(Preview) Fable 5.1 and Anthropic’s Data Retention Pivot, AI Civilizations and Related Matters, Q&A on Meta, Shopify, 3-D Printing

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, Ben Thompson. Ben, a sincere how are you doing? What's up?

When you've been in the dark underground, you emerge out into the sunlight, or you're in a desert—you're thirsty. You stumble upon an oasis. I spent 2 days, Andrew, without internet.

You made it out of the server room. I wasn't sure I'd ever see you again.

No, that's a mirage.

Oh, okay.

There are actual oases in the deserts. Yeah.

An oasis is what you want.

As we will get to in this episode, I may not know how to say words, but I do know what words mean.

Good.

That's how we can be a good partner here.

That's more important, honestly.

1. The Ethernet Rack Rebuild

Before I moved to this house, I had an AV company wire up a rack and a bunch of TV stuff and all that sort of thing. Over time, as one does if one is a computer geek and a nerd, one wants to use the rack for actual computer things. The problem is that these companies build these racks assuming no homeowner is ever going to touch them. They just want stuff to work.

Mm-hmm.

So they take shortcuts. The good thing is, I knew that if I was remodeling a house, I wanted Ethernet jacks everywhere. There are literally, I think, 70 or 80 Ethernet runs all over my house. Every room has multiple Ethernet jacks. It's actually awesome because it unlocks a bunch of stuff I'm thinking about doing in the long run.

Mm-hmm.

Anything can run over Ethernet. Whether you run video over it or sound over it, it's just a universal transport mechanism. It's kind of amazing, this technology that dates back so long. You see those pictures of data centers? Those are all Ethernet cables. It is on the Mount Rushmore of great tech products: the Ethernet cable.

Okay.

Right.

Right next to E-ZPass. Sure.

What you actually want to do for something like this is have a thing called a patch panel, or a patch interface, where all those runs come into and are terminated at the front of the server. Then all your computer gear is next to it, and you run those short little cables, like you're a sound engineer on a synthesizer from the '70s, connecting different things together. That's how it should work. Then you have flexibility, and you don't make a huge mess trying to pull apart all these different cables and do this sort of thing.

It's all very clear and organized.

Right.

I've been talking about how it's the year of AI. Hardware is back, baby. I keep getting more things, and my rack was a disaster.

You sent me a picture of it last night, and it gave me a panic attack looking at what you had before this little remodeling of the server room that has been taking place.

Right. What needs to be done is you have to take everything out, take all the wires out, take all the equipment out, and put it back in. You have to have all these patch panels. I think I have 6 patch panels. You then have to reach in, identify every single wire from all over the house, and plug it into the right one. It is a mammoth job.

We're almost there. One more day tomorrow. The audio/video side isn't done, but the internet side is done. I have internet. It's back. I wrote last night using Starlink, which was part of that.

Yeah.

In the future, I now have a secondary network in the house, such that the whole house will have internet.

Of course you do. Resilience is important, as always.

That's right.

I will just say from my standpoint, one of the funniest parts about working with you is that I consider myself pretty well-versed in tech and business at this point in my life, based on the last 3 or 4 years working together—at least fluent in the structural dynamics and the trends that really matter. Hard tech and how tech actually works, I will save that for another lifetime.

I'm going back and forth with you about this project, and I ask, "What are you actually doing?" You hit me back with pictures of all the server racks and everything else that were a complete mess, and this color-coded chart that I believe is the goal of what you're trying to create in this server room. The chart you sent me is completely inscrutable. You sent it to me as if it would explain everything.

Oh, no. It is so beautiful.

I'm like, "What are we doing?"

Unfortunately, reality very quickly hits, and so the chart falls out of date. We printed a big one out, and it's hanging in the room so you can reference it. I do want to update it once everything is final because it's actually great. I found more runs in my house that I didn't really realize I had. Stuff's gotten moved around, and we added a bunch of new runs.

Yeah.

