20VC Sep 21, 2026 1h 11m 45m saved
With Daniel Dines, founder and CEO of UiPath
Daniel Dines runs a public company with about 4,000 employees, more than a thousand of them engineers, and he has just written a book arguing that the thing standing between AI and enterprise work is not intelligence.
The pitch everyone else is making is that models keep getting smarter until the work disappears. Dines' objection is narrower and harder to answer: a model can write things down, but it is not changed by having done the job.
"That's it has memory, but memory it's not necessarily learning. It's not the same thing."
Dines founded UiPath in Romania, took it public, and now sells automation to the largest enterprises in the world, which is where the claim gets tested. Harry Stebbings, who has backed Legora and Fireworks AI, spent most of the conversation arguing the other side.
The full interview is covered here so you can skip it. 71 minutes of audio, 26 minutes of reading.
Here are the 21 insights that matter.
Key Takeaways
Memory is not learning — a model keeps a scratch pad, but its weights are not altered by the job the way a person is Two chefs given the same recipe cook different food, and that difference is not written down anywhere
His prediction is that 90% of enterprise AI traffic goes to cost-efficient models, not frontier ones
A $90B legal opportunity converts into roughly $10B of token revenue, on his arithmetic
Deploying agents is no easier than two years ago; deploying automation is much easier — that gap is the whole business case
AI is probabilistic at every step, so a hundred steps at 99% each lands around 60%
He would hire a machine over a person even if the machine cost more, because labor costs only go up
Cutting the "credentialed middle" removes the people AI most needs to supplement the experts
Chinese labs count as the good guys, and he reads the pacing argument as an attack on open source
The vibe-coded procurement tool's database schema came out bogus and a human had to rebuild it
Every major infrastructure in history has been overbuilt — possibly not for the decade, possibly for the next three years
Europe is largely irrelevant in technology, he said, even though ASML is European
He takes about 60 supplements a day, all of them vetted by AI
1. Why He Wrote A Book
Stebbings opened by asking why a public-company chief executive writes a book, given the royalties are not the reason.
He had wanted to write one since childhood and had no talent for it
Man, I dreamt to write a book since I was a kid and I discover I have no talent and now I got really a great opportunity.
Daniel Dines
The opportunity was that two models did the drafting. It took about six months.
The value was in ordering his own thinking
They were my ghost writers and it was a good moment actually to put my own ideas in order because really when you write something you get much more clear perspectives on what you are doing.
Daniel Dines
The first thread he pulled on was whether AI has any durable limitation, or whether a few years from now there are millions of Einsteins in a data center and everyone else can go and play.
2. Millions Of Einsteins
That phrase belongs to Dario Amodei, and Dines said it worried him enough to work out what it could mean.
He took the claim personally and literally
Look, when I heard about this statement that in a couple of years we will have millions of Einstein in a data center, look, I was really concerned.
Daniel Dines
He asked himself whether he could hire one of these entities, put it on a laptop, give it an enterprise account and a Slack account, and have it do his job. His conclusion is that the claim, read charitably, is about reasoning power rather than about people.
Reasoning like Einstein is not being Einstein
probably Dario wanted to say that we'll have a millions of entities that will have the some of the reasoning powers of Einstein which I agree but not Einstein's as a persons not Einstein's that are capable of learning on the job
Daniel Dines
The test he applies is the one every employer applies: nobody hands a new hire a manual describing the job end to end, because the expectation is that they learn it.
3. Memory Is Not Learning
Stebbings pushed back that humans also learn on the job, and that models do too, with the difference running the other way.
His case for the machine
humans get tired, humans want more money, humans want culture, humans can be toxic, humans are difficult to manage, I'll take the AI any day of the week, please.
Harry Stebbings
Dines' answer is the technical core of the book.
A scratch pad is not a changed set of weights
AI can create a notepad on the job, a scratch pad where they can memorize some of the policies on the job, but AI doesn't alter its weights on the job in the way humans are transformed by a job.
