0:19 [Music] [Applause]
0:22 [Applause]
0:26 So first question for you uh what is
0:29 definitely happening by the end of 2026
0:32 AI agents ship code directly to prod in
0:34 your environment, right? Not in like
0:38 some uh playground. Uh voice AI replaces
0:41 text for most business communication.
0:44 Inference cost dropped below a cent per
0:48 million tokens or wall-ally like we're
0:51 all chilling.
0:56 first one ship ship code directly to
0:58 prod. Okay, this is a hopeful set of engineers.
1:01 engineers.
1:03 All of you want to get rid of your own
1:10 The good thing is I also don't have
1:11 internet so I can't look at my next question.
1:17 No, it's going to be good. It's going to
1:19 be good. Um
1:25 I present from your phone. Uh, no. I was
1:27 going to go through poll questions while
1:44 While this is happening, I'm actually
1:45 just going to introduce myself so we're
1:47 not wasting the time. Um, my name is
1:51 Sarah Goa. I, uh, helped start a AI
1:52 native venture fund. It's called
1:55 Conviction. And we got going about two
1:57 and a half almost three years ago now
1:59 just before the starting gun of chat
2:02 GPT. Um as always in technology
2:04 investing most of life it's better to be
2:06 lucky than right. Hopefully you can be a
2:10 little of both. Um uh and and the point
2:13 of having a new venture firm I I worked
2:14 at Greylock. It's kind of a
2:16 traditionalist venture firm a great one.
2:17 My partner Mike Vernal used to work at
2:19 Sequoia. You guys have probably heard of
2:23 them. uh was that we think like actually
2:24 you know at risk of sounding like those
2:26 people this time it's different right um
2:28 that this is the largest technology
2:30 revolution that we get to be a part of
2:33 and that there's so much change in the
2:35 technology the types of businesses you
2:37 can build the product decisions you make
2:39 what challenges these startups and big
2:41 companies face that you know maybe
2:43 there's opportunity for like a startup
2:47 VC as well and so um you know I'm I'm
2:49 thrilled to be working with like really
2:50 interesting people in the industry so
2:53 far. Uh Mike and I are investors in
2:55 companies like cursor, cognition,
2:58 mistral, thinking machines, Harvey, open
3:01 evidence. So a mix of um base 10 like a
3:04 mix of uh infrastructure model and
3:06 application level companies and you know
3:09 one more are my kids coming up yet?
3:13 Okay, cool. Um one more uh just
3:15 observation from the last two and a half
3:17 three years of doing venture. I I was an
3:20 investor for about 10 years before that
3:23 is I have never seen the like just the
3:26 uptake from users that has been possible
3:28 in the last couple years. I'm sure all
3:30 of you have experienced that it is not
3:33 trivial. Um you know AI product and AI
3:36 engineering uh and this is kind of the
3:37 theme of my talk so I'm sorry to give
3:39 away the punch line but it's quite a bit
3:42 harder than people had hoped. Um but the
3:44 the value creation is massive. Um, we
3:48 see companies going from 0 to 10, 50,
3:50 100 million in run rate very, very
3:51 quickly, faster than we've ever seen in
3:54 any technology revolution before. Um,
3:57 and I get asked a lot like where are we
3:59 in the AI hype cycle? Is the winter
4:02 coming? Is this like infinite AI summer?
4:05 And I would say um having actually been
4:07 an investor or an operator through a
4:10 macro cycle at this point like I try to
4:12 pay very little attention to what the
4:13 marketing world is saying or even what
4:16 the markets are saying, right? Because
4:17 you know if you're if you're an operator
4:19 or an investor
4:21 maybe you care about what the stock
4:23 price does every day, but really you
4:24 want to figure out if the company you're
4:26 working for or starting is going to work
4:28 long term, right? And if the products
4:29 are going to work long term. And the
4:31 things that I get most excited about are
4:37 seeing like crazy usage numbers. Okay.
4:41 Thank you, amazing AV team.
4:43 Okay, I'm gonna I'm gonna go real quick. Um,
4:46 Um,
4:50 where are my presenter notes?
