The 87-Kilobyte Program Two Billion People Open Every Month

Dave Plummer's first computer was not a machine he owned. Around 1979, at eleven, he rode his bike to the local RadioShack and found a stack of boxes labeled "computer." The staff had just received a TRS-80 Model 1, Level 1, with 4K of memory, and had not set it up. He offered to do it. He had no idea what he was doing. Tandy had used the same five-pin DIN connector for power, video, and cassette, so a wrong plug would blow the machine. He read the labels, got it working, and then did something that still sounds like him: he typed English into the BASIC interpreter. `PRINT 2+2` worked. Ordinary sentences did not.

The big three of that moment were the TRS-80, the Commodore PET 2001, and the Apple II. Asked for the greatest computer ever built, Plummer does not pick the Apple. He gives the vote to the Commodore 64, on sales rather than design. A machine, in his accounting, is as great as the number of people it reached. Lex Fridman, whose first computer was also a 64, agrees. The criterion is blunt and consistent with the rest of the conversation: impact first, elegance second.

His own first Commodore 64 arrived without a real warranty. The floppy drive overheated, and because the family had bought it from a small shop rather than a major retailer, the machine had to go back to Germany to be fixed. He was left with a cassette deck and a machine-language monitor loaded from tape. The monitor could disassemble 6502 code but had no assembler, so he entered opcodes in hex. The first program was a clone of Galaga: enemies attacked over time, the logic was complete, and all of it was hand-coded machine language. 6502 code does not relocate. To insert anything in the middle he had to jump out, do the work, and jump back. It was hideous spaghetti. It ran.

Then he made a backup. He copied a blank floppy onto the data floppy. Weeks of work vanished in one reversed operation, and the guilt stuck: he knew he had rushed, or made a stupid choice, and he had destroyed the only copy. That was his first lesson in data management. He fell in love with the machines in two stages. The first was immediate fascination with the TRS-80. The second came in university, when he realized he loved programming and had no idea whether that meant making a VCR clock blink or working on an operating system. The consolation was the same either way. He was finally in the right area.

Plummer dropped out of high school, and he is careful about the verb. There was no moment when he decided to drop out. He went less and less until showing up would have been embarrassing, and then he stopped going. The mechanism is the part he wants people to watch. Each day the decision is tiny: I am just not going to class today. Repeat it enough times and it becomes a large decision you never made while fully awake. If you want a non-standard life, he and Fridman agree, make the large decision explicitly. Do not let a hundred absences make it for you.

Out of school he worked nights at 7-Eleven and in a paint warehouse. The 7-Eleven shift, as he describes it, is a rotation kept for people smart enough to handle overnight accounting and administration who also have a personal reason to be there. He had no diploma, so he qualified. Being held up at knifepoint was "entertaining." The sucky part was the gas dips: a fifteen-to-twenty-foot wooden stick, marked in inches and feet, dropped into underground tanks because there was no electronic sensor. The first time, he dropped the pole and grabbed it again. Splinters, forty below. That was the job.

Standing there, he had a revelation that next time he would hold the stick differently, and three seconds later he heard how ludicrous that was. He did not want to do this for a whole life. At twenty-one he went to the principal of the local high school and asked to come back. Too old, no room, the principal said. Plummer's counter was that someone would drop out before the term started, so there would be a seat; assuming the seat, could he return? The principal let him. He finished the three or four classes he needed.

College did not snap into focus immediately. He was a terrible high-school student, and in his first university semester he still was not taking it seriously. Geometry 90 was the makeup class for the twelfth-grade course he had missed. He mercy-passed by about one percent, which terrified him: a professor's grace had nearly been the end of the entire degree. Remedial marks did not count on the transcript, so the next semester was a fresh start. He did it for himself, and that, he says, made all the difference. His advice to students is unromantic. Find the part of school you actually enjoy, then put up with the hard parts, because doing them later is worse.

The jump from 7-Eleven to Microsoft still sounds unlikely when he tells it. In his third year of university he spent a summer at the phone company converting UBNet to TCP/IP, which in practice meant swapping cards and then figuring out why a customer's `config.sys` no longer let Lotus run because the machine had 10K less free memory. It was, he says, a horrible time to be working in computers. At lunch in the food court he read a book on how Microsoft came to be — he remembers the title as something like *Microsoft*, or *Bill Gates and the Making of Microsoft Hard Drive* — and became convinced that was where he wanted to work. He was in Saskatchewan. The book did not explain how to get there from there.