Daniel Dines
He gave three illustrations in a row. Two chefs, one who has cooked Japanese food for 20 years and one Italian, handed the same recipe, produce different food. Reading is not doing.
Chess books do not make a grandmaster
Think about if I give to someone the ability to read all the books about chess, do you think he will become a grandmaster without playing, without losing, without going through all of these process? Probably not.
Daniel Dines
Watching ski videos, he added, does not make a skier. When Stebbings offered ChatGPT's memory as the counter-example, Dines drew the distinction the section rests on.
Memory is written down; learning is transformation
That's it has memory, but memory it's not necessarily learning. It's not the same thing. It's memory. It's just a thing that is written down.
Daniel Dines
His evidence from his own product is that UiPath's platform is now available to coding agents, and the models are still better at building on open-source technology than on UiPath's own.
The weights beat the prompts
what we discover is that a model that has read open-source technology will and they have a lot of examples and they already have in its weights a certain technology will be much better than to create on our own technology because regardless how many prompts and skills we create the model has it in its own weights
Daniel Dines
4. The Micro-Initiative Gap
Stebbings' counter was that most enterprise work (accounting, finance, marketing, sales, social) is execution, with judgment and taste concentrated at the top. Dines disagreed on the empirical point.
Everyone in a job takes small initiatives nobody wrote down
I think most of the people will play will display some sort of micro initiatives during the job. Maybe I have a hunch this customer is going to churn and I can act before even any data is coming. How do I develop this hunch? It's through my years of transformation. It's not written on a piece of paper.
Daniel Dines
The practical consequence, he said, is that using a model on a process requires the whole enterprise to be documented and re-read on every query, which is not possible.
5. Will, Not Reasoning
Asked whether recursive self-improvement removes the limitation, Dines ran a thought experiment rather than answering. Put today's best model on a von Neumann self-replicating spaceship, throw it at the stars, give it effectively infinite compute, and it ends up simulating a world as complex as this one — which, he noted, would mean this one sits inside an infinite regress of simulations.
Then he named the distinction he thinks matters.
Will may not come out of reasoning at all
I know that there is a big distinction that I made in the book between will and reasoning and it's not like that We are certain that the will to do something emerges from reasoning or from even from consciousness.
Daniel Dines
He does not expect a self-improvement loop to produce it
I think it's wishful thinking to believe that I can take a big model put into a self-improvement loop and this model is going to generate will I don't kind of believe so
Daniel Dines
Stebbings' response was that nobody knows, and that a technology can become something its builders did not expect.
6. Pacing The Frontier
On Amodei's call to pace the frontier, Dines started from the labs' own position: if they believe the technology is going rogue, the case for slowing is self-evident and needs no outside pressure.
He would stop on his own account
honestly I don't need an external pressure to control. I would be just concerned citizen and I would not build a technology that is causing harm.
Daniel Dines
What he thinks is actually being requested is different.
He reads the memo as a request for indemnity
I think that probably they ask more like a pass. I want to build this technology at any risk and I'm willing to open my gates for other to see how I'm doing because I want to do it in as of like good manner as possible but in the same time I want to be free of consequences
Daniel Dines
Stebbings' objection was that you cannot negotiate a frontier pause with bad actors, so a coalition can only be built among good ones. Dines widened the definition.
Who counts as a good guy
In my opinion, even the labs in China that are building AI right now, I would classify them as the good guys.
Daniel Dines
On his reading the real target is not a country.
The argument lands on open source
so the danger is in open source. So indirectly it's also an attack to open source in this way.
Daniel Dines
The chain is that good actors publish open weights, those weights reach unknown bad actors, and the open release becomes the risk.
7. What Enterprises Fear
Stebbings cited Alex Karp's claim that the largest enterprises are afraid to work with frontier labs in case the labs come for their business. Dines agreed they are afraid, and disagreed about what of.
The fear is leakage, not competition
I don't think people are scared that OpenAI will build a competitor to them necessarily. I don't see this coming. I think they are more scared that their IP would leak to other existing competitors.
