back to index

Startups building new products with Claude | Code w/ Claude


Whisper Transcript | Transcript Only Page

00:00:00.000 | welcome welcome everyone i'm joe i'm from the startups team here at anthropic and we're so
00:00:10.720 | thrilled to have all of you here today we have a breakout of six amazing founders that have really
00:00:18.640 | found product market fit building on cloud across a bunch of different industries across agentic
00:00:24.080 | coding across design across music generation and a lot of these startups have really seen that
00:00:29.840 | zero to multi-million user journey within the past 12 months and i know in this crowd there's a lot of
00:00:35.840 | you know really great founders too and so we want all of you to stay until the end of the session
00:00:41.280 | we have a little surprise for all of you to be able to spin up quicker uh on claude four um in
00:00:47.760 | our new family of models so without further ado i want to bring up kevin uh who is the ceo and
00:00:56.160 | co-founder of tempo labs awesome thanks joe everyone nice to meet you as she mentioned
00:01:05.120 | my name is kevin one of the co-founders of tempo and it's an honor to be here i'm to kick off i
00:01:09.600 | have a video uh that'll give you a little bit of a taste of what temple is all about
00:01:59.600 | awesome so uh what is tempo at tempo we're building cursor for pms and designers uh tempo is a new type of
00:02:21.600 | ide that feels more like a design tool like figma than it does vs code and the whole idea behind it is to
00:02:27.360 | to give designers and pms the ability to collaborate with claude uh and to create you know the first
00:02:33.440 | draft of pull requests and in many cases uh the actual end-to-end pull request uh kind of pulling the
00:02:39.040 | engineer out of the loop in some cases but if anything making the process of working on code
00:02:43.360 | collaborative as it stands today it's really non-engineers don't you know collaborate too much
00:02:47.280 | on code so i'll show you a bit of a live demo of what that looks like uh this is tempo uh tempo as you can tell
00:02:54.000 | uh you know has three tabs product uh a prd tab a design tab and a code tab so imagine you're a pm right
00:03:01.360 | you generate uh airbnb app with uh claude and you're like okay great this looks you know it's a it's a it's
00:03:07.360 | a quick and dirty prototype demo uh and then you pass it off to your designer and your designer looks at
00:03:12.480 | uh the real airbnb and it's kind of like hmm this is kind of off i need to get it pixel perfect so uh you come in
00:03:18.560 | in here and you realize that uh you know you spent some time working on this new animated uh tabs header
00:03:26.160 | so you can see you know just like the it's nice and pretty just like the core uh airbnb uh so what i
00:03:31.440 | can do now is i can uh drag and drop this right into my header over here uh give that a second there you
00:03:39.120 | go we've got it in the header uh and then what i can do is just rearrange this to get it in the middle
00:03:46.640 | um and again every change that i'm making is actually you know you can see editing the original
00:03:51.120 | source code in the correct position um so okay so we're kind of like starting to make some progress
00:03:55.920 | now clearly these little new badges are not great you know whenever whoever designed or created that
00:04:01.760 | didn't do a great job so why don't we fix that up really quick um so i'm going to open up my uh dom
00:04:08.720 | tree over here which shows me everything that's in the dom i'll find the new badge and i'll just click
00:04:13.680 | delete just like i would boom okay we're getting somewhere close looks you know half decent now
00:04:19.440 | let's compare to airbnb okay we've got some work to do clearly claude thought we need this little
00:04:25.040 | text header i'm just going to delete that like i normally would in a design tool traditional design
00:04:30.800 | tool like figma i'm going to delete this over here and then i am going to let's compare again
00:04:36.960 | i'm going to move my uh the ellen to my property grid here to the center i'm going to add a little bit
00:04:47.280 | of gap between them and then maybe here i'm going to remove this uh tailwind padding and then i'm going
00:04:53.440 | to compare to core airbnb and then the site that i built in tempo and now you get an idea of how you know
00:04:59.520 | pms designers can collaborate together with claude and then again i can enter a commit message push
00:05:03.920 | this to uh to github and actually the really cool thing is that this actually this code isn't running
00:05:09.680 | locally on my machine it's actually running on a docker container in the cloud so this whole thing is
00:05:13.520 | collaborative i can send a link to someone just like figma and they can come and collaboratively code
00:05:17.280 | with me and with claude there you go tempo is again it's an ide for designers and developers uh to
