The best programming language is the one you already use

John Carmack's first program was not Hello World. In a Radio Shack, on a TRS-80, he typed `10 PRINT "JOHN CARMACK"` and learned `GOTO 10` so the name would scroll off the screen. He had no contact with Unix culture; Kernighan and Ritchie had not yet standardized the ritual. What hooked him was simpler and stranger: a machine that would do exactly what you told it, with a genie's monkey-paw catch that you had better be careful what you asked.

Asked for the best language of his career, he gives caveats instead of a crown. He writes a fair amount of Python for machine-learning work, but he did not grow up in it and does not dream in it. A triply nested Python loop, he says, can be thousands to a million times slower than a proper GPU tensor operation — numbers so large they keep hitting him "upside the face." The first reflex is always to reshape the work into one big vector that a C++ library can swallow.

Serious programming still happens in C++, and a C-flavored C++ at that. He keeps his distance from modern template metaprogramming, having watched over-abstraction produce train wrecks. Years in Lisp and Haskell changed how he writes C: less random mutable state that you lose track of. The hard part is not the week you write the code. It is whether the program can bend as the world changes, and whether a revolving door of maintainers can pick it up. On that score C has an underappreciated virtue. It is so simple that anyone who knows it can jump into OpenBSD's kernel and understand what is happening without first learning a private dialect.

Go, he notes, was mocked as crude when it shipped; it caught on for the same reason C did. The best language, he concludes, is generally the one you are currently using. In almost every mixed-language project he has seen, the mix was a mistake. At Meta that means JavaScript for React, C++ for the real work, and Java against Android — "horrible things." He once told Andrew Bosworth he wished the company would hire only C++ programmers. Bosworth's reply was that they could not find enough of them.

What makes a good modern programmer? Carmack's answer, delivered as a lecture to Meta's TPM organization, is blunt. Everything should flow from user value. Code golf is a fun puzzle; it should not be the motive. You are solving a problem or providing entertainment that displaces something else in someone's life. If you have given people more value than it cost you and your team to create, the world is better. If you have shoveled cash into a wood chipper, you should not feel so good about the architecture.

Pride in a specific sequence of instructions is allowed a tiny dopamine hit. The top-level metric is whether someone chose your product and their life felt better. At companies with nearly infinite money — Meta, Google, Apple, Microsoft, Amazon — the trap is refusing to choose. Feature A or feature B cannot both ship just because both seem good. Without a fierce ranking you are not making a value judgment; you are avoiding the job of telling team B that A matters more.

When data exists, listen to it. He has watched a fancy telemetry system cause the performance regression it was supposed to diagnose. When you are doing something truly new you have to guess, but Meta already has more than enough VR users for any experiment someone wants to run. Do not invent Alice, Bob, and Charlie. Talk to the thousands of people who already own headsets. A passionate leader who can hold the whole product in their head is valuable — Steve Jobs stories include him being technically wrong in the room — and so are independent teams that surface hits the dark cabal would never have approved. In VR, the applications nobody predicted were the ones that worked.

Carmack was never an all-nighter programmer. His brain, he says, turns to mush after about twelve hours. What he did instead, for decades, was work sixty hours a week: ten-hour days, six days a week, and try to be productive at that. Without kids he ran a cycling schedule, starting an hour later each day, sometimes arriving at two in the afternoon and working past midnight. School drop-off yanked him onto a civil clock. He still experiments on himself. A four-mile walk at Texas sunset, plus a shower, was punching a hole in the day; he moved the miles to morning so the night hours would stay intact.

He always tries to get eight hours of sleep. You can work a hundred hours a week and still sleep eight if you prioritize. He gets into online arguments about this constantly. People who say you are less productive past forty hours, he thinks, misread the data: marginal output after eight hours is lower than at the peak, but you are not literally getting less done. The point where you start breaking things is not eight or ten or twelve hours. His fictional test is an asteroid on a collision course. Do you want the people at SpaceX building the interceptor to clock out at five?

