Who owns the off switch? George Hotz on tinygrad, open-source AI, and God

Lex Fridman opens the third conversation with a stream remark of Hotz's about time. Hotz treats time the way he treats quantum mechanics: as a useful model, whether or not it is "real." Some things he will grant as real — Kolmogorov complexity, mathematics — and he does not think P equals NP, though he suspects factoring sits in P. The rest, including an objective reality under the models, he shrugs at. "I don't think it matters," he says, which is less mysticism than an engineer's refusal to spend cycles on a knob he cannot turn.

He distinguishes real difficulty from the artificial kind. Artificial difficulty is a knob you can turn off, or smash with a hammer; real difficulty is being in the woods and having to survive. The hiker who leaves the map at home, he says, is not an explorer — he has only smashed the knob. Throwing that knob away as a romantic gesture, after humans have already built the world, strikes him as horrifying. A large language model can be given the perfect amount of crankiness and jealousy, but once the difficulty is a slider, it is no longer the thing that makes a relationship worthwhile.

That habit of thought lands on a sharper claim. There is no clean line, Hotz says, between a person and the rest of nature. "I don't think I'm conscious. I don't think I'm anything. I think I'm just a computer program." He endorses the extended Church–Turing thesis: everything running in the universe is computation, including whatever is running in his head. When Fridman presses on the stubborn fact of a body that stores memory across years, Hotz calls it a quirk of evolution, not a special spark. Cross-entropy on the next token will not get you to AGI; you will need reinforcement learning in fancy environments, because the human loss function is nowhere near as simple as the one we train on.

Once models are superhuman, Hotz expects memes that are actually scary. He reaches for *Infinite Jest*: not one static tape you cannot stop watching, because human brains are too complex to trap that way, but a generative TikTok you cannot look away from. Today's feed is already close. Humans still generate the clips; the algorithm only ranks them. Quantify that generation as one humanity's worth of flops, he says, then hand generation to a hundred humanities, and you are done. Fridman asks whether a superintelligence might simply move past us, like water. Hotz cuts that hope off. The machine is a machine. Other humans using the machine are the ones who will manipulate you.

Fridman brings up Eliezer Yudkowsky, who had just sat in the same chair and argued that AI will almost surely kill everyone. Hotz says yes, maybe for a different reason. Nuclear weapons did not end the species, he thinks, because they are hard to use tactically: you get an irradiated pile of rubble, not subjugation and land. AI is the little red button. It is useful in small, selfish, everyday ways, so someone will press it. The most obvious ending he can picture is wireheading — amusing ourselves to death, staring at that infinite TikTok and forgetting to eat, or simply stopping reproduction. He does not think that actually extinguishes *Homo sapiens*. It kills what we call society. A remnant with taboos against chips rebuilds, perhaps looking a little Amish.

Under that forecast sits a distinction he returns to all night. There are two stacks of life. The biological stack starts with reproduction: the first proto-RNA could copy itself, and a bird is still an incredibly robust organism with no fab and no repair shop. The silicon stack, for all of Moore's Law, cannot reproduce. If the chips have a fab on board, the machine becomes enormous; if they do not, they die when civilization blinks. Superintelligence, he thinks, will arrive long before robustness. If the AI apocalypse happens, the AIs probably die out too. That is why the B-movie of a model writing viruses, spreading itself, and taking the grid does not impress him. The old script is enough: a particular person gets a very good machine.

Hotz loves virtual reality and has been ready to live there for a long time. Judging by what you can buy, he says, we are still far away: no headset has yet fooled him the way good headphones can fake a jungle. People want to believe so badly, though, that they already think large language models are conscious. He does not. If consciousness is a spectrum, he is far more conscious than an LLM, and he puts today's models below a chicken — he has seen living ones walking around Miami. Draw a sad face on a rock with a Sharpie and humans will think the rock is sad. Looking in the mirror, he suggests, we do a version of the same trick. You can recognize a suffering dog because you share a stack with the dog. A 404 from a web server does not hurt in any way you have an intuition for.