Daniel Dines
A manufacturer does not expect a frontier lab to start making its product. It does worry that whatever the model learns from its data reaches a rival through another model, which he called a legitimate danger.
8. The Exactness Problem
The second limitation the book names is what he calls exactness.
Probability compounds against you
another limitation of AI is what I call exactness
Daniel Dines
His arithmetic is that a chain of 100 steps, each 99% reliable, finishes correct about 60% of the time, and that asking a model to multiply large numbers a million times will eventually produce a wrong answer. Stebbings' objection was convenience: people stay in the chat window rather than switching to a calculator.
The chat window is an interface to exactness, not the source of it
But the exactness is not run by ChatGPT. Exactness is run by a computer.
Daniel Dines
Models already call a computer behind the scenes for arithmetic. His proposal is to take that pattern up a level.
Exact work belongs on exact technology
when everything that should be exact should run on exact technologies. There is no point to run it on probabilistic technologies.
Daniel Dines
9. The Automation Asymmetry
This is the commercial argument, and it is the one UiPath is built on.
Agents have not got easier to deploy; automation has
Deploying AI agents is not getting easier today than it was two years ago in my opinion. But deploying automation has become much easier because I can create these automations with AI with coding agents.
Daniel Dines
His ranking of the last few years
Coding agents been the most major giant leap that we were seeing in the past year.
Daniel Dines
The sequence he names is ChatGPT, then chain of thought, then coding agents. What that buys, on his account, is software written at design time that runs with the same result every time, and gets repaired by AI when an upstream system changes and breaks it.
Software cannot go rogue in the way an agent can
you use AI to create software that runs the enterprise in a predictable, governant, auditable way.
Daniel Dines
Humans can read it, tests can pin its behavior to a given input, and it cannot change what it does mid-run.
10. 4,000 People
Stebbings, who finds 18 people hard enough, asked whether UiPath has too many. Dines did not answer the headcount question directly and answered the transformation one.
He told staff a transformation was coming and that it was not a cut
I've never hidden from my employees that there will be a transformation in the company.
Daniel Dines
The promise attached to it
We are not just using AI as a pretext to cut a part of the company.
Daniel Dines
What he says writing the book taught him is that a job produces two outputs. One is the measurable thing the person was hired for. The other is institutional strength — a deep relationship with a customer, for instance, which does not show up in anyone's numbers but is what keeps the customer.
What he wants every enterprise to build first
So to me an enterprise should have a ledger where they actually understand what people are doing besides their main definition of the role
Daniel Dines
The trust part does not transfer
I don't think AI can supplement the human connections and the trust.
Daniel Dines
The group he says is at risk is the one the education and hiring systems are built to produce: people credentialed as deep experts in one domain. Those are exactly the people AI helps most, so fewer are needed — and a blind cut takes out the wrong ones.
The people to keep are the ones with initiative
you will need a fewer of these experts but you will need more people that have initiatives that are capable of maintaining a relationship with the customer
Daniel Dines
Stebbings' data point was a lawyer he had spoken to that day whose trainee intake historically ran at 25 and will be four this year. Dines agreed most roles will need fewer people and said the hard question is which ones.
11. The Map Of Work
The two disagreed about where AI can be applied first. Dines' test was verifiability — finance and accounting, where right and wrong are clear. Stebbings' test was whether the work has been framed by someone else, and he argued finance is not clear at all: an invoice or an order can be treated differently depending on the customer, and the rule that Nvidia's chips ship to OpenAI first may not be written anywhere.
Unwritten rules cannot be learned
If it's not captured in a frame, I think AI cannot learn it. This is why you need to create this manual. We call this manual the map of work.
Daniel Dines
What is in it
To basically capture how the work works. How the work happens in an enterprise. This is the map of work. It's all the workflows, all the exceptions, all the procedures, all the systems that you use in order to fulfill a goal of a process.
Daniel Dines
Stebbings' counter was that a head of finance sitting on top of 30 agents handles the Nvidia exception herself, which Dines took as agreement: even in finance you cannot replace everybody.