4:51 Okay, we're we're just going to keep
4:54 going. It's cool. It's cool. Um, so I
4:56 want to talk really quickly about uh
4:58 just a few things today. I think we lost
4:59 a little bit of time, but let's let's
5:01 say let's talk about capabilities, what
5:03 we're seeing work in the market, and
5:07 then um uh maybe some advice on like
5:09 what to build if those are, you know, a
5:11 question you're considering. Uh I think
5:13 the shorthand that we're going to use in
5:16 this presentation is like cursor for X,
5:17 right? Uh and I do think that's a really
5:20 massive opportunity. Uh the first thing
5:22 in capability for this past year is
5:25 clearly reasoning. Um, reasoning is a
5:26 new vector for scaling intelligence with
5:28 more compute. The labs are really
5:29 excited about this because they get to
5:31 spend more money and get more output.
5:34 Um, but we should also be really excited
5:36 about this in terms of unlocking new
5:38 capabilities. Right? If you just put
5:40 aside how it works, it's a confidence
5:43 boosting implementation detail. Um, but
5:45 we should expect more capability. You're
5:48 unlocking a new set of use cases like
5:51 transparent highstakes decisions where
5:53 showing the work matters. uh sequential
5:55 problems, problems where you need to do
5:57 systematic search. I I think this looks
5:58 like a lot of problems that we're
6:01 excited about and um face in knowledge
6:04 work every day. Uh as you have just seen
6:06 demos of and I'm sure are working on
6:08 given reasoning, people are really
6:12 excited about agents. um to put a you
6:14 know I want to do like the Steve Balmer
6:15 impression that's like agents agents
6:19 agents agents agents agents but uh I um
6:21 you have to give me more than 12 minutes
6:24 to like get that sweaty
6:26 uh but but like the non-marketing
6:28 definition that I think of is it's
6:33 software that um uh it takes some set of
6:36 steps it like plans it includes AI it
6:38 takes ownership of a task and it can
6:39 hold a goal in memory
6:41 you know, try different hypotheses,
6:43 backtrack. It ranges from super
6:45 sophisticated to super simple. Um, some
6:47 of the tools that might use to
6:49 accomplish a task include other models
6:52 or search. And largely, it's just like
6:54 AI systems that do something. Um, and
6:56 that's not a chatbot that looks more
6:59 like a colleague. Uh, and you know, one
7:00 thing that I think we have a really
7:03 unique vantage point on is, uh, we back
7:04 a small number of companies at
7:06 conviction, but we also run a grant
7:08 program for AI startups. It's called
7:09 Embed. We get thousands of applications
7:12 every year. Um, and includes like user
7:13 data and revenue data and like really
7:16 amazing people and the number of agent
7:18 startups has gone up 50% over the last
7:20 year and a lot of them are working like
7:22 we do see stuff that's working in the
7:24 real world and uh that's super exciting.
7:26 Uh, other modalities are progressing
7:28 too. I'm sure a lot of people are using
7:31 voice, video, image generation um, even
7:34 beyond you know studio gibli. But you
7:36 have companies like Hey Genen and 11 and
7:38 Midjourney that are rocketing past 50
7:40 million of AR. These are real businesses
7:43 now. Um, I want to see if I can quickly
7:45 play for you. They told me to express
7:47 myself, so I did. They told me to
7:50 express myself, so I did. Now I'm banned
7:53 from three coffee shops. Hands can hurt
7:55 or heal. That's the difference between
7:57 chaos and creation. So if you're
7:59 wondering where Q3 is headed, So if
8:01 you're wondering where Q3 is headed,
8:03 here's the thing. Consistency always
8:06 beats urgency. We've got the projections
8:08 ready and let's just say it's looking
8:10 solid. I would definitely recommend it
8:12 to anyone. I would definitely recommend
8:15 it to So I I think like if you just are
8:17 looking for artifacts of improvement,
8:19 this is from a company called Hey Jen.