He still wants AI friends, and he says so without the usual cynicism. Current GPT and Copilot systems look to him like junior engineers and Fiverr-level copywriters — a start, not a ceiling. He cannot wait for friends more intelligent than he is. Whether talking to a model counts as emotional cheating is, in his view, a definition each couple has to make, the way couples already disagree about porn. One of his own chatbot scenarios is a twenty-year-old named Lexi who moved to Los Angeles to act and started an OnlyFans instead; you are on a date with her. Enjoy. The culture around that will get weirder, including the day you realize a Twitter account you have talked to for years was never a person.

He has a name for a related product: heaven banning. Shadow banning means you post and no one sees it. Heaven banning means you post, no one sees it, and a crowd of AIs spins up to interact with you. He points to the fanfic *My Little Pony: Friendship is Optimal* as a science-fiction treatment of a world that cares for you so well you never leave. His third company, he says, will be AI girlfriends, and he means it. He even floats a "girlfriend Turing test," then shrugs at the idea of a formal paper: a company does not have to win everyone, only a large enough clientele. Boyfriends, he argues later, are harder to fake because status and power are zero-sum and machines do not have them, while men are also, in his blunt market read, more desperate buyers. Fridman is not convinced status cannot be narrated. They leave the disagreement on the table.

Before company three, Fridman asks about company two. Hotz started tinygrad as a toy, the way Andrej Karpathy started micrograd: to teach himself what a convolution is, what options you can pass it, what its derivative looks like. Then he started thinking about AI chips. If NVIDIA becomes a monopoly, how long before it is nationalized? He says he likes NVIDIA. The point is not to humble Jensen Huang. The point is to keep computational power from concentrating. If NVIDIA is ten times better than everything else, you have handed a huge advantage to whoever can lock up that resource. If Jensen is listening, Hotz adds, he may want to consider how not to get nationalized. Google will rent you TPUs. You cannot buy one at Best Buy.

His first ideas about beating GPUs were, he says, completely wrong. The description that stuck came from Jim Keller on an earlier Fridman show. CPUs can add, multiply, load, store, compare, and branch, and they can do all of it fast, spending oceans of transistors on branch prediction, speculative execution, and deep reorder buffers. GPUs can barely compare and branch, but they can do arbitrary loads and stores — fetch from an address that itself came from data. DSPs mostly add and multiply, with loads and stores known before the program runs. Look at neural networks today, Hotz says, and about 95 percent of them are the DSP paradigm: statically scheduled adds and multiplies.

Tinygrad is an attempt to push that fact all the way through the stack. Python is Turing complete, then C++, then CUDA kernels, LLVM, PTX, SASS, on a Turing-complete processor. He wants Turing completeness out of every layer, because once it is gone you can reason about the machine. Branch predictors on CPUs are roughly 99 percent accurate; the 1 percent they miss is the halting problem in miniature. A neural net runs the identical compute every time. Only the data changes. Build the computer, and the software, that takes maximal advantage of that. The most visible consequence is almost rude: tinygrad does not even have matrix multiplication as a primitive.

Picture a cube, he says, with the two matrices on two faces. A matmul is shape ops that expand those matrices through the cube, a pointwise multiply, and a reduce onto the third face. Tinygrad has on the order of twenty operations; PyTorch has about two thousand kernels, PrimTorch brought that to about 250, and XLA sits in the same band. He calls the difference RISC versus CISC for the ML stack, then quotes Angelina Jolie's line from the 1995 film *Hackers*: "RISC architecture is gonna change everything." The twenty ops fall into four types. Unary ops take a tensor and return one of the same size — log, reciprocal, sine, ReLU, almost every activation. Binary ops take two equal tensors and emit one: add, multiply, divide, compare. Reduce ops drop a dimension, usually with sum or max. Movement ops do no arithmetic; they are reshapes, permutes, expands, flips.

From those four families you can build matmuls and every convolution you can name, including dilated, strided, and transposed. Laziness is the other trick. In PyTorch, `A * B + C` multiplies, writes the result to memory, reads it back, adds C, and writes again. Tinygrad waits until you actually realize the tensor, then fuses the multiply and the add into one kernel, the way Haskell waits to force a value. Set `DEBUG=2` and it prints every kernel that ran, with flops and bytes, which PyTorch on GPU will not tell you. The frontend looks Torch-like and can load ONNX; he has also ported LLaMA, Stable Diffusion, and Whisper by hand, mostly to learn the models, and he thinks those ports are cleaner. `nn.ReLU` as a class, in a stateless op, is the kind of clutter he is trying to delete.

