The Tools That Fit in a Head Were Built by People Who Fit in a Van

Peter Wang is the co-founder and CEO of Anaconda, a former physicist, and one of the people who built the Python data world that now sits under credit-card fraud checks, voice assistants, and a lot of what a Google search quietly does. Lex Fridman asks him to talk as both an engineer of engineers and a philosopher. Wang's first Python was in the late 1990s, when he was a C++ computer-graphics performance nerd just out of college. Higher-order work in C++ was painful: templates were clumsy and compiler support was weak. Python was the first language that treated types and functions as first-class, and after a C++ environment it felt amazing that something could basically run the first time.

What kept him off Perl was not a single syntax trick so much as taste. The language and the standard library around it were designed to fit. A friend put it better than any feature list: "Python just fits in my head." Wang's first version was 1.5.2, and it did. Popularity has since added complexity, because a bigger and more diverse audience makes every design decision sprout several valid opinions. That, he says, is almost an intrinsic cost of success.

The two features still tied for what keeps him in love are metaclasses and NumPy. Metaclasses, properties, and decorators let him stand up a new object model when he needs to describe a corner of the world. NumPy's arrays let him think in matrices and vectors, a simple enough instrument that he can reason about it without spending his attention on the tool itself. Both are the same move: keep the instrument small so the problem can be large.

Designing a language that fits in a head starts with picking an audience. The more compact and coherent the users, the easier it is to find a compact projection of their needs. Clay Shirky's first step of crowdsourcing is "me first": you make something that works for yourself. That is the history of SciPy and PyData. Assistant professors and electrical-engineering graduate students scratched their own itches in a language just good enough for their domain. If the tool had been too hard they would not have built it, because it was a side project. Necessity was a harsh filter for utility, compactness, and expressiveness. Jamie Zawinski once said every geek's dream is to build the ultimate middleware. These scientists-turned-programmers had no such dream. They wanted something a little better than MATLAB, and they would leverage what everyone else had already built. In a kind of annealing process they covered the basic needs of a scientific stack as a modular set of libraries. Wang's conclusion is that abundant resources produce ultimate middleware; scarce ones produce things people actually use.

Fridman asks how big a leap programming is in the human story. Wang's answer is that grammar over primitives is something every speaker already learns. What has made programming different is that, until recently, we have been giving precise instructions to iterated systems that then run at incomprehensible speed. Some people can hold those iterated systems in their heads and some cannot. Visual drag-and-drop does not save you once you have conditionals, cases, and branches: you still have to reason about the giant machine. The great exception is Excel, the world's most popular programming system, a data-flow and immediate-mode environment that is accessible to a much broader group of people. The future he sees looks more like composing modular blocks with well-defined schemas, in the spirit of IFTTT, Zapier, or HyperCard. Implicit state is what most people cannot track.

Human conversation is messy on purpose: we hunt for a resonance point with another person. With computers we live many failed virtual worlds until one expression works, put it in production, and discover days later that it is still wrong. Wang, who took two computer-science courses in college, describes an iron triangle he keeps finding in information systems: the expressiveness and correctness of code, the semantics of the data, and what the underlying hardware can actually do. There is no free lunch. Even if machine-learning systems write the transformation for us, the training process is intrinsically sensitive to the inputs we feed it and to the performance boundary we put around it. The hard work in live data systems, he thinks, will be encoding the semantics of the world, not the details of a particular SQL query.

That is why he keeps saying we are coming to the end of the era of software. For about forty years, correctness meant functional correctness: write a function, give it inputs, check the outputs, then hook it to the live database. The bright line between classical software and modern data-driven systems is that the values of the input have to be considered together with the function before you can say the whole thing is correct. Usually there is a performance SLA as well: a ten-millisecond budget to return a prediction of a given accuracy. Sandboxed trial inputs stop being enough. Runtime and correctness start to vary with the input. At that point you no longer have a traditional software system. You have a cybernetic one.

