Shirin Ghaffary on Anthropic's safety culture, Meta's Scale AI deal, and the new CEO pick

they pick the machine. More often than not, we've crossed the threshold. Uh, we've been talking about the economic touring test for a while. I'm glad that he wrapped it in a bow. Maybe it needs its own coinage. Maybe we need to dig into anthropic uh more deeply, but we have the perfect guest for that.

We have Sharon from Bloomberg joining us right now. Welcome to the stream, Sharon. How are you doing? Good. Thanks for having me. Thanks so much for being here. Uh I wanted to have you on because you've been on an absolute tear uh writing fantastic pieces.

Maybe it would be best to start with Anthropic, which I was just talking about. Um what did you learn from uh the the piece that you wrote about anthropic and Daario? We haven't had a chance to have him on the show. We talked to Schultto. Uh they have an interesting culture, interesting history.

Um can you give us your kind of 50,000 foot foot view on the company, what they stand for today and maybe how that's changed?

Yeah, I think Anthropic is a company that probably the most right now is um facing pressure both to live up to its really strong kind of um ideological ethos which is around building AI safely as an and as an alternative to labs that they felt maybe weren't doing it as responsibly as they were and then also to deliver on the business value.

Right? So, I was kind of um I learned just how sort of savvy actually Anthropic has been with Dario at the helm. Even though he has this very deep research science background, he's been able to cut some major deals, grow the enterprise business quite quickly.

Um double and now there's other reporting about tripling revenue. So, you know, that's that's sort of where they're at. But as you get more money, as you get more you get bigger and bigger funding rounds, you know, how does does that translate into pressure to move really fast?

Um, at the same time that you're trying to be responsible. Did you get a sense that they have a they have a specific strategy around diversification in terms of hyperscaler partnerships because they've they've done a deal with Amazon. I believe they were just mentioned at WWDC is plugging into Xcode.

it feels like they're trying to uh maybe go more of as an infrastructure provider. Um they seem to be doing very well in in coding and with uh on the enterprise side. Uh how much of that is deliberate or just like uh it got pulled out of them because of where their position in the market was?

I mean so am I reporting like it was very deliberate their strategy to first of all not lock in with any one hyperscaler right so they have both Google and Amazon as partners on that front.

um that kind of hedging I think was very calculated and also um they you know Dario in my interviews he's been pretty dismissive our reporting about Stargate um you know saying it's not clear exactly what it is and I I think his stance is that sort of you can you can scale up and not need to make such a big deal about it at least that's what he sort of said publicly to me right um but but as as you said like I think the pressure is on right for them to to deliver and show that they can he invented scaling laws but can you build up the compute and pay for it to do that.

Yeah. Can you talk a little bit about the rhetoric around safety?

Um, I was very I I never really got caught up in the paper clipping and the AI doom narrative, but after seeing uh Deep Seek that's what every human says right before they get but but uh but after seeing Deep Seek uh drop and the anthropic paper about uh Manurion candidates or what was it secret hidden agents?

I forget the term that they use but essentially sleeper agents embedding uh something nefarious within the weights of the model that maybe only comes out when it interacts with a certain endpoint or a certain web page. It it discovers that hey, I'm not just running, you know, in some random solo dev's cursor instance.

I'm actually inside the NSA. I should go around and do something else. Um and and that really flipped a bit in my mind to say, okay, safety research is actually maybe extremely important and maybe even extremely commercializable because people will want to do safety evaluations on the models that they use.

Uh how has the safety rhetoric evolved at uh at Anthropic over the last couple years in your in your in your reporting? Yeah.

So I think you know AI doomers or people worried about AI safety have been doing these sort of um hypothetical thought experiments like the paperclip experiment for a very long time like predating even Dario Emmed is right entry onto the scene.

Um however I think at Anthropic what I've seen evolve is they're starting to to outline and try they're trying to be as concrete as possible to explain like what the risks are what the different levels of risks are.

or they made a whole um responsible scaling uh kind of document framework to say okay it's it's kind of modeled after how um biological labs are that are very high risk and how for okay if if what you're working on is this level of biological risk then you have to have this corresponding level of actual physical lab safety so that's how they viewed it anthropic and so you know when I was reporting this out they had recently had this kind of false alarm where they thought they were crossing into a level that was too dangerous for them to be able to handle And then they did extra tests and they're like, "Okay, no, we're all right for this release, but we better get up to par.

" What time period was that? Was that like an hour or like 10 days or a month?

So, you know, Daria told me they basically had to delay um not this release, but the the last bike one um by something like nearly a week because at the la sort of in the final days leading up to the release, um their safety team, you know, one of the red teaming uh group came to them and were like, listen, like we this model actually may be more capable than we thought, right?

Um so I do think that's that's a challenge, right? is taking these really hypothetical, seemingly kind of sci-fi scenarios and translating it to like, well, when is that actually going to happen at our lab and what do you have to do to protect yourself?