8:22 Um you can make clones of yourself of
8:24 fake people and like you have gestures
8:27 and expressions that uh reflect emotion
8:30 and content now, right? So these models
8:31 work together and like I don't know
8:33 about you guys but looking at that last
8:34 gal like I feel influenced. I don't know
8:37 what the bunny is but I would buy it. Um
8:39 and and and so I think like huge swaths
8:40 of the economy are going to be affected
8:43 by this sort of multimodality. Um some
8:45 investors or operators would say
8:47 multimodality would just be for niche
8:49 verticals that enterprises don't have
8:51 you know your average enterprise doesn't
8:53 have that much voice video image data
8:55 today. Um, but I think that changes,
8:56 right? When you can do stuff with this
8:58 data, when it is structured and
9:00 understood, there's more reason to
9:03 capture it. And I think of like how much
9:04 video do all of us watch every day? It's
9:06 one of the highest bandwidth
9:07 communication methods, and we're just
9:09 going to use more of it. Um, we think
9:11 voice is where we're going to see uh
9:14 applications first in business workflows
9:16 um because it's already a very natural
9:18 communication mode. So, uh, everything
9:20 from medical consults to lead
9:23 generation, places you already had
9:24 business voice, you just couldn't scale
9:26 it before. Uh, I I think that's where
9:27 we're going to see it first. But as
9:30 these other modalities become more
9:32 controllable and also less costly, we
9:34 should see all of them. Uh, I I think
9:35 it's safe to say you can expect
9:39 capability improvement in every part of
9:40 the model layer, which is really
9:41 exciting. A lot of people were talking
9:44 about the uh the data wall or like the
9:46 end of AI summer, but for anybody who's
9:49 building applications, I I'm at least to
9:51 tell you one person's opinion is uh it's
9:55 not coming. Um and and then usefully for
9:58 all of us, uh that market for model
10:00 capabilities is getting more
10:03 competitive, not less. Um Sam Alman
10:05 himself, I think, said it best. Last
10:07 year's model is a commodity, which is a
10:09 scary thing for a model provider to say,
10:10 because last year's model is now pretty
10:12 damn good, right? The numbers tell the
10:15 story. GPT4 went from $30 per million
10:18 tokens to $2 in about 18 months. The
10:20 distilled versions of that are like now
10:21 10 cents. So, we can really use them
10:24 very broadly. Um, if you look at this
10:27 chart, uh, green is Google, yellow is
10:29 anthropic. So, you see, you know, it's a
10:31 real mix. This is data from Open Router.
10:34 So, thank you Open Router for that. But
10:36 um you really saw Claude cut into
10:39 OpenAI's market share and Google come
10:41 roaring back with Gemini. Uh this data
10:42 is obviously a little biased because a
10:44 lot of people just go direct to OpenAI,
10:45 but if you're into multimodel that there
10:47 really is a mix and you do have credible
10:49 new players like SSI and thinking
10:51 machines, some of the best researchers
10:53 in the business with orthogonal
10:55 technical approaches um entering the
10:57 frey as well. And I'm sure many of you
10:59 have experimented with DeepSeek uh
11:02 coming out with releases of you know
11:05 both base and reasoning models that are
11:07 uh reasonably competitive with a claimed
11:09 fraction of the training cost like we
11:11 should just assume that open source will
11:13 do as open source does and we can rely
11:15 on the model market to compete for our
11:16 business which is really exciting. Um
11:18 and so the view is plan for a world that
11:21 is multimodel. um tools like open router
11:23 or inference platforms like base 10 help
11:25 that uh and uh I think like be
11:28 comfortable with that I I am okay so we
11:30 have all this capability let's ship uh
11:32 shift quickly to the application layer
11:34 we have to start with cursor uh a
11:36 million to 100 million of AR in 12
11:38 months and half a million developers I
11:41 assume all of you uh zero sales people
11:43 to start that's not growth that is a
11:45 killer application um cognition which
11:47 started with more autonomy is already
11:49 the top committer in many companies
11:51 feeling a little threatened but also
11:52 excited because recruiting is hard. And
11:55 then Windsurf who's on a tear itself and
11:56 really beloved is being acquired by
11:59 OpenAI for $3 billion. So we know for
12:02 sure that the labs don't think that they
12:04 can just you know steamroll everyone
12:07 right lovable and bolt hit 30 million of
12:11 AR each in a handful of weeks uh helping
12:14 non-engineers vibe as well. So you know
12:17 our our our ranks are expanding. Um and
12:19 I think it's useful to just like analyze
12:21 a little bit why code was first. Uh
12:25 fundamentally it is text with it's log
12:26 it's like logical language with
12:29 structure right so much of coding is
12:31 sophisticated boilerplate like we all
12:33 love engineering but some of it is like
12:36 craft work not new algorithm work um you
12:39 don't need AGI to write a like uh an API
12:42 endpoint or um a react component.
