Latest / The Learning Corner by Precursor / Episode #46: AI Troubles for Taco Bell, Is Non-Consensus Dangerous, Are Software Economics Going Off a Cliff with AI

Transcript

Episode #46: AI Troubles for Taco Bell, Is Non-Consensus Dangerous, Are Software Economics Going Off a Cliff with AI · The Learning Corner by Precursor

• 0:00
  Contrarian shouldn't be an identity or a strategy.
• 0:02
  I tell people you should focus on finding companies that are good.
• 0:05
  Sometimes companies that are good are misunderstood, so investing in them will feel,
• 0:09
  quote, unquote, contrarian because other people don't see.
• 0:12
  But if you just say, I'm just gonna do the opposite of what everybody else thinks as a
• 0:16
  strategy, that seems silly to me.
• 0:18
  It's the week of September 1.
• 0:21
  Welcome to the Learning Corner, a Precursor Ventures podcast.
• 0:28
  Welcome to the Learning Corner by Precursor.
• 0:30
  I'm Mia Farnam, and I'm joined by RGP and managing partner Charles Hudson where we will
• 0:35
  be talking through the articles and Goodreads that caught our team's attention over the past
• 0:39
  week.
• 0:40
  Our first topic is a recent Wall Street Journal article about Taco Bell's bumpy rollout of
• 0:44
  voice AI across stores that was kicked off over a year ago.
• 0:48
  Customers have reported glitches, long wait times, and even trolled the system with prank
• 0:53
  orders like thousands of cups of water.
• 0:55
  Taco Bell's chief digital officer admits the tech is still inconsistent even in his own
• 1:00
  experience and says humans are often still faster in busy stores.
• 1:04
  The company is now rethinking its strategy leaning into a hybrid model where staff decide
• 1:08
  when to use AI versus step in themselves.
• 1:11
  They're not alone in this decision.
• 1:13
  In fact, McDonald's and Wendy's have both had to revise or pause similar AI pilots.
• 1:18
  Taco Bell's parent company, Yum Brands, is now working with NVIDIA to improve their strategy.
• 1:22
  But across the industry, everyone's still figuring out where voice AI can thread the
• 1:27
  needle on value add.
• 1:28
  Similar to last week's episode where we talked about the high failure rates of AI pilots,
• 1:32
  stories like these suggest we're starting to make progress in figuring out what works and
• 1:37
  what doesn't when it comes to enterprise AI rollouts.
• 1:40
  Anything else that you wanted to note here from your perspective?
• 1:43
  I love a good anecdote, and the 18,000 cups of water cracked me up.
• 1:47
  I continue to think this is a positive development.
• 1:50
  Yeah.
• 1:50
  Because my sense is what's happened in a lot of companies, talking to my friends who work in
• 1:54
  bigger public companies is the message from the board and the CEO was like, roll out AI.
• 1:59
  Get on it ASAP.
• 2:01
  And vendors said, we have an AI product that can do everything you needed to do.
• 2:06
  Just like turn the whole operation over to AI.
• 2:08
  And many of those vendors had no real world experience.
• 2:11
  But I think either naively or in a salesy way had confidence in their product's ability to
• 2:15
  deliver.
• 2:16
  And what's happening is, like with any new technology, once companies actually use it,
• 2:20
  they can figure out the limitations.
• 2:22
  So I think for those who poke fun at Taco Bell, I think what Taco Bell probably learned is that
• 2:27
  AI can do a handful of things super well today on its own, a handful of things pretty well
• 2:32
  with human supervision, and there's probably a set of tasks and things they hadn't thought
• 2:36
  about, like these prank orders that a human would be able to detect that it's not ready to
• 2:40
  handle.
• 2:41
  And so my hunch is that Taco Bell's conclusion will not be we should not use AI at all.
• 2:46
  It will be we should be more judicious and thoughtful about, as their chief digital
• 2:50
  officer said, where humans put their hands on the wheel or where we let AI do its thing.
• 2:54
  And I think there are going to be more stories like this over time as companies and look.
• 2:59
  I think not to belittle Taco Bell's business, drive thru is a pretty safe domain in which to
• 3:05
  test voice AI.
