🚀 Add to Chrome – It’s Free - YouTube Summarizer
Category: Entrepreneurship
Tags: AIentrepreneurshipinnovationstartupstechnology
Entities: AnthropicArtifactClaudeInstagramKevin SystromMike KriegerStanford
00:00
when is it time to call it? When is the chapter closed?
I think one big piece of advice is I think we're often as entrepreneurs, we have a lot of weight on our shoulders and between Instagram and Artifact, I did some angel investment and it's never positive news
00:16
when an entrepreneur tells you that a company doesn't work. But I've seen it before, you know, and so it's not a surprise that some companies don't work out.
That's baked into the model. And so I guess the advice I would give is really check in with your investors and don't I've seen entrepreneurs get stuck
00:32
in this idea like I have to keep going because my investor expects me to and I think what they'll find is sometimes investors are like you're trying you've tried it all in this space. I think it's time to to call it and that's okay.
And you know it's never it's never a celebration but it doesn't have to be
00:47
you know a tragedy either. I'm Mike Kger.
I'm currently the chief product officer um at Anthropic which is an AI company based here in San Francisco. I co-founded Instagram and was its chief technology officer from
01:04
about 2010 to 2018. So nice crazy journey that we had there too.
Started one other company with the same founder as Instagram called Artifact which was a AI news startup and then uh joined Anthropic just under a year ago. So it's been the kind of third chapter in my professional career and it's been it's
01:19
been a lot of fun so far. Credit my dad for a lot of this.
He brought home a computer for us when I was about four. I remember it had it around Microsoft Daws still.
It didn't even have Windows. But one of the things that was very cool is you can just type
01:35
edit and you would be able to open up the files and all the even the apps that came with the computer. And so I think I liked taking things apart and putting them back together.
And I think I really liked seeing how things worked. So it sparked very early for me when I was just four.
when it came time to decide
01:50
what my career was going to be. What was interesting was in Brazil at the time, there wasn't really the same kind of technology industry as there is today.
And so I didn't have any friends or parents friends who were working in tech. And so I didn't really see it as a career for me.
I just thought it was a
02:06
thing other people did. And so it actually only really took me until I came to California that I realized that this childhood interest that I had actually was something that I could do as a career and and make a whole living off of.
When I got to Stanford, I discovered
02:21
that they had this degree program called Symbolic Systems, which only Stanford has, and it's kind of this little strange program that when I was there, I think per year they had about 40 or 50 students, and now I think it's more than 200. So, it's really grown, which is great.
What I loved about it is that it
02:38
wasn't just computer science which I was interested in but it also included design it included philosophy included psychology and it was really you think about the idea of the program it's not just to study computing it's to study the the whole context around why we build software that was really powerful
02:54
for me because I knew I liked design and I liked making things useful for people but I also liked the actual building you know of those things so I think a few of the things I really came out of that program with one was this idea that everything you build should solve a
03:09
problem for somebody, right? It's like kind of the core of design thinking that you identify problems, you figure out how you can best solve them after doing the research and then you validate what you've built to make sure you actually solve those problems.
I think that is really important and that's a big piece of what I what I came out with. The
03:25
other is the value of prototyping. So, you know, instead of just working for 6 months on a project and then showing it to somebody for the first time and then it turns out you got something really fundamentally wrong, willing to build prototypes, show them along the way and open up the design process a lot more.
And maybe the last one is the value of a
03:42
good team um and having the right partners. So, person I co-founded Instagram with is somebody I actually met back at Stanford.
At the time, we didn't know we were going to start a company together, but I knew that partnering with the right person, especially that somebody that has similar skills in some ways, but different skills in other ways, and they
03:58
can complement you, um, can make the difference as well for the company succeeding. For example, there's a brand new built-in camera.
Now, the camera is another way that the iPhone 3G has really changed how people use their cell phones. The new iPhone 3GS, there's a
04:13
brand new 3 megapixel autofocus camera. It takes pictures that are even larger, even more beautiful.
You have to kind of rewind your uh time machine back to 2009. Now we take for
04:28
granted that everybody uses a lot of apps on their phone, that we take a lot of photos on our phone and that we network or we, you know, use social media on our phones. But in 2009, that wasn't the case quite yet.
So the cameras were getting better, but they were still quite bad. Facebook had a mobile app, but um it was kind of that
04:44
early one where it was kind of just trying to recreate the website on the mobile side and it wasn't quite as a native and there were very few sort of new social products coming out and the context is important because the reason I got really excited to collaborate with Kevin who ended up being my co-founder is that he wanted to change all three of
05:01
those things. So he was building time as a location sharing site a little bit like Foursquare if you remember Foursquare but with photos and videos attached as well and really creating a social experience when you're out and about.
That was really exciting to me because I saw the potential for mobile device to create a much more personal
05:17
connecting experience. So first moment was reconnecting with Kevin and seeing that he was interested in these ideas.
Second important moment was me getting my work visa so I could go work with him. So I was first had to transfer my work visa and that actually took four months.
And then the third important moment was us realizing that product
05:33
that we were working on it was called Bourbon Burbn which was a that sort of location-based check-in location sharing app was good but it was not product that was really going to where we wanted to go and so we had a moment where we had to take a pause really peel back all of what was not working about the product
05:48
and realized that at the core the piece that was really working was this uh aspect about taking photos sharing what you were doing and connecting with people that way. And that was really the the moment where we took a lot of distracting things, stripped them all away and then focused on what would
06:04
eventually become Instagram. I think a lot of the art and it's a part both art and science of building great products especially with emerging technologies.
So Instagram it was the rise of the mobile phone. finding the technologies that are ready
06:20
for broader adoption than they currently have and then finding how do you build the product around it to make it usable for many many people that's a common sort of through line and I think the kinds of things you can look for in terms of that early energy is are there some people already starting to do
06:36
something interesting in the space but it's all the early adopters that's an interesting kind of area um the other thing I look at often is the rate of change so if you look at the iPhone 3 to the 3GS to the for and you look at the quality of the photos and the quality of the networking um stack and the ability
06:52
to connect, you can see the jumps. So you're like, okay, if it goes one or two more jumps, it's going to be incredible.
Same with LLMs. I joined Anthropic.
We had just come out with Cloud 3 and Cloud 3 was our first model that really started to get more usage. It was still, you know, limited in a lot of ways, but
07:08
then you see the next we did cloud 3.5 and that was a big leap. So you can imagine that you're going to continue to get these large jumps.
And so as a product builder, you have to start thinking, I want to build a product that's useful today, but is also ready to catch the wave of what the next big
07:23
leap is. You've closed uh the Instagram acquisition.
Why did you do it and what are you going to do with it? Yeah.
So Instagram is great, right? I mean, they're um they're this super talented group of of engineers.
They're building
07:39
this amazing product. Um they just crossed 100 million registered users and they're killing it.
They started off building on top of our platform. You know, you can take pictures with Instagram and you can share them to Facebook.
Um, and it's it's really first class, right? So, sharing a picture from Instagram, it basically appears exactly
07:54
the same as if you shared a photo on Facebook. So, it's great.
They did a really good job with that. Our mission around Instagram is we think Instagram is amazing and we want to help it grow to hundreds of millions of users.
We want to help them out um with with whatever we can. Um, but we we have no agenda in terms of making them go onto our infrastructure or something.
A lot
08:10
of times companies um force companies that they're integrating to do stuff like that. I think it's primarily a waste of time.
We're not going to do any of that. We're going to just try to do the things that that we would have done if they were an open graph partner, but now we can prioritize them more highly.
I mean, they can do a lot of them directly because they have access to our code directly.
08:27
I like to say that at Instagram, you know, before the acquisition, we didn't yet have a company. Um and what I mean by that was we were 13 people had some ideas around revenue, but we hadn't built any of them out yet.
still very much that like early startup. Joining Facebook was really interesting to me
08:42
because they were at the time thousands of people. They had just crossed a billion users, so much bigger than than us in terms of company size.
But they were really focused on preserving as much of the startup culture as they could. So for example, uh every 3 or 4
08:58
months they would do a whole company hackathon where everybody would get to focus on what they were interested in for a couple days instead of the the ordinary road map. They really prized experimentation.
They really tried to, you know, famously had the move fast and break things kind of slogan. So I think
09:14
it was good that we went to a company that was bigger but not super big company or slow. And it taught us that you can grow your team and grow your ambition but still focus on the things that keep you moving fast.
