（预告）关于前沿 AI 实验室、如何向持怀疑态度的公众讲清 AI，以及 Amazon（还有 Apple？）加码与 Elon 竞争的6个问题

Hello, and welcome to a free preview of Sharp Tech. Hello, and welcome back to another episode of Sharp Tech. I'm Andrew Sharp, and on the other line is Ben Thompson. Ben, how are you doing?

I'm doing okay, Andrew. How are you?

I'm doing all right. I'm pretty jealous after the Nico Rosberg interview this week on Stratechery. You interviewed the absolute GOAT of F1 broadcasting. It was a tremendous hang. I enjoyed it.

I usually don't interview VCs. They're always wanting to talk their book. But I guess I have an exception: if you have driven an F1 car in RVC, then I'll give you the time of day.

His life story is super interesting, and the way he's leveraging F1 into his venture career is also really smart and interesting.

I actually thought there was an interesting point here. You can look at someone like Nico Rosberg—an absolute life of privilege, right? His dad was also an F1 champion, one of those father-son combinations. He grows up—

He was also conducting the interview from Ibiza, where he owns an ice cream store chain.

Two ice cream shops, in fact, because they're doing so well. “I've been coming here to Ibiza my whole life”—that's a line that occurred in the interview. He speaks 5 languages, went to very high-end international schools, and obviously came up to succeed in F1. It's always striking: you have to be in a kart by the time you're 5 years old.

Mm-hmm.

Five years old might be too late. That speaks to something I always find fascinating when talking to anyone at the absolute top of their game. There's so much that goes into it that you don't appreciate. How do you actually incorporate all the inputs of an F1 driver? It's not just your eyes or your hands. It's your butt, your sense of what the car is doing.

And your sleep. There are all sorts of things you have to account for to excel at that level.

Right. But this applies to lots of fields. When you dig into anyone who is excellent, it turns out there's way more involved than you might think. It's easy to see with sports who is excellent and who isn't, but when you dive into things like business, there are very few people who just suddenly show up out of nowhere and are awesome at something. When you dig in, it turns out it's been years and years and years and years and years, right?

I think you're a great podcast host, and you can be a bit of a contrarian. Then I hear about you as a little Andrew, and I've had the good fortune of meeting your dad.

I've been like this for years.

Yeah, no, you've been training to have terrible AirPod takes literally since you came out of the crib.

Anyhow, what I appreciated about Nico is that it's easy to dismiss someone like that and say, “Oh, well, they had all these advantages in life.” What is striking is how he doesn't deny that he's had advantages. He just relentlessly identifies his advantages and then leverages them to do the next thing.

Yeah.

That's what distinguishes people who start out in a good place. Not everyone who starts out in a good place actually accomplishes big things, because you have to take advantage of your circumstances. I thought it was interesting to tie that all together.

There was a Justin Huang interview in the news this week that I might write about next week. This was a bit of a softball, but it was actually more interesting to me than I expected for that reason: this linkage that you wouldn't think exists but does.

100%. What I appreciate Nico for is his uncomfortably frank F1 broadcasting and his questions to F1 drivers. You can tell that he doesn't need that job.

Yes, exactly.

He has fun with it and is going to speak his mind the entire time. Shout out to him.

I was disappointed—

Welcome as the third chair on this podcast anytime he wants.

No, I was disappointed that he was not as critical of the current F1 regulations as I was.

Well, Mercedes is cutting him a check, after all.

He is. He is still associated with Team Mercedes.

Gotta play it diplomatically for now. As for the show, Ben, we are going to start with a 6-pack of AI questions that kind of run the gamut here, just bouncing all over the place. Daniel writes:

1. Consumer AI Hits Compute Limits

“Longtime reader and listener here. About Monday's article from Ben on AI and aggregation theory, isn't the central question, which I felt went unaddressed, whether having the best product in the consumer space specifically is going to go from zero marginal cost to high marginal costs? In the latter scenario, the products that consumers will expect in the future are thinking and agentic and would therefore always benefit from more compute. That means the primary unit of quality goes from upfront design and implementation to continuous build-out of ever more capacity to serve ever-rising per-user token demand.”

What do you think of that question?

It's a good question. It puts its finger on something that is somewhat unresolved. I tried to put forward that it was unresolved, particularly at the end: let's assume the best product wins, but what if the best product means the most compute, and that's actually what matters?