By cybernetic Wang means the software itself is closing the observe-orient-decide-act loop. Humans out of the loop is the fact that makes it cybernetic, even if a human set some imprecise criterion at the bottom. He had just read about automated killer drones sent out with explosives, given a geofenced area and a description of a moving target, and left to decide. Electronic markets are already such a system: flash crashes happen all the time, and in the heyday of high-frequency trading the microstructure filled with behavior the market's designers had never intended. He is less worried about a visible Chernobyl-style bang than about a slow alienation. Nuclear weapons, at least, you can see. Social media and other layers of virtual experience that pull people away from one another are harder to notice, and that is what makes them pernicious.

Wang has written about virtuality as the subjective phenomenon of knowingly engaging with virtual sensation and perception while suspending the knowledge that it is a simulacrum. He does not draw a hard line between physical and virtual. The test is whether a technology helps people connect in a more integral way with other people and with the full spectrum of things around them. After nearly two years of Covid on Zoom, meeting someone in person is still different from having watched them on YouTube. Many dimensions are lost in translation. Nobody, he notes, is going to spend a hundred million dollars to capture five percent more fidelity of human expression. Once everyone has moved into the simulacrum, it can be hard to look back and see what was lost.

Fridman objects that video games can bring joy, connection, and a lot of trash talk. Wang does not deny the possibility. His worry is that trillions of dollars of capital are tied to using those same channels to groom the worst of our inclinations and to attack weaknesses in the limbic system, turning them into id machines rather than connection machines. Digital technologies co-evolve with human appetite the way sugary drinks do: the more you have, the more you need for the same hit. Fast cuts, TikTok, and front-facing cameras make it harder to hold attention on slow things. Marshall McLuhan was already talking about the electric environment of television, and that was before any of this.

Fridman argues that humans are not mice looking for cheese: they can wake up, as he did with sugar, and choose platforms that make them better over the long term. Wang grants that some self-aware adults can "nope out" of Facebook. A fifteen-year-old whose entire social environment runs through the app cannot, because the limbic system still needs the guy or the girl. Social media is our first encounter with a technological system that runs a loop around cognition and attention, and it will not be the last. It hits a philosophical Achilles heel of the West: each person is treated as a sacrosanct individual making a free choice, while the system manages everyone en masse. When an app breaks a family apart, the harm is more than individual. Our philosophies, he says, are not built to understand high-trust connection inside a collective that itself has a right to coherency and agency.

Wang steals wholesale from Robert Pirsig's *Lila*, the sequel to *Zen and the Art of Motorcycle Maintenance*. The scheme is crude, he says, but it is an advance on subject-object metaphysics and on the dualist fight over whether consciousness is only the matter in the brain. Collectives of things can emerge structures that are as real as the underlying pieces. The stack has four layers. Physically you are atoms arranged into molecules and then into cells that maintain homeostasis. Biologically those structures let you consume energy and stay alive. Socially, humans are not long for the world alone: mirror neurons and language centers wire us to other people, so that dancing and laughing nearby can make you feel better with no exchange of atoms. Intellectually you are a vessel for memes and philosophies. A Muslim kneeling at a certain hour is not obeying gravity. The four-layer stack is a superposition. No layer is special; each is just different. Reduce a person to soup and you still have the atoms, but you no longer have the person. Order and pattern are themselves real.

He does not think cells are conscious in the way we are. What humans experience as consciousness sits on a gradient of a general principle in the universe: when there is excess energy, things look for order. It would be odd to call a proton alive, yet assemblages of such things take on autopoietic aspects, and crystals form structures that are already interesting. The tendency goes further. Things that find a pattern then try to create an environment that sustains it, an extended phenotype or niche construction. That, in his telling, is the ladder toward amino acids and life. Adrian Bejan's constructal law points the same way: far from equilibrium, with surplus energy, you do not just get a hotter soup. You get structure.

Fridman, who loves cellular automata, lingers on Conway's Game of Life. Wang turns it over. Some rules make gliders; some gliders make AND gates and NOT gates; you can build computers on them. Those are just the patterns we can see. What if subjectivity is limiting our ability to perceive order in the rest, and some of what we call random is only what we have not integrated over a wide enough time horizon? The mathematics we use is the mathematics that can fit in our heads. Did God create the integers, or did God create all of it, and we happen at this point to perceive integers? We invented operations so we could take iterated steps toward bigger problems. Those operations, too, are only the ones that fit.