He had already built the thing that would open the door. HyperCache was a filesystem cache for the Amiga, which shipped with none. The tight path that moved blocks in and out of the cache was 68000 assembly; everything else was C. It was an N-way associative cache that tried to match drive geometry and prefetch whole tracks. He sold a couple of thousand copies at twenty or forty dollars, enough to put himself through school. Distribution started on Usenet and Amiga forums with a thirty-day trial, then retailers started calling. One order for fifty copies at twenty dollars — a thousand dollars in the late 1980s — arrived on his future wife's car phone, antenna on the roof, box in the trunk.

Shareware in that era meant a four-by-six registration card: name, address, email if you had one, mailed back for update notices. He went through the whole stack looking for Microsoft suffixes, found three or four, and cold-emailed them. He was an operating-system student in Saskatchewan looking for an opportunity. One reply came from Alasdair Banks, who passed him to Ben Slifka. Slifka did a phone interview and hired him as a summer intern on MS-DOS. The specific path is almost impossible to copy now. The logic underneath it is not: first make something you can show, then walk into the room holding that thing.

Why did Microsoft dominate personal computing for so long? Plummer's first answer is the room. It was the single most potent assemblage of smart people he has ever been part of, in academia or industry. At a regular computer company, he says, the one person who actually knows what he is doing has a smarter friend, and that friend probably works at Microsoft. You arrive as the big cheese from a small town and suddenly decide not to talk, in case you sound stupid. Bill Gates, in this telling, is not a myth. He was relentless about one slogan, a computer in every home and on every desk. It was his special interest. He hired people as smart as he was, or smarter, and built an almost unstoppable machine of intellect to go forth and make very simple products.

MS-DOS was one of those products, and it is the team Plummer joined. Before DOS, Microsoft was largely a language company: BASIC for many machines, a Fortran compiler, a Pascal compiler. The deal to ship MS-DOS with every IBM PC set a standard they leveraged for decades. They did not charge IBM a lot. Making it a standard paid out over time. Plummer will not pretend that was pure genius. To a certain extent they lucked into it; on the other hand, they were smart to have done it. Luck and judgment, both admitted.

What was MS-DOS, stripped of nostalgia? Largely a command launcher. You typed a name; it looked in the current directory and on PATH; if it found the program it loaded it and ran it. That, he says, is ninety percent of the product. Environment variables, a small scripting language, access to disk and CPU, no multitasking, no graphical interface. DOS 4.0 had a graphical file manager; DOS 5.0 added a text-based UI for an editor and QuickBASIC. Underneath, it was still a prompt. It was also exactly what that market needed.

The 640K ceiling was not a Gates quote. Plummer knocks that one down on contact: it has been attributed to Gates, and it is not him. The ceiling itself, though, sat on every operating-system developer as a responsibility. Whatever you used, the user did not get. Waste 10K and every machine in the world had 10K less. His first real intern work was adding CD-ROM caching to SmartDrive, because discs were new and Microsoft Bookshelf was one of the few things you could run; a smart cache made them dozens of times faster. The harder job was moving SmartDrive and later DoubleSpace into high memory, playing the A20 line so segment-and-offset arithmetic could reach an extra 64K above 1MB, with stubs in low memory. He had never written a line of x86. He was an Amiga programmer. He learned HIMEM on the job.

Windows 95 was being built in parallel. Plummer spent three or four months on the COM/OLE team writing the presentation cache: enhanced metafiles so a Word document could show an Excel chart without loading Excel every time. That code shipped in 95. About six months after arriving he moved to the Shell team and worked on NT from there forward. Windows 95, as he describes it, was 16-bit Windows 3.1 plus 32-bit support, VxD drivers, and an entirely new user interface. People lined up at night to buy it. He does not want to reduce the leap to the Start menu, but he thinks the new shell was a big part of it. He could not quantify what was different and awesome. He knew he wanted in, which is why he started writing a shell extension that later became ZIP folders.