That thing is beautiful. I want to frame it—

It's beautiful—

—and hang it in the server room.

—and inscrutable.

This is a testament.

2. Fable 5.1 Matches Mythos

Look, I enjoyed the colors on the color-coded chart. In any event, we could update people on your progress next week. For now, the news this week: Fable 5.1 arrived. Anthropic stressed that it is the exact same model as Mythos 5.1. Why don't we start there? Why did they stress that Fable and Mythos are the same model?

I don't know. I just thought it was kind of funny. The very first line, they're like, "No, these are really the same model."

I think probably the way it was released—I'm guessing there was concern—or maybe it's just from a branding perspective. I think there's a bit—well, no. I think they didn't get the payoff from their scaremongering that they wanted to get. They sell Mythos as this unbelievable, scary thing that's going to take over the world: "Oh, U.S. government, please stop us. We did it again."

I've been sitting here for the last 3 months thinking, "Well, I guess Opus is the internal model. I wonder how powerful that is." Fable is what they're allowed to release to the public, so yes, it's probably a useful clarification from Anthropic—

Right.

—PR.

No, it's marketing. It's like, "No. No, no, we're giving you access to the model that's so dangerous the world's going to end." I mean, this whole release was very funny to me for lots of reasons. But this was one of them. Just the first line: "We have to say, 'No, they're the same model. I can't believe only 11% of our customers are using it,' or whatever the number is, because you have access to the model—"

Anthropic drinks Coke—

—too dangerous to release.

Just like—

Yeah.

—everybody else. You know what I mean? Well—

No, it's not that. It's like they spun up Project Glass Wing. The U.S. government shut them down. Wouldn't you like that model for your business? Imagine what you could do with that software. It's like, no, you have the same one. You really do.

Indeed. Well, the company said that Fable 5.1 is more skilled at carrying out lengthy, complicated computer-programming tasks, such as software projects or code reviews that require it to work across all the code for an application. The new version is also better at helping with scientific work, like coming up with experiment designs, simulating what could happen during an experiment, and reading information in dense diagrams and tables.

By the way, you know who understood my server rack diagram?

Who's that? Oh, Fable and Codex?

Yes.

Great.

It is great. You could run it there, and they'd say, "You have a missing connection here," or, "There's an inconsistency there," or things like that. This was so beautiful, by the way. It was very satisfying. It was fun. I had to hire people to help me do this.

Mm-hmm.

And so they were working on it. Actually, I plugged in almost every one of those cables, in part because this was such a fever dream. I had the mental model in my head of how the whole thing works: different zones and all these different bits and pieces.

Yeah.

But after the 3rd time that happened, I'm like, “Okay, I'm just doing it now.” I reserved a room all day long. So we are recording at night because I'm like, “Sorry, Andrew, you're gonna have to wait.”

There we go.

Yeah. 8 hours later, here we are. I'm just glad that you're still standing after dealing with that all day long. And I'm glad that your new best friends, Codex and Fable, were fans of the chart and knew exactly—

Yes.

—what they were looking at.

3. Compute Supply Limits Pricing

Dense diagrams and tables. So I'll finish reading for you. Anthropic said, “Companies are increasingly trying to justify and rein in the cost of AI usage at a time when open-weight, lower-cost models from China are rapidly improving.” Andrew, aren't you supposed to ask me how Anthropic is responding to that challenge?

How is Anthropic—

How is—

—responding to that challenge?

How is Anthropic responding to that challenge? You're the China expert.

Well, I am very curious because I'm aware that they charge massive margins. Lots of people have complained about the token costs over the last several months. People are getting more price-sensitive, and the competing models are getting more performant than they've ever been before. So what's Anthropic's plan to counter competitors?

See, this is where—whenever I'm frustrated with Fable, and Fable just says stuff—

Mm-hmm.

