Anthropic CPO Mike Krieger: Building AI Products From the Bottom Up

00:23:58
https://www.youtube.com/watch?v=Js1gU6L1Zi8

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.

Show more

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

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