A productive day in that era started with an hour downloading the MS Mail address book, because ten to fifteen thousand employees each had to pull a fixed list of ten to fifteen thousand people over 10-megabit networking. The useful part of that hour was Stress: unused machines ran tests all night, crashed into a debugger over a serial cable, and arrived triaged. A crash in the Start menu went to Dave. Then it was back to coding, which for him was about eighty percent fixing bugs, especially other people's, while porting. He wrote new code on the side to get it out of his system. The main job was taking the entire Windows 95 user interface and moving it to NT, which meant Unicode: every eight-bit character became sixteen. He compares porting to breaking into a house and going through the drawers. You find the pictures on the wall. You also find things in the nightstand. He saw 200-character lines with profanity in them. It all got cleaned up over the years. At the time it was not always professional. The other side of that tour is the programmer's fate: you write a line thinking nobody will see it, and a crowd of people will, including you five years later, looking for a bug.

NT itself was close to a clean sheet. Microsoft brought over Dave Cutler from Digital Equipment, where he had done VMS and RSX-11. After Prism and MICA were canceled at DEC West, a whole team showed up, Cutler and Mark Lucovsky among them. They had OS/2 as a starting point, but OS/2 was assembly and NT was going to be C; how much they actually reused, Plummer does not know. Cutler is the architect of the kernel. In the Linux world that role is Linus; in the Windows world it is Dave C. Other people contributed huge pieces. Cutler was the driving force, and at eighty-five, Plummer says, he still codes every day, still a Microsoft Fellow, still going in. Wicked smart, and also like a farmer: he followed people around to make sure work got done and got done right, and he would not tolerate crap in his operating system. The Digital Equipment style of leadership was a paradigm shift for Microsoft developers. Tools mattered too. They were using Diff and manual deltas. Git, Plummer says, would have been a luxury.

The intellectual fights could stop being about the code. Two smart people would compete until besting the other person mattered more than the technical merits. Plummer lost one of those fights and still thinks he was right. On a MIPS box, unaligned addresses are illegal. ID lists — path components glued together — became odd-length once characters went from eight bits to sixteen, so every later string sat on a bad address. Reading them meant exception handlers and byte copies, a hundred to a thousand times more work. He argued they should guarantee even lengths, even late in Windows 95. It became a shouting match. He lost. That code, he says today, is still thousands of times slower than it has to be. Nobody cares, because it is plenty fast. It could be a lot faster. Half the day was debugging in pure assembly anyway, no source-level debugger, on four instruction sets: Intel, MIPS, Alpha, and PowerPC. He puts his career at twenty percent creating and eighty percent debugging. When Task Manager sometimes showed more than 100 percent CPU, kernel people were unsympathetic. He filled the code with asserts, then put his phone number in the message. Stress finally caught it on a machine with a debugger attached. It was a kernel accounting bug. They commented the number out instead of deleting it, so it shipped, and the way he finds Task Manager in leaked NT sources is to Google his own phone number.

He is religious about asserts, and the Task Manager bug is why the sermon sounds lived-in. He does not assert things he hopes are not true. He asserts things he knows cannot be true. When one fires, it is a bug, not a warning. Add them while you write, not later: if a pointer cannot be null, assert it; if a count is always less than twice the byte width, assert that. Some people skip asserts because they are afraid they will fire, the way you might not look left at an intersection because you do not want a car to be there. The point of looking is to know.

Task Manager itself started as a side project he wanted for himself: see which apps are using CPU and memory, which are locked up, and kill them. He wrote the basics at home against the registry performance keys, because from home he would not call internal APIs. Once it came in-house he could call `NtQuerySystemInformation` and `NtQueryProcessInformation` and get real answers fast, which is why the app feels so responsive. He was not chasing features. He wanted a basic set, everything he personally needed, and then absolute robustness and absolute smallness. The original binary was about 87 kilobytes.

Getting to that size took a hardcore refusal to link the C runtime, which would have added something like 96K and nearly doubled the app. He called no runtime functions, then had to invoke C++ object constructors by hand from the dispatch table because nothing else would. The other requirement, not freezing, meant assuming almost nothing. If you ask the shell to run an app, the path might be a TCP/IP share that takes ninety seconds to time out, so every API that might block goes on another thread. The program got more complex because it had to stay up.