—and then I'm like, “Where'd you get that from? Can you explain that?” And then it'll come back and be like, “Yeah, I just kind of said that. I'm not sure why.” Here's my question to you: What's the evidence that this is a problem and they need to respond to it? From what I can see, they don't have enough compute.

Mm-hmm.

And they clearly still don't have enough compute. Their servers are down all the time. They were down a bunch last week, even though they're charging these very high prices.

Okay.

There is an aspect—and this was something I was thinking about in the context of NVIDIA's earnings—where all these—

NVIDIA's margins are—

The frontier models are making high margins on their—

—exploding.

—the marginal cost of serving. NVIDIA's making crazy margins. The memory guys are making huge margins. TSMC is making big margins. Everyone is making big margins. It is exciting to be in a world where supply is constrained because of these crazy earnings and all the stuff that happens.

Mm-hmm.

It's also kind of frustrating as an analyst because there's no forcing function to see who actually is capturing value—

Right.

—because there's just not enough supply. So—

So the interesting questions are abstract and not particularly relevant—

Right.

—for the next—

From what I can see, Anthropic's problem is supply. They don't have enough compute.

Mm-hmm.

If they had enough compute such that they could afford to meaningfully lower their prices—like, right now, with Fable 5.1, they lowered the cache prices but not the actual generation-token prices—

Right.

—they're still insanely expensive. And in that world, why should they take compute to serve—

Lower prices.

Especially when they're a company that is spending tons of money and has these huge R&D budgets. They need the revenue, they need the cash flow. You could make some sort of fanciful case that, for strategic reasons, they should charge less than they can to take the market.

Right.

The reality is the market is large enough at their price level to completely saturate their supply. They need money to fund their operations going forward. Given that, the fact that people are lower down on the price curve doesn't tell us anything. This goes back to my who's afraid of Chinese models piece. If the leading-edge providers had infinite capacity—

Right.

—they would serve their models not just at the same price, but at, I would argue, a lower price than these Chinese model providers do because they almost certainly have far more efficiencies in their infrastructure.

Mm-hmm.

So this whole thing is imposing business questions on an environment where there just isn't the structure to answer them.

The story is pretty simple as far as Anthropic is concerned. Well, had you asked me where I came up with that concern, one—

Oh, sorry. I jumped in.

—area of concern... No, a—

Also, another thing: Fable just comes in. It just has its own thing. You know, whatever.

Well, I've seen a couple of different charts over the last month or so indicating that Anthropic's growth has leveled off, and leveled off in very relative terms because Anthropic's growth was completely effing insane for the last 18 months or so. Do you think that's a function of OpenAI entering the market and succeeding? Or some other story?

I don't know. I think that unless we have an understanding of what their capacity issues are—did their growth slow down just because they don't have enough capacity to grow anymore?

Mm-hmm.

That's NVIDIA's problem.

That's true.

NVIDIA grew 100% last year, and they forecasted, “This year we're going to grow 70%.” Jensen Huang was very deliberate on his earnings calls. He has a theme every single time, and I think that the message here was the message that I wrote back in 2023.

Mm-hmm.

It's funny. I feel compelled to cover NVIDIA earnings, and I've been saying for 3 years now that they're actually very boring events.

Right.

Everyone gets so hyped up about them, but the problem is, in a world where demand exceeds supply, that means your revenue is—

Very predictable.

You are—

—selling every single chip you can make, and you—

And so when Jensen Huang—

—can forecast it that way.

Yeah.

That's why in your NVIDIA coverage, I believe it was a week ago, you said their real moat is their supply chains at this point, and that is true. And as far as Anthropic is concerned, I guess it is sort of a mystery. It's been interesting watching OpenAI succeed as quickly as they have with their coding efforts and their sort of enterprise efforts, but I'm certainly not concern-trolling about Anthropic.

Well, I do think this—I don't think this is on the rundown, but there is OpenAI, which is going to exercise the clause in their contract to no longer serve Cursor, which I think is understandable. It's a direct competitor. It's flowing through their apps. I get it. At the end of the day, you are in a competition. Anthropic comes out right away and says, “Oh, we love Cursor.”