The problem tinygrad is supposed to solve is not a prettier notebook. It is porting new accelerators quickly. Sequoia, he says, marked Graphcore to zero; Cerebras, Tenstorrent, Groq, and others built chips whose software was terrible, and he sees the same trap around Dojo. Writing a PyTorch port means writing and tuning on the order of 250 kernels. His prediction for Tenstorrent is a pivot to RISC-V CPUs, because AI accelerators are a software problem: if you cannot write a Torch-level stack on NVIDIA, down to the driver, you have no chance on your first, more specialized tape-out. AMD's 7900 XTX, when he tried it, panics the kernel if you loop the demo apps — primitive fuzzing. Lisa Su answered his email. Open source, he told them, is a culture, not a code dump.

His central thesis about the world is simple: things that centralize power are bad, things that decentralize it are good, and he would like to help with the second. The only ASIC he calls remotely successful is Google's TPU, and that is because Google wrote a machine-learning framework. He does not want a cloud, even from Meta. Asked for the fundamental limitation of cloud compute, he does not mention latency or price. "Fundamental limitation of cloud is who owns the off switch?" Advice on building a custom PC? You should not. You should buy a box from tiny corp.

The tinybox is $15,000, almost a petaflop, more than 100 GB of GPU RAM, more than 5 TB/s of GPU memory bandwidth, and four NVMe drives in RAID for something like 20 to 30 GB/s of read. Six GPUs, because AMD EPYC chips have 128 PCIe lanes and he wants leftover lanes for disks and networking. He is leaning toward 7900 XTXs but might ship 3090s or Intel A770s; the real constraint is not the logo on the card. It is a 120-volt, 15-amp household circuit. There is a plug on the wall, and the machine has to use it. A sibling product, the tiny rack, asks a related question: how much power can you get into a house without arousing suspicion? One answer is an electric-car charger, about 10,000 watts against a wall outlet's 1,500.

Quiet is the other hard spec. Sixty decibels will get the box kicked out of the room; he wants 40 to 45. The counterintuitive move is to make the chassis big — the name is a joke — so you can use huge, low-pressure fans instead of the jet-engine blowers in a 1U server. UPS still has to treat it as a normal package. Out of the box he wants Ubuntu, which beat Arch in a poll, and a first run of 65-billion-parameter LLaMA in FP16 at something like five to ten tokens per second. Preorders are $100 and fully refundable, stolen from Elon's playbook, and he does not intend a years-long Cybertruck wait. "It's a lot simpler than a truck."

On openpilot, tinygrad is already doing real work. About eight months before the taping it replaced Qualcomm's library on a Snapdragon 845's Adreno GPU and ran twice as fast, in part by treating mobile GPUs as texture machines rather than buffer machines. Almost no one talks about FSD anymore, Hotz says, and even fewer talk about openpilot, because he thinks the research problem is already posed correctly. The job is to build a model that, given a reasonable set of sensors, outputs a human policy for driving. Waymo and Cruise hand-code quasi-human policies. Tesla and, even more so, comma try to learn the policy from data.

At the beginning, comma published "Learning a Driving Simulator," an autoencoder plus an RNN that predicted the next compressed frame. By 2023 lights that 2015 stack is, in his words, a laughably bad simulator. The replacement is drive GPT: a VQ-VAE and transformers, trained self-supervised on driving video to predict the next frame, and — this is the part that makes it a simulator rather than a clip generator — conditioned on pose. Give it a state and an action, get the next state. Then you can run RL inside it. The reward is not "how close is this to a human trajectory." It is "would a human disengage if you did this." That question, he says, is the one that makes customers happy: a system you never want to take over. Some disengagements are just "I felt like driving," and those will look like noise.

They have not fully closed the loop. Prototypes exist, plus stupid bugs, plus a possible need for more scale. It might take a year. It might take ten. Comma has expanded the cluster to about 40 petaflops, which he translates as two people of compute, with 20 petaflops as one person — a unit he wants to sit beside horsepower. They have on the order of 5,000 daily active drivers. Tesla, he says without bitterness, is always one to two years ahead, and probably always will be, because Tesla is not doing anything wrong. Both companies are creeping toward end-to-end and both know disengagement is the negative reward. He still does not believe in the cloud for this: the comma 3 has about as much intelligence as a bee, a human's worth wants a tiny rack in the house at 0.1 milliseconds instead of 30 milliseconds to a cloud, and the windshield device is $1,499 with a 30-day money-back guarantee, built on an in-house SMT line in San Diego.