• 3:06
  Like, no one's gonna lose their life or lose their livelihood if you get my Chalupa order
• 3:11
  wrong.
• 3:12
  Yeah.
• 3:12
  And I think we're gonna continue to learn that there are limitations for what unsupervised AI
• 3:16
  can do in actual production deployment environments.
• 3:19
  That's a great segue into my next question.
• 3:21
  Because when I was looking through the comments across accents going through, there were lots
• 3:25
  of different quippy things coming in and out from frustrated customers.
• 3:29
  But it was all done in kind of in, like, light hearted fun.
• 3:31
  There were some people who kind of pointed out that the low stakes nature of this experiment
• 3:36
  is precisely what made it tolerable in this moment.
• 3:39
  So if it was applied to exactly what you mentioned, higher stakes industries like health
• 3:43
  care, fintech, similar failures could actually result in some serious consequences.
• 3:48
  It could range from financial fraud to loss of life.
• 3:52
  Curious to kind of blow that out a little bit more.
• 3:55
  Like, what kind of guardrails should exist in these critical industries, especially when
• 4:00
  others are able to experiment more freely?
• 4:03
  Like, Taco Bell can play around with this and get it wrong a 100 times, then it's okay.
• 4:07
  Can you get it wrong a 100 times when it is applied to health care and fintech or even
• 4:12
  broader industries, etcetera?
• 4:13
  I think it's harder in health care in particular because there are some clear and
• 4:17
  obvious use cases where AI seems to be as good or better than humans and more consistent, but
• 4:23
  also it makes mistakes and hallucinates.
• 4:25
  I think the challenge is I don't think there's a universal belief in what the trade off
• 4:29
  between those two things should be.
• 4:31
  I've met some people who are like, until AI is perfect, which is a higher standard than we
• 4:35
  hold humans to, it should never be used in these and I'm like, well, that's probably not
• 4:38
  where we're gonna end up.
• 4:39
  And I've had other people who've said, look, other people's pain and suffering is the cost
• 4:44
  of us getting this stuff right, and so we should use as much of it as we possibly can and
• 4:48
  live with the consequences also feels irresponsible.
• 4:50
  So I think socially, we're gonna have to work out what are the acceptable error rates or or
• 4:55
  if we're gonna use a lot of AI.
• 4:56
  Ultimately, somebody we talked about this last week.
• 4:58
  Somebody has to bear the risk and liability in the end.
• 5:01
  And with humans, it's I wouldn't say easy.
• 5:03
  It's straightforward.
• 5:03
  It's like the person who did the thing bears the responsibility.
• 5:06
  And what happens when the person who did the thing isn't a person but an AI agent or a bot
• 5:10
  or an algorithm?
• 5:11
  Like, who's gonna bear that liability?
• 5:13
  I think we'll we'll sort this stuff out.
• 5:15
  And that's why I think in health tech, we're seeing a lot more applications of voice AI or
• 5:18
  AI more in back office operations to kind of streamline note taking, transcribing, etcetera.
• 5:24
  I think it'll be a lot more of a discussion when it starts to bleed into, I don't know,
• 5:29
  like on a robotics nature of like taking part in surgeries, etcetera.
• 5:32
  All of which like already happened today.
• 5:34
  Yeah.
• 5:34
  Like there's a level of machine tech that do really in-depth surgeries that human hands
• 5:39
  can't specifically do.
• 5:41
  So that isn't necessarily new to everyone.
• 5:43
  The topic that we talked about last week around like mental health and therapy and chatbots,
• 5:47
  etcetera.
• 5:47
  OpenAI released an announcement this past week of trying to create some guardrails, especially
• 5:52
  with all the news that's kind of come out of difficult consequences of chatbots bleeding in
• 5:56
  a bit too freely with users in areas that maybe they're not quite ready to do so.
• 6:01
  Lots of things to be ironed out there for sure.
• 6:03
  Our next topic is a recent post from Martin Casado on X where he shared the brief but quite
• 6:08
  impactful tweet.