00:05:23.120 | collaborate together on code we already have a number of customers and have been growing extremely extremely
00:05:28.400 | quickly uh and what we're seeing for customers that adopt tempo is designers truly are turning into
00:05:34.080 | design engineers and we're seeing that you know approximately 10 to 15 percent of front-end pull
00:05:38.800 | requests are actually being able to be opened by designers directly they're often able to do pixel
00:05:43.520 | pushing themselves without involving an engineer they can just work with tempo and with claude directly
00:05:48.480 | and in in about 60 percent of pull requests we see that a a sizable chunk of the front-end code has
00:05:55.600 | been generated by designers by pms and by claude and it actually accelerates it's useful code that
00:06:00.800 | accelerates a true front-end engineer uh in their process of going to production and so that's tempo
00:06:05.920 | we are in the we're imminently about to release support for claude 4 so i invite you all to uh give it a
00:06:11.680 | shot uh and uh empower your pms and your designers to start to collaborate with your engineers on code thank you
00:06:20.400 | howdy everybody um my name is andrew file i will start uh with a little bit of a past history i build
00:06:27.840 | and sold software business for more than two billion dollars so i ran a team of more than thousand people
00:06:32.800 | and despite doing that i could tell that only about two to five percent of our ideas ever came to life
00:06:39.360 | because the majority of the time we spend on fairly routine work and the larger is the company the more of
00:06:45.280 | that routine is in your uh kind of uh the the bigger that routine makes part of your of your day job so
00:06:51.600 | i fundamentally believe that uh we can automate at least 90 percent of that routine work giving
00:06:57.600 | us opportunity to move 10 times faster and bring more of our ideas to life and um i personally uh believe
00:07:05.200 | that a lot of their of this transformation will happen this year and we're uh actually observing it
00:07:11.680 | as we speak including today i think that uh this week actually demark demarks their transition into the
00:07:19.120 | next generation so how only 12 months ago the whole industry was on the first generation of ai coding
00:07:24.080 | assistance so a code completion it kind of uh was a very nice convenience but not a game changer to to
00:07:30.240 | anybody and then october last year um thanks to entropic we got um cloud 3.5 and that was a game
00:07:38.720 | changer and that allowed um us and zencoder and uh cursor and windsurf and a lot of other players to
00:07:45.280 | build true coding agents inside of your id so that opened up a completely new uh use cases completely new
00:07:52.160 | scenarios we saw usage skyrocket uh 10x 100x and we saw people uh being able to do amazing things that they
00:07:59.360 | haven't been able to do before and part of that revolution was um the model support for tools
00:08:06.240 | and environment because behind a lot of those generational changes there are some technical
00:08:10.640 | capabilities that kind of unlock all of that right uh there was also move in the models from coding
00:08:17.200 | quote unquote like olympiad style stuff to true software engineering and that was uh very very very
00:08:22.800 | important and then uh obviously context matters um software repositories are quite large so it's good to
00:08:28.880 | to have a larger support and i think that uh in this year we'll actually see transition to the next
00:08:34.400 | generation uh for me one of the biggest enablers there is verification uh so i'm super excited about
00:08:40.240 | something that we haven't yet talked uh at this conference today is computer use we saw rapid progress
00:08:45.520 | in those models and they allow us to build uh proper verification and verification is key to scaling
00:08:51.760 | ai uh and kind of delivering more fully autonomous cycle and then the other thing is uh in real life in actual
00:08:58.560 | uh production work uh where you got multiple people engaged and multiple teams engaged uh typically
00:09:05.200 | a lot of the time is wasted doing what i call in between uh kind of passing the ball and so when we
00:09:11.280 | think about the problem we think about software development life cycle as a whole uh not just about the
00:09:16.800 | coding and that's a big part of our dna and so super happy to announce a couple of weeks ago
00:09:23.520 | uh came with zen agents so taking that concept of um coding agent and extending into custom agents that
00:09:30.400 | you can share across your whole organization with full support for mcp with specialized coding tools
00:09:35.360 | and whatnot but you can basically deploy them across your whole sdlc from their uh work on your prd
00:09:42.400 | to to the actual coding to verification to code reviews and whatnot so that's super exciting and i'll just