He also reinvented Hamming codes without knowing the name. Every row and column in Task Manager carries a dirty bit, so he can find which cells changed and, with the ListView control, repaint as little as a single cell. That is why it paints fast and resizes smoothly. Resizing was his personal north star: any size, even 32 CPUs, which was impossible at the time — it draws eight graphs and wraps, and it still works. As far as he knows the original code is mostly still there, under layers of drawing, dark mode, and XML that took it from 87K to four megabytes. "That's the world we live in," he says. Later in the conversation Fridman asks how many people have used it. Plummer has seen the metrics. About two billion a month. What he loves is not the vanity. It is that a plumbing diagram he had in his head is now running on billions of machines.

Someone walked into his office, asked what he wanted to do for the next three months, and offered Space Cadet Pinball. NT needed visual splash: high-speed graphics, or at least responsive ones. A lot of the original was assembly, and he was on a MIPS box, so he rewrote it in C to port across platforms. At the heart of the game is a huge state engine, a switch with something like fifty entries and an Easter egg inside. He treated that as a black box and wrote drawing and sound on top. People still ask why the physics differ slightly from the Windows 95 version. The answer is a bug. He drew as many frames as the machine would give him. On a modern PC that can be 5,000 frames a second, so physics interpolate 5,000 times instead of thirty. Arguably better. At least different. It was later fixed. He takes no credit for the design; that belongs to Cinematronics. The table, he notes, is close to Black Knight 2000, a physical machine he owns.

Solitaire and Minesweeper, he says, were not included as games. They were educational tools for the mouse. Solitaire teaches drag-and-drop. Minesweeper is probably right-click, the flag. A generation then spent uncountable hours on the tutorials. When he volunteered in his kids' school, none of the third-graders were impressed that dads worked at Microsoft — until one found out he had worked on Pinball. For a brief window, that was the credential that landed.

Windows activation is the feature people say like an accusation. He got it late in the XP ship, after whoever was supposed to adapt Office's activation had slipped it enough times that it was not going to happen. He had a reputation for fixing things quickly. With the DRM team and the researchers doing product-key math, they cranked it out in time. He does not know the revenue impact and imagines it was substantial. He also admits it is annoying, especially phone activation: hardware parameters and a product key encoded into letters and numbers someone will read aloud. If you think that is painful on your end, he says, imagine the person on the other end of the line listening to product keys for eight hours a day.

One of Plummer's more popular recent videos asked why you can no longer move the Windows taskbar. The outrage, he found, was palpable. Freedom to put the Start menu on the left, top, or right raises the complexity of rendering, layout, and tabs, widens the bug surface, and probably costs a developer, or part of one. The real question is at what point it is still worth satisfying a small set of users. Those decisions were not his. Fridman's reply is about reputation. VS Code lets you customize everything; the Start menu should not feel stingier. Love in the details, including details only a few thousand people notice, compounds into a place that takes care of power users, which is also a hiring advantage. Plummer's guess is scheduling. When they rewrote the Start menu, cutting left and top and right and two rows of tabs may have pulled delivery in by four months. He is not saying that was the right call. He is saying that is the kind of pressure that produces it. Craftsmanship, in his view, needs someone to own a component long enough for it to settle. If the UI churns every release, polish never arrives.

ZIP folders were another home project with a picture taped to the monitor. In 1993 he and his wife drove around looking at houses. One had three bedrooms, a torch-red '93 Corvette convertible in the driveway, and a $300,000 price he could not approach, even for a down payment. He cut the photo from the flyer and stuck it on the screen. At night he wanted two things the bug-fixing day job did not give him: creativity, and extra money. From an MSDN, then MSJ, sample on bringing up a folder, he added ZIP support incrementally, sold it as shareware at $19.95 or $29.95, and moved a few hundred to a few thousand copies. One morning a woman named Betsy called from Microsoft and asked if he was the Dave Plummer who wrote Visual ZIP; they wanted to talk about an acquisition. He asked what building she was in. It took a few minutes of talk about travel and legal before both of them realized she did not know he already worked there. They had cold-called the author. He took the offer because the alternatives were quitting to keep selling, or stopping the product to keep the job. Then he bought a used '93 red Corvette.

Inside Windows they stripped features to simplify. Encryption came out because algorithms still counted as munitions. Multi-volume support came out to reduce complexity. The format itself came from the BBS era, when people dialed in for trialware and needed compression; executables often shrank by about half. Phil Katz's PKZIP for MS-DOS became ubiquitous, so by the time Windows shipped there were a decade of ZIP files and no way to open them. That code is still in the system. It was written in the single-core days and does not multithread. On a 96-core 7995, unzipping uses one core.

