Anthropic CPO Mike Krieger: Building AI Products From the Bottom Up
Summary
TLDRIn a discussion led by Mike, the Chief Product Officer of Anthropic, the conversation revolves around the future of AI content creation and the evolving role of AI in storytelling. Mike emphasizes that regardless of the medium, the core of content creation lies in the story and the connection with the audience. He discusses the shift towards AI-generated content, suggesting that the distinction between AI and human-generated content will diminish as AI becomes more prevalent. Mike shares insights into product development at Anthropic, highlighting the importance of addressing real user problems and fostering creativity in product design. He also touches on the evolution of coding models and the integration of AI in various workflows, stressing the need for better user experiences and the potential for AI agents to interact autonomously. Overall, the discussion reflects a forward-looking perspective on the intersection of AI and content creation, with a focus on user engagement and innovative product development.
Takeaways
- 📖 Storytelling remains central to content creation.
- 🤖 AI-generated content will dominate the landscape.
- 🔍 Understanding models is crucial for user control.
- 🛠️ Product development should solve real user problems.
- 💡 Creativity in product design is essential.
- 👩💻 Coding models are evolving rapidly.
- 🔗 Integration of AI in workflows enhances productivity.
- 📊 Compute power is vital for AI advancements.
- 🌐 AI agents may interact autonomously in the future.
- 📈 User experience needs to improve for AI tools.
Timeline
- 00:00:00 - 00:05:00
Mike, the chief product officer of Anthropic, discusses his background, including his brief time as a founder at Sequoia and his involvement with Instagram. He emphasizes the importance of storytelling in AI content creation and the need for models to provide users with control and understanding of the content they generate.
- 00:05:00 - 00:10:00
Mike elaborates on the product development framework at Anthropic, highlighting the shift from a top-down planning approach to a more bottom-up creative process. He shares insights on the development of products like MCP, which emerged from recognizing commonalities in different integrations, and the importance of community involvement in product evolution.
- 00:10:00 - 00:15:00
The conversation shifts to the future of AI and the role of agents. Mike expresses excitement about the potential for models to work autonomously and interact with each other, emphasizing the need for better protocols and memory systems. He also discusses the challenges of scaling product organization and the impact of AI on internal workflows.
- 00:15:00 - 00:23:58
Finally, Mike addresses the balance between research and product development at Anthropic, stressing the importance of integrating research insights into product offerings. He reflects on the evolving landscape of AI applications and the need for products to be more AI-native, while also considering the implications of agent-to-agent interactions and the future of AI in various industries.
Mind Map
Video Q&A
What is Mike's role at Anthropic?
Mike is the Chief Product Officer at Anthropic.
What company did Mike co-found?
Mike co-founded Sequoia and was involved with Instagram.
What is the focus of Anthropic's product development?
Anthropic focuses on solving real user problems and fostering creativity in product design.
How does Mike view AI-generated content?
Mike believes that the distinction between AI-generated and human-generated content will become less relevant as most content will be AI-generated.
What is MCP in the context of Anthropic?
MCP refers to a protocol developed at Anthropic for better interaction with models.
What challenges does Mike see in AI product usage?
Mike notes that many users find AI products hard to use effectively, especially when first approaching them.
What does Mike think about the future of AI agents?
Mike is excited about the potential for AI agents to work autonomously and interact with each other.
How does Mike use AI in his work?
Mike uses AI as a thought partner for writing and planning tasks.
What is the importance of compute in AI development?
Compute is crucial for training models and balancing research and product development.
What does Mike think about the integration of AI in various workflows?
Mike sees value in integrating AI across different disciplines within organizations.
View more video summaries
Labor slammed for ‘lies’ and ‘broken promises’
I Made a Fortune Selling AI Systems and Here's How
Sindrom Koroner Akut (SKA/ACS) - EKG, Patofisiologi, Diagnosis, Tatalaksana
Kisah Nabi-Nabi #3: Nabi Idris عَلَیهِ السَّلام - Khalid Basalamah
Stocks Market Surge! Is This the Start of a Major Rally?
How to Customize Zen Browser for the Ultimate Experience!
- 00:00:02You all might know Mike of course as
- 00:00:04chief product officer of Anthropic, but
- 00:00:07you were also Sequoia founder at one
- 00:00:08point. Is that right? Yeah. And for a
- 00:00:11hot week. For a hot week. And what was
- 00:00:13that company? It was Instagram.
- 00:00:15Instagram. Thank you. Welcome everyone.