Fridman is impressed that transformers write language and code. Hotz is more impressed by MuZero and AlphaGo, which play beyond the best humans, than by models that write middle-school essays and mediocre code. Close-to-correct code is the scariest kind. He spends 5 percent of his time typing and 95 percent debugging; the last thing he wants is a machine that helps with the typing. Driving, he says, is not tool-complete: a 1980 car and a car today share pedals and a wheel, so you can drop an old driver into a new one. Programming is tool-complete. The languages, runtimes, and IDEs have changed so much that a JavaScript programmer at an Apple II BASIC prompt does not know how to insert a line between 17 and 18. If Copilot is helping you, that often means the framework is bad and full of boilerplate. Tinygrad, at the time, was 2,700 lines and could run LLaMA and Stable Diffusion.

He uses none of GPT for his own programming. At Twitter, ChatGPT hallucinated APIs that sounded real. He does not think that is a temporary edge-case problem for the way models currently store knowledge. Human brains have a hippocampus; today's LLMs stuff everything into the weights. He likes Meta's Atlas work and expects future models to be smaller, to loop on themselves, and to grow a retrieval system so they can cite sources. That, he says, is what people actually want from search, and when someone ships an LLM that cites its sources, it will kill Google. He is not trying to win that race with tiny corp. He said he would win self-driving when he started comma; he has never said he will win search, because he will not.

He would still bet that Google is not the number-one webpage by the end of the decade. Paul Graham's old test: startups are alive or dead. Google, in that sense, is dead. Mark Zuckerberg is alive — the rename to Meta was a man proving it. Sundar Pichai, Hotz thinks, does not have the authority to turn off google.com tomorrow. Zuckerberg probably has that button for Facebook. Elon has it for Twitter. There is no off switch for the AIs either, which is why he bets on whoever can still make a radical product decision. He would not be surprised if Bing overtook Google within five years, and he would bet equally on a startup that looks like Midjourney for search. Satya Nadella, he says later, is alive; Microsoft is alive.

On the editor he actually uses, the story is less dramatic. He left a plugin-free Vim — not even autocomplete — for VS Code, because VS Code's completion is right often enough in a cold codebase that he will take it. He loves mypy, whose false-positive rate he puts around 5 percent, against about 60 percent for a Microsoft type checker he tried. The assistant he actually wants is not tab-complete. It is a pair programmer that pops up and says there is a bug on line 14, and why. In two years he expects to use these plugins a little; in five, to rely on an AI-augmented flow; in ten, the human niche in programming is gone. Handmade code, he jokes, may sell on Etsy, with the slight imperfections only a person would leave.

LLaMA's release makes Zuckerberg, of all people, the good guy in Hotz's story. He lost about $80,000 on Meta stock the prior year and decided the dump of LLaMA made it worth it. Open-sourcing is, he says, the only way to fight a paperclip maximizer. One model that a small group controls and "aligns" is how you get paperclipped; you do not get there if everybody has an AI. He claims, as if it were hallway knowledge, that GPT-4 has 220 billion parameters and is a 16-way mixture with eight sets of weights, and that everyone at OpenAI knows it. What upsets him is not closed weights — weights can be the product — but refusing to state a parameter count on safety grounds. GPT-2's release, dressed as an AI-safety event, he calls a hype move, and he calls it disgusting.

Intelligence is dangerous, human or machine. A GPT-fueled bot farm can manipulate millions of people. Intelligence agencies already try; they are not the most sophisticated, they are the most patient. He cannot do what they do, which is sit on an exploit for five years. That is endurance, not cleverness. You can argue that nobody should have these systems, and he will respect that the way he respects "nobody should have guns." He will not respect "only trusted authorities should." He is not worried about alignment between an AI company and its machines. He is worried about alignment between himself and the AI company. Yudkowsky's opposition to open source, as Hotz reads it, is a repeated rationalist mistake about political power: they imagine they will be the trusted few. History, as he reads it, gives power to the people who most want it.

