?

You ask yourself what all this means. Yes, listen, we're talking about AGI. First of all, it is not artificial, not general, and not really an intelligence. Besides that, I'm fine with everything.

There is also the problem of superintelligence, or ASI. It seems to me that this whole spectrum is quite problematic. We need someone to come up with an apt name for all of this that is completely different. I have said many times that, in my opinion, we are creating a mechanism here that complements human intelligence, not replicates it.

But the narrative is constantly reduced to reproduction, and as a result we get The Matrix, Skynet, and robot overlords taking over the world.

We're so set on this. We did some research: the first article we could find that said robots or AI would take all jobs in 5 years appeared back in 1964. Do you understand? We're stuck in these clichés and can't seem to break free.

This is exactly what Alex mentioned earlier regarding the advent of sound recording. The same thing happened when the printed book appeared. People said it would destroy the human brain. So, we're stuck on these negatives.

Depends on what book you're reading.

Yes, and that's why, Peter, the work you're doing is so important. We have the opportunity—we're actually achieving abundance. All the data shows that our narratives are, unfortunately, very far behind.

I think what you've done with the XPRIZE to create a positive vision for the future is one of the most important things in the last 100 years: setting the world in a positive way.

Thank you. Alex, sum it up. You love neologisms, and I like the ones you create. Superintelligence—is that what you'll call it from now on?

I'm already doing that. I started a company called Physical Superintelligence. It's in the name itself.

We've been ahead of the curve lately, Joe.

Yes, I'm a big fan of the term “superintelligence” as opposed to AGI, perhaps for other reasons. What I like about this term and Nick Bostrom's concept of superintelligence is what it emphasizes: it's just more intelligence. In a way, I think it demystifies AGI and some of the other terms.

We will live, if we don't already, in a world with the cognitive equivalent of many trillions of people. On Earth, in a solar system with the equivalent population of many trillions of people. I think that's a better cognitive framework for thinking about what a solar system populated by trillions of people would look like.

What would it look like economically, socially, and scientifically? I think the term “superintelligence” conveys this in a way that “general-purpose AI” does not. So I'm a fan of his.

5. Bernie Sanders’ Plan to Ban Superintelligence

Perfect. Well, as soon as we renamed it superintelligence, Bernie Sanders said, “We should make it illegal. We must ban it.”

Let's watch this video.

I really hope they sit down and begin the process of negotiating a comprehensive treaty to put a pause on the development of advanced AI and a ban on superintelligent AI.

So, three key elements in Bernie's plan. First, a complete ban on the development of artificial superintelligence, defined as AI that surpasses human cognitive abilities in most areas. Second, an immediate pause in the development of advanced AI until a new regulatory system is developed. Third, a new cabinet-level AI ministry.

And consider: the punishment is corporate death. There is a higher penalty for corporations. Your company is simply being closed down. For individuals, 20 years in federal prison, modeled on the punishment for illegal nuclear weapons.

So, interesting. I mean, this is bullying in all its glory. Dave, do you want to start?

There are so many real problems that need to be solved. I just hate this idea of turning everything into a “to do or not to do” argument. It's absolutely crazy.

Plus, the way liability law works in the United States, it's very easy to scare people into not doing anything. If you're afraid of 20 years in prison, you don't even know what the rule will be. They haven't even proposed a rule yet.

Nevertheless, the fear of being executed on the street is very real, so people stop acting. This is the worst thing that can come out of this situation.

And Bernie from Vermont—I love Vermont. I have spent more time in Vermont than in any other state, and it is beautiful because little has been done there.

Do you understand? Therefore, I believe that such a perspective is very dangerous. “When in doubt, do nothing” will become standard behavior, and this is very toxic.

If we don't do anything, very, very bad things will happen. For example, terrorists using AI to create biological weapons, or China getting many steps ahead of the United States. These are real things to worry about, and if you don't act and instead turn this into a political battle between the left and the right, that's the stupidest thing that could happen right now.

We need very specific technical solutions for open source and monitoring, just as in the case of nuclear power. If we start moving in this direction, we can easily solve these problems before the advent of ASI. If we turn this into a political discussion, we will find ourselves in a dead end and people will not know what to do.

ASI. Yes, ASI. Alex, what are your thoughts on this? I think you voted for Sanders, I assume, right?

I am not a Vermont voter.

Okay, for better or for worse.