• 6:09
  The idea that nonconsensus investing is where the alpha is is actually quite dangerous in the
• 6:15
  early stage.
• 6:15
  Follow on capital tends to be more and more consensus aligned.
• 6:19
  Obviously, this calls into question the legitimacy of the long held belief that non
• 6:23
  consensus investments are the strongest ones and certainly brought people to the comment
• 6:28
  section to make their opinions heard.
• 6:30
  Martin later posted, quote, I'm much less interested in whether consensus based investing
• 6:34
  works and much more interested in the dangers of being non consensus.
• 6:38
  He says, I actually believe the following two statements are roughly true.
• 6:42
  A, VC consensus is often hype driven and totally disconnected from markets, and b, early
• 6:47
  markets are efficient and therefore good deals get identified early, and that's reflected in
• 6:52
  marks.
• 6:53
  In other words, deal hunting is the wrong approach in VC.
• 6:56
  Charles, I sense this is a topic that was widely discussed prior to Martin sharing his
• 7:01
  post.
• 7:01
  Yeah.
• 7:01
  Wanting to get your perspectives on this long standing discussion or debate.
• 7:05
  I think this is one of those evergreen topics.
• 7:08
  This and, like, should you send monthly updates seem to, like, bubble back up into the the
• 7:12
  discourse.
• 7:13
  I think people seem to forget the other half of being contrarian.
• 7:16
  It's the end right part.
• 7:18
  I've never understood this, like, we're a contrarian investor.
• 7:20
  I'm like, you can't short private companies.
• 7:23
  Yeah.
• 7:23
  You can only go long.
• 7:24
  You can go long or you cannot participate.
• 7:26
  Those are actually your your two options.
• 7:28
  Until you can short private companies, you can't really be contrarian shouldn't be an
• 7:33
  identity or a strategy.
• 7:34
  Tell people, you should focus on finding companies that are good.
• 7:37
  Sometimes companies that are good are misunderstood, and so investing in them will
• 7:41
  feel, quote, unquote, contrarian because other people don't see.
• 7:44
  But if you just say, I'm just gonna do the opposite of what everybody else thinks as a
• 7:48
  strategy, that seems silly to me.
• 7:50
  Yeah.
• 7:50
  Like, very silly.
• 7:51
  At the same time, I think VCs are very good at identifying the kind of people that they think
• 7:55
  will be successful, which is different from the kind of people who are, in fact, in the end
• 7:59
  successful.
• 8:00
  And I think right now, I think consensus good businesses there was a period of time where it
• 8:06
  was very clear that Stripe was a consensus good business.
• 8:09
  Investing at Stripe at the moment it became a consensus good investment was a good deal
• 8:13
  because the company has grown in value probably a 150 to 250 since the point it was obviously
• 8:18
  gonna work.
• 8:18
  I'll give you another example.
• 8:20
  OpenAI at a billion was probably a consensus good company.
• 8:23
  It's up 500 x as of today from then.
• 8:26
  So Martin's right.
• 8:27
  The funny thing now is that we have a relatively small set of venture firms who
• 8:32
  define taste.
• 8:34
  Like, they have all of the money for follow on capital.
• 8:37
  So consensus is whatever they like.
• 8:39
  So if Martin's statement is, you should invest in the kind of companies that I think are good
• 8:43
  if you want my follow on dollars, I think that's probably true.
• 8:46
  Yeah.
• 8:47
  It's unlikely that Martin's gonna take a 16 z's follow on dollars and put them into a company
• 8:52
  that he doesn't think is good.
• 8:53
  But I think by the time it gets to Martin, that's different than what it looked like or
• 8:57
  Martin, rather.
• 8:58
  It's different than what it looked like when the company started.
• 9:00
  So I think if you're gonna do things that are weird, a couple things have to be true.
• 9:04
  One, you have to believe eventually the company will be well understood.
• 9:08
  Two, that it will in fact be a good company.
• 9:10
  And three, that you and the other people around the table have the resources to finance that
• 9:14
  company through that dip where it's maybe hard to finance.
• 9:18
  So I don't know.