I think the value of the team is something that will stay with me forever because you know
09:31
you can have the right strategy and you can have a good product ultimate the details that go into those products the the pace and speed that you're able to execute on and like how much fun it is which is also really important ultimately comes down to having the right team around you and that's a
09:46
mixture of people that are talented but don't have a lot of ego which is a interesting and difficult combination to find sometimes people that are willing to be generalist lists rather than just be caught in their sort of individual silo. And so, you know, at Instagram, we had people who would start off, you
10:02
know, writing the back end for a feature, but would also then build the iOS or the Android part. And it's important, I think, for you to have that fluidity, especially on engineering or designers that would code or product managers that would design, and it's just not getting everybody boxed in.
So, that aspect is really important. And
10:18
then also a team that really cares about the product they're building, which again is easily more easily said than done. It's I've al I've seen companies where everybody is working and they could be working hard but they don't have a passion or really an attachment to the thing that they're working on.
You're never going to get great products
10:35
that way. It just never happens because the great product breakthroughs come from people being close to the details, understanding what could be better, coming up with the next idea themselves, not waiting for a, you know, strategy meeting.
So that value kind of throughout the right kind of people and the right sort of setup of team is
10:51
something that I think about all the time. In 2021, I founded Artifact of the same co-founders, Kevin.
Our goal, our observation was there's been a rise and
11:06
at the time it wasn't, you know, LLMs yet, but there was a lot of rise in machine learning and the beginning of some of these neural networks. For all of that interest and rise, there was still not a lot of products being built that felt very personal.
Because if you think about it, the promise of some of this machine learning was great with a
11:23
lot of signals, we'll be able to incorporate what's personal to you and also what's aggregated out, you know, among a broader group and be able to tailor the experience. But you looked around and there actually wasn't that many personal experiences.
So our bet with the company was by combining cutting edge machine learning and good
11:40
product design, we'd be able to build products that felt very personal to you. And we started with news uh and articles in general because that was something that we felt most people are readers even if they're not like book readers.
They like reading articles online. And there's an existing kind of ecosystem of
11:55
blogs and newsletters and news sites that was out there that we if we could just connect people to those sites be able to deliver a good product. And the idea behind the company was this is the first product, but we'll kind of take the same personalization technology and be able to map it to other products that
12:11
are also at this intersection of personalization and content. So we thought about shopping recommendations, we thought about local recommendations, all these ideas around how to personalize information using machine learning.
12:27
We launched um the product after about 2 years in private beta which I think was too long because it took us too long to get out to the market to start learning and it also meant that by the time we launched the team was already quite exhausted for having worked really really hard on this product for 2 years non-stop and then we let the product run
12:44
for about a year and after a year what we noticed was the energy was not there in the system. like we would work really hard to like improve a feature or add social features and comments and re-shares or posts and user generated content like we we were trying big ideas
12:59
like it wasn't like the product was standing still and it was very hard to shift the energy in the system and I think that was because there was not a fundamental fit around what we were doing and what people were wanting and I think there was probably two aspects of that the first one was even if our algorithms I think were very good the
13:16
mobile web kind of websites that we were sending people to for this news. They were often like full of ads or not formatted very well or full of like pop-up videos.
It was just not a very good experience once you actually clicked through. And that was like I think a hard experience to deliver.
And I think the second part was our product
13:32
got really good if you put enough data in there and you read enough articles. We'd get very personal to you at the level of all right, Mike is interested in Formula 1, but not just Formula 1.
is interested specifically in this driver and also he likes Brazilian modernist architecture but also when that's paired
13:48
with you know Scandinavian design like that level of like really specific knowledge but to get that you had to read a lot of articles and so most people would come in they would read a couple of articles they would say this isn't very different than Apple News and then they would bounce off so I think we bet too hard on personalization without
14:05
remembering that you also have to be good at the very beginning before you've really gotten a lot of personal content I think one big lesson is that users are not going to adopt a feature or a product just because of the the technology underneath or just because it has intelligence. It actually has to
14:21
still solve the problem. It goes all the way back to what I had learned at Stanford.
So as we build products at Enthropic, of course, the models are very intelligent and they can do a lot. But you have to do more than just say this model is so smart.
You have to go beyond that and say and here's what it can do for you and here's how it can connect to you and here's how it can be
14:37
useful to you even if you're just getting started with it. Right?
It goes back to that feeling an artifact of yeah the product is good if you put a lot of work into it. Like we need to design things with claude that are useful for somebody that's just discovering LLM kind of for the first time.
So a lot of the work that we're doing right now is
14:53
how do we lower the barrier for people who aren't yet familiar with all the ways that AI can help you in your daily life to just give them some really concrete things that they could do where it's useful to them today rather than in theory. So that's a big lesson lesson learned.
I think another one that's
15:08
important as well is how do you get user intent and the personalization more quickly? And so for us, it's, you know, when when you sign up for Claude, we ask you some questions.
And it's nice because, you know, you can actually chat start chatting with Claude and get a feeling of what it's like to talk to
15:23
Claude. Um, and hopefully, you know, the questions we ask in that onboarding process in your first day let us be able to be more tailored and personal to you quickly.
And that was kind of came out directly from the Artifact experience. That's a great question around like when is it time to call it?
15:40
When does the chapter close? The way that we did this with Artifact, um, and I think this helped, was we sat down, me and Kevin, we made a list of like what are the ideas that we still have in this space that we will feel really silly not having tried before shutting it down.
15:56
And we wrote them down and then we prioritized like three big ones that we said, we want to try this and after we try this, we'll take a step back and say, did it change? Did it change the the trajectory of the company?
Um, and we did that and then we kind of wrapped up 2023 and enter 2024 and we said we
16:14
tried them. It's still not, you know, it's still not changing the the trajectory.
Then, you know, it's time to to move on. But it is quite tricky.
I think it's I've seen entrepreneurs get stuck at companies for years that they feel, you know, they owe it to themselves or their investors to keep
16:29
going, but it's not likely to to shift the directions. So I think being concrete either with a date or with a set of projects helps you um know when to move on.
I think one big piece of advice is I think we're often as entrepreneurs uh we we have a lot of
16:45
weight on our shoulders and you know between Instagram and Artifact I did some angel investment and it's never positive news when an entrepreneur tells you that a company doesn't work but I've seen it before you know and so it's not a surprise that some companies don't
17:01
work out. that's baked into the model.
And so I guess I the advice I'll give is really check in with your investors and don't I've seen entrepreneurs get stuck in this idea like I have to keep going because my investor expects me to. And I think what they'll find is sometimes investors are like you're trying you've
17:16
tried it all in this space. I think it's time to to call it and that's okay.
And you know it's not it's never a celebration but it doesn't have to be you know a tragedy either. [Music]
17:32
I think what was interesting with the anthropic role is that it was both joining an existing company the company was about 3 years old at the time had already launched models had cloud.ai and AI but there was still a lot of zero to one to be done so I don't think we had our mobile apps yet um we hadn't built things like claude code which is our
17:48
agentic coding tool so there was a lot of empty space still in the product so it was kind of a combination of 0ero to one and established uh we have a claude character um effort and team and that team is really focused on what are claude's philosophies what's cla's vibes
18:05
how should claude respond and that's been uh an effort even from early days but It's something we've focused even more on recently in the rest of the AI landscape. You can often have these evaluations or evals right where you say all right how is you know the model doing at math or competition coding or
18:22
agent coding it's a challenge to say what kind of vibes does the model have we have some internal ways of doing that often you can try to use claude to listen to claude and then say you know what kind of vibe does it have but it also comes from just using claude a lot internally so every day there are
18:39
hundreds of conversations that you know people atanthropic are having with claude with the intention of understanding how is Claude's personality evolving uh when we're doing training how could it be different going forward even things like should claude be verbose like should it say a lot or
18:54
should it be concise and the answer varies on the depending on the situation I think one of the hardest ones is um we are building a fundamentally a dynamic model and a dynamic system and so as an example you know we have in cloud.io Okay, we have a a feature called
19:10
artifacts where uh you can sort of work on a document or even a website along with claude. You know, the model learns how to make artifacts, but the taste that it has in artifacts and how it changes and what language it uses and exactly what it does evolves from model to model.
And sometimes you don't know
19:27
how it's evolved until very late in the model training process. So, one of the biggest challenges with building products alongside doing model development is they're both moving targets.
Our product is evolving, the models are evolving. Often the models are evolving up to like a week before
19:42
launch and we're trying to build products alongside. It's a very interesting dynamic system.