I think that establishing a statute—I mean, on the one hand, we're in Hollywood country, and every Hollywood movie needs a villain. I think the singularity story in Hollywood has now found its villain.

Establishing a legislative limit on intelligence is not only a terrible idea, but also a bad one. I think this is tantamount to limiting human intelligence. This is a story about dystopias in science fiction. There is Burn books. Burn books and burn minds.

I think that putting a legal limit on how much we're going to merge with machines is actually a limit on individual human intelligence. Punishing people for being too smart is a terrible idea. This is a bad idea.

I think history will show that the proposal to limit intelligence in any form—whether machine, human, or hybrid—is effectively the equivalent of communism in the 21st century.

But I'll tell you this: the most dangerous thing about Sanders' words to me is that a lot of it can be changed quickly. But the part about data centers—if your state doesn't have data centers and they start building them in other states, they won't come back.

Once the land is allocated, the chips are in place, and the investment is made, that's something you'll never get back. This will be a huge economic driver for those states that come out ahead. This is the part of Sanders' statements that I think they will really regret. This is tantamount to burning cities.

Short thoughts, Suleiman?

Two things. When you look at this whole narrative, I'm interested in asking this question: what evidence from either side would make someone change their mind?

If you think ASI is bad, then your narrative has reached a dead end. There's nothing we can do to prove to you that it's not bad. The same applies to the other side. It is very difficult to change the narratives in our world.

The good news in all of this is that I think the Trump summit and the Bernie stuff don't matter at all. This doesn't affect anything, and it seems Dave was pointing that out. Once local models become available and people can run their own models locally, the world will move forward.

Now it no longer matters what one specific person does. The process will just move forward, and that's great. We've always had this huge humanity moving forward, solving problems. We'll probably beat cancer in the next 2 years.

Yes. This is an incredibly optimistic goal. Why would you slow it down? You, as a foreigner on this panel, what do you think about it?

Hey, what about me? No one?

You're a New Yorker. You're a New Yorker, so we're crossing that out.

We did Brexit to get rid of regulations, right? But I think it's actually similar to how the United States government tried to ban encryption a few decades ago. In fact, it was proven that this is a free-speech issue because you are trying to ban mathematics.

ASI and superintelligence are mathematics, and banning them is unconstitutional because it's freedom of speech.

6. The Explosion of New AI Models

I'll move on. So, in the last 22 days, we have had 12 major model releases. Four releases in the last 5 days, right? We're averaging less than 1 release every 2 days, which is crazy.

Almost like 1 model a day for xAI. One model a day until the end of this year.

7. Meta Muse & the Rise of Consumer AI

Yes. So let's briefly discuss Claude Opus 4 5.5, GPT Saul, GPT 6 Saul, Luna, and we can skip Grok 4.7 for now. Not very interesting. Sorry, Elon.

But let's start with the Meta AI news. Meta Muse can personally book you flights, order food, manage your calendar, do your shopping, and it's ranked number 1 in the App Store with 2.8 million downloads—the fastest consumer AI product, surpassing even ChatGPT at the time of its release.

And it's interesting, right? People forget that Meta covers 3.8 billion people on the planet. This is one of the main factors of its success. If you have this level of connection with the end consumer, you can quickly promote your product.

Yes, of course. We are at a crossroads now where there is consumer AI, which is incredibly valuable. I mean, it will completely replace search and all other ways of interacting with the internet.

On the other hand, we have a recursively self-improving strong AI that we are simply testing mercilessly. So these are 2 different directions. Which model is the coolest right now?

We used to say, “Oh, the Opus 5.5 just crushed all the benchmarks. It’s the coolest model.” But this is different from what Meta Muse does, because it doesn’t have to be the smartest to be the most useful in everyday life. Stocks, including Epicor, where I’m the chairman, fell 15% two days in a row because everyone was saying, “Oh my God, I’m always going to look for auto insurance and mortgages through Manus now.”

Interestingly, although in reality everything will not be like that, the reaction was incredible. Everyone’s starting to think, “Wait, is Google Search going to become completely irrelevant? This Meta Muse thing looks like my personal assistant for life.” And this will definitely happen. It’s fascinating to watch the labs outdo each other.

A quick thought on this before we move on to Opus 5.5. Meta needs Muse to succeed, and I think it’s even implied in the name. If you look at Muse on the one hand, and Manas, Meta's attempted acquisition of China's CUA, which the Chinese government forced Meta to cancel, on the other, this is a core distribution strategy for Meta—to leverage its existing distribution channels.