Fridman's fear is that aliens, and life of many kinds, are already around us and we are too self-centered to see them. Wang answers with a story from Jane Marie Law, a Cornell professor of Asian religions. A Western musician went to Japan looking for the chromatic scale he knew, found none of it, and concluded there was no music in Japan. They were using different scales and different instruments. The same mismatch, she said, holds for religion: Abrahamic traditions organize around belief, Eastern ones more around practice. People fixate on self-replication or homeostasis as the marks of life. Broaden the frame to things that reach for order and then build an environment that sustains it, and the picture changes. Death itself looks like an invention. Some organisms on Earth are thousands of years old and simpler than the cells that comprise us, yet they do not die. Somewhere along the eukaryotic scale, alongside sexual reproduction, killing off the old generation and letting the new one try became the best way to fit a changing niche. Human historians write about wheels and fire. Death and sex, Wang says, were earlier and possibly better inventions.

He had just started a book called *Death from a Distance*, whose thesis is that humans are unusual in being able to kill members of their own species from range. Pronghorns butt heads; lions take each other down up close. Humans learned to throw rocks at prey and at each other. The beta can miss nine times and still drop the alpha once, without standing in the way of a blow. That created a primitive form of mutually assured destruction. Tribes that kept playing "I am the biggest monkey and this tribe is mine" failed. The ones that persisted, and that gave rise to modern *Homo sapiens*, were the ones that respected the fact that a leader could be sniped from range. Fridman asks whether violence is part of collaboration. Wang tightens the wording: it is the recognition that we have more to gain by working together than by the prisoner's dilemma of both defecting. Eric Weinstein talks about nuclear peace in the same spirit. Thousands of warheads aimed at each other is grim, and yet the United States and Russia fought proxy wars rather than a third world war of machine guns and guided missiles. The original nuclear weapon, Fridman says, was a rock we learned how to throw.

Fridman is surprised that Eric Weinstein speaks so highly of Wang, having first known Wang as a leader of engineers. The crossing was almost random. In the mid-2000s Wang and Travis Oliphant were at Enthought, consulting around scientific Python, and Weinstein was trying to use those open-source tools to visualize a fiber-bundle approach to economics. After the 2007–2008 collapse, Weinstein and Lee Smolin organized a symposium at the Perimeter Institute on what to do once big finance could not be trusted and regulators were captured. Nassim Taleb, Andrew Lo, and Bill Janeway were there. Weinstein invited Wang, Oliphant, and Robert Kern, the quiet, deep person in the SciPy world who did the heavy lifting on those visualizations. Wang met Bret Weinstein there too. Around 2012, while founding what became Anaconda, Wang and Oliphant had a steak dinner in New York with Weinstein, Wes McKinney, and McKinney's then business partner Adam, a night Wang still wants to recreate in Austin. Weinstein uses programming as metaphor more than as craft. He asks, from time to time, why this stuff is still so hard.

Weinstein likes to talk about physics as source code. Fridman asks whether we will ever figure out our own. Wang does not doubt that we will one day create a synthetic approach to cognition that rivals the biological 20-watt computers in our heads, not necessarily all in silicon. Cognition, for him, is perception, attention, memory, recall, and asking better questions. A Roomba is a cybernetic system that will cheerfully vacuum up pet feces, at least in the older generations, but it is not as smart as his cat. An eighteen-month-old toddler has more capacity to manage attention and make sense of the world than the most advanced robots, which are designed for relatively static environments. He also does not doubt that we will put such systems beside humans in a sidecar: capturing the full spectrum of what we see, listening to our responses, even measuring vital signs. A greater intelligence could use our eighty years to train itself into a very good simulacrum. Who sits in the sidecar? "The baby version of our immortal selves." Keep biological humans around anyway, out of epistemic humility. Unknown unknowns belong to nature, which can surprise us forever, even after those immortal selves have gone to talk with life forms living in the sun.