- 00:00:17Mike Lauren, take it from here. Thank
- 00:00:20you for joining, Mike. Yeah, happy to be
- 00:00:21here. Hey, everyone. Um, so those you
- 00:00:23may not know, but Mike is actually a
- 00:00:24content nerd, so it's pretty fun to have
- 00:00:26the AI filmmaker ahead of us. Where do
- 00:00:28you think the world of AI content is
- 00:00:30going? I think regardless of like the
- 00:00:33medium or how much AI is being used to
- 00:00:35create things, I think you'll keep
- 00:00:37coming back to like is there a story
- 00:00:39being told? Is there a person behind the
- 00:00:40content that people can connect to and
- 00:00:42ultimately like react to over time? Um,
- 00:00:44and so it's like another tool in the
- 00:00:47toolbox of a storyteller. Um, and I'm
- 00:00:49curious how you guys think about as you
- 00:00:52build more content, as more pixels get
- 00:00:53generated, how do you help people build
- 00:00:54control? like Enthropic has done a
- 00:00:56really nice job of helping us understand
- 00:00:59models with mechanistic interpretability
- 00:01:01and where are models how do you make
- 00:01:03Golden Gate clawed how do you think
- 00:01:04about giving that option to your users
- 00:01:06and your customers yeah I think you know
- 00:01:09there's probably things that are useful
- 00:01:10at a point in time right now like
- 00:01:12there's you know talk about watermarking
- 00:01:14and like oh is this AI generated but you
- 00:01:16know and maybe this was in the
- 00:01:17conversation earlier today I wasn't here
- 00:01:18in the morning but the majority of
- 00:01:21content will be AI generated so the
- 00:01:22distinction of like was this made by AI
- 00:01:24or not I think is going to be a not
- 00:01:26useful one. Um, I think there will still
- 00:01:28be interesting questions of like
- 00:01:29derivation and providence and and those
- 00:01:31kinds of things. They can get easier uh
- 00:01:34with AI. I mean, it's funny to bring it
- 00:01:36back to blockchain, which I feel like is
- 00:01:37not a cool thing to talk about anymore,
- 00:01:38but like was like theoretically one of
- 00:01:40the problems that was being solved with
- 00:01:41blockchain is probably much more doable
- 00:01:43when like the entire N10 pipeline um um
- 00:01:46is bits. Um but yeah, I think like the
- 00:01:49the things that were important in the
- 00:01:51past world like what did you source like
- 00:01:53is there a citation like when I think
- 00:01:55about like documents is still important
- 00:01:56and more doable now. Um but like whether
- 00:01:59it's AI generated I think is like not
- 00:02:00the interesting question going forward.
- 00:02:02Interesting. And so let's dive into
- 00:02:05Enthropic a little bit and some of the
- 00:02:06products you guys are building there.
- 00:02:08You guys have done a really nice job
- 00:02:09with artifacts with the coding models
- 00:02:11with MCP. I'm curious for you as a
- 00:02:13product officer, chief product officer,
- 00:02:15what your framework is for building
- 00:02:17products and how do you make them how do
- 00:02:19you make them the products better than
- 00:02:21just the model itself? Yeah, I think um
- 00:02:24I guess two thoughts on this. Like one
- 00:02:25is um the things that were useful in the
- 00:02:28Instagram age are still useful now,
- 00:02:30right? Which is like are you solving a
- 00:02:32real problem for people? Like if you're
- 00:02:34creating a developer tool, are you
- 00:02:35enabling you to do something interesting
- 00:02:36and novel and quickly? If you're
- 00:02:38building an enduser product, like are
- 00:02:39you meeting the needs of people where
- 00:02:41they actually are? So I think that like
- 00:02:42remains as important now as it ever has
- 00:02:45been. I think what's different and a
- 00:02:46lesson I had to unlearn is like on
- 00:02:48Instagram we did a much more sort of
- 00:02:50tops down you know 3 to six month time
- 00:02:53frame you know of of planning. Thomas in
- 00:02:55the third row can relate to this. We
- 00:02:57were definitely much more like plan and
- 00:02:58deliver it. I think this is true at
- 00:03:00anthropic and in talking to kind of my
- 00:03:02counterparts in open at other places
- 00:03:04like you just have to allow for much
- 00:03:05more bottoms of creativity because most
- 00:03:07I think the best products are the ones
- 00:03:09that are built very close to the model
- 00:03:10and you can only kind of tell what
- 00:03:11they're capable of like pretty late in
- 00:03:13the process and so I've just learned to
- 00:03:15kind of invert the sort of part of the
- 00:03:18like creative process to be much more
- 00:03:20bottoms up which you know as like a bit
- 00:03:22of a control person it's like a little
- 00:03:24hard but I think it's also like opened
- 00:03:25up some really interesting things like
- 00:03:27artifacts was a research prototype that
- 00:03:30then got taken by like a designer and an
- 00:03:32engineer and then shipped to production.
- 00:03:33I think I've heard that story not just
- 00:03:35from us but from other like creators in
- 00:03:37the space as well. Yeah. Could you
- 00:03:38actually give us some examples? I'd be
- 00:03:39curious maybe MCP is one of the more
- 00:03:41interesting products the whole industry
- 00:03:43is starting to adopt. Where did that
- 00:03:44come from and what's the story? Yeah,
- 00:03:45it's funny for MCB because I actually
- 00:03:47like recently was like I was like, you
- 00:03:49know, half of my job is making memes and
- 00:03:51sharing them internally. And one of I
- 00:03:52was like making a meme around like when
- 00:03:54MCP was created. It was like a twinkle
- 00:03:56in like two people's eyes. And I was
- 00:03:57like went back and it really started
- 00:03:59from like watching us try to implement I
- 00:04:01think at the time we're implementing
- 00:04:02like Google Drive integration and then
- 00:04:03we were imp implementing GitHub
- 00:04:05integration and like those things should
- 00:04:06have more in common than not, right?