He has never felt burnout. He can turn his gaze to another interesting problem and have a great time. The ability to work hard, for him, is the ability to have several kinds of work ready — a research paper that needs a sharp hour, notes that need filing, a few hundred lines to bang out — and to pick the one that fits the state of his brain. Nutrition was never a religion. As a broke teenager he hated not being able to buy pizza whenever he wanted; once he could, he did. At id Software a pizza arrived every day. The delivery driver knew his name, and years later Carmack learned they had never raised the price. The real ritual is Diet Coke: drop the can, crack it, now we mean business. Eight or nine a day, still.

Game development and Silicon Valley, Carmack says, are different cultures, and the valley's hostility to debuggers still boggles him. Anyone who thinks you should just read the code and think about it is making an insane statement. You cannot even read all the code on a big system. You have to do experiments. Adding a log line, recompiling, and rerunning is an incredibly inefficient way to do that. When he writes C or C++, he always runs from the debugger. The first act after typing a function is to set a breakpoint and single-step it. Other people say they do that in their heads. "Your head is a faulty interpreter."

Id's codebase had a reputation for robustness. Microsoft told them Xbox crash counts were so low they were probably hardware errors. Then Carmack put every static analyzer he could find through the whole couple of million lines. It was humbling. Copy-paste that never changed a type. Printf format strings with the wrong data type. Missing null checks. Bugs in his own code that had been sitting there a year, submarines waiting for someone to step on them. His conclusion: anything syntactically allowed will eventually appear in a large enough codebase. Good intentions will not stop it. You need automated tools.

He likes asserts as active comments — they do nothing if the world is as you expected — and he often prefers a fixed-size array with an assert on the limit to a container that grows forever. If usage jumps ten times past what you imagined, you want to be told, not to suffer a silent quadratic slowdown. On the communications side, early id ran NeXT workstations. Unix had a `.plan` file that the `finger` command would dump to anyone on the internet. Carmack started keeping a to-do list in his, starring items as he finished them, then dropping in notes and short essays. Most Doom players had no Unix box, so websites scraped the files. A couple dozen Dallas developers were publishing the same way, years before blogging had a name.

Before id, Carmack was a hardscrabble contractor in Kansas City, bouncing between rent and groceries. The company that sent the most work was Softdisk Publishing, whose business model now sounds bizarre: a monthly software subscription. You paid, they mailed a disk of random programs. He would write a tiny game, sell it for three to five hundred dollars, and port the same title from Apple II to IIGS to IBM PC to stay afloat. Jay Wilbur and Tom Hall kept asking him to move to Shreveport. He refused until he met two programmers who knew more cool stuff than he did: John Romero and Lane Roth. Romero had been at Origin Systems. For the first time, programming was not a library of out-of-date books.

Softdisk spun up Gamer's Edge, a PC-game subscription that demanded a new title every month, start to finish, no slipping. Romero had once trained by making a game for every letter of the alphabet. Carmack compares the stretch to the Beatles' Hamburg years: the forgotten gigs that built the hands. Resource limits, he says, are surprisingly important. Give a team all the time and money in the world and you do not always get something better. Necessity is still the mother. Meanwhile Scott Miller of Apogee was sending Romero fake fan mail from the same return address, trying to recruit them into shareware. They thought he was lying about the money. He was not.

What PCs could not do, and consoles could, was a huge scrolling world. Redrawing the whole screen got you five frames a second. Carmack's first trick, adaptive tile refresh, compared the current frame to the next and redrew only the tiles that changed — often about ten percent in a Mario-like level of sky and repeated brick. The worst case was a checkerboard: every tile changes, and you are back to five frames. Designers had to keep levels visually plain. The first memorable all-nighter of his life was spent with Tom Hall cloning Super Mario Bros.'s opening stage, swapping in their character Dangerous Dave. They left a disk on the desk for Romero and Wilbur. By afternoon the room had decided they were going to start a company.

Commander Keen was Tom Hall's baby: a kid with a zap gun and a homemade rocket, built to look like Mario under the tile-refresh limit. They shipped the shareware trilogy in about three months while still making monthly Softdisk games, reusing code in ways that later caused legal trouble. Keen immediately did three times the business of Apogee's previous hit, some thirty thousand dollars a month. In those same months Carmack found a better scroll. EGA cards had 256 kilobytes in four bit planes of 64K. You could tell the card to start scan-out anywhere inside that window. Everyone assumed you could pan a 2×2 screen and no further.

He asked what actually happens if you start at the end of the 64K block. It wrapped. You could scroll forever and draw only the newly exposed strip of tiles. No more plain skies; no more designer self-restraint. The bill arrived later. Super VGA cards with 512K or a megabyte did not wrap. They scrolled into real, uninitialized memory. On those machines the game hitched every so often. Rely on a hardware behavior nobody intended to support, and you accept the compatibility tax. Carmack calls it the two-edged sword of unconventional technology.

Steven Levy's *Hackers* had been formative. Carmack wanted to live in the third section, with the game hackers. The ethic, as he took it, was that information is not zero-sum, and that you can take joy in someone else's amazing thing. Id eventually released the source for its seminal titles. Programmers inside the company mostly got it; the broader industry, and many modders, were more possessive, more like artists guarding credit. He will not fight for primacy. He saw Battlezone and Star Wars in arcades; he stands on those shoulders. The inverse-square-root hack often attributed to him was not his. The internet, he says, cannot be convinced.

Wolfenstein 3D reused the enemy AI, door logic, and pickups from an earlier overhead game, *Catacombs*, a Gauntlet-like crawl. Hover Tank, Catacombs 3D, Keen, and Wolfenstein even shared a map editor. The change was tiny and total: move the camera from God's eye into the eyes. Carmack calls it the purest A/B test he has ever run. Artist Adrian Carmack had drawn a troll, picked up a key, walked to a wall. The wall played a little vanish animation. A monster was standing there. He practically fell out of his chair. Games had not done that. You could be happy or sad for your little guy. You were not inside. Something reptile in the brain said *oh shit*.

Catacombs 3D had shipped six weeks late, still glitchy on edge-on walls. For Wolfenstein, Carmack abandoned that half-working span renderer and went to ray casting in a small amount of assembly. Three hundred twenty times a frame, a ray walks from the player until it hits a wall. In a tight jail cell it is cheap — three or four tiles. On a 64×64 map, an empty room can force dozens of steps and the frame rate falls over. Characters were not 3D models. They were eight cardboard views, scaled. A general scaler would have been slowest when a monster filled the screen, exactly when you needed to react. He wrote a program that wrote a hundred little programs, each hardwired to draw an image at one height. At maximum scale it was one read and three or four writes — faster than a straight copy. The idea came from Apple II sprite work: pre-shifted ships, compiled shapes, data turned into code.

After Wolfenstein, Keen, and Spear of Destiny, Carmack could see a hard creative wall. Every artist on earth could not make a masterpiece inside Wolfenstein's blocky grid; they could only make more Wolfenstein. Doom added variable floor and ceiling heights, arbitrary wall angles, richer lighting. Those increments, he says, pushed the design space near Turing-completeness. People still publish new Doom maps. Minecraft made the same point from the other direction: shrink the blocks enough and the box is no longer too small. Crossing the threshold matters more than how fancy the renderer is.

The technical spine was a binary space partitioning tree, learned from Foley and van Dam and eased in on a Super Nintendo port of Wolfenstein. BSP carves walls by walls until any viewpoint has a definite front-to-back order. Paint the nearest first and stop when the screen is full. Arbitrary angles brought epsilon problems: two sloped lines meet at a rational the machine cannot represent, so vertices snap and seams open. He wasted time chasing exact rationals; they explode. Doom still has a community name for the leftover case — the hall of mirrors, two frames of garbage when a surface fails to draw. Ken Silverman's Build engine, used in Duke Nukem 3D, did none of this. There is never only one true way.

Doom was also the first id game designed to be modified without hacking the executable. WAD and PWAD files overlay new data on a clean copy of the game. It still was not a fully general 3D engine. You could not build a pyramid or a bridge over a passage. You could not look up. Four degrees of freedom, not six. That unfinished list is what Quake tried to swallow in one bite: true six-degree-of-freedom rendering, a new surface-cache lighting model, QuakeC so game logic was data, and internet client-server instead of peer-to-peer. If he could go back, he would have put TCP/IP and a scripting language on the Doom renderer first, shipped a Doom-looking, Quake-feeling game eight months earlier, and given Pentium-class hardware time to become normal. A 486 could run Quake. It was not a great time.