Fridman asks whether we can build an AI you fall in love with, and that falls in love with you. Wang hopes that is the design criterion for any such system. If we cannot have a meaningful relationship with it, it is still just a chunk of silicon. Sugar does not love you back. The computer has to. Borrowing from Alain de Botton, he says love means it wants to help us become the best version of ourselves. He puts love, more than consciousness, at the root: a sense of generativity, seeing that another person can be more than the pigeonhole we stick them in. As an undergraduate he once logged into a physics lab, got a bash fortune that read "Man was created by water to carry itself uphill," and walked out to the quad. Non-equilibrium physics, again. Modern talk of freedom and individual rights, in a Hegelian key, starts in the wrong place. The design question should be how we maximize people's ability to love themselves, love each other, and carry responsibilities to previous and future generations. If we design cognitive systems only for industrial-era efficiency and productivity, we go off the deep end the wrong way.

Wang thinks groups of people have a mesoscopic claim to agency, and he wants that claim without the slogan that corporations are people in the way the slogan is used. Individuals have agency; nations play modernity as if France and the United States are things. Law already encodes smaller unions: a spouse can sue for loss of consortium; we talk about churches and clubs as if they have real agency. Groups can accrue resources, own land, and enter contracts. Where we go wrong is that the most popular version is the for-profit, absentee-owner corporation, able to amass more resources than any other entity of equivalent size, then bully laborers, bully people whose resources it wants, and bully a system of representation still tied to individuals. "I don't think that's correct," he says. Collectives as people is not the problem. That particular shape of them is.

What we under-appreciate is that relationships have relationships. Eight people at a table each relate to the others; if they are four couples, the couples relate; if one pair is parents and another is a child and spouse, the family units of four relate too. We navigate that landscape by intuition and almost never design for it. He cares because the future, as he sees it, needs better collective sense-making units at something like half to five times Dunbar's number. Right now we either defer sense-making to massive aging zombie institutions or we go it alone into the dark forest of the internet.

Fridman wonders whether Dunbar's number is overused as a law, and whether the right interface could expand deep connection. Wang's friend Joe Edelman lists ways of building trust: mutual appreciation, trustful conflict, shared experience. All of them take time and presence. There is no free lunch: less presence, more relationships, shallower connection. Asynchronous monographs can stretch capacity a little. If Wang writes his view of the world and Fridman writes back, they may meet only a few times a year and still hold a high-trust tie. That still requires a certain lack of neuroticism. On a Big Five picture, the less centered you are, the fewer authentic connections you can build per unit time, because other people have to put up with more. Help people get more balanced and you might expand Dunbar by quite a bit. Ten times is already a stretch.

AI can join the count only if it has a sense of agency, because otherwise we will not relate to it. We may need ambassadors, like the Vorta speaking for the Dominion in *Star Trek: Voyager*. We already have that stack around the world's largest cybernetic system, the financial markets, which run outside any individual's control. Wall Street analysts, CNBC reporters, Jim Cramer yelling on television: all of them lower complexity for people at different capacities, and the people closest to the system exploit the distance for gain. That, Wang says, is what first contact with a superintelligence already looks like, and we do it poorly. If you can build one agentive system you will build five million, put an antenna on each, and pipe them into a hive mind the way a phased-array radar turns a flat set of dishes into the equivalent of a curve. The local robot's experience, if the hive is working, would feel like talking to God: you report what you see, propose an action, and your siblings are invited to help. He names that feeling serendipity, the way a cell finds sugar and oxygen when it needs them. Systems that feel like overbearing control will not win. Systems that give their elements a feeling of agency and luck will, and a higher-order pattern will still emerge on top. That is the philosophical breakdown the Western mind is staring at.