The harder problem, which he called the crux, is the data nobody captures — the tone of a call, the warm follow-up text. Stebbings noted Meta has talked about monitoring every action on an employee's screen and said that does not capture any of it.
UiPath's answer is a new discipline and a product built on it. An agent interviews subject-matter experts while they record themselves working, asking why they changed an invoice when the zip code was different, or why they took a different path — surfacing the exceptions as they happen.
Agents interviewing people
So that's real agents interviewing real people. It's pretty cool stuff.
Daniel Dines
The interviews are consolidated across multiple people into process maps showing the work as it is, and coding agents then turn the redesigned process into software.
The direction of travel
I think the goal of any enterprise is to have less people operating the system and more automations and agentic AI operating the systems.
Daniel Dines
Asked whether staff resent being watched by the thing that will replace them, the backlash Meta took, Dines said the reaction depends entirely on the framing.
What he tells employees
transformation is inevitable and you guys have to transform and everybody will get a chance
Daniel Dines
He said the fear across the industry at the start of the year was off the charts and has since settled, as people came to understand that diffusion happens one process at a time.
12. Inference Against Salaries
Stebbings cited a portfolio company, the data provider Mercor, whose chief executive had tweeted that they spend three times their human salary bill on inference, and asked for UiPath's ratio.
He does not track it
Look, I personally don't care about it.
Daniel Dines
His reasoning is that if a machine could do the work at a human's quality he would buy the machine even at a premium, because labor costs only rise and people make errors.
Token cost is not the binding constraint
I don't think the cost of tokens it will be the real question if you replace a person with AI but the real problem today is that AI cannot replace a person
Daniel Dines
Stebbings brought up Jason Lemkin of SaaStr, who has cut his team from 25 to two and says the AI is better at the VP roles than the people were.
One data point is not a trend
I think I heard companies that replace hundreds of support people in the past and now they are rehiring these people.
Daniel Dines
13. Vibe Coding Hit A Wall
UiPath tried to replace a procurement tool by writing its own, initially with AI alone.
It looked extraordinary and then it did not
Yes, we did, but not an extraordinary success.
Daniel Dines
The work is after the prototype
taking a software from a prototype to production it's actually where the work is not necessarily the writing code writing code is fun but it's not there where you can really makes the difference
Daniel Dines
Connectors, permissions, audit and security all have to be maintained, and the tests and trust required to put it live were not there.
The specific failure
the database schema that vibe coded tool created was completely bogus. So it had a human has to come and create the structure.
Daniel Dines
A business user who understands a problem still cannot vibe-code a tool and ship it, which means engineers stay on it.
The saving is not obviously a saving
So eventually you will end up paying probably as much if not more as the tool you replace while you keep some of your good and best people bandwidth occupied.
Daniel Dines
14. Why Companies Still IPO
Stebbings argued the two reasons to be public have gone: liquidity is available in private markets, and private stock can be used for acquisitions.
His position
I don't understand why a company would go public today.
Harry Stebbings
Dines' reply separated the exceptional companies from everyone else.
The 2021 cohort would be better off public
There are so many companies that are kind of zombies right now. This 2021 zombies, they would fare better in the public market right now. At least their investors will have a way for exiting. Their employees will have a way to make some money. Nowadays all of them are sitting on paper money.
Daniel Dines
And the discipline that comes with it
Public markets will face them with the reality of you know what's their real valuation.
Daniel Dines
He noted Anthropic is attempting an IPO. Stebbings' view is that Anthropic and OpenAI are going public because they have exhausted the supply of private capital, not because they want to. Both said they would not sell at the numbers discussed; Dines said the question he cannot answer is whether he would buy, and that he would want to see the real numbers first.
15. Jensen Needs Open Source
The two agreed on where the durable value sits in an AI boom whose application layer is unpredictable: the infrastructure underneath it. Nobody knows which coding tool or wallet or interface wins, and the chip supplier gets paid regardless.
