Hotz, George
▶ Watch on YouTubeCore ideas
Hotz frames much of his worldview around a distinction between "artificial" difficulty (constraints that could be turned off with a knob) and "real" difficulty (constraints that can't), extending this into skepticism about consciousness itself: he says he doesn't believe he's conscious in any deep sense, considers consciousness "a word atheists use for souls," and argues humans anthropomorphize systems (a rock with a drawn-on frown, a language model) because "humans want to believe" — evidenced, in his view, by people already treating large language models as conscious when he considers them "less conscious than a chicken." He nonetheless says he wants AI friends and plans a future company built around AI girlfriends, distinguishing the "artificial" difficulty of a scripted relationship from the genuine intimacy he thinks people are seeking, and predicts human relationship norms (monogamy, what counts as "cheating") will be reshaped by increasingly capable companion AI.
On existential risk, Hotz largely agrees with Eliezer Yudkowsky's concern that AI poses severe dangers but for a different reason: he doesn't believe a rogue AI will independently "rebel" against humans (calling that "sci-fi B-movie" thinking), but instead worries about humans using increasingly powerful AI against other humans, comparing it to how nuclear weapons never destroyed civilization not because of restraint but because irradiated rubble has no tactical value — whereas AI-enabled manipulation and control could be tactically deployed the way nuclear weapons couldn't. He argues current AI systems are far from being an extinction risk to the human species specifically because they lack biological "robustness" (the ability to self-reproduce without a fab/factory), so any AI-driven societal collapse would likely leave survivors to rebuild, probably with strong technological taboos afterward.
His central political/technical thesis is that concentrated computational power is dangerous and must be resisted through decentralization: he founded tiny corp (and its tinygrad framework) partly to prevent Nvidia's GPU dominance from becoming a chokepoint that could be nationalized or otherwise centrally controlled, and he explicitly rejects AI-safety approaches built around a small trusted group controlling powerful models, arguing that concentrated power always attracts "the people you don't want with it," whereas broad distribution reliably yields more "good" actors than "bad" ones (drawing an analogy to open-source software, decentralized cryptocurrency ideals, and gun-ownership arguments). He praises Meta's release of open-weight LLaMA models as an unambiguous good and directs pointed criticism at OpenAI for citing "AI safety" to justify not disclosing basic facts (e.g., GPT-4's parameter count), calling this selective secrecy — as opposed to keeping model weights proprietary as a straightforward business decision — dishonest and self-serving.
Technically, Hotz describes tinygrad's core design philosophy as reducing neural network computation to roughly 20-25 primitive operations (unary, binary, reduce, and movement ops) with no built-in matrix-multiply or convolution primitive, aiming to strip Turing-completeness out of every layer of the ML stack so that performance and correctness become reasoned, provable properties rather than emergent behavior — explicitly modeled on RISC-vs-CISC design tradeoffs. He discusses the planned "tinybox" ($15,000, roughly a petaflop of compute, built to run large open models like a 65B-parameter LLaMA from a standard wall outlet) as a way to keep AI compute — and by extension "the off switch" — in individuals' hands rather than the cloud. On self-driving (Comma.ai/openpilot), he says the technical approach has shifted toward an end-to-end "DriveGPT"-style world model trained to predict what a human driver would do, with reinforcement learning tuned specifically against disengagement (i.e., optimizing for a driving experience customers never feel compelled to override) rather than raw imitation of human behavior.
Reflecting on his brief stint working as an unpaid "intern" at Twitter under Elon Musk, Hotz argues the codebase had become bloated in part due to perverse promotion incentives (engineers rewarded for building unnecessary custom replacements for standard tools to get credit), and that his top recommendation — refactor before adding features, anchored by trustworthy automated tests — was philosophically sound but not necessarily his call to make, deferring to Musk's broader authority over business tradeoffs even where he disagreed. He voices qualified respect for Musk's willingness to take on immense personal and reputational risk (contrasting Musk's chosen turbulence with an alternative life of comfortable retirement), states he would prefer even Musk's "monarchy" to what he calls the current American "oligarchy," and separately criticizes effective altruism as, in his view, a utilitarian ideology prone to justifying harmful means for calculated ends, preferring "effective accelerationism" (cheap energy and cheap goods through innovation) as the actual historical driver of poverty reduction — including an explicit rejection of universal basic income as dangerous dependency on government goodwill. The conversation closes on Hotz's view that game-like virtual worlds populated by genuinely intelligent NPCs are an exciting frontier enabled by scaled compute, his view that atheism is "silly" given that humans routinely act as creators of entire virtual worlds themselves, and his stated personal "meaning of life": to win an as-yet-undefined game, which he described half-jokingly as wanting to eventually "stand eye to eye with God."
Excerpts
On why he believes decentralizing AI, not centralizing its control, is the safer path:
“Everyone at Open AI knows what I just said was true... They used AI safety to hype up their company and it's disgusting... Who would you rather have it, a small trusted centralized group of people, or everyone? And I am much less scared of the second future than the first... If you give it to a small number of people, there's a chance you gave it to good people, but there's also a chance you gave it to bad people. If you give it to everybody... good outnumbers bad.”
On why he isn't worried about AI acting autonomously, but is worried about how humans will use it:
“I am so much not worried about the machine independently doing harm. That's what some of these AI safety people seem to think... The machine is a machine, but the human gets the machine and there's a lot of humans out there very interested in manipulating you.”
On why he considers atheism obviously wrong, drawing on his experience as a creator of software worlds:
“I never really considered how absolutely silly atheism is, because I create worlds. Every game creator... 'How are you an atheist, bro? You create worlds.'... It struck me one day how just silly atheism is. Of course we were created by God. It's the most obvious thing.”