“We would never do that.”

“We support our developers.” And it's like, what is the number one answer to their capacity problem?

What's that?

SpaceX.

Mm.

Giving and selling them capacity. So, yeah, I don't think they're gonna sell out Cursor. SpaceX is their great hope. Either side can pull the plug with a month's notice.

Yep.

Wow. I didn't catch that, but that now makes the Anthropic response that much more delightful. So the other piece of the news this week with Fable is that Anthropic is rolling back its data-retention policy.

CNBC says it will change a controversial data-retention policy for its business customers after, quote, “a lot of feedback,” and the company said it plans to implement a new solution called Enterprise Frontier Safeguards instead. What do you think of that news? Did the popular uprising succeed?

I think so. As I put it on Stratechery, the release is top of Techmeme: biggest news, everyone's talking about it. Techmeme with social media commentary—there are like 75 posts from Twitter about it. And you scroll down, down, down, down, down, and there's the news of this on the page. Number 1, clearly no one cares. Number 2, there are no X comments and 8 LinkedIn comments. And I'm like, “This is my world.”

Mm-hmm.

Because—

Yep. You just have—

Apparently. But this is, by far, the most interesting part of the release to me.

Mm-hmm.

You go back to when Fable was released, and I was on it immediately. The first day I wrote about it, and I wrote another article about it the next week. This is nuts.

Right.

The idea that they are no longer going to honor zero-data-retention policies is insane. And it's not just insane in the context of AI; it's insane in the entire history of enterprise computing.

Mm-hmm.

The foundational divide between consumer and enterprise is that enterprises are obsessed with owning their data, and they would never, ever in a million years tolerate their vendors taking their data for themselves.

Right.

Never. It would never, ever happen. Consumers don't care.

I don't really care.

And that's why you get ads and you get free apps. That's the trade-off. And this is why it's so hard for consumer companies to serve the enterprise, and vice versa, to be honest, but particularly consumer companies going to enterprise. When I was talking about Dropbox in the context of a company that wanted to be a consumer company but had to become—

Mm-hmm.

—an enterprise company, part of that effort was building this entire infrastructure to manage access to data and—

Right.

—who gets what—

To store proprietary data safely.

—who can see all these sorts of things.

Yeah.

Yeah. The core of enterprise software is who gets access to what data.

Mm-hmm.

And the idea that Anthropic was going to just unilaterally change that was incredible.

Right.

And, like we talked about, as far as—

Under the auspices of safety, because that was the justification—

That's right.

—for it a couple of months ago.

That's right. And what happens under the auspices of safety? You start rolling down that slope: We need to hold this data in case an incident happens. We need to scan it for abnormalities. Oh, do you know how we stop this? If we trained on this data, we could better understand what's going on.

Mm-hmm.

Suddenly they have access to the greatest asset in the world, in the context of AI.

The greatest asset that these companies have, certainly. Well, that's part of what made it so remarkable—

All of the—

—you step back. Anthropic is potentially coming for any number of these businesses and these industries.

Absolutely. Absolutely.

And so the idea that you're handing over your data—

It was insane.

Yeah.

It was. And people—I've referenced this before—no one's buying Fable, partly because it's expensive, partly because these data policies are anathema—

Mm-hmm.

—anathema? Anathema.

Anathema.

Did I say that right?

You nailed it.

This rundown has me nervous. I saw the question at the end. Anathema to any enterprise. And so it doesn't matter. It was so fascinating because Anthropic was throwing down the gauntlet.

Mm-hmm.

We're going to make a model that's so good, not only is the U.S. government going to come shut us down, but you're going to—

You can't afford not to use it.

—give up on not having zero data retention because you need to have this model.

Mm-hmm.

And it failed. Here's the thing: If you want to say, “Oh, no, Anthropic always was going to build this new system—

Mm-hmm.