- 00:04:08It's like you're bringing context into
- 00:04:09the model. Um and we had done like two
- 00:04:11completely different interp uh like uh
- 00:04:12implementations internally and the third
- 00:04:14one that we were like queuing up was
- 00:04:15going to be like yet another uh
- 00:04:17completely you know bespoke thing and
- 00:04:19you know usually like my general pattern
- 00:04:20is like do things three times and on the
- 00:04:22third time you can try to figure out
- 00:04:24what the abstractions are and this is
- 00:04:25definitely that case where it's like all
- 00:04:26right what is in common here and where
- 00:04:28are things going uh but it definitely
- 00:04:29was not like a top down like we need a
- 00:04:31protocol for like better interacting
- 00:04:33with models it was again two engineers
- 00:04:34being like yes I think this is a good
- 00:04:36idea let's go let's go prototype and
- 00:04:38build it um and then like just spending
- 00:04:40the time like let's make the protocol
- 00:04:42better. let's like make it truly open so
- 00:04:43it's going to get adopted beyond
- 00:04:44entropic because we think there's value
- 00:04:46in not just us like owning a protocol
- 00:04:48but instead it being much more
- 00:04:49standardized um and then iterated on
- 00:04:51from there and now it's gotten much more
- 00:04:53of a community flavor where you know we
- 00:04:56tropic you know we're over a thousand
- 00:04:57people but still feels very startupy
- 00:04:59like we're working with like places like
- 00:05:01Microsoft and Amazon that have like all
- 00:05:03sorts of four-letter acronyms like you
- 00:05:05know I actually was going to cite them
- 00:05:07but I don't even remember some of them
- 00:05:08but it's like deep you know
- 00:05:09authentication like identity management
- 00:05:11with Exchange servers. I'm like, these
- 00:05:13are not the considerations that we think
- 00:05:14of a priority, but they are when you
- 00:05:15actually open it up to a broader group.
- 00:05:17Yeah, that's awesome. And where do you
- 00:05:19think it goes from here? It's been
- 00:05:20interesting to see a lot of the people
- 00:05:21in this room adopt MCP. You guys had a
- 00:05:24new release I think yesterday around
- 00:05:25integrations. Um, so like once you have
- 00:05:27the seed that comes bottoms up, how do
- 00:05:29you nurture it and grow it? Yeah, I
- 00:05:31think the two areas like MCP adjacent I
- 00:05:33get most excited about. One is around
- 00:05:34just taking action. So a lot of like V1
- 00:05:36of these projects was around like how do
- 00:05:38you bring context into the models? um
- 00:05:41like we launch our integrations where
- 00:05:42you can pull in like GitHub, you can
- 00:05:44launch like Zapier actions, but I think
- 00:05:45like the the right mode or like the
- 00:05:47actually taking actions is going to be
- 00:05:49much more uh important going forward
- 00:05:51because ideally you want these things to
- 00:05:52actentically not just in retrieving but
- 00:05:54also being able to automate workflows.
- 00:05:56The second one is like when MCPs and
- 00:05:58just agents more generally interact with
- 00:06:00each other and what the right protocol
- 00:06:01is. It feels early to try to standardize
- 00:06:04this too much. Like I know Google's
- 00:06:05doing agent to agent like I think we're
- 00:06:06still exploring what like the right
- 00:06:08patterns are. Um but that I think is
- 00:06:10going to be very interesting like
- 00:06:11internally we talk about like at what
- 00:06:12point will your agents hire other agents
- 00:06:14and what does that economy of you know
- 00:06:17uh of things even look like. So that is
- 00:06:18what I get really excited about going
- 00:06:20forward. That's awesome. So at this
- 00:06:22point um you guys have done an amazing
- 00:06:23job with your coding products that feels
- 00:06:24like it's more than just bottoms up a
- 00:06:26couple people tinkering with it. I'm
- 00:06:28curious how you think about it as a
- 00:06:30focus and what you guys have gotten
- 00:06:31right so far. Yeah, I mean even coding
- 00:06:34like I have a lot of awe of watching our
- 00:06:36researchers like uh it's also you know
- 00:06:38you can have a top down sort of like
- 00:06:40idea of where to go but so much like
- 00:06:41research innovation comes from like a
- 00:06:43couple people you know pushing the
- 00:06:45boundaries of um RL like Dan was talking
- 00:06:47about earlier right like there's like
- 00:06:48like a lot of these things come from
- 00:06:49discovery and that process needs to be
- 00:06:51pretty much bottoms up. I think a thing
- 00:06:53that we've tried to do well on the
- 00:06:55coding side is um not just focused on
- 00:06:58the benchmarks but also really like is
- 00:07:00it generating code that people like
- 00:07:02working with or is it generating good
- 00:07:03outcomes as well and so that's like a
- 00:07:05thing that we'll we'll definitely
- 00:07:06continue to to push on as well but it's
- 00:07:08been interesting like you know we
- 00:07:10definitely did not coin like the term
- 00:07:11vibe coding I think that that has like
- 00:07:13its natural limit in terms of like but
- 00:07:15it can create interesting things but is
- 00:07:17that like the way you're going to want
- 00:07:18to do an entire like codebase with like
- 00:07:19a team of 100 like definitely not right
- 00:07:21and so I I think we're we're internally
- 00:07:24figuring out what the role of like
- 00:07:25generating code is within our code base.