The blue screen of death, he explains, is the kernel with no remaining legal option: a driver writing memory it does not own, or freeing the same block twice. The machine goes down to prevent further damage. Windows 3 had a blue screen unrelated to NT's. He asked Jon Viert, who wrote NT's, why white on blue. The folklore was that you could walk a lab of fifty Stress machines and spot the crashed ones. The real reason was plainer. MIPS firmware was blue on white, Visual SlickEdit was the same scheme, and you could code, boot, crash, and reboot in one palette. Reboots fix so much, he thinks, because memory leaks accumulate and because code wanders into states nobody tested; a restart is a fresh heap and no leftover goofiness. The most beautiful user-mode code he remembers reviewing was Bob Day's named-pipe replacement for the shared memory Windows 95 shells had used. The debugging performance that set his template walked in six weeks after he started, held a null-modem cable, blasted through WinDbg call stacks, said "Oh, I see," and left. That was Laura Butler, later a distinguished engineer. Fridman asks how many people Task Manager has touched. Billions, Plummer says. About two billion a month.

Asked how his programming has changed, Plummer says he now spends the care up front. He used to write and write and clean up later; too many projects, his and other people's, became unfixable. He still leans on C++ and assembly, though he writes two dialects: modern C++20 with no raw pointers or C strings, as safe as Rust in his view, or C with classes. The current detour is Lua and Python, for an AI that plays Atari Tempest, a 1980 3D vector game on which he holds the world record. He pulled the ROMs, reverse-engineered zero-page variables, and wrote Lua to stream features over a socket to Python running dueling Deep-Q. It reaches about level 36 of 96, better than most humans, not yet better than him. That is the goal.

After Microsoft he built a lot of software, including a company called Software Online that got into trouble for nagging users to upgrade. Other than family health scares, it was the scariest time of his life; the run-up was one of the most exciting. Banner ads on a site he remembers as Fast Click, no cap on the credit card, $10,000 spent over a weekend, $38,000 in sales. Scale that, he thought, and he did, for years. The attorney general arrived with about twelve claims. Ten, to him as an autistic person, were obviously false, which made it hard to even hear the two that might be true. A 30-day trial that reminded you once a day after expiry was, the AG said, too often, amounting to spam. They agreed to about once a week and a one-click uninstall. The other issue was a checked-by-default disc for $3.95 or $4.95, because in 2001–2003 people who bought a download still expected media in the mail, and skipping it produced cancellations. In Washington State that default is negative-option billing. He would be more careful now. David Ogilvy's Beetle "Lemon" ad was the pattern he tried to follow — clickbait, then information — and he aimed at greed (a faster PC) rather than fear. Other people, he has learned, bring different background knowledge and may hear a scare he did not intend.

The nagware lesson folds into a complaint both men have about Windows. Fridman does not want to log in with a Microsoft account; the installer makes skipping feel like being tricked, extra confirms and all. Plummer's own trap is "recommended settings" on an upgrade, which sounds like the thing you should pick and, he is pretty sure, resets you to Edge. Recommended by them, not for you. "You don't want your operating system to be an adversary," he says, and sometimes Windows feels as if it does not have your interests at heart. He would rather pay $199 a year for Windows Pro, or twenty dollars a month, and not be upsold or have his data monetized. Ranked as products, he is a fan of Windows 2000 Server — he ran businesses on it, salons VPNed together onto SQL Server. For consumers, longevity and completeness, he and Fridman both put XP at the top. Fridman wants the code-smell test: a design philosophy that is consistent, not dumbed down, not overcomplicated, not crashing. Plummer's word is crisp. Not a lot of rough edges. Snappy.

He plays slot machines even though he knows the bet loses, because the lights and high-contrast colors are a dopamine feast. Internally, a black box draws the next random number and the payout; the game then invents a movie to match. The reels do not land and then get scored. The outcome is decided first. He could not get anyone to talk about the code. They are basic Windows PCs on a secure enclave he does not know much about. In the 1970s or 1980s, a Vegas technician burned his own ROMs with a backdoor, swapped them in on service calls, and came back months later to collect. Plummer's other special interest lately is old hardware: 1960s muscle cars, and PDP-11s. He spent years gathering parts, then tried to max one out the way you max out a PC. Four megabytes of memory, astronomical for the era, a PDP-11/70 panel, two chassis, knobs to choose which bus the LEDs show, switches to halt and deposit into memory. The lights needed a rewritten BSD kernel. Add a disk controller, rebuild the kernel, find the driver in the source. He is pretty familiar with that kernel now.