00:09:50.880 | basically show you a quick video and we'll be good to go
00:09:58.560 | and we're looking forward for entropic to bring us mcp
00:10:28.320 | registry uh but while that's not available we offer you ours so we got a hundred um i believe
00:10:34.560 | about 100 mcp servers out there and obviously you can configure your own and also as we launched that
00:10:40.320 | zen agents we were excited to launch the community aspects of it so there's an uh mit licensed github
00:10:45.920 | repo where you can commit your agents and if we approve your pr uh everybody will will see them so
00:10:51.200 | we're excited and think that that can help kickstart the creativity process in a lot of organizations and
00:10:57.200 | kind of take it beyond the basic coding use uh truly across the whole software development life cycle so
00:11:02.560 | thank you all
00:11:05.440 | hi my name is jordan i'm the head of ai engineering at gamma and at gamma our mission is to help bring your
00:11:12.880 | ideas to life whether that's presentations documents websites or even social media carousels we do all this
00:11:20.400 | so we can help us through ai to help build these things for you
00:11:24.640 | and we love quad uh throughout gamma's kind of history of using ai there's been
00:11:29.680 | two moments where a model upgrade has made significant impact to our key metrics around user
00:11:36.480 | satisfaction for deck generation the first was when sonnet 3.5 came out and the second was when
00:11:42.560 | sonnet 3.7 came out we actually saw an eight percent increase in this metric which is something that we
00:11:48.320 | spent hundreds of hours trying to prompt engineer around and could not get such improvements
00:11:53.440 | and one of the big reasons for this was around web search so the built-in tool in sonnet that we can
00:12:00.080 | natively use web search for and not have to use the third-party service
00:12:03.840 | so i'm going to show you the difference what web search made in gamma by showing two decks generated
00:12:09.600 | with gamma one with web search and one without so this is gamma gamma used ai to generate this it used
00:12:16.480 | sonnet 3.7 but it did not use web search and as you can see got a lot of things wrong this conference
00:12:24.320 | isn't january 22nd 2025 it's not three days i don't think these are the right speakers
00:12:33.200 | so let's see what we can do with web search so in gamma you can create a entire presentation web page
00:12:41.440 | document anything from a single sentence so i'll type in code with claude conference 2025
00:12:49.120 | and it's going to first create an outline for me it's going to search the web to figure out what
00:12:56.800 | should be in this outline and it will inevitably find details about the code with claude conference
00:13:02.800 | and put them there so as you can see we're already off to a better start where we have the correct dates
00:13:08.240 | we have real technical sessions we have the right location so this looks much better for us
00:13:16.400 | and i've even gone ahead and created an an anthropic theme that we can use
00:13:30.560 | so now that we have the outline generated gamma will take that outline and it will take the other
00:13:34.560 | information it has and create an entire presentation for us with hopefully the correct details so let's check
00:13:49.280 | so i think already we're off to a better start it has the right date has the right location
00:13:53.440 | it tells us the schedule
00:13:56.720 | and overall this provides us a much better starting point if we were to say make a presentation or a
00:14:05.520 | website about this conference it's not going to be perfect but it'll get you to a point where you have
00:14:10.400 | the correct information and you're at a good starting point to actually finish your presentation
00:14:14.640 | so this one change from sonnet 3.5 to 3.7 with with web search was a huge lift for us
00:14:25.200 | and we're very excited to continue to use claude and to continue to see these improvements
00:14:32.640 | lastly we are hiring so if you are an engineer or product or designer and are interested in working
00:14:38.320 | at gamma you can go to careers.gamma.app thank you hi my name is amar goyal i'm the ceo and co-founder
00:14:44.880 | of bitto we're building the hopefully the world's best ai code review platform
00:14:51.440 | so as we all know we're using all these cursor windsurf cloud cloud code to write massive massive
00:15:00.000 | amounts of code the amount of code that developers are writing we think is going to go up 10x over the
00:15:04.240 | next year or two but what that means is that you know vibe coding which we all talk about does not
00:15:10.560 | equal vibe engineering how do you get code that's scalable reliable performant fits into your architectural
00:15:17.040 | design patterns and that's still left to the code review process but code review is not going to
00:15:21.920 | scale teams have trouble keeping up today it's not going to scale for 10x the amount of code