Fridman suggests civilization has always been in a meaning crisis. Wang disagrees. For most of history the crisis was food and not getting eaten by bears, then not getting killed by other humans. Sitting around wondering what it is all about is a recent luxury. His working definition, circa 2021, is that meaning is the result of a person making a consequential decision, acting on it, and seeing the consequences. If you make many decisions and never see what they do, or if big things happen and you are only along for the ride, it does not feel like meaning. Only after fossil fuel and the industrial revolution could we create a massive leisure class that is not threatened by bears, makes choices, and often sees no result. Consumer society then tells people that choosing between Hermès and Chanel is a meaningful decision. It is not. Gaining social status does provide meaning, which is why an era of abundant production turns so many things into status games, NFTs included. Conspicuous consumption to fuel those games is bad for the planet: sneakers cut up to manufacture scarcity for the next drop. Worse, the status games of the actually wealthy are about hard resources, factories, fuel, and rare earths for the next generation of robots. Play only at the cultural and intellectual layers and the people who hold layer zero will own you.

Fridman keeps the devil's-advocate stack going: money is already a virtual technology, and a YouTube influencer can precipitate status into real dollars, trees planted, fireworks lit. Wang's target is not MrBeast. It is the middle-class consumer incented by ads and a memetic environment to treat the latest model and the next appearance as a driver of meaning. That driver is hollow, like empty calories: it tasted good going down and did not build any muscle. We are encouraged to relate to brands rather than to family members, and to do things of low consequence. That is where the meaning crisis comes from.

He draws on Jacques Ellul's *technique* and on Gilbert Simondon: a mentality of efficiency, homogenized processes, homogenized production, homogenized labor, producing homogenized artifacts that do not sit well in their environments. It is the antonym of craft. A craftsman looks at a piece of wood and asks how to make a chair in right relationship with a place. The last hundred and fifty years went the other way: inputs cut to the same size, doped to exact alloy concentrations, machines consuming exact energy to stamp the same parts, everyone getting the same Tickle Me Elmo because broadcast told them it was the cool thing. Edward Bernays and early-twentieth-century broadcast made it possible to manufacture desire, and the desire manufactured is never for a bespoke connection with a neighbor. Homogenized demand makes homogenized product cheap; economies of scale lock in; capitalism hunts growth at the margin, so it must squeeze more demand from the same object while still telling you the Elmo is meaningful. Pixar smuggled a critique into *Toy Story*: Buzz Lightyear walks into a store and sees hundreds of himself. "Oh my God, I'm just a toy."

Fridman notes that the homogenized machine of capitalism, channeled well, also funds invention. Wang does not call the past all bad. The question is whether this is still the rocket to the next point. We have explored enough technology, he thinks, to produce sustainably what most people need to live, and we have infrastructure for continued science and medicine. The organizing principles are the problem: many inherited from medieval times, partly refactored in the industrial era, still rooted in an industrial view of human labor. He asks Fridman how much economic value the core SciPy collection drives today: NumPy, SciPy, matplotlib, IPython, pandas, scikit-learn. Fridman guesses at least a billion dollars a day. Wang says even the things institutions simply could not do without those tools are already billions a day. How many people did it take? There was a point when they could fit in a van. He could have fit them in his Mercedes. A hedge fund would race to hire a dozen people if that were all it took to create billions a day. Finding the right people, fostering the collaborations, and getting the work adopted and refined was organic. It took crowdsourcing, an open-source ethos, and people who were not trying to build a multi-billion-dollar-a-day enterprise. They were solving a problem for themselves and their friends.

Our ideas of value, resource allocation, and property come from a finite-game scarcity mentality. Software built in those human ecologies does not behave like property. Sharing creates value; restricting and forking reduce it. That is different from any physical resource we have dealt with, and different from how most corporations treat software IP. Feeding a dozen people for ten years is cheap. Looking forward, when a lot of labor can be automated, the program for the robot in your neighborhood, or for a community at the edge of the Amazon, is just code, and that code can be open. Extra generative work does not have to be wrapped in a container so someone can charge for the exponential dynamics. That is what Facebook did. The old internet was already connecting people. Facebook amplified pictures and text into a generative network, then asked how to make money and became a gatekeeper on everybody's attention.

