20VC with Harry StebbingsDaniel Dines on Why Work Processes Not Models Will Be The Most Valuable Asset in AI
20VC with Harry Stebbings
Why Daniel Dines wrote the book 1:31
Daniel Dines explains that he dreamed of writing a book since childhood but never thought he had the talent, until Claude and ChatGPT helped him as ghostwriters. Writing forced him to organize his own scattered thoughts about AI over roughly six months, starting with a core question about the limitations of artificial intelligence.
Millions of Einsteins in a data center 2:30
Dines reacts to Dario Amodei's claim that we will soon have millions of Einsteins in a data center. He argues this is misleading because such AI would have Einstein's reasoning power but not his ability to learn on the job the way a real employee does. He illustrates this with two chefs who each spent twenty years cooking different cuisines, or someone who reads every chess book but never plays a game, to show that reading and watching are not the same as becoming transformed by experience.
Memory is not the same as learning 8:01
Dines pushes back on the idea that AI memory equals learning. He says a model's memory is just written notes it can consult, not a change to its underlying weights, whereas a human is genuinely transformed by every conversation and carries that change forward. He gives the example of UiPath's own platform, where coding agents perform better on open-source technologies already baked into their training weights than on UiPath's own technology, even with many prompts and examples, showing the gap between lived experience and read information.
Will, reasoning, and self-improvement 10:01
Asked whether recursive self-improvement could remove this limitation, Dines offers a thought experiment of a self-replicating spaceship endlessly gathering energy and improving, which might eventually simulate a world as complex as our own. He distinguishes will from reasoning, arguing in his book that will is a separate part of the fabric of the universe, and that it is wishful thinking to assume a self-improving model would simply generate will.
Pacing the frontier and the good guys 12:00
On Dario Amodei's call to pace AI development, Dines says if a lab truly believes its technology is becoming rogue and uncontrollable, it should slow down regardless of external pressure, since laws already exist against causing harm. He suspects labs frame this as protecting against bad actors while really seeking freedom from consequences. He states he would classify even Chinese AI labs as good guys, and suggests the memo's real target is open source, since open technology could fall into unknown bad hands.
Enterprises fear leaking IP, not competition 15:03
Dines addresses concerns raised about big enterprises fearing frontier labs. He does not think companies fear OpenAI itself becoming a direct competitor, but they do worry their proprietary data could indirectly help train models that benefit existing rivals, which he calls a legitimate danger everyone is trying to guard against.
AI's exactness problem 16:30
Dines describes another limitation he calls exactness. Because AI is probabilistic at every step, a 99 percent success rate compounded over 100 steps drops to around 60 percent accuracy, which is why AI can fail at tasks like multiplying large numbers repeatedly. He argues that just because a tool like AI can attempt a task doesn't mean it should, and that exact tasks should run on exact technologies like computers rather than probabilistic models.
The asymmetry between AI agents and automation 19:32
Dines identifies what he calls an asymmetry: deploying AI agents directly hasn't gotten easier in enterprises, but deploying automation has become far easier because coding agents can build automations at design time that then run with exactness. He calls coding agents the biggest leap since ChatGPT and chain of thought, and explains that when an automation breaks due to upstream changes, AI can fix the automation itself, creating auditable, governable software rather than unpredictable, rogue behavior.
UiPath's scale and workforce transformation 21:30
Dines shares that UiPath has over a thousand engineers and around four thousand employees total. He says he has always been open with employees that AI will bring transformation, but insists the company won't use AI as a pretext to simply cut staff. Instead, workforce transformation must happen alongside AI adoption.
The hidden outcomes that matter 25:01
Dines explains that every job has a measurable outcome someone is hired for, but also a second, less visible outcome tied to institutional strength, like the deep trust an employee builds with a customer. He warns that cutting staff based purely on visible output risks hollowing out an enterprise of exactly the talent needed to thrive alongside AI, using the example of a credentialed middle class of workers whose deep domain expertise is becoming less essential than their initiative and relationship-building.
Choosing who stays: the map of work 27:32
Responding to the idea that fewer trainees will be needed everywhere, such as a law firm cutting its trainee program from twenty-five to four, Dines argues the key isn't verifiability of tasks but whether the work has been captured in a clear frame. He introduces the concept of the map of work, a manual capturing all the workflows, exceptions, procedures and systems in an enterprise, arguing this map must be handed to AI for it to succeed, since unwritten exceptions, like an informal rule about which customer gets priority shipping, cannot be learned otherwise.
Ontology and interviewing employees 31:00
Asked how to capture illegible, hard-to-measure data like the warmth of a customer call, Dines describes a new discipline UiPath calls ontology, aimed at surfacing how work actually gets done. He describes an ontographer agent that interviews subject matter experts in real time while they work, asking why they made specific decisions, such as changing an invoice over a mismatched zip code, in order to surface hidden exceptions and consolidate them into process maps.
Pitching transformation without fear 33:00
Dines closes this stretch by addressing the fear that monitoring desktops to build these maps will scare employees into thinking they are being watched for replacement. He says the key is how a company frames the message: transformation is inevitable, but employees who become more literate in AI will have a better chance, not just at UiPath but in any future job.
Fear Gives Way to Realism 34:30
Daniel Dines says that at the start of the year, fear of being replaced by AI was extremely high across the industry, but people are now grasping that adoption inside big companies happens slowly, one process at a time, because what he calls "millions of Einsteins" are not yet hireable. He argues that inference cost is not really the deciding factor in replacing a person with AI. Even if a machine costs more than a human, he would still hire it if the quality matched, since human costs only rise while machine costs fall, but the real obstacle today is that AI still cannot do the job at that quality.
Vibe Coding Hits Production Limits 37:00
Dines describes trying to build a procurement tool almost entirely through AI-generated code, calling the initial results amazing but not an extraordinary success once it reached production. Problems showed up around connectors, permissions, security, and testing, and the database schema the AI created turned out to be unusable, requiring a human to rebuild it. He concludes that turning a prototype into a working tool is where the real effort lies, not in writing the code itself, and that companies may end up spending as much as before while still tying up their best engineers.
Public Markets and Overvalued Zombies 40:00
Asked whether he would buy Salesforce, Dines says only as a system of record, not because AI will replace it through vibe coding. He questions why companies go public today given that private markets already offer liquidity and stock-based acquisitions, pointing to Stripe's plan to buy PayPal with private stock for 60 billion. He separates exceptional companies from what he calls "2021 zombies," many private companies sitting on paper valuations that would benefit from the reality check public markets provide.
Anthropic, OpenAI, and Concentration of Value 42:30
Dines recalls once picking Anthropic as his top bet when it was worth 60 billion, and says he still would not sell if he held it now, though he is unsure he would buy in at a two trillion valuation without seeing real numbers. He describes the market as one where the big keep getting bigger and value keeps concentrating. On Nvidia, he argues Jensen Huang's fortunes depend on open source succeeding, since if Anthropic and OpenAI become a duopoly controlling all intelligence, they will eventually print their own chips, which he says is not that hard to do.
Infrastructure Overbuilding and Its Limits 46:30
Dines explains that infrastructure is historically always overbuilt because companies want to claim the largest share of a big opportunity, guaranteeing losers when the market settles. He resists the idea that current AI infrastructure spending is straightforwardly justified, saying it depends on how fast AI actually replaces human work, whether that happens in two years or ten. He suspects current buildout may not be excessive for the next decade but could be excessive for the next three years.
Where AI Replaces Law Work 49:00
On the legal industry, Dines argues AI performs extremely well when a task is clearly defined with documented rules, but struggles where ambiguity and custom exceptions dominate. He estimates the U.S. legal industry at 300 billion dollars, and while Harry Stebbings suggests 30 percent of labor could be replaced, Dines argues the actual token revenue captured might be closer to 10 billion, since frontier models are becoming interchangeable and the real value sits in the workflow and process mapping built around them, which is what he sees Harvey and Lorra actually doing.
Cost-Efficient Models Will Dominate Traffic 52:00
Dines predicts that within twelve months, roughly 90 percent of enterprise token traffic will run on cost-efficient models rather than top-tier frontier models, since most operational work doesn't need that level of quality. He still expects to use Anthropic and OpenAI's efficient models while keeping an open-source backup, arguing that a responsible enterprise cannot risk being locked into a single provider.
Owning a Map of Work 53:30
Dines agrees that mid to large companies will eventually want to own their own models rather than rent intelligence, but stresses that the real asset is the documented "map of work," the detailed record of how a company actually operates. Without it, transferring learning from an old model to a new base model causes serious losses. He says this map, not the model itself, is the company's real intellectual property, and he speaks positively of Fireworks as an open-source provider he already uses, suggesting he would consider investing at a 15 billion valuation if they secure their own compute.
Data Providers and a Change of Mind 58:00
Dines and Stebbings discuss data providers like Mccor and Surge, arguing they are undervalued because raw storage is worthless without the intelligence to extract the right context for a model at the right moment. Dines then names the biggest thing he has changed his mind on this year: realizing that AI needs a detailed manual of a company's processes to work well, and that this points to a deeper distinction between memory and true learning, a theme he says runs through the whole conversation.
Transformation Versus Memory 59:30
Reflecting on writing his book with AI assistance, Dines notes that AI has no style of its own because style requires individuality shaped by lived choices. He argues that humans are changed by the act of doing something, the way Einstein was gradually transformed by years of thinking about the speed of light before arriving at relativity, whereas AI models simply record information without being altered by the process. He suggests this is why AI can solve difficult existing math problems but has not yet created an entirely new framework like relativity, and that closing this gap may take a model that can be transformed on the job over a long stretch of time, something he speculates could take another twenty years.
Optimism, Job Loss, and Europe's Position 1:03:00
Dines says he is optimistic for himself and therefore for his children, pushing back on being called an AI doomer, while Stebbings argues job losses will come faster than Dines expects. Turning to Europe, Dines admits that from a technology standpoint the continent is largely irrelevant despite having the talent and the chipmaking equipment, through ASML, to lead. He contrasts UK teams leaving work at 5pm for the pub with a more driven American culture, and credits his own success to what he calls an American school of entrepreneurship, though he still believes strong AI application companies, even if not frontier labs, can come out of Europe, and he affirms that sovereignty over AI infrastructure matters as an argument.
European Demand For On-Prem AI 1:07:30
European customers overwhelmingly prefer on-premises software with model optionality rather than being locked into one provider. Daniel Dines mentions trying to convince the team at Fireworks to make their software available on-prem, but they wanted proof of revenue first. He believes this on-prem market is a big business that is only just beginning to take shape in Europe, where trust and sovereignty concerns run deep.
UiPath Growth And Market Perception 1:08:00
UiPath currently does about 1.6 billion dollars in revenue, growing 14% last year. Dines acknowledges the harsh public market reality that companies not growing 20% or more get treated as losers, regardless of the underlying business. He argues public markets are confused about who the real AI winners and losers are, and with hundreds of software companies to evaluate, investors rarely look deep enough into any single one.
The Bull And Bear Case 1:09:00
Dines lays out the bull case for UiPath reaching 50 billion dollars in value: Gartner recently moved the company from challenger to leader in its Magic Quadrant for business orchestration and automation, validating its shift from pure RPA into a broader orchestration layer. He describes this as building a "map and rails" system, where the map captures enterprise context and the rails handle orchestration, letting agents work within controlled boundaries rather than running wild. The bear case, he says, is that AI could become genuinely brilliant and token costs could fall near zero, producing millions of true reasoning "Einsteins" that need no such scaffolding.
Quickfire Personal Reflections 1:12:30
In a rapid round of questions, Dines says the hardest part of being CEO is aligning people despite differing personalities and egos. AI has changed his routine so much that he now spends half his day working alone in Visual Studio Code with Claude and ChatGPT, reviewing markdown strategy documents instead of decks. He predicts Nvidia could see a 40% run over three years and would bet on that over Anthropic reaching a 7 billion dollar valuation, adding that "billions are nothing today."
Loneliness And Longevity 1:15:01
Asked about the loneliness of leadership, Dines advises founders to stay close to old childhood friends who can relate to them beyond the transformation of success. On longevity, he says he takes around 60 supplements a day, all vetted by AI, and has cut back heavily on drinking, saying he feels better now than ten years ago. The conversation closes with him giving Harry a copy of his new book, which is being distributed at UiPath's upcoming Fusion event.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
