Latest / THE DAILY BRIEF / Anthropic Targets Enterprise Growth with New AI Model

Transcript

Anthropic Targets Enterprise Growth with New AI Model · THE DAILY BRIEF

• 0:00
  Speaker 1: Thirty hours straight, is how long that it can code
• 0:04
  Speaker 1: on its own? What are the technical feats needed to
• 0:06
  Speaker 1: be able to go that long where humans can well
• 0:09
  Speaker 1: definitely not survive that unless as a whole load of
• 0:12
  Speaker 1: caffeine involved.
• 0:13
  Speaker 2: Good morning. I think one of the main advances we
• 0:15
  Speaker 2: made was around memory and what we call context management.
• 0:18
  Speaker 2: So if you think about how you know human works
• 0:20
  Speaker 2: for longer periods of time, you're writing things down, You're
• 0:22
  Speaker 2: making sure you can always pick up where you're left
• 0:24
  Speaker 2: off if you're coming back the next day. So with
• 0:26
  Speaker 2: clouds on at four point five, we did a lot
• 0:28
  Speaker 2: of work on that memory management. So the model you
• 0:30
  Speaker 2: think about, it sort of writes down what it's doing,
• 0:33
  Speaker 2: keeps track of its state, and then if it needs
• 0:35
  Speaker 2: to sort of backtrack a table to then keep going.
• 0:37
  Speaker 2: And that's how it's able to stay coherent for a
• 0:39
  Speaker 2: much longer period of time than any other of our models.
• 0:42
  Speaker 1: How much have you managed to lower therefore those ideas
• 0:46
  Speaker 1: of inaccuracies or more broadly, that well they're making things up.
• 0:50
  Speaker 1: Hallucination has always been the key issue here and been
• 0:53
  Speaker 1: something that's limited EDGENTICAI adoption.
• 0:56
  Speaker 2: Yeah, this is the model that we have that is
• 0:57
  Speaker 2: also besides being our most powerful, is also are safest
• 1:01
  Speaker 2: and most coherent, so as the lowest hallucination rate, and
• 1:04
  Speaker 2: is the least susceptible to things like jail breaks. And
• 1:06
  Speaker 2: I think that matters a lot. I think I tell
• 1:08
  Speaker 2: you my product team all the time. It's no use
• 1:10
  Speaker 2: going for twenty thirty hours if you're making mistakes along
• 1:12
  Speaker 2: the way, and so having it be both accurate producing
• 1:15
  Speaker 2: good code that's the prerequisite, and then you can focus
• 1:17
  Speaker 2: on scaling up on the time horizon.
• 1:19
  Speaker 3: Mike, can we talk a bit about the audience for
• 1:21
  Speaker 3: Claudes so on IT four point five. The real emphasis
• 1:24
  Speaker 3: from Entropic from the early days was enterprise customer as
• 1:28
  Speaker 3: opposed to a sort of direct consumer. But this field
• 1:31
  Speaker 3: of tools for the developer is expanding. It's probably more competitive.
• 1:35
  Speaker 3: Who are you hoping uses this?
• 1:39
  Speaker 2: It's really we've taken a business focus, but that also
• 1:41
  Speaker 2: manifests kind of in the prosumer space, and so we
• 1:43
  Speaker 2: have a lot of what we call power users who
• 1:46
  Speaker 2: might be developers or might just be early adopters who
• 1:48
  Speaker 2: want to bring AI to their work. So one of
• 1:50
  Speaker 2: the things that Sonnet four point five can do, along
• 1:52
  Speaker 2: with writing code is also creating really professional looking word
• 1:57
  Speaker 2: and PowerPoint and Excel documents. It actually uses the same
• 2:00
  Speaker 2: coding capabilities under the hood, but not to write code,
• 2:02
  Speaker 2: but instead to produce documents. And that sort of capability
• 2:05
  Speaker 2: means that we're starting to see adoption in the enterprise
• 2:08
  Speaker 2: as well.
• 2:09
  Speaker 3: Mike, you and I've discussed this in the past. We
• 2:11
  Speaker 3: know Mike Krieger as Mike Krieger CTO Instagram, And what
• 2:16
  Speaker 3: I'm seeing right now in the field of your peers
• 2:18
  Speaker 3: is the reports on open Ai and what Meta is
• 2:22
  Speaker 3: doing in social media and video. Is that a direction
• 2:26
  Speaker 3: you want the product team to take, Claudin.
• 2:29
  Speaker 2: We're focused much more on the on the productivity use case,
• 2:32
  Speaker 2: and so when I think about our roadmap, it's very
• 2:34
  Speaker 2: much how do we take workoff people's hands, or how
• 2:36
  Speaker 2: do we accelerate books and make them, you know, the
• 2:38
  Speaker 2: work the best that it can be. How do we
• 2:40
  Speaker 2: automate your work in the browser. So much more on
• 2:42
  Speaker 2: that productivity side of things, and I don't think you'll
• 2:44
  Speaker 2: see us play very much in that entertainment space, Mike.