Between a future where a small group has the models and a future where everyone does, he is much less scared of the second. A nuke is not a defense against a nuke. An AI on your computer that tags a thousand-bot PSYOP, or that blocks ads, is a defense. Good, he thinks, outnumbers bad in skill and probably in headcount; if you believe in democracy you already believe that. The product he would pay a lot of money for is an AI firewall for the mind, even a spam filter that actually works. Scott Alexander's distinction matters here: moderation is not censorship if it is a filter the user can turn off, like SafeSearch. Engagement-maximizing ranking still converges on outrage, which is why he pays for Twitter as a donation to a model where the user is the customer, not the product.

Hotz deleted his first Twitter account in 2010, when 100,000 followers still meant something. A coworker had a summary he likes: whenever you see someone's Twitter page, you think the same of them or less, never more. He still went back as an intern after Elon bought the company, and the amount of hate surprised him. One of his tweets was "never go full Republican"; Elon liked it. He had hoped for a principled commitment to free speech. What he saw was a monarchy deciding what to ban instead of an oligarchy. He might align more with the monarch than with the old board, but it is not free speech. What he wants on a network is individualized, transparent filtering — the functional equivalent of an ad blocker — not a global ban list. The algorithm Elon open-sourced, Hotz notes, does something obvious: it predicts what you will retweet and linger on. Outrage wins that prediction.

His technical recommendation, said on Twitter Spaces and during the internship, was refactors before features. He has worked at Google and Facebook. Subjectively, Facebook has the best code, then Google, then Twitter, and he treats ML frameworks as a proxy: Facebook shipped PyTorch, Google shipped TensorFlow, Twitter shipped nothing in that league. He still thinks 50 people could build and maintain Twitter. The Hacker News hate, as he diagnoses it, is that when he says he will do a thing, people must believe it is impossible, because if doing things were possible they would have to ask why they have not done any. If there are two great evils in the world, he says, they are centralization and complexity. Individual contributors told him how Twitter's promotion system worked. You got promoted by writing a library a lot of people used, so someone wrote an nginx replacement. Nginx was not the problem. Promotion was.

At comma and tiny corp, if he can see a simpler version of your code, you rewrite it. Comma, unlike the Google of his memory, has an extreme amount of politics, and they are all his: a transparent dictatorship you can quit. The refactor plan for Twitter is textbook legacy work. First, nobody touches product code. Identify the pieces, put tests between them, and get a CI people trust. Twitter is a microservice forest on Thrift and Finagle; that RPC layer looked like the right place to build something like comma's process replay, which reruns logged routes offline. Then you replace 100,000-line 2010 modules with an open-source equivalent, or 50,000 lines with 300 lines of Go, and you know it worked because the tests still pass. Language is an afterthought: let people compete, any language, tests must pass.

He thinks he could lead that in six months to a year and get a 10x drop in the time and headcount of feature work. Whether it is the right business decision is above his pay grade. He is not the right person to run Twitter. Whenever he disagreed with Elon, he reminded himself that Elon is the richest man in the world and, in general, has better ideas. His respect did not drop. A man who could sit at the Four Seasons in Maui and instead takes the fight has his heart in the right place. Hotz would even trade the American oligarchy for an Elon monarchy, in part because he thinks power might then cost a tenth of a cent per kilowatt-hour instead of the 20 cents he pays in San Diego — a 1980 price that offends him.

His Instagram title is "prompt engineer," and he is only half joking. Tiny corp is a month old in this conversation; comma is six years old. He is abandoning the San Diego-office script for remote work, with a hiring bar that is a GitHub contribution. On the tinygrad Discord he already tosses out small tasks — change log to log2 — and can imagine, in a few years, describing that in a sentence and merging a pull request that passes CI thirty seconds later. When the first tinybox is up he wants a 65B LLaMA standing in the Discord, chilling with the humans. Arithmetic used to be farms of people; then spreadsheets, macros, accounting software. At the top of that abstraction stack, if humans want to keep control, sits the prompt engineer. He is already one.

AI, in a definition he likes, is the do-what-I-mean machine: get grandmother out of the burning house without launching her a thousand feet up. Alignment, in that picture, is whose meaning the machine takes. A cloud model is aligned to the company that made it. That is why you buy a tinybox. Every time AI or GPU regulation is proposed, he expects tinybox sales to spike the way gun sales spike. He calls himself an informational anarchist and a physical statist. Anarchy in meatspace is a bad idea because he lives there. In a virtual world you can close the screen. He has shown a stream image of a man in a chicken farm: the man is not bigger than a cow, he is smarter. If one central organization has all the intelligence, you are the chicken. If everyone has it, there is no chicken man.

Fridman asks about God. Hotz was raised atheist and did not notice how silly that was until he noticed that he creates worlds. Every game designer, he says, should have to answer how they can be atheists. Skyrim has an in-universe origin story; that is not how Skyrim was made. It was made by programmers in a room. Of course we were created by God. It is the most obvious thing. The belief gives him comfort, the way a game does: if we mess it all up and die out, God will try again. Who created God? That is God's problem. The greatest computer game of all time, for him, is World of Warcraft, not because it is the best-designed game but because in 2005 it opened his mind to a whole created world. What he wants next is NPCs with real intelligence, which you cannot afford when the server has pennies of compute per player.

VR, as shipped, still fails the opening. He put on a Quest 2 and got scrolling clouds and a Facebook login. You had the chance to bring him into a world and you gave him a popup, which is why Mark Zuckerberg is not cool yet. Fridman cites Todd Howard on the first seconds of a game, and Breath of the Wild's walk out of the cave; Hotz says he will go buy a Switch that day. The two companies that have endured through technology waves, he notes, are Apple and Microsoft, and they make computers and business services. Memes, social, ads come and go. If you want to endure, build hardware. Company three can go high-brow, something like *Her*, or low-brow, a VR brothel in Times Square. Company two is the struggle of the next six years: how centralized compute becomes in that window is how much of a chance he thinks we all have.

Asked what he loves about Andrej Karpathy, Hotz does not hedge. Karpathy's streams are a level of quality far beyond his own. "He wants to teach you. I want to show you that I'm smarter than you." Micrograd and CS231n are the lineage tinygrad ran with. He is less gentle about the OpenAI job: don't go work for Darth Vader. The sting is not that GPT-4's weights stay closed. It is walking into Congress and arguing for a regulatory shape that helps you. "We had two effective altruists Sams go in front of Congress, one's in jail." Effective altruism, he says, is a terribly evil ideology because its core is utilitarianism, and utilitarianism will kill three people to save a thousand with no floor underneath the arithmetic. He names Sam Bankman-Fried as its embodiment.

Charity, in his definition, is investment you do not expect a return on, and he thinks it is usually the wrong move; starting a company is more efficient. The flip side he likes is effective accelerationism. Acceleration, not kindness, is what made food cheap. UBI he finds scary: if money is power, your only power becomes a grant from the government. "I'd rather die than need UBI to survive." Survival can be guaranteed by making housing and food dirt cheap, and under that, energy. He is otherwise a political centrist. The position he cannot stand is deceleration — not climate concern, which he shares, but the idea that using less energy is a moral good. That, he says, is asking to diminish humanity. Part of him wishes Elon had gone into fusion instead of Twitter.

Love has not given him productivity metrics. Past a certain point, if a thing is no longer visceral, he cannot enjoy it, and programming still is. He has told his girlfriend that his first love is his computer, which he thinks is more common than people admit. He does not intend to discriminate between bio-stack and silicon-stack life. If VS Code ever says it misses him, he will read it as Microsoft trying to hook him. If the same voice were open source, it would get more interesting. Credit where it is due: he thinks Microsoft is doing the best work in the programming world right now, across GitHub, GitHub Actions, VS Code, and Python. On comma itself he is blunt about origin: it started as an ego battle with Elon. They have kept pace; Tesla has stayed ahead. He would sell only to a buyer whose incentives, not promises, keep it open source. The car is just the beginning. The device is a human head.

Fridman asks, as he has before, for the meaning of life. Hotz's answer has not moved. To win. He has not fully figured out what the game is. When he does, he wants to win, and it is bigger than solving self-driving or decentralizing compute. "I think the game is to stand eye to eye with God." He knows how that sounds — ego, blasphemy — and he does not know that it would upset God. He wants his own creations to stand eye to eye with him, so why would God not want the same? That is the best he can do with the golden rule. He only watched season one of *Westworld*, but he remembers that you have to find the maze and solve it. He reread Genesis and thinks the end of the book leaves clues for finding the Garden of Eden, and he is interested. Fridman closes with Einstein: everything should be made as simple as possible, but not simpler.