I think there are a few problems with the premise of this assertion. Number 1, it sort of assumes that these agents will never be good enough—that there's basically an infinite capacity for how good they could be—and that might be the case. I did write an article about one of the brilliant insights of building a product focused on consumer satisfaction. I think I was talking about this in the context of Amazon: no one is ever actually fully satisfied.

It's a goal that you can never reach.

It can always be more convenient.

Right.

Sure.

This fits in that framing: can agents always be smarter to the extent that those improvements can continually be absorbed, just like the improvements in the user experience I've put forth previously can be continually absorbed? That might be the case, but it's worth—

It might be the case, but as a normie, can I just interject to say that the differences between models are already very difficult to perceive as a user? If we're talking about the consumer space specifically, which is what he's focused on, I'm not convinced that the performance delta is going to be what decides who wins in that space over the long term, because everything is so performant already, and it's only going to become more so over the next couple of years.

I completely agree. Implicit in this, this is actually a very optimistic framing in terms of the doomer narrative around AI, because you're basically stating AI is never going to be smart enough for humans.

Right.

It can continually get smarter and smarter. Maybe that's the case, but it's worth pointing out that this is contrary to the narrative that AI is going to surpass human capabilities. To the extent that it surpasses human capabilities, then it is going to be good enough, which means the increase in quality and capability, to your point, is not going to be noticed or appreciated because it's already—

Yeah.

—good enough. In which case, we are reverting back to a zero-marginal-cost world where the computing is effectively free because, oh, yeah, the AI is up there in their planning meetings saying, “Oh, yeah, the humans can run on the model from 13 years ago, because they don't know any different, so it's good enough,” right?

Running on those chips from 2035? Those don't matter. Like, what? Super—That's—

Mm-hmm.

There's this overall attitude toward AI of always choosing the pessimistic interpretation of everything, even when those interpretations are totally in conflict with each other, and this is a good example. This insistence that compute is going to be consumed infinitely is in direct conflict with the idea that AI is going to get smarter than humans. Once it's smarter than humans, it's smarter than humans, right? At least to your point, if we're talking about the consumer market.

Yeah.

And you made the point: What do consumers actually want? If it's just an answer bot, we're already in a pretty good position, to your point, right?

Mm-hmm.

We saturated the space remarkably quickly. Now, I think there's a bit where one reason it makes sense to focus on enterprise is that right now AI is clearly a productivity booster, and I've been making this point repeatedly: enterprises pay for productivity and consumers don't.

Yeah.

Consumers don't want to be productive. They want to enjoy themselves and have fun. Watching Reels is not productive, but it is enjoyable.

It's a great business.

Right.

Yeah.

And it's a very good business, right? Now, of course, there's this idea that we're going to have assistants that help us do things and whatever it might be, but at the end of the day, a lot of stuff—the point is the friction.

Mm-hmm.

What do people do for fun? I saw a Reel, and I'm going to badly interpret a comedy bit: The way to identify an activity that rich people enjoy doing is that it's a reinterpretation of what poor people already do. So—

Hmm.

For example, let's go hiking: walk very long distances and tire ourselves out. Guess what? You can walk for free all the time—

As if we don't have cars.

Wherever. That's right. Right? There was a whole list of things. I'm not going to reinterpret it. I think some of them were fairly objectionable, as any good comedy bit should be. But—

2. Enterprise Pays Consumers Do Not

Okay.

There is an aspect here of this being maybe a reason why enterprise markets are just always different from consumer markets.

Hmm.

And maybe a mistake companies make again and again and again is forgetting this distinction. That kind of gets into my bit. It probably makes sense for OpenAI to double down on enterprise because that's where the money is for—

Right, that's the money for the next 5 years.

—the clear-cut AI use case, which is increasing productivity.

Mm-hmm.

That, though, is a reason to be quite optimistic about Meta because they're not in the enterprise space, and they already have a functioning advertising business—more than functioning—and OpenAI just can't do both. It is probably a mistake to try to do both. Sure, ChatGPT is still the biggest today, but every little bit of time—

And when you say they can't do both, are we basically talking about optimizing the ads engine on the consumer side to monetize that over the next couple of years, while also focusing on the enterprise side and dedicating compute to the enterprise side? Because I don't really see why they can't do both.

Well, it depends on what. I guess you're the AI optimist here. They could just have AI write all the code to build up their ad engine on the consumer side and—

That's right.

Whatever.

Bitter-lesson-pilled, you know what I mean?

It might be. Maybe you're right. My general view is still based in a view of talent scarcity. Actually—

Yeah.