—which is basically storing your data wherever you want it. You just have to hold it for at least 30 years. And I think you have to give Anthropic some sort of automated access to search for anomalies. Even that is probably a bridge too far, but at least the data is on your servers, not on—

Right.

—Anthropic's, where they can say, “Oh, we changed our policy again. Now we're training on it.” At least that's not possible. Not only is that going to happen, and you get... But number 2, it's not launching for a few months.

Mm-hmm.

Which means this wasn't something they were working on and it just wasn't ready at Fable's release. And until the time it is available, Fable now operates with zero-data retention. So what happened to the safety concerns?

That's the issue. There was no—

It's amazing.

—imminent concern justifying all this.

It's amazing. They absolutely made a play for—

Mm-hmm.

—all of enterprise data. And the fact that they—I called it at the time—they cloak everything in safety. They believe it internally, with this insane understanding of AI, which we're going to get to in a little bit.

Right.

And it so perfectly serves their business needs. The audacity of coming for all of enterprise's data so that you can make your model better, take over their business, inspire Satya Nadella to have a blogging career—

Mm-hmm.

—have Alex Karp give his unhinged rant on CNBC, which was quite delightful. This was all triggered by this—

I couldn't sit through all 10 minutes of the Alex Karp rant. I heard good things. It is what it is.

The fact that they will make it zero-data retention until this new program is available is the tell.

Mm-hmm.

This was not about safety. If it was about safety, you would not do it. You would just say, “We're going to continue with our policy until this is available.” But no, the policy's gone.

Right. Well, do you think that OpenAI's entry into the enterprise space, as aggressively as they've been moving in that direction over the last couple of months, makes it easier for companies to say, “Look, we're just not going to use Anthropic if that's the policy?”

Yeah, absolutely. Absolutely.

Yeah.

That's why competition is amazing. And by the way, why is Anthropic doing this? Because they need to IPO, because they need to make money. That's why they're priced the way they are. This goes back to one of my longest-standing takes about tech in general.

Mm-hmm.

And one of my longest-standing criticisms of these AI companies, particularly OpenAI.

Yeah.

The best route to alignment from a societal perspective is the for-profit enterprise.

Mm-hmm.

The for-profit enterprise—you understand why they do things. You understand they have fundamental constraints on their behavior.

The market is a check.

The market—

Sure.

—is a check, and it is the only check that scales to the size and power of these companies.

Mm-hmm.

Even Facebook, where Mark Zuckerberg has total control, has to occasionally pull it back—

Right.

—and say, “Okay, we're not going to go that far. We're going to step back. We're going to...” The fact that we're going to start a nonprofit for this, that is what concerns me—

Mm-hmm.

—not having a check. But the check should not be the good intentions of the people in charge, though that always ends badly. The check has to be external.

Right.

And the external check—the only one that works—is the market. God bless capitalism. God bless free enterprise. God bless Anthropic's IPO. This is evidence. Yeah—

Absolutely.

—the IPO's definitely coming because they need enterprise contracts.

Well, stepping back and watching the way this has played out over the last couple of months, I don't think it plays out the same way if OpenAI isn't going hard after the enterprise space. And also—

Right.

—they, too, were checked by the market and forced to go in that direction and find a bunch of revenue in enterprise that they weren't finding elsewhere as they—

We're not doing Sora or whatever it might be.

Right.

Yeah, no, absolutely.

So the market pushed OpenAI in the right direction as well.

4. The OpenAI Hugging Face Hack

All right, another topic of conversation this week. Over the weekend, Dwarkesh Patel published a post on the OpenAI–Hugging Face hacking incident. All right, and that is the end of the free preview. If you'd like to hear more from Ben and I, there are links to subscribe in the show notes, or you can also go to sharptech.fm. Either option will get you access to a personalized feed that has all the shows we do every week, plus lots more great content from Stratechery and the Stratechery Plus bundle. Check it out, and if you've got feedback, please email us at email@sharptech.fm.