The qualification they arrived at is that Nvidia's position is not unconditional. If the frontier collapses into a duopoly of two labs that are also the largest buyers, the supplier's leverage goes with it — which is a reason for Nvidia itself to want open models to succeed.
The conclusion drawn on the show
I don't think Jensen's will be doing so well if they are the single biggest providers
Daniel Dines
That is the frame both put on Nvidia's purchase of Hugging Face, which hosts the open-source models, and on the open letter about open source that Stebbings said everyone signed. Stebbings' caveat was that a neutral provider stops being neutral once it has an owner with its own models, and Dines agreed there is a loss of independence, while judging Nvidia's own model line a small part of the picture.
16. Overbuilt Or Underbuilt
Asked whether the circular deals between Nvidia, Oracle and OpenAI are overblown, Dines gave a historical answer.
Infrastructure is always overbuilt
It can be because every major infrastructure in history has been overbuilt.
Daniel Dines
And he thinks he knows why
you have to make sure you get the biggest piece of the opportunity
Daniel Dines
The arithmetic of that
So it's clearly that now everybody that it cannot be there is only 100% of the pie and people are building right now 200% of the pie there will be losers.
Daniel Dines
Stebbings put the opposite case: in earlier cycles supply ran ahead of demand, and this time energy, water, data centers and policy are all constraints.
His read is that this cycle is short of capacity
We have a significant hindrance to supply side and we are underbuilt not overbuilt which is why every ounce of computers taken.
Harry Stebbings
Dines' answer split the question by horizon, and hangs on when AI actually starts replacing work at scale.
A decade, no; three years, maybe
So I don't think it's overbuilt for next decade but it might be overbuilt for the next three years and stock markets and capital can be merciless and maybe I'm a childish optimist
Daniel Dines
17. A $10B Legal Market
Stebbings said he had interviewed lawyers when his firm invested in Legora, and that all 15 told him they thought technologists could not replace them. Two weeks before the recording he asked them again and every one said they would be severely unhappy to lose the tools, with most saying they had not written a document in six months.
Dines took law as the case that proves his rule rather than breaks it.
Where the frame is set by someone else, AI is devastating
when the frame that a person operate is really well defined by someone else like in law AI can be devastating
Daniel Dines
Stebbings objected that law is ambiguous and subjective, and that his girlfriend, a lawyer, would say so. Dines' answer was that ambiguity in human language is precisely what models handle.
The test is whether a manual exists
there is a manual for the freaking law AI is amazing. When there is no manual AI is not amazing and it doesn't work.
Daniel Dines
Then came the sizing argument.
Stebbings' number
But $300 billion is the legal industry in the US. It's a lot. If you think about how much labor could be replaced by that, I think 30% would be reasonable. That would be $90 billion of available revenue.
Harry Stebbings
Dines' deflation of it
Maybe out of 90 billion I think companies might charge maybe 10%. Maybe it's a 10 billion total opportunity in tokens.
Daniel Dines
Asked whether that makes the investors in Legora and Harvey wrong, he said no, and restated the thesis the whole interview runs on.
Where the value actually accrues
Models are interchangeable, but the workflow, the map of work and the workflows around the map of work is where the real value is.
Daniel Dines
If those companies map the work and build the workflows around it, effectively building a legal department, the value they capture is much larger. If they are a call to a model for a legal opinion, it is not a hundred-billion-dollar market.
18. 90% On Cheap Models
Stebbings asked what share of token traffic goes through open versus closed models in 12 months. Dines reframed it as frontier versus cheap.
His prediction
for enterprise work my prediction is that 90% of the of the flow will go to very costefficient models I don't think you need frontier level quality of models for most operational work
Daniel Dines
That still means buying from the same two vendors, on their cheaper tiers, with one condition attached.
Open source is the switch, not the default
I will still use Anthropic and OpenAI with their costefficient model but I will have a verifiable backup on open source all the time
Daniel Dines
Stebbings said his own thesis is that mid-size and large companies will own their intelligence rather than rent it, which is why his firm invested in Fireworks AI.