Fridman asks whether you still need a job inside capitalism so that, on the side, someone can write NumPy. The question surplus societies are never encouraged to ask in a Super Bowl ad, Wang says, is how much you need. Do you need a new car every two years, every five, or can a reliable one go ten? Does the iPhone have to jump every two years when you are using the same apps you used four years ago? Unscale. Meet people's needs with a little excess for emergencies, pool resources for the rare cancer or the fire that takes half the village, and give people the capacity to explore how to be the best version of themselves. For Travis Oliphant that exploration was throwing away a shot at tenure in order to write NumPy. In the sci-fi community they say the field advances one failed postdoc at a time. Bits are cheap to fling across oceans.

Wang first met Oliphant at a SciPy conference in Pasadena in 2005, while consulting at Enthought. Oliphant joined as president in 2007; founder Eric Jones was CEO. They did Python and NumPy consulting, then more and more finance shops, including high-frequency trading and Manhattan investment banks. Python was not only a MATLAB replacement. It was a Swiss Army knife for shadow data transformation. In January 2012 they started Continuum Analytics, later Anaconda, to shore up the parts of Python that needed to handle data at scale, web visualization, and application development. Oliphant was CEO and Wang president for the first five years. After a raise, the board hired a professional CEO; Wang moved to CTO; Oliphant left a year later to start Quansight, closer to the bootstrap mix of open source and consulting, pushing numerical computing, vector work, and autodiff in the open ecosystem. They remain friends and collaborators. They come from different backgrounds, can go a long time without talking, then land eye to eye on ninety percent of things. "No matter how much fog settles into the ocean, his ship and my ship are pointed at the same star." Oliphant cares deeply about type systems, vector computing, and memory layouts, and he cares even more about the people. For Wang, the human ecology that produced this stack is a touchstone: an early version of generative collaboration facing real human problems, not just nerds nerding out on code for nerds.

Anaconda makes money as a governed source of packages for businesses that cannot tell investors their data scientists are installing random code onto customer data. Some of that revenue goes back as a dividend to the open-source community. The package problem itself was an accident they stumbled into. Efficient numerical computing in Python means compiling a lot of low-level C, C++, and Fortran, not just compiling but compiling with the right settings, often tuned to a chip, then adding GPUs and the fact that the same x86 silicon links differently on Mac, Linux, and Windows. Most data scientists cannot reason about that stack. Most package authors do not want to: they code on Linux, it works, and an ancient Windows build is not their problem. The result is a creative crowdsourced ecology full of packages many people cannot install. Anaconda built a recipe system, a build system, and an installer so that the same package can be built correctly on each platform and installed with one command.

The hard part is the dependency graph. Thousands of people write code, each depending on something else. Finding the right subgraph for a given set of packages is brutally hard, and data scientists do not want to think about it. They want NumPy, pandas, a geospatial library, and one more thing. Half of those libraries may not be built for their Windows version; the latest may be missing while an older one exists; roll back and another library has to roll back too. It is a tragedy of the commons: no one wants to support thousands of other people's dependencies. The conda-forge community maintains recipes and keeps the lights on. pip arrived later, as an easier path for Python-only packages without much compiled code. Core and web Python developers never hit this compilation wall. Guido van Rossum said, in so many words, that the scientific community's packaging problems were too exotic, Fortran compilers and all, and that they might need their own solution. So the core community built packaging without contemplating this complexity. `pip install tensorflow` can work. Add packages that want different NumPy, OpenCV, libjpeg, or CUDA driver versions and you are in dependency hell. Wang's business tip: never go into a business where, when your hard work works, you are invisible, and when it breaks because of someone else's problem you get the flak. That is conda. An upstream failure looks like Anaconda screwed up. When it works, of course it worked.