12:44 Second, you have deterministic
12:46 validation. You can automatically check
12:49 if code works, run tests, compile,
12:51 execute, do things developers would do.
12:54 And third, researchers believe code is
12:56 crucial for AGI, right? So, they poured
13:00 resources into it. Um, and uh code
13:02 became a key benchmark and a training
13:03 priority and an area for data
13:05 collection. But I think the last point
13:08 is um the money point to me. Uh
13:11 engineers built tools for engineers.
13:12 They understood the workflow intimately
13:14 and that made all the difference. And
13:16 that last part is the playbook for every
13:17 other industry. I'm sure people are
13:19 building things that serve beyond
13:21 engineers. And I don't think the winners
13:24 will just be AI experts learning those
13:26 domains. They'll be customer centric
13:29 like problem centric builders who
13:30 understand AI and then redesign
13:32 workflows from first principles around
13:34 manipulating those models. Um and so I
13:35 think that's really the opportunity to
13:38 build cursor for X. Um let's think a
13:41 little bit about what that means. Cursor
13:44 is not a single model. Uh you know one
13:46 model's doing diffs, one's doing merge,
13:47 one's embedding the files. They
13:50 manipulate and package up the context.
13:52 They prompt the models very skillfully.
13:54 They let engineers avoid repetitive
13:56 tasks and standardize with things like
13:58 um cursor rules. And then if you're
14:00 using cursor in a team or even yourself
14:02 regularly, retrieval accuracy gets
14:04 better the more you use it with coverage
14:06 and freshness. And so all of this
14:08 happens in a UX that makes sense, right?
14:10 Like I, you know, I use VS Code. I'm
14:12 familiar with it. My shortcuts work. Um,
14:15 and I make it safe to say yes, right?
14:18 Like green for add and red for subtract
14:20 makes sense. I can scroll through it.
14:21 Um, and it's fast enough that I don't
14:24 get frustrated. So my my view is cursor
14:26 if it's a wrapper, it's like a very nice
14:29 thick perhaps 14 or 15 billion dollar
14:30 wrapper, right? It's like if your
14:35 burrito was 80% wrap and 20% fill, but
14:36 you got to choose the fill and there's
14:38 like an empty like an open market for
14:41 fill, right? Um, and so where's the pro
14:43 where's the value now? It may not be in
14:44 the protein. It's kind of in the
14:47 company. Um, so like if we try to
14:50 generalize that recipe a little bit, if
14:54 you are building a generic text box like
14:55 unless you're just like learning to do
14:58 this, please don't like OpenAI already
15:00 one that or it's just not very valuable
15:02 to do. So your domain knowledge, your
15:04 workflow knowledge can be the bootstrap.
15:07 If you already know what users in your
15:09 industry need, don't make them explain
15:11 it. Uh, build products that show up
15:13 informed. They collect and package
15:14 context automatically including from
15:17 other sources not just natural language
15:18 presented to the models use the right
15:20 models at the right time now known as
15:23 orchestration and present the outputs to
15:25 the users thoughtfully right um so I do
15:27 not think this is the end of the guey uh
15:29 I I think you can capture and enable
15:31 workflow with these models and all this
15:33 requires taste and a ton of work I' I'd
15:35 argue that like some version of this
15:37 recipe is much of the work each of us is
15:40 going to do so don't listen to the labs
15:42 from a user experience perspective The
15:44 prompt is a bug, not a feature. I think
15:45 it's like a stepping stone. Don't make
15:48 me think as a user. The best AI
15:50 products, they feel like mind readading
15:52 because they are. Um, there's enormous
15:54 headroom in building these products. And
15:55 I I think that's really exciting because
15:56 that's what most of us in this room have
15:59 alpha on. Uh, what is a software company
16:02 if not a very thick like workflow
16:04 wrapper most of the time? That's true in
16:08 2015. It's true in 2025.
16:11 Um, besides code, where might you go
16:14 apply this? We think the opportunities
16:16 to build value around the LLMs exist in
16:19 every vertical and profession. Uh, but
16:21 here's something counterintuitive.
16:23 Beyond coding, one of the things that
16:25 I've been surprised by is that the most
16:27 conservative low tech industries seem to
16:29 be adopting AI fastest. We call this the
16:32 AI leaprog effect internally. Um, these
16:33 are three portfolio companies. Um,
16:37 they're working. Sierra resolves 70% of
16:39 uh customer service queries for their
16:41 customers. They serve people that you
16:44 know you guys use like SiriusXM or ADT.
16:47 Harvey is you know two years in well
16:50 over 70 million of ARR. It's AI is
16:52 essential now to being competitive in
16:54 the legal industry. Um there's a company
16:56 called Open Evidence uh which helps
16:58 doctors stay upto-date with medical
17:00 research. You have to be a clinician to
17:02 use it but you know you give it your
17:03 medical ID number and you can do
17:06 intelligent search against um uh medical
17:09 research uh at the point of clinical
17:11 decisionmaking. Today it reaches a third
17:14 of doctors in the US weekly and the
17:16 average user uses it daily, right? And
17:19 so I think there's just examples of, you
17:22 know, huge value beyond chatbt. These
17:24 are companies that know their customer
17:26 and are solving real problems. As a as a
17:28 piece of trivia that you may or may not
17:30 know, um Brett at Sierra is the chairman
17:34 of the board at OpenAI. Um OpenAI was
17:38 Harvey's uh seed investor. And if you
17:40 know these people are not fretting about
17:41 thin rappers like I suggest you don't
17:44 either. Okay. Finally, I'll make an
17:46 observation. A lot of people are excited
17:48 about full automation. Now I'm sweaty
17:50 enough. So agents agents agents agents
17:53 agents agents. Um but when we analyze
17:56 the applications to embed I said you
17:58 know it's gone up to 50% you know
18:01 doubling a applications for agentic
18:04 startups in the last year. Um I I think
18:06 some people think co-pilots are
18:08 yesterday's news. They want to get to
18:09 the endgame, right? Like you know your
18:12 colleague and AGI. But in terms of what
18:14 works, like the data on what's driving
18:16 revenue, uh I think co-pilots are still
18:18 really underrated. We see a whole
18:21 spectrum of how much automation and I
18:23 think the uh Iron Man analogy is still
18:26 really great here. Tony Stark's Iron Man
18:28 suit augments him, right? He can do all
18:30 these amazing things, but could also fly
18:32 around on command, could do some basic
18:35 tasks without Tony. And my experience
18:36 with these companies has been that human
18:39 tolerance for failure or hallucinations
18:41 or lack of reliability, it just reduces
18:43 dramatically as latency increases,
18:45 right? Um, so the path of least
18:47 frustration today for many domains is to
18:49 build great augmentation and then just
18:51 ride the wave of capability because we
18:54 know it's coming. And so my advice for
18:56 many domains would think about like you
18:58 know build the suit and you can extend
19:00 out to the suit that flies on its own
19:04 once Tony or any of us is wearing it. Um
19:05 Um
19:06 I'm not going to go through each of
19:09 these mostly because I lost time but um
19:10 there are a ton of opportunities. We put
19:12 requests for startups on our website.
19:14 We're interested in a couple different
19:18 categories of things. They go from uh um
19:21 like just good fit for purpose like the
19:23 law is a space of lots of text
19:26 generation, right? Um to things that
19:28 weren't possible before AI. My partner
19:30 Mike will say like this is a really
19:32 interesting era of machines
19:34 interrogating humans. What can you do if
19:36 you can go like collect data on demand
19:38 from people? Um we could talk to every
19:40 customer, not just the top 5% by
19:44 contract value. Um, we could root cause
19:47 every alert proactively, right? Versus
19:49 like just firefight. Um, and the mental
19:50 model is how can you build as if you had
19:53 an army of compliant, infinitely patient
19:56 knowledge workers. Um,
19:58 you know, one aside here is I think
20:01 there are many hard problems where like
20:03 the basic premise is the answer to them
20:05 is not in common crawl, right? The
20:07 reasoning around them is not in common
20:09 crawl. So um this would be robotics,
20:11 biology, material science, physics,
20:15 simulation. Um they require clever data
20:17 collection. Um probably interaction with
20:20 atoms, not just bits. Super scary uh for
20:21 a software person, but I think the juice
20:23 is worth the squeeze, right? The same
20:25 reasoning that crushes math olympiads
20:27 can seemingly navigate molecular space.