It's what makes it exciting because the model can be creative and it can surprise you in a lot of ways. Uh but it's also um a lot more challenging than try classic product development.
19:59
I think it's really important to as a product development team to be really upfront about the capabilities and the limitations and the risks. So you'll see when you sign up for cloud and you go through the the first conversation, we really emphasize here's what models are
20:15
good at and here's where they can still make mistakes. Um and here's the limitations.
And I think it's important to lead at the beginning because I want people to have the right mental model when using these models. They're not perfect.
They don't know everything. um but they can be very helpful in for a given task.
Similarly, I think goes back
20:31
to the kind of that quad character piece. It's important for the model to also be aware of its limitations.
And so often you'll find if you ask cla about medical questions, it might say, "All right, I can tell you more about this, but first I'm not a doctor." And if you're worried about this, then you might want to consult a professional.
20:47
And it's not just a thing we put in for liability. It's a thing we put in because we think it's really important to have Claude recognize its own limitations as much as possible.
Yeah. So it's an interesting question for us which is how do we gather user feedback on claude in a way
21:03
that is like most helpful and the way we've currently found although I think there are probably other ways that we can continue to evolve it is for every answer in in claude we have you know thumbs up thumbs down how did we do for both thumbs up and thumbs down you can write a little paragraph about why it was good or was it bad and it's so
21:20
interesting because um and that extra bit of signal around why is the most important part and so we have a product manager and she's She spends a lot of her time aggregating and evaluating and looking at how the model is doing out in the wild. And then you start seeing themes.
So an individual answer maybe
21:36
doesn't give you the clue about how you need to improve the model. But then in aggregate you say, okay, Claude is being too too verbose here in this particular case.
We're going to change it for the next model. Or Claude is changing its mind too quickly.
So when the user disagrees with Claude, maybe the user is
21:53
actually looking for a debate. Sometimes Claude is like, "It's fine.
Like you're right. I'm sorry I was wrong, but the user was actually looking for more of a back and forth.
So that kind of conversational feedback is incredibly valuable. And we basically for every new model training will first look at all
22:10
the feedback in aggregate for the previous one and say what do we want to keep the same and what do we want to change or what do we want to double down on and what do we want to make different. So that feedback process is very very important and it's actually the main feedback we get.
So we don't train on any user data um including
22:26
conversations for anybody but we do use the the thumbs up and thumbs down as a really helpful signal to figure out what we need to improve. I think there's two areas of consolidation that I expect to see.
One is on the model development side of things as these models they're
22:43
already very expensive to to build and train at scale. as they reach even higher levels of intelligence they're going to require that same kind of jump in compute and resources and I think we'll probably come down to three four five companies doing that not as many
23:00
you're already seeing some kind of consolidation and and in that side of things I think anthropic is well positioned there we've got really great partnerships with Amazon with Google that I think will unlock a lot of compute for us and the other side is um consolidation in more of the apps side
23:16
of things because I spoke at the Y Combinator batch of startups recently and everybody's working on something that is AI or directly uh adjacent to AI and that's great there's a lot of products still to be built in this space but not all of them are going to work so
23:31
in the same way as you saw that with mobile and social media where it's useful to have that explosion cuz no one company is going to map the whole space out then you start naturally seeing you know what's getting traction what's going to get consolidated what are good idea what were good design ideas but
23:46
wrong sort of business ideas that then will evolve into better business ideas. That evolution I think is very healthy.
Uh I saw that process with mobile, I saw it with social media. I think now is the time to start seeing it.
Well, maybe you know in a 6 month time you'll start seeing more of that within the app space
24:02
as well. I think there's when I think about where we are and where we need to get to, there's a couple of avenues I think are really important.
The first one is understanding people for more than just a single conversation or even a couple of conversations. So we're not static,
24:19
right? We are dynamic.
We have relationships. We have moods.
We have challenges. We have wins.
And I think for models to be feel really like they can work alongside us to really be part of our lives. They're going to need to develop a sort of empathy beyond a single conversation and really have that
24:34
sort of longer horizon interaction with somebody. I think that's really important.
I think that comes from contextual understanding and memory. For example, the second part is the models need to learn when they can be proactive and when they should sit back.