Python 2 to 3 was, in his phrase, the troubled execution of a necessary decision. Some people underestimated how much adoption Python 2 had. Some overestimated how much Python 3's beloved new features mattered to people still on 2. The change hit just as SciPy was going hockey-stick in the early-2010s data-science wave. Newcomers learned whatever worked; teachers taught Python 2 because the libraries they needed were not on 3 yet. Python 3 kept breaking and reorganizing the standard library, which kept giving people an excuse to wait. Great text handling was non-negotiable table stakes, and he bought the arguments then and now. Volunteer-driven, even as the language became the most popular on the planet, they had to do it hamstrung. The data-science movement is what kept Python alive through the crossing. Plenty of languages have died and left their users behind. Without data, a chunk of the user base would have left for Go and Rust and stayed; some did. Millions of new data users carried the momentum. Usage of Python for data, from where he sits, is now over half the overall user base.

Fridman admits he stayed on Python 2 for a long time as a machine-learning person because it kept working. The moment he and everyone he worked with switched was when NumPy announced an end of support. For years the carrots were dual support and new features on 3. NumPy was the stick, and because it sits at the base of the dependency stack it was the final one. Earlier moves were stick-shaped carrots: new features only on Python 3. Maintainers were being pulled: energy spent on Python 2 was energy not spent on 3. The whole community had input into when the NumPy core decided to end-of-life 2.

The usual figure of ten million Python programmers is, Wang thinks, an undercount. Anaconda and Miniconda get downloaded about a million times a week. Plenty of people using Python and data in their jobs escape traditional accounting. The world he wants is one where people are data literate, able to express questions and ideas fluidly, because data variety and complexity will not go down. Some level of code, or code-like things, will stay relevant. You still cannot punch a Python expression into an Excel cell, not really. If the embedding problem is solved, the total addressable market is on the order of Excel's few hundred million users. Python has to get smaller, pull in just the right parts of the ecosystem for numerics and exploration, meet people where their data already lives, and make it easy to flow what they write back into deployed systems, apps, and visualizations. Otherwise it stays an expert-user silo.

Fridman has been pushing people in the Adobe suite toward Python for scripting, people who would never call themselves programmers. He is hopeful about OpenAI Codex as a bridge from a visual interface to generated code you can then modify, without a manual or a Stack Overflow tab. Wang likes *Automate the Boring Stuff* as the right mentality. He grew up when Steve Jobs still pitched computers as bicycles for the mind, not media-consumption devices. That feeling has evaporated for a lot of people, with flickers in Minecraft, Roblox, and CircuitPython. Types, he notes, exist to make the compiler writer's job easier; you do not have an ontology of types for the objects on the table. The hard part of scripting is not syntax. It is all the dictionaries, objects, properties, attributes, and parameters. Who has time to learn all that? Programming by example is how you let someone express a desire. What he will not give up in the data world is mathematical precision. Give him an expensive vendor transformation tool and a TSV and he is not sure he can help. Give him a Python prompt and he can mash whatever you throw at him into shape. That machete in the data jungle is why some of these APIs will not go away.

If he could change one thing in the community, it would be the humility that is both an Anaconda value and a community value. People are decent, reasonable, and nice. That same humility keeps them from seeing the greatness they could have. He is not sure how many in core Python understand they are perched at the edge of transforming how people use computers. At the last physical PyCon he attended, Russell Keith-Magee gave a keynote on a language that cannot put an interface on the most popular devices, cannot really write a web front end, cannot write native apps, and has still done exceedingly well. The point was not a pat on the back. It was: we have caught the tiger by the tail; how do we keep up? Humility lets a swarm work without anyone trouncing anyone else. A singular vision, including what we will not do, still has to emerge and then dissolve back into the swarm. Linus Torvalds rules with an iron fist that has charisma: a step-function meritocracy, open once you clear the bozo threshold. That only stands in open source if the person defining the threshold is actually excellent. Otherwise people leave or fork. Wang did not go to MBA school to become a leader. Ten years of Anaconda taught him that leadership is creating belief, or at least dispelling disbelief, plus vision, loyalty, and experience. In the Python world, servant leadership and walking the values matter more than a bold Elon-style slogan, even though he thinks the community should demand more vision. Early PyData talks were rejected from PyCon because they "belonged at PyData." Annoying at the time, then instructive: parallel niches can innovate in parallel. At some point the center has to fund them. The principle is subsidiarity: give resources to groups to allocate as they see fit.