And it's both physical security, like making sure no one can break into your lab or something like that, but it's also um you know, actually making sure that there are safeguards on the software side to stop someone from say and I think I think the most tangible threat that I heard them talk about was the bi biological one, right?

And that was a specific concern with that release I mentioned that could someone use this to come up with a novel boweapon. Sure.

There there's an there's an interesting dynamic with safety teams where if you're sitting there with a team of let's call it 20some people and you're costing the company tens of million you know I don't know$10 million a year.

It's not a you're not going to feel that like you're you're your job your job's that safe if you're if you're never sounding the alarm.

there's kind of an oh this feels a little dangerous let's look into it you know there's also just so many other dynamics at play and I think people always dig into this like uh you know our models too powerful that's a great marketing line oh it's too good like you know you could never give it to the the wild it's too powerful uh and then there's also um it's also maybe you just haven't put together all the inference capabilities and scaled up your data centers to really have a have a productive launch you need an extra two weeks in the oven.

Oh, yeah. It's a safety issue. So, there's a million different ways that the safety stuff can be kind of abused. I'm not accusing them of abusing it, but it does feel like uh you know, we we noticed this with uh with with the Llama 4 launch.

uh they haven't been able to release Behemoth yet and they had some some issues with uh benchmarking but notably they didn't say it was a safety issue and that would have almost been an easy out and it's one that I think most foundation model companies would have taken a year ago but the the the community the developer community just doesn't seem to buy that line of reasoning as much anymore.

I don't know if you had any reactions to the to the long I I have you know in my own interviews with Daario I've asked him about this and um you know he has been upfront like we we reported on you know there were delays to um just this latest big release right the the latest uh sonnet they're kind of or sorry is it opus whatever the biggest one is I'm forgetting all the specific names right now and um you know we we have report on on these delays we were we wrote you know um a big article about how um scaling laws at least with the pre-training side of these models were starting to slow down and you know anthropic in my in my reporting with Dario he was pretty upfront that like that wasn't a safety like you know that big delay was not that big model being um later than some of their early projections that's because they were facing kind of the same technological problems that I think many labs are facing right and ultimately they did release it so but I think people you know there it's absolutely reasonable to question any company's motivation when they say something slowed down I think I saw no evidence for that in that this specific case sure Sure.

Yeah, that makes sense. Um, uh, let's talk about the most recent deal. Uh, Meta acquiring Scale AI. Uh, 49% of the company is changing hands, I believe, at around 14 or 15 billion dollars. Spiritually, the whole thing. Yes.

But technically, uh, what what what was what has been your reaction and what uh what have you learned from reporting on that? Yeah, I think you just covered that that uh they they have a new CEO. So, so break. We just reported that. Yeah.

per um you know for our sources that they have picked uh the new CEO to be Jason Droi uh former VP he's currently uh you know at scale being promoted and and he was um formerly at Uber a VP and formerly also a partner uh in VC at benchmark so but you know that's just that's just the latest right we we kind of we have really been breaking the news on this front we had a scoop this Saturday night um that we were the first ones to report that these uh talks were happening for Meta to make this multi-billion investment in scale I think it really caught you know a lot of people by surprise in the industry.

Um this is a huge investment planned um on Meta Front, right?

probably their biggest ever um in a company and I think it's you know I think it's a sign that Zuckerberg is ready to go full on founder mode and try to catch up on AI and this last llama 4 release was you know widely considered a disappointment and so how do you how do you change course bring this big right big investment and this super intelligence lab and he's recruiting engineers we've reported on that they're making crazy high offers, right?

Um they they will open Mazak will open the the checkbook for the top talent. Yeah. I think the interesting thing that stood out about some of you know people were just reporting the headline numbers of saying, "Oh, you could make, you know, 10 nine figures was the number. " Yeah. Nine figures.

But you have to look if if you're trying to recruit somebody that joined OpenAI 2, three years ago, they're they're probably run rating at $10 million a year. Yeah. already on on a on a full um fully loaded comp basis. So I think it makes sense. What have you learned about the future of the scale business, right?

Is Jason going to be, you know, pounding the pavement, trying to get, you know, more customers for the core business, continue to adapt the core business. Yeah, there's a lot of scale competitors right now that are saying like, "Hey, we're going to take a lot of they're foaming at the mouth.

" Yeah, we, you know, we're we're going to use this to our advantage when we go to the other hyperscalers and say, "Do you want to do business with a company that's owned by Facebook or Meta? " Yeah, absolutely. I think um it's really unclear what's going to happen with the future of scales business.

um you know obviously the meta investment you know that's expected you could see partnerships happening there on that front but what will happen with the other customers I think we don't know um and and what will happen with um yeah other competitors in the space could they start to right gain ground um as scale deepens the ties with Meta we'll see I think what we do know you know we we have reported that um Alex Wang plans to to join Meta um I think for Meta he is seen as someone who is just sort reinvigorating right their AI team and this renewed AI effort.