• 2:47
  Speaker 1: That productivity perspective, it's come under some concerns recently. Think
• 2:52
  Speaker 1: about the MIT report everyone suddenly talking about this is
• 2:55
  Speaker 1: only well ninety five percent of the test. We're basically
• 2:59
  Speaker 1: failing out there in the wild. How are you making
• 3:02
  Speaker 1: sure that enterprises adopt your products but actually see the
• 3:04
  Speaker 1: productivity gains from them.
• 3:07
  Speaker 2: I think this is really important, where if you know,
• 3:09
  Speaker 2: AI gets brought into the workplace without the right either
• 3:12
  Speaker 2: tools around it or enablement, what you end up with
• 3:14
  Speaker 2: is this disillusionment a couple of months later around well,
• 3:17
  Speaker 2: folks aren't adopting it, or yeah, it helped me a
• 3:19
  Speaker 2: little bit, but not enough. And so we have a
• 3:20
  Speaker 2: lot of emphasis on let's make sure the work is
• 3:23
  Speaker 2: actually good. You know, you might hear this word online
• 3:25
  Speaker 2: like slop, where AI is creating work that actually just
• 3:28
  Speaker 2: is not very good. And I think of us, we're
• 3:30
  Speaker 2: trying to produce the anti slop work that actually, you know,
• 3:32
  Speaker 2: maybe gets you eighty percent of the way there, but
• 3:34
  Speaker 2: it's eighty percent that then lets you complete the work
• 3:36
  Speaker 2: in a way that you're proud of rather than you know, ops,
• 3:39
  Speaker 2: it did something, but now I feel like I have
• 3:40
  Speaker 2: to start over because it didn't really help. So I
• 3:42
  Speaker 2: think that's the really key piece for enterprise adoption, and.
• 3:45
  Speaker 1: Therefore, does it remove the need for so many people
• 3:49
  Speaker 1: or ultimately there's still this argument this can augment but
• 3:52
  Speaker 1: will it start.
• 3:53
  Speaker 2: To replace min We think a lot about what, you know,
• 3:56
  Speaker 2: the comparative advantages are of people, you know, as relates
• 3:59
  Speaker 2: to AI, there's a lot of still relationship building and trust,
• 4:03
  Speaker 2: critical analysis and strategy that really comes on the human
• 4:07
  Speaker 2: side of things, and so we really try to design
• 4:09
  Speaker 2: tools that as much as possible play up those parts
• 4:12
  Speaker 2: of that human AI interaction, knowing that there will be,
• 4:15
  Speaker 2: you know, laborships that are almost inevitable, but if we
• 4:17
  Speaker 2: can design our products along the way to maximize both
• 4:20
  Speaker 2: people's understanding of AI but also their use in a
• 4:22
  Speaker 2: complementary way.
• 4:24
  Speaker 3: Mike open a eyes holding dev day next Monday. It's
• 4:27
  Speaker 3: probably on your calendar for peripheral awareness, but I'm very
• 4:32
  Speaker 3: conscious that you're kind of speaking to us twenty four
• 4:34
  Speaker 3: hours after the news of clouds on IT four point
• 4:36
  Speaker 3: five came out. Have you any data on this sort
• 4:39
  Speaker 3: of reaction to it, early demand and where that's coming from.
• 4:45
  Speaker 2: It's been really interesting how quick people are to adopt
• 4:48
  Speaker 2: a new model. So by I think about one pm yesterday,
• 4:50
  Speaker 2: we already had more usage of clouds on at four
• 4:52
  Speaker 2: point five than all of our other models combined, which
• 4:55
  Speaker 2: really speaks to the eagerness of a lot of these
• 4:58
  Speaker 2: startups that are building on top of our models, as
• 5:00
  Speaker 2: well as early adopters to on day one you saw
• 5:03
  Speaker 2: GitHub and Cursor and Windsurf and many of these products
• 5:06
  Speaker 2: that build on top of our models want to incorporate
• 5:09
  Speaker 2: clouds on at four point five, and so we had
• 5:10
  Speaker 2: this really early crossover moment as well.
• 5:12
  Speaker 3: That is interesting data. Correct me if I'm wrong, But
• 5:17
  Speaker 3: this model is running on Project Rainier, right, is that
• 5:20
  Speaker 3: sort of operationally and infrastructure wise where the training and
• 5:25
  Speaker 3: now inference of it is being done.
• 5:27
  Speaker 2: So we do our both training and inference across You know,
• 5:30
  Speaker 2: we have partnerships with Google and Amazon, but we have
• 5:33
  Speaker 2: a significant part of this model being served now from
• 5:38
  Speaker 2: Amazon as well, and we're seeing a lot of growth
• 5:40
  Speaker 2: on a tobospdrock as well.
• 5:42
  Speaker 1: Just going to that infrastructure layer. You're obviously the product
• 5:45
  Speaker 1: visionary here, but you need to have the energy the
• 5:48
  Speaker 1: compute to bring your products to life. Many worrying that
• 5:51
  Speaker 1: we're in some sort of bubble cycle around AI. How
• 5:54
  Speaker 1: do you think about that as you drive this business forward?
• 5:56
  Speaker 2: Might I think there's this combined need to both scale
• 5:59
  Speaker 2: up for the training side but also on inference. And
• 6:02
  Speaker 2: as we've scaled, especially with our business arrangements and the
• 6:04
  Speaker 2: companies building on top of son It, I think we
• 6:06
  Speaker 2: now have a sort of forward looking perspective on what
• 6:10
  Speaker 2: our inference needs will be, and I think that'll let
• 6:12
  Speaker 2: us go out and also secure the kinds of compute
• 6:14
  Speaker 2: deals that we need to both feel the training but
• 6:16
  Speaker 2: also have that sort of revenue generating inference side as well.
• 6:20
  Speaker 1: And we get so focused on the compute needs of
• 6:22
  Speaker 1: the United States, but we've been talking a lot about
• 6:25
  Speaker 1: that in Europe. How it's scaling in the UK. From
• 6:27
  Speaker 1: your adoption, how are you seeing things different globally MIC,
• 6:30
  Speaker 1: Because those that are actually deploying what's happened with Sonnet
• 6:33
  Speaker 1: four point five and the latest models, we see this.
• 6:36
  Speaker 2: A lot in terms of our global footprint. There's something
• 6:38
  Speaker 2: we started expanding earlier this year, and so for our
• 6:41
  Speaker 2: rollout of trainum too, which is the chip that Amazon
• 6:44
  Speaker 2: has built for AWS that we use pretty extensively for
• 6:47
  Speaker 2: our cloud models. A lot of that deployment is actually international,
• 6:50
  Speaker 2: and when I go to Europe, for example, I hear
• 6:52
  Speaker 2: a lot of questions about data locality and making sure
• 6:55
  Speaker 2: that inference is happening in local data centers. And the
• 6:57
  Speaker 2: only way we're going to be able to do that
• 6:58
  Speaker 2: is to have that international footprint of these chips. And
• 7:01
  Speaker 2: so you've seen the same in APAC.
• 7:04
  Speaker 3: Mike. We are going to ask you a question about
• 7:06
  Speaker 3: talent wars, but I'm just going to make an appeal
• 7:09
  Speaker 3: to you to just be honest with me on this.
• 7:12
  Speaker 3: How big a factor it is or isn't for you
• 7:15
  Speaker 3: right now in the product team Anthropic in particular, I'm
• 7:18
  Speaker 3: looking at the pace at which open ai is putting
• 7:21
  Speaker 3: stuff out, Meta is putting stuff out. Just through your experience,
• 7:25
  Speaker 3: what's the talent situation right now?
• 7:29
  Speaker 2: I'm seeing much more of that talent sort of you know,
• 7:32
  Speaker 2: back and forth happen within the research side in general,
• 7:35
  Speaker 2: a little bit less on the product, although I think
• 7:36
  Speaker 2: there's some key hires where that's been the case. One
• 7:39
  Speaker 2: thing that's been a positive sort of maybe surprise or
• 7:43
  Speaker 2: just outcome of how mission oriented a lot of Anthropic,
• 7:46
  Speaker 2: a lot of the Anthropic team really is, has been
• 7:48
  Speaker 2: It's affected us very minimally in terms of that back
• 7:50
  Speaker 2: and forth that you're seeing maybe among some of the
• 7:52
  Speaker 2: other frontier labs, which is very encouraging. Of course, you
• 7:55
  Speaker 2: have to continue to make sure we build a great
• 7:56
  Speaker 2: culture and maintain that mission alignment. But so far it's
• 7:59
  Speaker 2: been it's been minimally affecting us.
• 8:01
  Speaker 3: If we take son It four point five is the
• 8:03
  Speaker 3: case study, what were the types of roles that you
• 8:06
  Speaker 3: needed to bring in to roll out the release.
• 8:10
  Speaker 2: I think people think of, you know, research and model
• 8:13
  Speaker 2: science as being fairly cut and dry. I actually think
• 8:15
  Speaker 2: that there's a lot of art and taste to it
• 8:16
  Speaker 2: as well. You are making a lot of decisions from
• 8:19
  Speaker 2: a research engineering perspective around what are the task the
• 8:22
  Speaker 2: model needs to improve on, how will it improve on that,
• 8:24
  Speaker 2: how will we know that it's improving on that? And
• 8:26
  Speaker 2: so a lot of that reinforcement learning post training piece
• 8:30
  Speaker 2: is the key shape of what we really thought about
• 8:33
  Speaker 2: in so on It four point five