The core theory of autistic thought, Plummer says, is monotropism. His brain does one thing, very intensely, and when it is done it can move on. He is not a multitasker. He is a serial single-tasker. Autism usually brings sensory sensitivities and repetitive behaviors; when a person cannot moderate those in daily life, it becomes a disorder, maybe one to two percent of the population. The biggest benefit is the amount of focus and dedication he can bring, if the task is something he loves, something rewarding, something he can make progress on. It is the same feeling as a kid with trains. The catch is ADHD. Acquiring the focus is the hard part. He falls asleep in noise-canceling headphones, or he cannot fall asleep. Once he is locked in, he is very hard to distract. A high threshold, and a room that is hard to leave.

The biggest challenge is that he does not know what anyone else is thinking. He knows what he would think in your place. That is the best he can do. Neurotypical people, he believes, get a running sense from posture, reaction, and face. That channel is missing, so he runs "a little proxy NPC game" for everybody he deals with. Telephone is especially bad, because only voice is left. FaceTime is better: after a joke, the other person might smile. On a phone he cannot tell sarcasm from seriousness. Literalism is part of the same package. His grandfather, building a kitchen planter, said the brackets would hold a horse. Five-year-old Dave was confused about why you would bring a horse into the kitchen and put it on a planter. Figures of speech were, for a long time, just speech.

The Bill Gates story is the social cost in miniature. His first summer, Gates had the interns — twenty, maybe twenty-five — over for burgers and beer. Ben took him across to introduce him: intern from Canada, in four months he had done these features, plus a copy of SmartDrive, and then the list. Plummer interrupted. It was three months. Both men looked at each other. He realized, too late, that this was the wrong moment to correct a guy. Little inaccuracies still drive him crazy. He is better than he used to be. He did not used to know that injecting a meaningless correction could land on the other person. Now he has a better sense of it.

Advice for flourishing, at least financially: sell what you can do, not yourself. A personality audition may or may not go well. A portfolio — GitHub history, projects, the actual algorithm — travels. Large-team software still has a social component, and that was a liability, especially once he managed people. He assumed everyone wanted to be motivated, led, and paid exactly as he would. Some people need affirmation, some want money, some want the important meetings. He had to learn to ask, because he is not good at the nudge. He also offers a rough two-question screen. What matters more to society, cooperation or creativity? Neurotypical people tend to pick cooperation; people on the spectrum tend to pick creativity, the value of the individual solving the problem. For more precision: a room has ten chairs, six people sit down, how many chairs are left? Most people say four. He says ten, because the chairs are still there. He is not being a dick, he says. That is how his mind works.

Emotional post-processing is the substitute for reading a room in real time. After meeting someone new — including Fridman, that day — he goes home and replays the moments where he had a choice, worst ones first. What did he miss? What was the other person thinking? How does he handle that next time, and does he need to call and fix it? Twenty years ago he had his father's 1969 car restored. Craftsmen spent three years and thousands of hours. He inspected, said this was nice and that was great, wrote the check, flew home. Ten years later it hit him that he should have told them what a great job they did. He had not, because he did not need those words in the moment and had not imagined that they might. He emailed the shop owner: you may remember me looking unimpressed; I was not. The owner wrote back that he had thought of that moment often.

That delay is a key to how he talks about empathy. There is a perception that autistic people lack it. Speaking only for himself, he feels fairly empathetic. The problem is communication, in both directions. He does not know how you feel, so it is hard to meet you there until you say what you are experiencing. Once he knows, he can feel incredibly sorry. Until then he assumes you will handle it as he would. Advice for the other side is accordingly plain. Be specific. Do not assume he will pick up clues, nuance, or subtlety. If you are trying to nudge him into a behavior, say, "Dave, this is what you need to do." Fridman asks, on the air, whether he has failed in any way today. Not yet. Score out of ten? 7.5. Floating point.