Quake was the first project where Carmack felt he gave it his all and still came up a little short on the when. The company ran hot. QuakeC taught a generation of game programmers and is, he says, nothing to write home about. You do not fall in love with a language's BNF. You fall in love because you can open the shotgun, set damage to 200, and walk around with a grin. He later decided a C interpreter would have been the smarter host. Michael Abrash, whose *Dr. Dobb's Journal* columns Carmack had treasured at Softdisk, left a senior role at Microsoft to write the inner assembly loops. The working pattern was awkward in a useful way: Carmack would rough out a span rasterizer; Abrash would spend a month making it beautiful; then Carmack would invent a span buffer that cut 30 or 40 percent of the pixels and the interface had to be rewritten. He knew it wore on Abrash. Globally it was right.

His real leverage, he argues, is not the tightest SIMD. It is seeing an angle where you trade away something you thought you needed and unlock a much larger gain. Knowing the whole stack — chip, compiler, architecture, people — is how you notice the trade. Doom was smoke and mirrors people mistook for a more general engine. They picked the right mirrors. He pushed Romero out of the company in those years, a decision he now ties to a naive equal-partnership structure and a buy-sell agreement designed to keep outsiders out. He was in his twenties, angry that not everyone wanted to grind at his level forever. Romero landed on his feet, raised some twenty million from Eidos for Ion Storm, and later was early to mobile. They still talk. Carmack was glad to see Tom Hall shipping a VR RPG, Demeo — the same Dungeons & Dragons they used to play on Sundays, with Carmack as dungeon master.

Carmack wants the metaverse built the way the web actually won: make something people love, then expand it. Minecraft and Fortnite, not VRML, are the existence proofs. Capability-first worlds such as Horizon Worlds are harder to love than a great game that slowly grows social persistence. History will name firsts; he thinks those names will be a poor map of what was actually a pile of overlapping S-curves. The modern VR era, in his telling, starts with his E3 demo of Doom 3 on the Rift prototype. Everyone assumed first-person shooters were a match made in heaven. They have some of the worst comfort problems. The most popular app became Beat Saber, which nobody predicted.

In hindsight the engineering is obvious. He had underestimated controllers. A gamepad plus head tracking felt amazing until you tried to pick up a pair of glasses: hover, press grip, watch them snap into your hand — an action that keeps reminding you the world is fake. Beat Saber never lets go of a physical handle. A saber is supposed to slice through things, so there is no missing collision. Music isolates you and rebuilds space in audio, VR's natural strength. Sessions last three or four minutes, which fits headset comfort, mobile batteries, and the fact that you stand still. Fitness was an accidental bonus. His slogan, widely read as dystopian, is that it should be better inside the headset than outside. Not everyone can have Richard Branson's island. Everyone can have a VR one.

Workrooms on Quest already beats Zoom on voice latency and gives a hint of presence. The product is still five times too many clicks. Replacing Zoom means everyone has to own one, which means cheap and light — a virtuous cycle he has always favored over feature-heavy PC headsets. Reality Labs' ten billion dollars a year still makes him sick to his stomach, even though Quest 2 is roughly the standalone, inside-out-tracked, 4K, PC-tetherable headset he wanted a decade ago. Small teams do analogous work with a tenth of the people. He will not believe a giant company is within a factor of two or four of the efficiency it could hit. The people he treasures are the few, among thousands, willing to hop into a kernel, a distro, or a firmware image to make one thing right.

Carmack and Elon Musk are almost the same age, with similar early computer books and a shared habit of calling bullshit on the same waste. He met Musk at the start of Armadillo Aerospace, talking propellants and flight computers. Years later he realized his own resources were roughly what Musk had when he went all-in on SpaceX and Tesla — a bet that could have made him the cautionary rich-guy story. Carmack had kept a finite "John's crazy money" box and a software day job. He has a ten-thousand-dollar bet that boots will not be on Mars by 2030, though infrastructure might; the room of brilliant people at an Intel dinner folded once real money was on the line. Self-driving by 2030 is a different bet, made with Jeff Atwood, because a dozen companies are pushing, not one man's passion. Nuclear fission, he thinks, looks like aerospace before SpaceX: the physics did not get harder. Fuel is a couple percent of the electricity price. A plant one-fifth as efficient and ten times cheaper would still be a breakthrough. Fusion is an absurd amount of machinery to make something hot. Fission is two rocks that get hot by themselves.

After VR he chose between cheap fission and AGI. Fission meant politics, a decent-sized team, and cross-functional work he does not love. AGI looked like the highest-leverage moment a single person might ever get. He is not a madman, he says, for claiming the code is likely tens of thousands of lines, not millions — something one individual could write, unlike a browser or an OS. The remaining unknowns are probably simple. His bet is fewer than six key insights, each writable on the back of an envelope. Combined with GPUs at scale and the data everyone already has, that could make something that behaves like a living creature. In 2022 he gave "signs of life" by 2030 a 50 percent chance. Talking to Fridman he had nudged it to 55 or 60. The antecedents, he suspects, are already buried in the literature, including clever 1990s work abandoned when it was tried with twelve neurons.

Signs of life means a learning-disabled toddler: a system that can engage a favorite show and a game, and learn in some recognizably human way. At that point you can deploy engineers, cognitive scientists, and educators, with rollbacks and A/B tests no real childhood allows. He does not think a robot body is required to get there. Simulated worlds were his career; tying yourself to real-time physics is a weight on the ankle. Reality, aerospace taught him, is a bitch. He does want real-time operation — a Zoom meeting, not a typing bot — because a stateful conversation will convince people faster than a humanoid, even though he does not much care about convincing people. Fast takeoff, in his view, is not credible. A human-level system starts on thousands of GPUs. You cannot casually rent that. Data centers take years. You cannot open TCP connections above a certain rate no matter how smart you are, and a model carved for one chip and interconnect will not just hop onto the open internet. The magic might fit on a thumb drive. The thief still has to build the data center and the childhood.

Most arguments about consciousness, he thinks, have little merit. He favors a pandemonium picture: sparse, slightly randomized memory fragments drifting until consensus bubbles up as a thought. The stochasticity is imagination. Death as a deadline is an interesting angle and probably not load-bearing; we will set the reward signals, and one of those six insights might simply be what the master reward has to be. Ethics and safety debates can wait until the toddler exists. He had tried the Victorian gentleman-scientist route, then noticed he was not spinning up cloud GPUs for no good psychological reason. He bought a quarter-million-dollar DGX so the idle fans would mock him, then took investor money because he feels responsibility for other people's cash. Andrej Karpathy, whose Stanford course and hacker's-guide posts were part of Carmack's larval phase, is the programmer's programmer in the field — the person he asked, among a hundred others, to come work with him.

Carmack was a competent club player in judo and gi grappling, never a star, often the old man on the mat in his mid-thirties. It mixed him with people who had nothing to do with games; once in a while someone would say "you're the Doom guy." He hung it up around forty after a separated rib, still debates going back, and has been tempted by a small private club Carlos Machado floated after a Joe Rogan appearance. The deeper gift, he and Fridman agree, is getting humbled. Nobody who trains thinks they are the best in the world, because everybody taps. That lesson travels to work, where it is harder to get humbled.

Invited to give a commencement speech at the college he dropped out of after two semesters, he told the room his path is not for everyone. Depth — knowing a system through and through so you can see which inefficiencies can be bypassed — is a minority taste. Most engineers need to do the parcel of work they were handed, reliably, so people can depend on them. For the ones who want the layers, the world has never been better. Embrace the grind. Prepare a broad set of tools, then watch for the place where current practice is not what is possible. You cannot know everything. You should convince yourself you can know any particular thing, because people made it. He calls that weaponized curiosity. He thinks he could have been absorbed by operating systems or databases the same way he was absorbed by games.

He almost never thinks about aging, death, or legacy. The one way finiteness bites is in the size of the problems he picks. Fission and AGI are swings that could affect a large fraction of the world, and he figures he may have a couple of full-strength at-bats left. After that, recreational retro programming — the old man gardening in code worlds. The meaning of life does not worry him. Humans are biological creatures with drives evolution happened to install. People who will not look at what is in front of them, he has found, tend to be less effective. The magic is gradient descent: the smartest people in the world spent decades refusing to believe that tiny local steps, not a grand vision, lead to the best answers.