20:28 And I think there are some really
20:30 fundamental questions for um human
20:32 society that can be answered when people
20:34 work on these problems. And uh it's it's
20:36 really cool as a machine learning person
20:39 to meet people in their at the top of
20:41 their field at the intersection of
20:42 machine learning and all of these other
20:45 areas because like you guys would also
20:47 the same architectures apply right and
20:49 and that's just um that's really exciting.
20:51 exciting. Um
20:53 Um
20:56 how should we think about defensibility?
20:59 Did this advance?
21:01 Okay. So, um, one last point and then
21:04 I'll conclude here. Uh, some would say
21:06 stay out of the weight of the labs.
21:07 Don't pick up pennies in front of the
21:10 steamroller, right? But I would offer,
21:11 um, what I think is an uncomfortable
21:14 truth. Execution is the moat in AI. Um,
21:16 and that's available to all of us.
21:18 Cursor arguably did not invent code
21:20 completion. They did not invent the
21:21 model. They didn't invent their product
21:24 surface area, right? They just
21:26 outexecuted on every dimension of this.
21:28 They shipped a great experience faster
21:30 than their competitors could copy. and
21:31 they capture the hearts and minds of
21:34 developers at least in this term. Um I
21:35 don't I don't mean this to be cruel but
21:37 I often get asked about like counter
21:39 cases and the importance of first mover
21:42 advantage. Let's be brutally honest. In
21:44 contrast, like Jasper had first mover
21:46 advantage brand. They raised $125
21:49 million, but its first product was a
21:51 series of prompts and a text box and
21:53 like very good SEO. And like you have to
21:55 keep running like ChatBT, you know,
21:57 crushed the first iteration pretty
21:59 quickly. And so, uh, I I don't think
22:00 this is satisfying advice, but I think
22:02 it is like real from the trenches. Build
22:04 something thick and stay ahead. And like
22:06 no domains are out of question. Um,
22:08 magical AI experiences, they build
22:11 customer trust and drive adoption. And a
22:14 lot of the data we need to improve these
22:16 experiences and the context we need it
22:18 is not easily available today. And that
22:22 advantage is you know uh open for the
22:25 taking and not for the labs.
22:27 So I I guess in conclusion I think the
22:30 opportunity is early and really massive
22:32 like I've made a career bet on it. Um I
22:33 I think many of you are. We're in the
22:36 dialup era of AI and we're moving pretty
22:38 quickly to to broadband. Um, Instagram
22:40 came four years after the iPhone. Like I
22:42 was was there when Greylock made that
22:44 investment. Um, Uber five years. Uh,
22:46 Door Dash six, right? So, the truly
22:49 transformative companies. They weren't
22:50 necessarily the first people to
22:53 recognize the changes or the opportunity
22:55 is those who reimagine the experiences.
22:57 Um, and the game board keeps getting
22:58 shaken up. That's the thing that's
23:00 different this time, right? It's like
23:02 getting a new iPhone that's actually
23:05 different every 12 months. And um so you
23:07 have like new model release, new
23:08 capability breakthrough, you know,
23:11 onetenth the cost. And every time the
23:13 game board turns, I think there are like
23:16 there's an opportunity to to win again.
23:18 Okay. Um so I I'll give you one last
23:20 sentence and be chased off the stage.
23:22 This was not my fault. Um here's what I
23:24 really want you to remember. Uh you as
23:26 the engineers got the magic first. Um
23:28 the anthropic like economic index said
23:31 that 40% of use was still coding. That's
23:34 not like 40% of the economic opportunity
23:35 in the world, right? And so it is the
23:38 job of everyone in this room and you
23:39 know globally online to be the
23:41 translators for the rest of the world.
23:42 So I encourage you to build something
23:45 revolutionary. Thanks. [Music]
Ekstensi Chrome