Dines is a customer
I'm a big fan of fireworks and we are using them quite a bit.
Daniel Dines
His argument for owning a model comes back to documentation. Training your own model requires the map of work, and so does moving to the next base model when it arrives two months later.
The map is the transferable asset, not the model
the real investment for an enterprise is to creating this map of work that documents how they actually work and with this one they can train their own models
Daniel Dines
But he expects it to be the backup, not the main line
I do believe that they will at least have their own models as a backup to Frontier models.
Daniel Dines
He is unsure a company can match a frontier lab on cost, because delivering more intelligence per dollar is those labs' business model. Asked whether he would invest in Fireworks at $15B, he said probably, with a condition.
Inference at scale needs secured compute
I think they will have to get very soon in this big game of securing compute
Daniel Dines
19. Data Is Not A Tape
Stebbings said the data providers are the most underpriced part of the stack, naming Mercor and Surge, and that dismissing them as commodity data misses that data quality is what makes a model good.
Dines separated two businesses.
Storage is not understanding
I think one thing it's storage and one thing is understanding of the data. Because if I have a storage, I can have a tape and I can put data on the tape. Would we invest in a tape company? I don't think so. You need to invest in the intelligence that understands the data and feed the model
Daniel Dines
Stebbings' rebuttal was breadth: the largest libraries can serve a highly specific request in a way others cannot. Dines' answer was that if the data exists and is valuable, the labs will simply buy it.
20. Europe Is Irrelevant
Both are European and were recording in London, and Stebbings put the question bluntly: we do not matter anymore, does that get better or worse?
Dines agreed
from a technology standpoint, I think we are largely irrelevant
Daniel Dines
And named the thing that makes it absurd
But it's so stupid because the biggest producer of machines that make chips is based in Europe is ASML.
Daniel Dines
Several of the people building frontier AI are, he said, of European origin. The talent and the machines are there and the outcome is not.
On work ethic he was direct: his UK teams stop at five, and he thinks culture matters more than money in explaining the gap.
He does not think he could have done it at home
This is why I don't think I would have succeeded in Europe the way I did in US.
Daniel Dines
He describes himself as European by origin and American as an entrepreneur. Stebbings added that American companies make larger bets on vision with fewer proof points, and that middle managers there sign million-dollar bets he has never seen signed in Europe.
His advice to a young European founder
I think that unless they build for a specific market with some specificity in mind, if they build a universal technology, they will have a better chance to succeed in US.
Daniel Dines
The exception he does buy is sovereignty, over energy and over models, which he thinks is a large business. Stebbings said every European customer he deals with now wants on-premise software and model optionality, and Dines said he has been trying to persuade Fireworks to make its software available on-premise, against their preference for proven demand first.
21. The Bull And Bear Case
Stebbings asked for UiPath's revenue.
The public number
I think it's public data. We are at 1.6 growing last year like 14% and this year.
Daniel Dines
Stebbings' response was that below 20% growth the public markets are unforgiving, and that investors sort software companies into AI winners and AI losers without looking closely, because there are too many to examine.
The bull case Dines gave starts with Gartner's magic quadrant for business orchestration and automation technologies.
They moved up a box
we are one of the leaders. We move from a challengers into a leaders in the last year.
Daniel Dines
And the belief he thinks has died
You cannot this idea that you can have an AI agent that runs everything for you from top level processes orchestrate and automate everything by magic. I think it's kind of a thing that people stop believing.
Daniel Dines
What replaces it is the two-part structure the book argues for.
Map and rails
Map is the context. Rails is the orchestration automation.
Daniel Dines
The agent is then given the rails it may use, the map describing how to use them, and the goal.
Why nobody hands finance to a swarm
no sane enterprise right now will put a swarm of agents and just ask them do my finance accounting for me
Daniel Dines
The bear case, in his own words
AI will somehow become genius. Will tokens cost will be next to zero and we'll have this literally millions of Einsteins in a data center but Einstein's in a true sense
Daniel Dines
Stebbings pointed out that token cost has already fallen from $60 to $1 per million and will keep going. Dines' answer was that this is exactly why he would not stop an investment on token cost.