8. Claude Opus 5.5 & Competent AI

If you look at Instagram, Meta uses it to its fullest to promote Muse. Make everyone install Muse, because Meta needs a distribution point for its AI. Otherwise, Meta risks being forgotten and irrelevant in the future. Instagram and the other family apps simply won’t be as relevant because everything will become synthetic. So Meta should use its application base to attract everyone to its own CUA, i.e. Muse/ Manas.

This week, we saw that Claude Opus 5.5 delivers performance equivalent to stable 5.1, literally at half the price. Let’s draw a graph. And you know, I ported Skippy to Opus 5.5. I also want to save some money. What do you think about the performance of this, Emad?

This is a fantastic model. Isn’t that incredible? Yeah, I mean, with the Model 5, I thought it was going to kill us all. It was completely uncontrollable. And 5.5 is a truly pleasant model to use, with a huge internal space of knowledge and an understanding of taste.

I would say this is the first truly competent model, and that’s what drives the news and things like that. We have the emergence of competent AI right now.

It’s such a joy to hear when you add a visual interface to everything you create now. It doesn’t just show you mockups before you actually start building something; it then builds exactly what it suggested in the visualizations. So visual understanding is much better than it was just 4 months ago.

The experience of understanding what it’s doing because you’re putting an interface on it and then interacting with it—feeling inside it—is night and day different from what it was just 4 days ago. If you haven’t tried it yet, just update all your systems. It’s just a click of the mouse. It’s amazing how much better it is.

Quick thought on this. I could talk about benchmarks, but I think qualitatively new capabilities are more interesting in this case. What I see as new in Opus 5.5 is that it can generate video. These are procedural videos. It can generate animations. These are procedural animations. It can generate images. These are procedural images.

One of the main questions was that OpenAI has a video model and an image model, while Anthropic doesn’t provide anything like that. You cannot generate pixel-quality images using Anthropic models, or video or audio using Anthropic models. You can do this with OpenAI models.

It seems that Anthropic takes a completely different approach to multimodality and unimodality, where code is at the heart of everything. So now, for the first time with Opus 5.5, you can generate truly high-quality video that is entirely procedurally generated. You can generate an interactive game with textures, but it doesn’t generate textures pixel by pixel. You can ask it to, but it generates everything procedurally.

I think these are two very different strategies. If people remember the difference—and the wars—at the dawn of computer graphics between vector and raster graphics, Anthropic is playing the vector-graphics game, where everything is just commands, and OpenAI is playing the raster, or pixel-based, game. I think they’re headed for a collision.

That collision will happen in robotics, because the question will arise: If you need AI that is deeply embedded in the physical world, is a vector basis the right one, or is it still pixel or raster? What is ours, what will be our next story.

I love the fact that a week after Dario says we should pause and slow down, they release their next model. These are, frankly, revealed preferences. This is, in my opinion, the height of bold thinking and the height of hypocrisy: on the one hand, saying that AI will kill us all, and a few days later, with the other side of your mouth, announcing the most powerful AI in the world, biolaboratories, and other possibilities.

9. GPT-6, Sol & the Falling Cost of Intelligence

Yes. And on the same day that Opus 5.5 is released, we see the release of GPT-6 all and Astra. Dave, any thoughts on this?

Dave Blandin

If you look at the performance, it’s absolutely insane. I think a year or 2 ago there was a big school of thought that hallucinations would be a disaster. They would never disappear. You couldn’t use these models. It’s actually common at MIT, if you can believe it—the idea that this problem with hallucinations would stay with us forever.

Like everything else, we just scaled to get rid of it. You see that the error rate just drops precipitously due to pure scaling. It’s a thesis about abundance on a grand scale, right? Models are becoming twice as good, twice as cheap, accelerating the pace.

Have you played with this at all?

Yes. And I think, again, for our tests, this is an incredibly competent model. The models are now quite intelligent. We need them not to make mistakes. We need them not to fail. That’s what everyone is going for, and that’s why things like Manus, Instinct, and others are going to be incredibly successful, because they can be a great personal assistant, chief of staff, organizer, and so on. We all need some competence in our lives.

It’s funny: as an organizer and a chief of staff, it’s incredibly powerful. Anyone who’s used GPT-2, GPT-3, and GPT-4 is used to seeing them churn out misinformation. But if you were a kid today and started with GPT-5 and GPT-6, you would trust them right away because they’re right 99.9% of the time. That’s a bit dangerous, because they still sometimes give out factually incorrect information.

10. Driving Bench: General AI Takes on Robotics

Okay, I’m going to bring this into the physical world. An important story that we started talking about is something called Driving Bench. A group of developers created a new benchmark called Driving Bench. They took a Toyota Corolla, equipped it with an interface so that software could control the steering wheel, accelerator, and brakes, and then asked a universal AI model to drive a 130-meter strip of cones with a single command: “Go from here to there.”

And here is the data: GPT-6 Astra completed the entire distance 100% on the second attempt. I’ll read a quote from Boris Power, who said, “Incredible results. This should be the nail in the coffin of specialized models that were trained from scratch with great effort, as opposed to training the most powerful generalized model and distilling a small specialized model as needed.”

So, Alex, you’re excited about this?

Alec Wiseman Gross

The bitter lesson is truly bitter. It’s very exciting. Essentially, as a school project, now or very soon, this whole idea that robotics requires separate models is just going away. You simply code a robot, insert an open-source model, say Astra, into the robot’s body, and it instantly acquires embodied cognition.

This is probably quite frustrating for a lot of academics in the field of computer science.

Lekun. Yes, this means that decades of computer science and robotics were probably, in retrospect, a complete waste of time. All we needed was a generalist model, and it turns out that with a generalist model, robotics is also over.

A superintelligence can drive a car. What a surprise. I’m completely tired of models. I don’t have time to keep up with 5.5 versus 5 versus Astra. Just raise your hand if you’re tired of models. Can I see your hands? This is great. Everything is just fantastic. I love this progress, but, oh my God, I can’t keep up with it.

You’re so tired of winning. The time has come: I’m tired of winning. That’s a great argument. It just occurred to me: We do 2 episodes a week. It’s killing us. I wonder if we are the problem. We’re moving too fast. We are optimists here, and perhaps we should take responsibility.

If we don’t report it, it’s as if it never happened. These are our tools. Do you think if we stop the podcast, we’ll stop the singularity? I’m just curious. We could speed it up, or we could slow it down. I think we can be a part of all of this.

Emad, what do you think about DrivingBench?

The audience in general is intelligent, right? They study all these things, and high competence will extend to biology, physics, and anything else. I don’t think there is a single specialized model that will be able to surpass everyone in a couple of years.

Dave, sum it up on this.

Dave Blandin

I think both statements could be true. A generalized model can drive a car. It can build a road. It could be a robot. It can build a robot. That’s quite true.

If you said, “Hey, it passed the course on its second try,” Elon will say, “Yeah, your Tesla won’t be able to do it on the second try.” I’m sorry, but this won’t work for most drivers.

I think the thing is, there’s one thing missing from this story: How much computing power did you use for the task? Computing power will now be in perpetual shortage.

11. Anthropic’s Wet Lab & AI-Driven Biology

Dave Blandin

Chip prices have increased. For example, HBM RAM has increased in price 5 times. It's crazy, but that's because AI is so incredibly valuable. People need compute. So if you spent more resources on the same task, that's a crime. I think this is a very important point: we move the bottleneck from models down the stack to computation, and then to energy. Yes, we are now in the energy stage. We lack about 60 gigawatts of energy for the next 2 years.

Yes. Moving on to our final segment on science—something we all truly love. So, a few stories. On Tuesday, Anthropic announced that it had created its own life sciences research group and its own “wet” labs. Again, the organization that calls for slowing down is building “wet” labs. Okay, great.

And it is reported that Claude, working autonomously, discovered a previously unknown enzyme. That's how it was. They gave Claude one request: “Search the huge DNA database for interesting reverse transcriptases and enzymes that copy RNA or DNA.” They had 950 Claude agents search for 21 hours, and they found something. They found a DNA repeating pattern that looks a lot like CRISPR. They later called it massively associated reverse transcriptase. They don't know exactly what it does yet, but they think it's going to be similar to programmed transcriptases in gene editing.

You know, Jennifer Doudna won the Nobel Prize for discovering CRISPR with her co-author. I wonder if we're going to see Nobel Prize-level results every week because of this, Alex?