00:15:26.320 | bitto's built an ai code review that plugs into github gitlab bitbucket works with over 50 languages
00:15:34.240 | our focus is really especially with sonnet having a really high quality model that provides a human-like
00:15:39.920 | code review focuses on the critical things our biggest thing is really about more signal and less
00:15:45.360 | noise how do we provide you actionable important suggestions let me show it to you so we've got a
00:15:50.640 | pr here this is gitlab that works kind of the same in all the platforms so once you create a pr
00:15:58.240 | bitto's fired bitto gets called and we first start by summarizing your pr this doesn't require
00:16:05.360 | any documentation or any comments we just look at the diffs we look at your code and we summarize it
00:16:10.640 | then we give you a quick kind of overview of what we found in the pr so here we found three
00:16:16.400 | actionable suggestions some things about missing resources a cache implementation that needs to be
00:16:22.880 | impacted and then a class cast exception error which we'll talk about next we provide a change list this
00:16:29.120 | change list gives you it's kind of a double click on the summary summarizes the key changes in your pr and
00:16:34.400 | then looks at what are the diffs that comprise that so you as a reviewer you get a quick
00:16:39.120 | double click on what actually matters in this pr now let's go through a couple of these suggestions
00:16:43.520 | so this first one is really about this resource management resource manager class is talking about a
00:16:50.080 | permanent resource leak if you don't fix this so we give you a detailed code suggestion
00:16:54.960 | which you can then one click apply as a commit or you can batch those commits
00:16:59.120 | the next suggestion that we'll look at really briefly is this non-thread safe static cache implementation
00:17:07.920 | and so we again we've given you a code suggestion we're suggesting a concurrent hash map instead of a
00:17:14.320 | hash map to avoid race conditions now let's talk a little bit about this third suggestion which i think
00:17:19.760 | is really important and highlights how we use sonnet so we really use sonnet to provide that human-like
00:17:26.080 | code review and those reasoning capabilities we've done a lot of work to really understand your code base
00:17:31.120 | and that's where sonnet kind of comes in so if you look here what we've got is this network data fetcher
00:17:36.480 | class it has been cast to a link list this is java if you're not familiar with java i'll just kind of
00:17:43.600 | walk you through it but really what's happening here is that this data uh you can see there's a data
00:17:49.280 | processor constructor that then cast something to an array list which is going to cause a class cast
00:17:55.120 | exception in production now how did the how did we find this well what bido has done is that it looked
00:18:01.520 | through the code and said hey for this network data fetcher you were trying to instantiate this data
00:18:06.400 | processor object this data processor object then well what is that we went and looked crawled your code
00:18:13.200 | understood it through abstract syntax trees a symbol index and sonnets reasoning and said hey this is cast
00:18:18.080 | as an array list and so now this array list you know is going to collide with this link list and going
00:18:23.120 | to create an exception in production this is probably an error that most humans wouldn't even find let
00:18:28.160 | alone you know most systems now if we want to shift this left we can even bring this to your ide so here's
00:18:33.120 | the same changes in your ide you can easily kind of right click and say hey i want to review my local
00:18:38.800 | changes stage commits etc run review those local changes the same agent runs in your ide again being driven by sonnet
00:18:47.600 | um and we provide the the output and then you can one click apply those suggestions
00:18:52.800 | uh so that's uh that's a little bit about how it how it works uh maybe i'll just mention one last thing
00:19:01.040 | so we have hundreds of customers and you know the question is well what impact is this really having
00:19:09.680 | so prs are closing in one-tenth a time so that pr that used to close take 50 hours to go from open
00:19:14.880 | to merge now goes open to merge in five hours these are three companies hundreds of engineers and bito's
00:19:22.640 | providing 80-ish percent of the feedback that a pr receives so we're taking care of the vast majority of
00:19:28.720 | the work and it's really being driven by the fact that ai provides that feedback in three minutes or four
00:19:33.520 | four minutes instead of one or two days thank you
00:19:36.560 | let's talk about something different so my name is hike and i want to talk about refusion so refusion
00:19:47.680 | is a generative music startup here in san francisco and we work on training frontier music models