• 9:19
  I think whenever we do things that are not well understood or kinda hard to get your arms
• 9:23
  around, I just assume there might be one or two financing rounds for that company that are
• 9:28
  really difficult.
• 9:28
  And either we or the other folks in the syndicate have to have a plan for what we're
• 9:32
  gonna do if we hit that speed bump.
• 9:34
  I mean, we just had this conversation, you and I, before we started recording this podcast of
• 9:37
  evaluating a deal and wondering.
• 9:39
  I felt like it was nonconsensus, but I did question how long will it take for folks to
• 9:45
  realize that it is good, and that's exactly to your point of, like, how many runs of funding?
• 9:49
  Is it gonna be a difficult journey?
• 9:50
  And then it also brought me to this question, like, when you're a later stage investor, I
• 9:54
  assume consensus is the only thing that you want to invest in at that point.
• 9:58
  Because early stage, that's the whole way you can make your money in alpha.
• 10:01
  But at series b and beyond, you shouldn't be investing in truly non consensus things, or do
• 10:07
  I have that upside down from your perspective?
• 10:09
  My sense is there's like two different jobs in venture.
• 10:13
  There's the early stage job, which is identifying at a very early stage the companies
• 10:18
  that are likely to be good, important companies five to ten years from now.
• 10:22
  And I think in later stage investing, the job is how do you find these special companies
• 10:27
  separate from the good ones.
• 10:28
  Yep.
• 10:29
  Because, basically, like, most companies that can get to 20 or 30,000,000 in ARR, they're
• 10:32
  good companies, but they're not all great companies.
• 10:34
  Yeah.
• 10:35
  And the question's like, how do you identify the truly great ones?
• 10:37
  Because you only get a few shots on goal, and the job of that late stage investor is really
• 10:42
  to find the special companies.
• 10:44
  And so I don't know.
• 10:45
  I think it's I think this is the hard thing about venture today is what's consensus is
• 10:49
  really the decision of a handful of very senior partners at a very small number of firms.
• 10:54
  And, like, the tastes of those people have a really disproportionate impact on what gets
• 10:59
  funded in terms of follow ons because they have all of the money.
• 11:02
  I looked in the comments section.
• 11:04
  Was like, guys, what Martina is really telling you is he knows what he's looking for.
• 11:08
  Yeah.
• 11:08
  If you want his money, it should look like things that he thinks are good.
• 11:11
  And if you show him stuff that you think is good that he doesn't think is good, probably
• 11:14
  not gonna get funded.
• 11:15
  The last piece on this, I know Yoni from Slow Ventures, he shared a post on his blog where he
• 11:18
  dove a little bit deeper into the post in more depth.
• 11:21
  Because I think there was a lot of comments in the comment section of, oh, Martin has it all
• 11:25
  wrong.
• 11:26
  This is so, like, anti venture, what he's saying.
• 11:28
  And Yoni was kind of like, you're misunderstanding the purpose of what he's
• 11:31
  trying to deliver here.
• 11:33
  He said, quote, I think it's better to think about nonconsensus narratives that are
• 11:37
  overlooked rather than betting against the crowd outright or finding undervalued
• 11:40
  companies.
• 11:41
  That's why I distinguish between contrarians and nonconsensus.
• 11:45
  I feel like what you just mentioned was pretty aligned with where Yoni landed.
• 11:49
  Yeah.
• 11:49
  I I like Yoni's writing.
• 11:51
  I often find that he puts the word things I've been thinking about.
• 11:54
  I think our job at Precursor is to find founders that we think are good working on
• 11:59
  problems that we think are interesting that are maybe not well understood.
• 12:02
  And it's usually because those problems are a year or two away from being obvious to
• 12:05
  everybody else.
• 12:06
  Yeah.
• 12:06
  But again, the founders have to turn out to be good.
• 12:09
  And so it's funny.
• 12:10
  Sometimes I read those comments as like, oh, you guys want venture.
• 12:13
  It's like one of my favorite quotes from The Wire when Marlow's in the convenience store and
• 12:18
  he's talking to the security guard and he's like, come on, man.
• 12:21
  I'm trying to do my job.
• 12:22
  And Marlow says, you want it to be one way, but it's the other way.