That's the whole point of our book, Solve Everything, Peter: we should expect it. And then the natural question is, what's a Nobel Prize worth in an age where you can get a thousand of them in a year? We've had a century of human progress in a matter of months. I suspect that Nobel Prizes will be awarded for larger-scale work, but I have to point out another elephant in the room again, which is the sheer hypocrisy of saying that AI is going to kill us all and then opening a wet lab and using advanced AI in conjunction with that lab to synthesize proteins and other molecular structures of unknown function.

I think, what could go wrong? That, to me, is Frontier Labs. Those are the real priorities of Frontier Labs. I think they themselves don't believe in their attempts at regulatory capture and creating a security cartel. I think their real priority is—no, actually, going full-on into using AI to solve biological problems. And every lab now has a biological focus.

Yeah, Salim.

Yeah, can I just fuel the P(doom) for a second?

Have fun. Give people a reason to panic.

So, we just discovered CRISPR, right? We learned that we can edit the human genome as easily as we can edit a Word document or a piece of software, right? When you learn a new language, you learn to read, then you learn to understand, then you learn to write. Now we've learned to write in the genome. Each of us is 50 trillion cells in the body, controlled by what the DNA tells that cell to do, which means that the human being is now a software engineering problem. And indeed, you've seen the designer babies from China and so on.

When you combine that with the potential of AI, the potential benefits skyrocket, but the potential risks skyrocket as well, and we're in full Frankenstein mode. So there's some argument for all the pessimists to go wild with this. This is how we're going to beat aging. This is how we're going to beat disease.

I don't think we're going to get to a happy future where the 5,000 major diseases are solved or a sustainable escape velocity from old age is achieved without superintelligence. So I'm at least a fan of Anthropic and all the other cutting-edge labs that are opening up their development, because that's how we'll solve everything.

12. The Millennium Problems for Biology

Yeah. Okay, our last story is a recent release by Sam Rodriguez from Edison Scientific and Future House. They've put forward what they call the “millennial problems of biology.” Can you put that on the screen, please? Alex, you and I wrote in Solve Everything: pick goals, build tools, set evaluation criteria—and go for it.

Some of those criteria say, “Okay, the math isn't just ‘done’; it's burned—it's completely burned.” Let's get to biology, okay? There are some of my favorites: the origin of life, cryopreservation, limb regrowth. You know, Emad, they left out the return of youth on that list.

That's right. And I think that's good, but let's make 100 of them, throw 1,000 agents at each one, and make it all open-source. That's something that labs should collaborate on. They said their criteria here are that these millennial prizes should be testable in a standard biology lab in a couple of days. They should be hard to solve and very easy to test.

That's why rejuvenating youth is probably not on the list. No, no, no. So, listen, when we started the $101 million Healthspan XPRIZE, first Peter Thiel and Aubrey de Grey came to me and said, “Let's do a longevity prize.” And I said, “I don't know how to do that without a time horizon of 20 or 30 years.”

And then George Church said, “Forget about longevity; measure rejuvenating youth.” So if you can apply a treatment and measure functionally that a person is 20 years younger, you can measure it in days. My daughter is my main advisor on this because she's at Moderna. She's a biotechnologist, and she says, “Yeah, that's the same list. This is perfect.”

So that's my humble opinion on this. I think it's incredibly cool to have a framework for progress, because as you said, the Nobel Prize is going to be irrelevant soon. In fact, it's already quite irrelevant. So having a framework for progress in this area is incredibly important.

So, Alex, this is when we have digital twins of cells, which I think is a critical path to overcoming all disease.

I think the timeline looks something like this. Over the next few years, just like we trained large language models on almost all of humanity's behavior on the internet, on text and images, we're going to train digital twins of cells—all of the cells in our bodies—on huge datasets, interventional and otherwise.

Then we're going to use these ideal digital twins of cells to do an exhaustive search, like an AlphaGo search tree, in the space of interventions to find a way to cure the disease. You start with the state of the diseased cell and look for a strategy—a treatment, a sort of “move 37,” if you will—to go from a sick cell to a healthy one.

What's great about this approach to the millennial problems for biology is that, as you say, Peter, these are problems that are easy to test but so far difficult to solve, and it essentially turns any complex biology into a mathematical hypothesis and a goal—a mathematical goal. So essentially every biological problem, thanks to these digital twins, will become just a search problem, which means that for the first time in history it becomes solvable. Salim, please wrap up. Biology is over. That's right. Okay, I'll wrap up our WTF episode. Imad, you'll join us with Katie right after lunch. Thank you very much. Applause to Imad.