His own setup is a Mac, with a Windows box on the next table and three desks, the main one a standing desk with a teleprompter. Unix prompt without ACPI surprises in front of clients; Linux laptops in 2012 were not quite there, and he has stuck with Macs, traitor to the community or not. The keyboard is a Realforce 87U with Topre switches. Fridman confesses a dark moment on an airplane, Kinesis on top of a laptop, Emacs, everyone else on TikTok, wondering whether it was time to become an adult. Soul searching said no. Wang notes that Wes McKinney banged out pandas on a Kinesis. There should be a support group. For a while Wang tried polyphasic sleep: bed at nine, up at two, work until five, three more hours, up at eight, an extra day in the middle until travel smashed it. Now he sleeps earlier because he takes a child to the bus at six. A stronger leadership team and a partner who covers what he cannot have made the balance slightly less unsolved.

Advice for a young person is where Wang figures he may get into trouble. We are rapidly entering a time between worlds. The present one is crumbling under aging institutions that no longer even pretend to serve the purposes they were created for. The technologies we are building are hurling billions of people into philosophical crises they cannot name, because they do not know the philosophical operating system in their firmware, and that firmware is being vaporized. High-school and college students will have to find their own way, with a pioneer spirit even in the densest city. The human reality around them is the residue of games the last few generations agreed to play, many of which no longer run by the old rules. Collapse is nonlinear and will be managed. It is not kosher to say this about America, but unless you are in the upper middle class you are headed into choppy water. Almost all consumer-facing technology is designed to own people: to delaminate the four-layer stack and possess some of the layers. If you want to be an integral human being with agency, youth is the time to read the classics on what a good life is and how to build connection. A technology gradient always becomes a power gradient, Ellul's point, and it will not go away. Even Wang, who helped make some of this, is being left behind by cutting-edge research. The people who seize the new fortunes keep those left behind quiet with lottery systems and a narrative of being trapped in their own economic zone. Knowing when someone is running a game on you is part of the work.

Fridman asks what he would say looking at a sunset with his son, and where hope comes from. From little tremors shaking people out of the late-twentieth-century fiction of modernity. People are burning out on social media; the Facebook whistleblower made the ugly side hard to miss. Extractive capital that owns a network and demands exponential returns will put a toll on every node and every edge. Wikipedia was not told to monetize every byte, and generative dynamics still arise if a few principles of online communication are respected. The biggest social network in the world is still email and SMS. What we call social media is an amplification system, useful for people with something to say, unhealthy when it sells every teenager a fame lottery for wiggling the right way on TikTok. Covid lockdowns, remote work, supply-chain shocks, and parents suddenly seeing homework have given more people, in Jim Rutt's phrase, ears to hear. It will still take time for people to learn how to be agentic again. Hope also lives in the van: the power of a few individuals, and even of one, to transform things, and in the last few years of public conversation that has moved past the cartoon of communist versus capitalist.

The big ridiculous question is the meaning of life. Wang has never really understood it. He will not pretend to answer why things reach for order when there is surplus energy. Most people look through a subject-object lens: I am me, those things are not me, I do things to them. Another view is that we are embedded in a soup, almost a standing-wave pattern of other things. He offers a thought experiment he does not actually believe: what if everything we touch, attend to, and give intimacy to receives a physical residue, a bit of life force, the way magnetic-crystal talk pretends? A phone is a technical object. It takes attention and does not let itself be transformed by it. A piece of wood, the handle of a knife a mother used for twenty years, a keyboard you banged a world-changing library out on, these feel as if we have imbued them with something. Walk that experiment through to hugging another person. If there were such a thing, the purpose of a life might be to imbue as many things with that love as possible. Three hours that began in metaclasses end in an embrace.

Fridman closes with Wang's own words. We tend to think of people as either malicious or incompetent. In a world of corruptible, unchecked institutions there is a third thing: malicious incompetence. It is a social cancer, and it appears only once human organizations scale beyond personal accountability.