• 12:24
  And I think there's just a lot of things in Adventurer.
• 12:27
  People are like, I wish it worked this way.
• 12:28
  I'm like, me too, but it doesn't.
• 12:29
  Right.
• 12:30
  And you have to recognize that.
• 12:32
  Especially if you wanna do different things, you have to actually understand how the
• 12:35
  business works today and what it would take for a founder with a different background to be
• 12:40
  interesting and fundable by a large venture fund.
• 12:42
  It's a lot to get those people to a place where those businesses are consensus interesting.
• 12:46
  Well, our last topic is a substack written by Sam Schillis titled, has the economics of
• 12:51
  software run off a cliff with AI?
• 12:53
  He explains that as AI tools make software faster to build, inference costs, so the cost
• 12:58
  every time a user interacts with an AI, are driving margins down.
• 13:02
  The old build once, sell infinitely model is completely eroding, and software is starting to
• 13:08
  feel more like a cost heavy operational business.
• 13:11
  Throughout the piece, he even compares the future of engineering to manufacturing where a
• 13:16
  small group of top engineer designs high leverage systems and the rest become lower paid
• 13:22
  interchangeable line workers.
• 13:24
  Sam shares a word of caution, is product and engineering leaders who don't embrace AI risk
• 13:29
  falling behind.
• 13:30
  The edge will go to creative hybrid builders, those who can understand both code and
• 13:35
  experimentation, not those who cling to what was historically the norm.
• 13:39
  This was the first time I had heard the manufacturing comparison when talking about a
• 13:44
  future for the engineering workforce powered by AI.
• 13:47
  I wanted to get any perspectives from you on on this piece.
• 13:50
  I like Sam's writing too.
• 13:52
  Anyway, he's seen a lot from, you know, Google Sheets to now.
• 13:55
  And I think most talent markets that we're in end up with this highly skewed star system.
• 14:02
  Sports has gone that way.
• 14:03
  Music has gone that way.
• 14:05
  Art has largely gone that way where the people at the top make extraordinary amount Yeah.
• 14:11
  Of money relative to the average person in their field.
• 14:14
  And I think it's interesting to think about, well, what would software development look like
• 14:18
  with a similar star theory?
• 14:20
  You would end up with a handful of really highly regarded, highly compensated, hopefully
• 14:25
  highly talented software engineers who make a lot of money for some special level of
• 14:30
  expertise.
• 14:31
  Maybe it's system design.
• 14:32
  Maybe it's understanding how to get the most out of AI coding.
• 14:35
  And then you have a lot of people who do a similar job and make a lot less money.
• 14:38
  You know, I think we talk about the extremes are that, like, everyone's gonna become a
• 14:42
  software developer or there'll be no coders, there'll be only AI, or there'll be one person
• 14:46
  in the world who writes all the software.
• 14:48
  All of these weird extremes.
• 14:49
  But I I do think the software job market is gonna look different, but I don't actually know
• 14:55
  that I know the answer.
• 14:57
  The one thing he said that I do think is definitely gonna happen is that you're gonna
• 15:00
  have developers who will end up spending more in tokens than they're paid in salary.
• 15:04
  Like, their their compute consumption will exceed their because I think their compute
• 15:09
  consumption will grow faster than their wages.
• 15:11
  So at some point, those two lines will those two lines will cross.
• 15:14
  And I think that will be interesting when you have a person who's paid, let's say, $250,000 a
• 15:19
  year, who's generating a $500,000 compute bill.
• 15:22
  But that person might be doing the work of what would have previously taken 10 people.
• 15:27
  And, like, how much of revenue is attributed to that token spend, I guess?
• 15:30
  This $250,000 that I'm spending on them is, like, a drop in the bucket for what they're
• 15:34
  outputting.
• 15:35
  Cost.
• 15:35
  I continue to believe the hardest thing VCs who've grown up in the SaaS era, including
• 15:40
  myself, and I try to wrestle with this a lot, is we grew up in a world where the marginal
• 15:45
  cost of delivering software to the marginal user was effectively zero, which is how you got
• 15:51
  the Skyheim.