- 00:07:27We use it a ton. Over half of our pull
- 00:07:29requests are cloud code generated.
- 00:07:30Probably at this point it's probably
- 00:07:31over 70%. But what does that mean for
- 00:07:33code review is something that we're
- 00:07:34figuring like you can get then you can
- 00:07:35get cloud code review or PR but then
- 00:07:37it's like turtles all the way down and
- 00:07:38like at what point do you have that that
- 00:07:40like oversight around like is this like
- 00:07:42going to lead us to an architectural
- 00:07:43dead end? Does that matter if you can
- 00:07:45like overpower the usual like tech debt
- 00:07:47rewrite with um AI coding like we're and
- 00:07:50I think probably other folks like in
- 00:07:51labs working on like coding models like
- 00:07:53kind of patient zero for some of these
- 00:07:55for better and for worse. I was actually
- 00:07:57very curious to hear about some of the
- 00:07:58second order effects of coding agents
- 00:08:00getting much better like code reviews is
- 00:08:03one like I'm curious as more everyone
- 00:08:05can write software where do we go? I
- 00:08:07mean internally I think like what I'm
- 00:08:08realizing is like it makes all of your
- 00:08:10other inefficiencies as a product
- 00:08:11organization like extremely painful
- 00:08:13because now it's like the alignment
- 00:08:15meeting is like it's not just standing
- 00:08:17in the way of like an hour of
- 00:08:18engineering work that would happen. it's
- 00:08:19like standing in the way of like the
- 00:08:20equivalent of like four or eight hours,
- 00:08:22you know, and so I think it's made made
- 00:08:23it like I think our product organization
- 00:08:26is going to break very much so with like
- 00:08:28with faster uh with code jet it just
- 00:08:30makes it very painful when you're like
- 00:08:32like it's even more wasted time with
- 00:08:34like driving alignment and the models
- 00:08:36are not helpful with that really. I mean
- 00:08:38they can synthesize meetings they can
- 00:08:39maybe like tee up the next conversation
- 00:08:41but they're not like they're not yet at
- 00:08:42the point where they're like
- 00:08:44organizationally driving uh decision-m
- 00:08:46interesting. Um I mean you guys are
- 00:08:48using a lot of enthropic at anthropic.
- 00:08:50Um here these are a couple examples. I'm
- 00:08:52curious what are the things that you're
- 00:08:54doing or that you've tried in the last
- 00:08:55six 12 months that everyone here should
- 00:08:57be using with your models or others to
- 00:08:59make them work better. I think what's
- 00:09:00been cool has been seeing like different
- 00:09:02disciplines inside the company that are
- 00:09:04not technical start using the models a
- 00:09:06lot. um and whether that's you know
- 00:09:07people in sales using it for meeting
- 00:09:09prep and you know they start from just
- 00:09:11using like what's available and then
- 00:09:13like some blocker becomes really
- 00:09:14apparent then maybe we'll build
- 00:09:16something bespoke in there. Um so that's
- 00:09:18been interesting but it's still less
- 00:09:20evenly distributed than you might expect
- 00:09:22even at an AI lab. I think there's uh
- 00:09:24even within a team like the salesperson
- 00:09:26that knows how to use it really well and
- 00:09:27the people are like doing it more
- 00:09:28traditionally and like the former person
- 00:09:31might be more effective or more like you
- 00:09:33know hit fewer blockers but it's not yet
- 00:09:35like yes it's like a requirement that
- 00:09:37everybody's going to use. Um myself like
- 00:09:40I just use it as a thought partner. So
- 00:09:42whenever I write anything whether it's
- 00:09:43like a strategy doc or a planning thing
- 00:09:45or performance review like uh I kind of
- 00:09:48re it's almost like in the same way that
- 00:09:50I started feeling weird trying to code
- 00:09:52on flights like after co-pilot where
- 00:09:54you're like oh wait I really feel like
- 00:09:55I'm half the engineer I usually am
- 00:09:57because this thing is not uh helping me
- 00:09:58go through like I feel that way now
- 00:10:00about if I write something and I don't
- 00:10:01have that extra sort of cycle through
- 00:10:04claude I'm like ah this is probably not
- 00:10:05like getting fully vetted. Um earlier
- 00:10:08Sam talked about how people in their 20s
- 00:10:10are the ones they're using these models
- 00:10:11the best. You're definitely closer to
- 00:10:12being in your 20s in your model usage
- 00:10:14which is fun to see. Yeah. Although it's
- 00:10:16like also surprising seeing like how
- 00:10:18people enter the workplace like we've
- 00:10:20been doing more work with universities
- 00:10:21and like you know they'll come into work
- 00:10:24in a very different way in terms of like
- 00:10:26the expectation of how much they're
- 00:10:27going to use Genai and like there not
- 00:10:28being a stigma for it. This is a big
- 00:10:30piece like some of our most successful
- 00:10:31internal products are ones that have
- 00:10:33shared visibility. like we like do a lot
- 00:10:35of things within Slack with with cloud
- 00:10:37integrated with internal tooling and
- 00:10:39I've learned that's really helpful for
- 00:10:40breaking down even at anthropic this
- 00:10:42like ooh did you make that with AI
- 00:10:43versus like yeah I did like it saved me
- 00:10:45like two hours like of course like I did
- 00:10:46other like better things to do than like
- 00:10:48write this performance review or
- 00:10:49something right and so like uh like even
- 00:10:51watching my time in this like last year
- 00:10:53and a half of like oh I don't know about
- 00:10:54like cloud and performance reviews to
- 00:10:56like now it being encouraged is is I I
- 00:10:58think a positive development uh of
- 00:10:59course you should read the result and
- 00:11:01make sure it actually acts but the thing
- 00:11:02that was really wacky was um our we have
- 00:11:05like a internal thing that can do you
- 00:11:06know go across all of Slack and all
- 00:11:08internal documents and but it's a public
- 00:11:10it's either a public or a private
- 00:11:11channel depending on how you want to use
- 00:11:12it but most people use the public
- 00:11:14version and what was happening around
- 00:11:15performance review season just a couple
- 00:11:17weeks ago is people using it to like
- 00:11:18generate their first drafts which was
- 00:11:20like very interesting in public so I
- 00:11:22don't know like I I wonder how much
- 00:11:23people who come up with just the
- 00:11:25assumption that you're going to use AI
- 00:11:26for a lot of what you're doing are just
- 00:11:28going to be more comfortable and not
- 00:11:29have that stigma around usage it kind of
- 00:11:31reminds me of the early midjourney days
- 00:11:33Yeah. Yeah. Yeah. Exactly. Like that
- 00:11:34shared visibility of usage is I still I
- 00:11:36still think very important. I think
- 00:11:37we're still at like the very beginning
- 00:11:39of how people even understand how to use
- 00:11:40this in their work. Yeah. It feels like
- 00:11:42there's a bunch of social opportunities
- 00:11:43which we haven't seen a lot of yet
- 00:11:44actually. Yeah. Um I'm curious to hear
- 00:11:47what's next for anthropic. Like you guys
- 00:11:49have done a lot on code. You've been
- 00:11:51thinking about the enterprise. Maybe
- 00:11:53there's more models coming up. Whatever
- 00:11:54you can share, we would love to hear.
- 00:11:56And then while he's answering that,
- 00:11:57we're going to do audience questions
- 00:11:58after this one. So start thinking what
- 00:12:00other people might have to ask and we'll
- 00:12:02jump to that next. Yeah, I think for us
- 00:12:04like on the both the model and the
- 00:12:05product side, it's like I know the word
- 00:12:06agent is like you know I'm looking at
- 00:12:09you know David and Robo and it's like
- 00:12:10top of mind for a lot of people. I think
- 00:12:12we we want to be as much as possible
- 00:12:14like powering a lot of that use case. So
- 00:12:16a lot of the like coding is the fir I
- 00:12:18think of as the first uh example of a
- 00:12:20broader theme which is can the models
- 00:12:22work for hours at a time like there was
- 00:12:24the meta chart from earlier and I think
- 00:12:25that that like is like I'm not going to
- 00:12:28call it our road map but it is like our
- 00:12:29goal which is like can the models work
- 00:12:31autonomously for longer and they're
- 00:12:32going to need things like memory they're
- 00:12:34going to need like advanced tool use
- 00:12:35they're going to need to onboard
- 00:12:37themselves organizationally like it's
- 00:12:38not stops being just about the model
- 00:12:40also is like the kind of full complement
- 00:12:42of things that you build around it like
- 00:12:44is it verifiable is like what does
- 00:12:46logging look like when you have a
- 00:12:46hundred agents working in your company
- 00:12:48rather than just one and like I don't
- 00:12:49think we'll like we won't play all the
- 00:12:52parts of that stack but hopefully we can
- 00:12:53enable a lot of that through the models
- 00:12:55and some of the building blocks nice and
- 00:12:56do you have any new models coming soon
- 00:12:58soon maybe maybe soon we always have new
- 00:13:01models coming soon yeah I look forward
- 00:13:03to seeing them I mean it's hilarious
- 00:13:05people like oh cloud 37 is still like
- 00:13:07377 is still the most popular cursor
- 00:13:09model and it's so old I'm like dude we
- 00:13:11released that in February it's like it's
- 00:13:12like the pace is very crazy uh And uh
- 00:13:16we'll have something cool soon. I'm
- 00:13:18excited for it. Do we have any questions
- 00:13:21from folks in the audience for
- 00:13:25There's one behind you, Daria, too. It's
- 00:13:27a comically large microphone. I like it.
- 00:13:29It's throwable. Wait me now. It's
- 00:13:31throwable. We need that. The one behind
- 00:13:33you or I'll go. Okay, I'll go. What's
- 00:13:35the You're a product person. What's the
- 00:13:37like what what keeps you up at night
- 00:13:39from a product perspective? Like what's
- 00:13:40the hardest product question you're
- 00:13:41dealing with right now?
- 00:13:47I still think I'll speak for our
- 00:13:49products, but I think this is generally
- 00:13:50true. Like these products are really
- 00:13:51like hard to use effectively for most
- 00:13:54people approaching it for the first
- 00:13:56time. Um like we'll build things that I
- 00:13:58think are useful and then like they'll
- 00:13:59be good workflows, but it's still a
- 00:14:00little bit too much of like if you hold
- 00:14:02it the right way, you can have like
- 00:14:03incredible results, but like a little
- 00:14:05bit off the beaten path or like if you
- 00:14:06don't have the insight of like, oh, you
- 00:14:08know, bring data this way or like this
- 00:14:10is what you can do and do these
- 00:14:11workflows. that still feels very like
- 00:14:13we're very far from like the first time
- 00:14:15you open Instagram it's like what do you
- 00:14:16do this thing you take a photo and like
- 00:14:17it's like it's definitely not that and
- 00:14:19part of that's being you know primarily
- 00:14:20more like work oriented than than like
- 00:14:23pure sort of like you know personal use
- 00:14:25case oriented but that keeps me up at
- 00:14:27night which is like I like there's still
- 00:14:29a huge overhang of like how models are
- 00:14:31useful to people and their capabilities
- 00:14:33today
- 00:14:38um so there's a certain future in AI27
- 00:14:40that is being predicted did and how does
- 00:14:42your world view like match or differ
- 00:14:45from that or where do things go? What's
- 00:14:48your commentary on that post in general?
- 00:14:50I think like maybe two reactions. One um
- 00:14:54like the importance of compute it's like
- 00:14:56it's not a novel or particularly
- 00:14:57profound statement but like like is I
- 00:15:00imagine a top topic of conversation open
- 00:15:03it's one at anthropic as well. So like
- 00:15:04what is our current compute story?
- 00:15:06What's like the next generation of
- 00:15:07compute like who do we partner with etc.
- 00:15:09So like that emphasis and the numbers in
- 00:15:12there are like pretty directionally
- 00:15:13correct overall. So I thought that was
- 00:15:14interesting. Um the one that I think is
- 00:15:17the most interesting like open question
- 00:15:19about whether it will play out this way
- 00:15:22is the like holding models back from
- 00:15:24release because they're going to be more
- 00:15:26useful and deploy internally. I even
- 00:15:27just saw there's an interview with um
- 00:15:31with Mark Zuckerberg like this week with
- 00:15:33at strategy and he was talking about
- 00:15:35like offering an API for llama and like
- 00:15:37the trade-off around like using some of
- 00:15:38the compute like that conversation is
- 00:15:40happening at every lab right which is
- 00:15:42incrementally do you spend the extra
- 00:15:43time on RL or do you spend that time
- 00:15:46like with a customer use case or do you
- 00:15:48spend it on you know your next pre-train
- 00:15:50and like um that allocation of relative
- 00:15:53compute is going to be incredibly more
- 00:15:55important and then at one point you're
- 00:15:56like wow Like if we have a very large
- 00:15:58product that is going to take a lot of
- 00:15:59inference and like that's highly
- 00:16:01profitable and that's useful but it is
- 00:16:03like directly taking time from from
- 00:16:04capacity for research you know and
- 00:16:06that's not even like research for the
- 00:16:08known runs it's also research for your
- 00:16:10wacky ideas from the two people in a
- 00:16:13room that like have an interesting new
- 00:16:14idea about how to scale that could
- 00:16:15become the next test time compute. So,
- 00:16:18um, that was very like closely matched
- 00:16:21and, um, it'll be this fascinating like
- 00:16:23we're kind of getting into this natural
- 00:16:24experiment with like Ilia's like SSI not
- 00:16:27commercializing from the beginning and
- 00:16:28like will they be in an advantage and
- 00:16:29that they can throw all of their compute
- 00:16:31towards training. I don't know. I feel
- 00:16:33like we've learned a lot from having our
- 00:16:34models out in the wild. I don't think we
- 00:16:36would have built like 37 sonnet the way
- 00:16:38we did it if it hadn't been for the like
- 00:16:41market feedback and seeing real use
- 00:16:42cases. So, I'm a big believer in having
- 00:16:44like an offering in the market. So, like
- 00:16:47that's probably the least plausible but
- 00:16:49will be interesting to watch over the
- 00:16:51next few years.
- 00:16:53I'm curious in a giant uh research plus
- 00:16:56product or how you balance um either you
- 00:17:00could imagine the product defines what
- 00:17:02sort of research happens and everything
- 00:17:04is vertically integrated and maybe
- 00:17:05that's the best product experience
- 00:17:06versus research which might want to just
- 00:17:09make the smartest models possible to
- 00:17:10push the frontier and then product sort
- 00:17:12of gets whatever happens and and makes
- 00:17:14do with it. like how do you uh how do
- 00:17:16you think about that? Yeah, that's such
- 00:17:18a good question. Um I think uh I I push
- 00:17:22our product teams and like in
- 00:17:23combination with research to be like if
- 00:17:25we are shipping things that could have
- 00:17:27easily been built just on top of our API
- 00:17:28and have like no other like way in which
- 00:17:30like at least their initial version
- 00:17:32wasn't better than what could be done
- 00:17:33like what are we doing? It's like we do
- 00:17:35have like these incredible researches on
- 00:17:36the other side. I would not say we were
- 00:17:38doing an like artifact is probably the
- 00:17:40best example of that where like you know
- 00:17:42that was fine tuned into the model it's
- 00:17:43useful etc. Um but then I think there
- 00:17:45was a a play like a phase where we
- 00:17:48weren't doing that as much of that and
- 00:17:49like I think we're getting back to now
- 00:17:50being like all right a full fully
- 00:17:52functioning product pod entropic should
- 00:17:54include applied AI should include
- 00:17:55somebody from like we have our cloud
- 00:17:57skills team which is basically like our
- 00:17:58finetuning team um to actually to
- 00:18:01succeed there but that's still probably
- 00:18:03only like what 10% of the research team
- 00:18:05is doing. And then hopefully the other
- 00:18:06things they're doing are generally
- 00:18:08useful like better instruction following
- 00:18:09is useful because then we can like do
- 00:18:11these things overall. Um, but I've
- 00:18:12always been interested with open air how
- 00:18:14they have like you all have like the
- 00:18:15chatbt model that is in the API that
- 00:18:17presumably not very many people use
- 00:18:18through the API but is like available
- 00:18:20there. Um, and whether we should like
- 00:18:22also have a like more fine-tuned like
- 00:18:24product oriented version. We've gotten
- 00:18:25away without that so far which is useful
- 00:18:27and mostly in compute preservation
- 00:18:29reasons but might actually be holding us
- 00:18:31back for some more differentiated
- 00:18:32product experiences.
- 00:18:35Um, thank you for taking the time. Uh, I
- 00:18:38was curious ho how you see um kind of we
- 00:18:43talked we heard Sam talk about being the
- 00:18:45one subscription for all things AI kind
- 00:18:47of integrating into all aspects of life
- 00:18:49and being that one-stop shop. How do you
- 00:18:51see anthropics positioning relative to
- 00:18:54that? I guess specifically I come from a
- 00:18:56world where you know I work on windsurf
- 00:18:58where we consume a ton of wind of
- 00:19:00enthropic but then I also use chatbt as
- 00:19:02like my app right and so do you draw a
- 00:19:04distinction when you're thinking about
- 00:19:06product strategy and what are you
- 00:19:07thinking long term in terms of those two
- 00:19:09things converging diverging
- 00:19:11I think in like there's a there it's a a
- 00:19:15question I think about a lot
- 00:19:17um there's what we find is like a lot of
- 00:19:20people at least at this phase in the
- 00:19:21product evolution like are comfortable
- 00:19:23like moving across or paying for
- 00:19:25multiple, right? And I'm sure you guys
- 00:19:26have seen this as well where it's like
- 00:19:27they're not replacement, right? Like
- 00:19:29people will pay for windsurf but also be
- 00:19:30in uh like might still subscribe to
- 00:19:32cloud or chatgpt in order uh in order to
- 00:19:35get something else, right? Or like a
- 00:19:36different workflow. I think that's
- 00:19:37sustainable in the short to midun and
- 00:19:39maybe in the long run there's going to
- 00:19:40be like maybe more desire for
- 00:19:43consolidation or maybe we end up with
- 00:19:44some maybe this sounds really dystopian
- 00:19:46like some version of like the like cable
- 00:19:48bundle of some of these things where you
- 00:19:50know there there's a little bit more of
- 00:19:51these. you can probably come up with a
- 00:19:52sexier name than the cable bundle for
- 00:19:54AI. Um, but there's probably something
- 00:19:56something to that regards. And then
- 00:19:58there's like the power users for whom
- 00:19:59like moving across things is valuable.
- 00:20:01Like we launched Cloud Max and like the
- 00:20:03top user request was like, can I use
- 00:20:04this for cloud code tokens? So, we
- 00:20:06launched that yesterday because it
- 00:20:06seemed like a natural evolution of yeah,
- 00:20:08if you're paying $200 a month for cloud,
- 00:20:10you're not going to probably be able to
- 00:20:11consume all of it using cloud AI. And
- 00:20:13that's where the bundle starts being
- 00:20:14useful. I thought it was interesting
- 00:20:15with with chatbt plus is the highest
- 00:20:18tier, right? where there's like yeah you
- 00:20:19can burn that down off of video gen or
- 00:20:21you can do it off you know on coding use
- 00:20:23case etc. I think that that at least
- 00:20:24feels valuable a product idea or concept
- 00:20:27that we've been thinking about is like
- 00:20:28it might be useful to be able to bring
- 00:20:29your tokens to other products as well
- 00:20:32which especially if you're bootstrapping
- 00:20:33a product and you might not be ready to
- 00:20:35pay get somebody to pay $ 20 to $200 a
- 00:20:37month like but they're already paying
- 00:20:39$200 somewhere else like maybe that's a
- 00:20:41useful way where they can get started
- 00:20:45right there.
- 00:20:47Hi Mike, thanks for being here. Um, uh,
- 00:20:50what's your take on how agent to agents,
- 00:20:52the new standard will play out over time
- 00:20:55and should we be waiting for something a
- 00:20:58new standard from anthropic? Yeah, we're
- 00:21:00like we have a lot of uh sort of wacky
- 00:21:03internal prototypes of agents talking to
- 00:21:04each other which I think will help
- 00:21:05inform like what are the right
- 00:21:06primitives that we want to have in
- 00:21:08there. Um, a question that I don't think
- 00:21:11anybody has solved yet from a research
- 00:21:13perspective, at least nothing that I've
- 00:21:14seen publicly, that is going to be very
- 00:21:16important, especially when agents start
- 00:21:17being sort of like your avatar out in
- 00:21:19the world representative of you or your
- 00:21:21company is like better discernment
- 00:21:24around like what you reveal and what you
- 00:21:26engage in, right? It's like
- 00:21:28um what is like if you're transacting
- 00:21:31with a vendor, sure you can re reveal a
- 00:21:33credit card information, but it's just
- 00:21:34like some other random agent you're
- 00:21:35talking to, probably not, right? if it's
- 00:21:37company to company what gets revealed
- 00:21:38and what gets um sort of withheld. So
- 00:21:41that is both a protocol but I think it's
- 00:21:43actually like a research question like
- 00:21:44models sometimes like they want to
- 00:21:46please so they want they'll want to tell
- 00:21:47you information but like how do we or
- 00:21:48they want to be too they're going to be
- 00:21:50too refusally if you like like never
- 00:21:52reveal any company information right so
- 00:21:54that sort of nuance and discernment
- 00:21:56feels unsolved um and then the other one
- 00:21:58that I like alluded to is just like
- 00:22:00auditability at scale is something
- 00:22:02that's going to be really interesting
- 00:22:03again I think like there will be
- 00:22:04products built on top of of that uh to
- 00:22:06solve that need but I was having a
- 00:22:07conversation with the founder last week
- 00:22:09around like what is identity management
- 00:22:11for agents and like what is you know do
- 00:22:13they have names like I don't know that
- 00:22:15feels kind of a little bit too
- 00:22:16anthropomorphic but maybe that's useful
- 00:22:18but feels like an agent should be better
- 00:22:20at doing the hundth task in the first
- 00:22:22which implies some kind of like
- 00:22:23longitudinal memory right um and there's
- 00:22:26going to be ones that are more like your
- 00:22:28extension of your work versus ones that
- 00:22:30are like wholly like an entire other
- 00:22:32employee right so I think those feel
- 00:22:34less like protocol questions and more
- 00:22:36like both like product and research
- 00:22:37questions to
- 00:22:40Hi Mike, thanks for your time. What do
- 00:22:42you think most people building in the
- 00:22:45application layer are getting wrong?
- 00:22:49I don't know about getting it wrong, but
- 00:22:50I think a a thing I've observed is like
- 00:22:53uh products that start sort of AI light
- 00:22:56and go AI heavy like tend to put AI like
- 00:22:59either in like a sidebar or like a it it
- 00:23:01ends up feeling like a secondary sort of
- 00:23:04surface and then especially as you move
- 00:23:06more and more agentically it's like
- 00:23:07harder and harder to make that like as
- 00:23:09full fullfeatured as you would want it
- 00:23:11to be. And so that's one thing which is
- 00:23:13like at what point do you rethink the
- 00:23:15kind of core sort of building blocks of
- 00:23:16your of your product to actually be more
- 00:23:18AI native. I think that's one. The other
- 00:23:20one is a shocking number of AI like
- 00:23:23native products don't expose the
- 00:23:26primitives of the application to the
- 00:23:28models enough. And what I mean by that
- 00:23:29is like you ask it something and you're
- 00:23:31like oh I can't sorry I can't do that
- 00:23:33for you Dave because like it hasn't been
- 00:23:34built that way. Maybe those two points
- 00:23:36are linked right when you build like
- 00:23:37I've built a guey and then like I've
- 00:23:39stapled a model on top. you don't
- 00:23:40necessarily think that like that model
- 00:23:42should actually be your like primary
- 00:23:44user of your application in a lot of
- 00:23:46ways.
- 00:23:48All right, Mike, thank you so much for
- 00:23:49joining us. Yeah, thank you all.
- AI
- Content Creation
- Anthropic
- Product Development
- MCP
- User Experience
- AI Agents
- Storytelling
- Coding Models
- Innovation