Masking, as he uses the word, is the act of acting normal. Every wave, expression, tone, and posture in a social interaction is a contrivance, and it is work. It comes naturally to most people; cool people do it really well. On the spectrum you fake it all. He points at Rush's "Limelight," written by Neil Peart: the world as a stage, performers and portrayers, not treating strangers as friends, faking an affect. He only speculates about people who have passed on, but he reads the song as someone struggling with the same labor. A meltdown, by contrast, is not a mood. It is a switch. Emotion overwhelms the ability to manage behavior — Rain Man at the airport, or a TV surgeon banned from the OR. Fridman calls it a blue screen for the brain. Plummer says kind of. The frontal cortex shuts down to an extent, filtering goes, and you are in a reptilian panic, fight or flight against a present you cannot tolerate. Usually it is accumulation: a bad travel day, an argument at work, a randomized schedule, then one more thing.

He has been married thirty-one years and together thirty-seven. The first sign they were different was sitting in the car at night, looking at brick pillars linked by anchor chains. He wondered whether the two-foot-square posts were hollow or backfilled with concrete. She asked what was wrong with him, why he had a place in his head that cared about that. He has never been good with people, and he does not always notice when someone likes him. He bought his son a shirt that says if you are hitting on me, be specific, because I am clueless. His only option was to be himself, which works for some people. In the last few years he and his wife ping each other: "You good?" If he only says yeah, she knows something is up. There is no reading it off a face, so they built an explicit check. Plenty of people, he thinks, feel "a little bit autistic" — maybe ten to twenty percent on the spectrum without it rising to a disorder. *Secrets of the Autistic Millionaire* was written mostly as a manual for his kids, who do not have ASD diagnoses as far as he knows, but who carry related traits. Why so many great programmers sit on that spectrum is, to him, the same three habits: single-minded focus, reducing a problem, and an obsession with what is inside. He took his mother's oven apart to see how the flip clock worked. That, he thinks, is a good habit if you are going to write software.

GitHub Primes is the same prime sieve, expressed in about a hundred languages, with rules strict enough to make the comparison fair. At most one bit per integer, no byte-sized cheat; allocate inside the timed loop; each implementation in its own Docker container; all cores; as many passes to 100 million as a five-second window allows. 6502 and other cramped machines run as exhibition pieces because 64K cannot hold the sieve. He started with Python, C#, and C++, published a video, and then other people piled in until two maintainers in Europe, Rucker and Tudor, took it over. Zig is on top as he speaks. Someone improves the C++, it leads for a while, the Zig people get angry and come back. It is a relay with no finish line. The machine that runs the nightly bench has a terabyte of RAM. The happiest moment in programming, for him, is still when a complex thing he pictured in his head actually runs that way. He never trusts it the first time. He assumes he missed something.

On the Python side of Tempest he has done a ton of LLM generation, because he is new to the language. Watching what the model writes has taught him a lot. His own from-scratch Python would be about four times as long, because good Python is terse. He is, he admits, vibe coding the project. Fridman presses the condition that this only works if you already know how to program. Plummer agrees. You cannot vibe-code yourself into a working system from zero. If you are already a good programmer, AI can make you incredibly powerful, because you can read what it emits and know what to tell it next. The future he sketches is less prompt engineer than architect. Bridge builders once welded beams; now they drag pre-formed sections around in AutoCAD. Programming will look like that: moving components and interfaces, describing interactions, letting the model build the parts. We are still a long way from "give me a Linux-compatible kernel" and having it crank one out.

Does climbing the abstractions that fast make him sad, from machine code through assembly and C++ to natural language? He thinks he came up at a fortunate time. Technology stacked over thirty or forty years, so he understands TTL logic and can also use AI to write code, and he more or less knows the pieces in between. There are holes. The completeness of the picture, he suspects, is only available to people who lived it. Duplicating that for someone starting now will be hard. Fridman's last question is the large one: the meaning of the whole thing. Plummer does not wander. "Making cool stuff." What he cares about is making complex things other people can use, which lets him be creative in a way painting or sculpture never would. Then maybe raise a couple of good kids and hand the baton off. Fridman closes with Bjarne Stroustrup: there are only two kinds of languages, the ones people complain about and the ones nobody uses. It fits. Task Manager, activation, and ZIP folders have all been complained about. They have also been used, by a crowd large enough to show up in monthly metrics with a B.