• 15:51
  And the products were hard to build, and we've kind of inverted that now.
• 15:55
  The marginal cost of supporting a user is not zero.
• 15:57
  There's compute that goes with it.
• 15:59
  It was, god, compute will go down.
• 16:00
  I'm like, but what if people just use more of it?
• 16:01
  The old Stevens paradox kind of thing.
• 16:03
  Like, there's nothing that says that the compute cost per user will go down just because
• 16:07
  the cost of compute goes down.
• 16:09
  Usage could go up faster than prices go down.
• 16:12
  So we don't we don't know what that is, and we don't know what's gonna happen to customer
• 16:15
  willingness to pay when they know that these products are, like, less hard to build.
• 16:19
  So we're, like, tinkering with cost, and we're tinkering with value.
• 16:23
  We're tinkering with all the assumptions of software, but we're still funding companies as
• 16:27
  if the old rules are true.
• 16:28
  I think that's dangerous.
• 16:29
  I wonder how many investors are just assuming that eventually the margins will end up looking
• 16:33
  like SaaS.
• 16:34
  Like, eventually, they'll flip to 80%, which
• 16:36
  of people believe that.
• 16:37
  Yeah.
• 16:37
  I don't.
• 16:37
  How many years will that take if it happens at all?
• 16:41
  Say margins will never flip to be as fat as SaaS.
• 16:45
  Are there other KPIs and metrics that you should be leaning on more heavily to kind of
• 16:50
  show you?
• 16:51
  I think this is the hardest thing for software investors.
• 16:54
  I I deeply believe this, and I think I'm right, but I'm willing to accept them wrong.
• 16:59
  I think software for most of my life has been a business where the producer of software
• 17:04
  captures all of the value.
• 17:06
  Yeah.
• 17:06
  90% margins really means the customer has no leverage.
• 17:10
  That's what that tells me.
• 17:11
  Yeah.
• 17:12
  And, of course, they don't because software products took an army of people to build.
• 17:16
  They were really complicated.
• 17:17
  They were expensive to maintain.
• 17:19
  And if you built an ERP system, your customer wasn't gonna do that on their and there weren't
• 17:24
  cheap alternatives unless you're gonna do some weird open source thing and support it
• 17:27
  yourself.
• 17:27
  So you could charge people 90% margins and get away with it because they didn't really have
• 17:32
  good alternatives.
• 17:33
  And it's not that the customers didn't get value.
• 17:35
  The the value captured by the provider was aggressive.
• 17:38
  Yeah.
• 17:38
  If you think about the total opposite in the spectrum is open source, where the economic
• 17:43
  value mostly goes to the person implementing the software, not the person who wrote it.
• 17:47
  The person who wrote it usually writes it for other reasons.
• 17:49
  I think the software business is gonna look more like open source, where a $10,000 piece of
• 17:54
  software in three years is gonna probably be the equivalent of a million dollar piece of
• 17:57
  software five years ago.
• 17:59
  It's just deflationary.
• 18:00
  You're gonna pay a lot less, which to me means margins will go down because the consumer of
• 18:05
  software is gonna capture much more of the value, and that's good for society.
• 18:09
  If you everyone's like, software eats the world.
• 18:10
  They're like, well, not at 90% margins.
• 18:12
  It's hard.
• 18:12
  Software's cheap.
• 18:13
  Software can go a lot of places that it doesn't go today and solve a lot of problems.
• 18:17
  So there will be more money made in software in the aggregate, but it might be a less
• 18:21
  attractive business from a margin standpoint.
• 18:23
  But I think you have to I was like, but you're gonna need fewer people.
• 18:25
  I'm like, but not none.
• 18:26
  And some of this is trading human workers for compute.
• 18:30
  So I I actually think software is going to become a less attractive business, but not an
• 18:34
  unattractive business, if that makes sense.
• 18:36
  Well, we are all out of time today.
• 18:37
  Thank you, Charles, for stopping by.
• 18:39
  Thank you, our listeners, for jumping in.
• 18:41
  Please like and subscribe, and we will talk with you all next week.
• 18:44
  Thank you.
• 18:45
  Bye.