Building that out is going to take a lot of time and focus and energy—

I see.

You should probably be building out more of a business in a place that actually pays—

Hard to win in 2 worlds that are ultra-competitive, and Meta already has—

—and very different from each other.

—all those muscle groups—

Yes.

—pretty well developed.

Well, and Google—

Yeah.

—as well. I don't mean to dismiss Google in this sort of story here. The Google story's pretty interesting, just because December wasn't that long ago, when everyone was like, "Oh, Google's won. It's all over."

3. Google Misses the Coding Wave

On the 6-pack of questions here, I have this question from Robin, who says, "Do you think the future narrative for this year might be Gemini falling behind by not milking the compute scaling law via Blackwell? Because I do." And I included that note from Robin mainly because it was the first time in months that someone has mentioned Gemini in our emails. So do you have an answer for Robin, and what do we make of good old friendly Google right now?

I don't know. Is Google fundamentally hindered by not using NVIDIA? I don't know that that's the case. It's hard to know any of this stuff for sure. The feedback I generally get, and I'm not an expert in this space, so this is mostly second- or thirdhand, is that Gemini's not very good at coding.

Mm-hmm.

Facebook got a lot of grief, and I think had a very traumatic event, to the extent that they basically fired an entire team.

Fired everyone. Yeah.

Because they cheated on benchmarks, by and large. And Gemini—I don't know that it was done intentionally—feels like a very benchmark-optimized model that people actually use and find fairly unsatisfying to use. I'm not necessarily the right person to judge this. I think Gemini would say, "Look, in real-world use cases in the enterprise, people are using Gemini up and down, left and right. And just because we're not releasing half-baked products and getting on the hype cycle doesn't mean it's not a real thing. Going forward, we have this integrated advantage, we have all this compute, et cetera, et cetera."

Yeah.

And by the way, it's worth noting—

That's true.

—Google Cloud numbers are awesome, right? The business results are there. Is that Gemini? Is that because they're renting out the cloud to other folks? Who can tell for sure? Your takability index almost needs a separate section for AI specifically because Anthropic is clearly top of the world right now.

Yeah.

Is it going to be the case that in a month the narrative is going to be totally different and totally flipped on its head? It sure seems like that's possible.

Right, and a company like Google—what can you say after the last couple of months? All the advantages that we talked about in December are still there, and they're just not necessarily winning the hype cycle every couple of weeks here. But by the same token, they have the funnel to consumers with Google, and they have gobs and gobs of money to continue throwing at the problem. On an infrastructure basis, they're well-positioned to own this space over the next 10 years. It's hard to get too down on them.

Yeah.

But it's also kind of curious that they're just never really mentioned in these conversations.

I don't really know anyone that uses them for coding. That seems to be a real—

Yeah.

—weak point as far as development goes, and we've talked about that being—

And coding is what's driving all the hype—

That's right.

—right now.

That's right.

So Google's sitting that game out. All right. Well—

I don't think they're sitting it out.

Continuing on.

They're certainly trying, but it doesn't—

Well, they're losing.

Yeah.

They're on the sidelines. They're benched. Thomas says, "Ben and Andrew, I was thinking about the current conversation around Anthropic and OpenAI, and in particular, how their beliefs about the path to AGI seem to be shaping their current positions. Anthropic is extremely strong at coding but constrained by limited compute, while OpenAI seems to have more reliable access to compute, yet may be slightly behind in coding performance and adoption.

"I think it all comes back to OpenAI having always been the scaling laws company and Anthropic focusing on recursive self-improvement. Dario talked about why he didn't buy more compute on Dwarkesh from an economic standpoint, but I also believe that he sees the path to AGI as the recursive self-improvement from the AI being able to train the next AI. When you believe this, it doesn't make sense to waste time on image or video models that just take away resources from the specific type of model needed to build AGI. It also means you can't overspend on compute because you need to make sure you hit the takeoff just right.

"OpenAI is, and always has been, about scaling laws, and if scale is all that matters, then you can't possibly overspend on compute. Codex was a late focus because they saw the success of Claude Code, but they have always been the AI company that takes scaling laws literally, and if you believe in scaling laws, there's no such thing as too much compute."

Both companies are learning from the other in these scenarios, leading to larger compute spend by Anthropic and more focus on coding by OpenAI. But I think the winner will be determined by how you can get to AGI. So, Ben, grade the theory there. What do you think of that breakdown?

I think it’s a pretty good one. One way to think about why Anthropic has been doing so well and OpenAI was scuffling a bit comes down to some combination of focus and alignment.