As people, we often know if somebody's
24:50
head's down working, we're trying to we're not going to go interrupt them or if they come to you for help, we can answer a question and maybe get back to work ourselves. So, how do we have that collaboration with models in a way that feels really natural?
As an example, we have Claude as a participant in our Slack channels, anthropic, which is
25:07
great. it can chime in and and be part of the conversation, but we're finding it still either will not participate enough or say too much.
So, how do we get it to feel like a more natural participant in these conversations is important. And then the third one is around agency and independence.
So, how
25:22
do we get the models to be able to take direction from people but then go off and either do work or do research or be waiting for additional input and do that over a much longer time horizon in the background. I think that's very exciting.
There's a lot of things that will be enabled by that capability, but
25:39
it's going beyond the sort of chat box with a single interaction to more of a longunning agent in the background. One of the big changes that's going to happen is the models will start being able to give you insights about your own self, which I think is very powerful, right?
Um I I have worked with a coach
25:56
for many years. I think it's one of the most important ways in which I improve as a leader.
I want that for everybody, right? Everybody has can benefit from reflection and learning and iterating on the way that they you know approach the world.
Right now you can go to Claude and tell it about how things are going
26:12
and it will probably be able to give you some coaching but it's very point in time. When I think about the potential it's to be much more of an ongoing sort of uh improver or you know conversational um you know coach.
I use this right now with Claude a lot. Anything that I write, I will put, you know, if I'm writing a strategy
26:28
document, I'll put it into cloud and say, "What am I missing? What did I forget?
Like, what poke some holes in my argument?" And it's great to have that partner. Um, I think that'll be one big evolutionary step.
The other one that's, this is more speculative, but if uh
26:43
models go from being single conversations, you know, single chats to being more of, you know, long horizon almost personalities, I'm curious what kind of relationship we'll form with the models. is like will it feel more like a friend uh in some ways or a co-orker in
26:59
some other ways? I think that remains to be seen.
I think in your early 20s you have a unique place in the world because you've you're pretty in touch with the trends that like teenagers have which is important especially if you're building
27:15
in consumer but you're now able to like take those insights and make them into something like a project or maybe even a company. So, it's a very special time in your life where you're like at the intersection of being a teenager, being really tuned into that, but being able to do those things.
I'd say the two
27:30
things I would I would I would add is one, not worrying so much that every step needs to make logical sense. There's definitely things that side projects I did or explorations I did where at the time I was like, am I wasting my time here?
But actually later, I find that it was connected to the next thing I did. So just trust that
27:47
if you're driving towards a direction, it might take some very strange routes to get there, but as long as you're learning, I think that that's fine. And then the second one is advice that I got when I was 21 and I still think about to this day is the companies change, the projects change, you know, the economy
28:03
changes, but the relationships you build in your career will be the relationships that you have over and over again. And so I was working at a startup as an intern and it was full of people who were in their late 30s who had worked together probably three or four times since they were 20 and it really stayed
28:18
with me. So like remembering that those relationships build on each other and recur which means that it's you know it's worth investing in the in that time.
I think for me I was talking to to Daniela who's one of the co-founders of Anthropic. I I told her and I meant it I
28:35
hope Anthropic is the last job that I have. I think I see the opportunity here to be paired with research and build interesting products for many many years because I think every year will be different.
When I was 18, I had this like promise to myself that I wanted every year to feel different. And this year has felt very different than any
28:51
other year that I've had being an anthropic. And I think that the year ahead is going to be quite different as well.
And so that's how I really think about the next 5 years is just making sure I'm set up for consistently learning, evolving, working with interesting people on interesting products. And then taking the space at
29:06
least once a year to step back and say, am I still learning? Am I still, you know, eventually at Instagram, you know, 8 years in, there was a time where I said, okay, I think I've I've gotten what I need to get out of this experience, and now it's time to try something else.
To me, entrepreneurship is about finding ways in which the world could be different and better and in
29:23
feeling empowered to go make that change. It's something that really drove me from a pretty early age.
And as I've seen that manifest both in the ways you can change your city, the way you can change, you know, the way people relate, of course, with Instagram having global
29:38
impact, large or small, it's really about identifying like what could be different like as getting curious about the world and then also getting curious about how you or you and a small team can see if there's a a change to be made there. Um, and I'm very excited about the way AI will empower people to ask
29:55
that question and also answer it.