Um he's a big name. He's incredibly savvy. Um you know very well connected and uh built a real business with scale. So you know I think that's where we're in the middle of you know of the deal. We're waiting you know to see the all the official details sort of as this plays out.

Uh earlier this week, Bill Gurley was talking about the M&A markets and why we're still seeing these kind of odd zombie acquisitions happening. Uh tech was laying the blame at the foot of Lena Khan for a long time. Lena Khan's obviously out. Why do you think the big tech companies are still using that?

She's still posting though. She's still she's still reposting stuff about being anti- monopolistic. So she's she's, you know, she's around.

Is it is this more just like this is just a more efficient acquisition vehicle generally even without even in a different administration and so it's just going to be continually used. I don't think this is an asset that Meta would want to own just generally, right?

You wouldn't want to fully own because you're you're you're you're dealing with tens of thousands of offshore contractors, right? Is that something that Meta wants? But they already do that with those. Sure. But but but they've had issu sure they but they've also had issues with it, right? It's a liability.

Anyway, sorry. So, of course, you know, um Lena was is known to be, you know, aggressive on going after, you know, monopolistic behavior and especially in the tech market. Um doesn't necessarily mean that the current FTC is not going to scrutinize these kinds of deals, right?

I would say it's way too soon to have any definitive take on how this FTC will act around these kinds of deals.

I think you know companies never want to um raise too many eyebrows with regulators and um even if people are from different political parties we have seen Republican controlled FTC sometimes you know I mean that is part of their job right is to review these sorts of large transactions so um that may be why you know we're seeing we're continuing to still see you know companies not want to go too far on that front.

Yeah. Yeah. Uh c can we move over to Microsoft? It feels like Sachi Nadella has carved out a ton of territory in the AI space and uh now the game is about holding on to it. Uh it seems like they might be back to training their own language models.

At the same time at the last Microsoft build keynote he was distinctly highlighting how model agnostic Azure was and how they would be happy to route you to any OpenAI model, any DeepSeek model, any Llama model. they they they want to be Switzerland, but at the same time uh they want to have a horse in the race.

So what's the latest thinking around Microsoft? I mean I report on Microsoft more in the context of OpenAI. Um so I'm not the Microsoft expert. I will say that yeah obviously we've seen um we've seen more attention around the Microsoft OpenAI relationship lately and what that means.

I think OpenAI is also diversifying, right? They're working with other um hyperscalers now for Stargate. They have Oracle on board. they have soft bank financing. Um, so this is no longer just like an exclusive relationship between Microsoft and OpenAI and I think that goes both ways. Yeah.

What is the dynamic right now in the reporter uh in the reporter talent space? It seems like the value of just being a pure play writer is maybe going down, but the value of being able to acquire, you know, new net new information and actually get scoops is is higher than ever. What is the dynamic right now?

We've seen, I think, a lot of different shakeups and and layoffs at different companies, but I'm curious what you're seeing.

I mean, I I think that the value of reporting, like a real original reporting, and going, getting scoops, getting people to share with you information that isn't public, um, that takes a lot of human skill that at least in my job, I, you know, I don't think that's replaceable by AI uh, for now. Yeah. Yeah.

That that that that's what I was getting at. Um, I do think at some point maybe you would have an army of agents that would just reach out to, you know, 200 employees at a startup at at, you know, a single given time. Um, but, uh, but yeah, that that that tracks. Yeah.

I in in terms of the Stargate uh project, uh, were you able to go visit as part of the reporting? Um, can you give us kind of the lay of the land on on on what uh who who the who the major players are and what the most interesting angles to dig into in the pro in the progress in Abene uh has been. Yeah, for sure.

Well, shout out to my colleague Brody Ford. I don't know if he's listening, but I'll send this to him after. He actually went out to Abalene for me because I was I was unavailable the one day that we could go or whatever. Um, and it's astounding. I mean, I saw the pictures. I got the live feed. Um, it is expansive.

Like I think it's hard to understand just the vast scale of it until you really see pictures, video or go there.

Um and you know I think that uh there was a lot of like there was sort of a lot of people doubting if Stargate was even real and I think there's still a lot we don't know about the future of this project like exactly you know where the other sites will be. They've said they plan 10 more at least in the US.

We don't know exactly like how the financing will work.

Um Stargate we talked to Masasan for this who's obviously you know Soft Bank's leading the financing they said it's project by project we'll go step by step um but I think what we learn can learn from abene is like at least that first site that abene site is very real like it is happening there it is being built out we've seen the pace of it um so you know that could end up being kind of a template or a test case right for how how um how much target can can execute in the future on all of its promises that makes a ton of sense anything else makes a ton of sense No, thank you so much for hopping on.

This was fantastic. Reach out when you got your next scoop.