Mm-hmm.

Anthropic is super focused on coding, as Thomas notes, because it sees coding as the way to—once the AI can program itself, then yes, recursive self-improvement, and that’s the actual sort of takeoff. Humans aren’t going to program the AI to AGI. The AI is going to level itself up into AGI.

Hmm. So is it correct to say that Anthropic is the most Bitter Lesson-pilled of the frontier labs?

No, I think this is kind of the opposite of Bitter Lesson-pilled. For them, algorithms matter. The thing is—

Mm-hmm.

You just have to get to the algorithms writing their own algorithms. But, yeah, it’s interesting to put that in a Bitter Lesson sort of framework. My initial response would be no; it’s actually slightly different.

And so the reason why that is great for them from a business perspective is, as we’ve discussed repeatedly, coding is a great application for AI.

Yeah.

You’re generating a lot of text, and it’s verifiable, which solves the problem of coming up with a bunch of text that might be hallucinating. You can systematically find issues and mistakes, go back and fix them, and get this sort of recursive loop going on. But this is very powerful from a business perspective because they are motivated by building God, as it were.

Mm-hmm.

Their way to build God is, it turns out, to be a product that everyone wants to buy just to make business applications, right?

Right.

4. OpenAI Trades Speed for Reasoning

OpenAI has had more of a challenge. OpenAI, to your point, when Thomas is talking about scaling, he’s talking about the Bitter Lesson. The Bitter Lesson is about scaling laws: just get more data, make it bigger, do more and more, and more compute and more data are going to solve all your problems for you.

So, yeah, I think that’s OpenAI’s overarching concept. Maybe. I think that’s not a bad way to put it. You also have a bit where OpenAI’s biggest advantage is that they’re really good at reinforcement learning and reasoning.

Mm-hmm.

OpenAI’s models today are still significantly smaller. They’re more like GPT-4-class models than, say, a Gemini, which is much larger. But the reason why it feels better is, if you use plain ChatGPT, it’s terrible. You could easily tell it’s not as good.

But where it’s good is if you’re in Thinking or in Pro, where the reasoning is just really good. It’s doing multiple things—it’s comparing them, deciding which one’s better—and it also makes it very slow.

Yeah.

One of the things with ChatGPT is, maybe this is another reason why the consumer might end up not working out for them: the best consumer models, I suspect, are actually larger models that just turn out a close-to-right answer the first time.

Decent answers.

Right.

Yeah.

Sitting around for ChatGPT—if you want to get the best out of ChatGPT, you have to use the modes that are super slow, and you have to sit around and wait for it, and it’s kind of a crappy experience, honestly. But it does give you really good answers.

There are a lot of people who actually think their coding capabilities are much better than Claude’s in many respects. But a lot of this is downstream from the fact that they’re just really good at reasoning, and it’s slow and it takes a long time.

But that’s a misalignment. It’s not aligned—

Yeah.

—necessarily with what they’re doing. They’ve done 47 different things, no focus. They have a research team that also wants to get AGI. They see reasoning as a way to get there. That’s not necessarily aligned with what they’re trying to build from a business perspective.

Mm-hmm. And the scaling concept—those are 2 totally different things, right?

Yeah. Well, I think their new model—I think they’re going to have a new class of model, like this Spud model. What’s going to make Spud unique is that it’s truly the next generation. And I actually think it’s going to be really interesting to see how Spud is, because if they can have a much better base model with their sort of RL layer and their reasoning capabilities, it could definitely be very, very capable.

Hmm.

So maybe that’s going to drive the next change in narrative, or whatever it might be.

I was going to say, continually updating the takability rankings in the AI space.

Right.

And who knows where OpenAI will be in 3 weeks.

But I do think what Thomas definitely is right on is Anthropic’s focus on coding—

Mm.

—has been hugely beneficial. And for OpenAI, just internal alignment has been a challenge. It’s been all the upheaval inside the team, and Anthropic has sort of gotten that for free. That’s a huge factor in their success.

5. Compute Scarcity Changes Prices

Fair enough. Okay. Well, a different Thomas wrote in and said, “Several times Ben has complained that TSMC left money on the table when they had the best process but were production-constrained. Surely they should just raise prices until the demand balanced. Well, likely, they worried that some of that balance would come from customers getting used to Samsung’s process or even Intel’s, and those losses could be sticky. In the AI world, if we’re now compute-constrained, should the AI companies be raising rates? What do you think?”