00:19:55.120 | and then create an incredible product experience for crafting and exploring the art of music and our
00:20:00.480 | song lyrics writing pipeline is powered by claude so our core thing that we train is a
00:20:06.960 | diffusion transformer we train this from scratch and we think it's the most creative music model
00:20:13.040 | in the world we train it for quality for diversity for speed for controllability
00:20:18.080 | and as a fun anecdote this square of pixels here represents 30 seconds of music in our latent space
00:20:24.560 | um it's kind of crazy to think about how compressed that is but uh writing fantastic song lyrics is a
00:20:33.120 | crucial part of great songs um and current llms are um good at a very many things but writing good
00:20:40.560 | song lyrics they're they're still pretty cringy um but claude is the best for sure and we built an agent
00:20:46.640 | called ghostwriter that's essentially meant to help you write and refine song lyrics as an artist
00:20:52.720 | and basically focusing on you know diversity humor taste what flowing with the music itself and this
00:20:59.280 | is a crucial tool um in our product uh it's been used tens of millions of times now to to write song
00:21:05.680 | lyrics so let's go demo some stuff if we pop over to my computer so this is our home page refusion.com
00:21:13.600 | check it out it's really fun there's a lot of amazing music to explore on here um i will
00:21:20.720 | just punch in something here uh let's say experimental indie trip hop about the feeling of getting better
00:21:32.640 | after being really sick there's a song concept pop in a couple of those um in the meantime let me just
00:21:41.520 | play a song and there's a there's a ton of really deep editing workflows here for getting in stuff but also
00:21:47.040 | just getting started you can just type something in and start listening to music and go from there
00:21:51.120 | so let me just play a song from the the home page and then we'll go back and listen to what we made
00:22:03.520 | don't downplay my brine i've been clocking kills since the ink dried on that line you ain't copping
00:22:08.240 | these souls now limited supply too many laws left cold in the breeze then they mama selling gumbo just
00:22:13.760 | to cover the fees please flip the corner cash the equity that's the creed but most is crashing out
00:22:19.120 | living life on the speed in the dry cool breeze thick hair like alicia keys told god your manifest me
00:22:25.280 | queen take it out fill the closet let us stun on the scene but i'm cut different ain't a rapper who
00:22:30.160 | i run my steez news all right so let's listen to some of the songs here we've got the reawakening
00:22:41.200 | let's see what the lyrics look like here written by a ghostwriter with claude
00:22:46.640 | my muscles forgot what they're meant to do
00:22:53.680 | the ceiling fan spins in slow motion while i gather strength to move
00:23:00.480 | fingertips press against the mattress testing if i can hold my way pretty good let's see dawn fade
00:23:23.120 | so whenever ghostwriter generates we have uh kind of an iterative process of thinking about the
00:23:28.240 | concept of a song ideating about actually the context of the genre you're writing for because
00:23:33.280 | say the lyrics of a drum and bass song are super different the lyrics of a you know a folk song
00:23:38.800 | that might be storytelling um and that process of just trying to get something that actually fits with
00:23:45.200 | the music is actually really hard and subtle and requires a lot of iteration and refinement so an example
00:23:52.160 | of some of the kind of deep editing stuff we can do here so we can do remixing to create variations to
00:23:59.040 | extend from a certain section to replace a section actually to swap the stems like the vocals and the
00:24:05.520 | sound while keeping the other one the same and even capturing this idea of a vibe which is like a short
00:24:11.760 | audio snippet that you can then use the prompt instead of text so you can mix and match these things
00:24:17.040 | and just get really deep in your refinement and then with ghostwriter let's say if we were to use a
00:24:22.480 | prompt here um we were to use the lyrics here actually throw them up then we could type in here like
00:24:31.040 | add a spoken word intro in french and just pop that in and start start iterating so yeah that's the that's the
00:24:43.680 | app and the product and it's it's a really fun experience and um claude's claude's definitely the
00:24:48.960 | best for this kind of stuff so yeah come talk to me come talk to henry if you're interested in music and uh
00:24:53.920 | thanks um hey everyone i'm drew from the co-founder and ceo of create uh we're an ai text to app builder
00:25:04.960 | that lets anybody build working software products um on the internet and uh the exciting thing about
00:25:12.160 | create is it's an ai agent that can take in just natural language prompts and go end to end uh on
00:25:18.560 | building things so we got our start in web apps and claude is one of the base models that powers a lot
00:25:23.520 | of our code writing for the agent um but we're really excited about um our new mobile app builder
00:25:30.320 | that is for the first time letting people actually build mobile apps and so today i thought i'd just
00:25:34.240 | go ahead and show it off so i'll switch over to the live demo um and we're in beta but uh somebody just
00:25:42.640 | actually emailed me an idea a few weeks ago for a family memory app that lets people uh basically
00:25:50.240 | store their memories on their phones uh upload their uh different images and oops live demos
00:26:09.600 | uh store their different images and uh get going on building their family memory app so let me actually
00:26:16.080 | go create a new project and type this in and say make an ios app for this
00:26:22.800 | and as you type creates agent starts getting to work um it's going to start building out all the core
00:26:30.800 | pages um as well as the back-end functionality for what you need um and so it will also pull in other
00:26:38.560 | integrations you can you might need to use um and go ahead and get set up i uh i am actually going
00:26:45.440 | to go ahead and switch to some pre-loaded things so the basic idea is you prompt uh and you're able
00:26:51.040 | to actually generate full applications um the cool thing about create oh there we go starting to build a
00:26:56.880 | little bit um is that um it also comes built in with back-ends and front-ends and everything you kind of need
00:27:03.920 | from the uh from the database to the actual core off um and so here it's actually building out my family
00:27:10.160 | memory's front-end but then in a second you'll see that also issues the like schemas uh for the database
00:27:16.960 | and goes ahead and deploys a full database and hooks up all the functions needed to talk to this front-end
00:27:21.760 | we are really excited about the fact that non-technical users in the hundreds of thousands are starting to
00:27:27.840 | pour into create to actually fully build apps um without needing to go to the id and truly just
00:27:33.520 | work from prompt land um and so one of the biggest things that we've been working on is actually the
00:27:38.720 | ability to also fully submit from create as well um create after your app is done then goes ahead and
00:27:45.520 | submits your app straight to the app store and builds a build for you um there's actually an app we
00:27:50.560 | built uh in a day um you can just in one click publish it and submit to the app store and this
00:27:56.560 | will go ahead and kick off a build um we actually already started uh got this app in the app store
00:28:02.160 | so if anyone here wants to download it and play with it it's an ai app that lets you very quickly take
00:28:06.480 | drawings and turn it into uh ai images called draw daily um and then in the beta the thing that's been the
00:28:12.720 | most exciting uh for us is just seeing all the cool apps that are being created so we just hosted a demo day
00:28:18.560 | and this is william and he is making an app to help you memorize meaningful connections
00:28:24.320 | and details of people's lives and pull it up and so you just talk to the app and it pulls up
00:28:31.840 | special information about your contact info and and helps you remember anything for sales calls or
00:28:39.520 | anything else you want or we also had a student at berkeley who's building the scholarship app that he's
00:28:47.280 | always wanted um where he's constantly filling out grants um and unable to kind of do it automatically
00:28:53.440 | and so he's building scholar gpt and a lot of these like the core agent of how they built we're using
00:28:59.520 | you know prompt caching tool calling a lot of the core primitives that anthropic makes available to
00:29:03.840 | actually get them to success and just to show a few more um blaze is actually a basketball coach
00:29:10.000 | and for the first time ever he's using create to get rid of all of these spreadsheets and drills he does for
00:29:15.840 | his uh coaches and instead uh you build a full player coach app that lets them uh download uh drills uh put
00:29:25.840 | their lesson plan out and then uh see animations on in the app of how they uh should be uh getting used
00:29:34.240 | and then finally uh we've also had personal finance apps being created um so like a personal ai money
00:29:40.240 | coach for gen z with full rag um that also has a claude powered uh assistant that can figure out your
00:29:47.600 | monthly income and give you personalized financial recommendations and so part of what we were like
00:29:52.560 | most excited about is just the explosion of who and how uh software will be made um when you give people
00:29:58.480 | of these tools and very excited to partner with uh claude to power all of this thank you
00:30:07.120 | Thank you.
00:30:08.120 | Thank you.