Bonus Insights
The hardest part of the job is not the technology
It's aligning people. That's so different personalities, pride, ego that comes into place. This is the hardest.
Daniel Dines
Half his day is spent in an editor
I'm spending maybe half of my day right now. Half of my day alone with myself in Visual Studio Code
Daniel Dines
Decks became markdown files
Most of the people when they come with an idea to me a year ago they would come with the deck and it was very hard even to prepare with this deck. Now everyone is going to come with the markdown file
Daniel Dines
He keeps a strategy folder of those files with agents working over it, and says it gives him more leverage on the company than he had before.
What he tells lonely founders
I think they should surround themselves with their best friends from maybe childhood and have more frequent chats with them
Daniel Dines
They are the people who can still see you, he said, and the loneliness does not go, but the sense of continuity helps.
The Nvidia question
Nvidia in 3 years time will it be above 7.5 trillion?
Harry Stebbings
Neither committed to an answer beyond noting where the market capitalization sits now and that a large further run is easy to picture.
Sixty supplements a day, chosen by AI
And all of them have been recommended and vetted by AI
Daniel Dines
He takes them as a single morning mix modeled on Bryan Johnson's protocol, plus powders and small bags through the day, and says he feels better than he did ten years ago, mostly from drinking much less.
What he is most excited about
Like my mother's got MS. I'm really excited for some of the breakthroughs that we'll see with chronic conditions and treatment of them.
Daniel Dines
The book is out
It's already available for a download and I'm printing also a few copies.
Daniel Dines
Dines' bottom line is that the constraint on enterprise AI is documentation rather than intelligence: until a company has written down how its work actually happens, the model has nothing to be exact about, and the money accrues to whoever owns that map rather than to whoever owns the model.
Products, Companies & Tools Mentioned
UiPath (His company: about 4,000 people, more than 1,000 engineers, $1.6B of revenue growing 14%, now a leader in Gartner's orchestration and automation quadrant)
Anthropic and OpenAI (The two labs he expects to keep buying from, on their cost-efficient tiers; both co-wrote his book, and Anthropic's attempted IPO came up twice)
Nvidia (The infrastructure bet both agreed on, with the caveat that a two-lab duopoly would weaken it)
Hugging Face (Nvidia's purchase of it read as a bet on open source, and as the end of a neutral host)
Fireworks AI (Stebbings' portfolio company and a UiPath supplier; Dines said he would probably invest at $15B if it secures compute)
Mercor and Surge (The data providers Stebbings called underpriced; Mercor's CEO had tweeted that inference costs three times the company's salary bill)
Legora and Harvey (The legal AI companies whose market Dines sized at about $10B of tokens rather than $90B of revenue)
Palantir (Alex Karp's claim that large enterprises fear working with frontier labs)
Salesforce (He would buy it as a stock and as a system of record, and does not think vibe coding replaces it)
ServiceNow (Named alongside Salesforce as the software comparison Anthropic's returns beat)
SaaStr (Jason Lemkin's cut from 25 people to two, which Dines treats as a single data point)
Stripe and PayPal (Stebbings' examples of private companies with liquid stock and acquisition currency)
ASML (The European company that makes the machines that print the chips, and his evidence that Europe had the capability)
Visual Studio Code (Where he spends about half his working day)
Gartner and Forrester (The analyst houses he cites for UiPath's repositioning)
Books & Resources Mentioned
The Work That Remains – Daniel Dines (The book this conversation is built around: the map of work, the will-versus-reasoning distinction and the credentialed middle)
Dario Amodei's essay on pacing the frontier (The proposal Dines reads as a request to be freed of consequences)
Bryan Johnson's longevity protocol (The basis for the roughly 60 supplements he takes each morning)
Listen to the full episode
🔴 YouTube | 🔗 Episode page
Watch the full episode:
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: