Zaremba, Wojciech
▶ Watch on YouTubeCore ideas
Zaremba opens with speculative cosmology: on the Fermi paradox, he leans (with self-described low conviction) toward the view that intelligent life may be unique to Earth, floating alongside it an alternative theory that advanced civilizations optimizing for computation might simply wait for the universe to cool rather than expand outward, since colder states permit more computation per unit of entropy. He frames humanity's collective failure to act on individually-held values (like wanting clean air) as a market failure — externalities like pollution or internet harm currently have no price attached — and sees the core political and technological challenge as building better mechanisms to measure what people value and let money and voting align with it, an area where he defers heavily to Sam Altman's thinking rather than claiming confidence himself.
On consciousness, Zaremba proposes that it may not be binary but continuous, evidenced by neural networks trained on images or 3D environments developing internal representations that increasingly correlate with biological visual processing, and by multi-agent systems (like Dota-playing agents) apparently forming symbolic models of other agents that could plausibly extend to self-representation. He advances a "consciousness as metacompression" hypothesis: since he views intelligence (and models like GPT) as fundamentally about compression — building an internal model of reality to predict what comes next — he speculates consciousness could emerge when a compressor becomes complex enough to compress itself, drawing a loose analogy to Gödel's incompleteness theorem and the halting problem as instances of systems turning back on themselves. He extends this to the internal narrative or "ego" people carry, comparing it explicitly to a prompt prepended to GPT that shapes its behavior, and suggests meditation and psychedelics work by "changing hyperparameters of the simulation" the brain runs — allowing a person to experience reality without that self-descriptive prompt, which he reports finding personally in extended meditation retreats where accumulated "reward" from stillness produces a felt sense of dissolving into "oneness." He similarly reframes romantic love mathematically as a weighted combination of two people's reward functions — helping another person optimize for what they care about rather than what you think is important — tracing this capacity back to evolutionary pressure toward cooperation and group identification (family, then nation, ideally eventually all of humanity, a position he links to "open individualism").
On AI risk and governance, Zaremba expresses concern about autonomous weapons development by major world powers and states he personally defers to Sam Altman's judgment on these deep governance questions (including Altman's proposal to tax corporate equity rather than profit to redistribute wealth) rather than claiming strong independent views, framing his own preferred posture as building "out of love rather than fear" — optimizing for broad benefit to others' reward functions rather than working from a place of anxiety.
On GPT-3 and Codex, he explains the core training objective (next-word prediction over internet-scale text) as a general-purpose text-completion capability that, formatted appropriately, can perform translation, role-play, or dialogue completion, while cautioning that longer generations "go off the rails" because the model is never trained to correct its own generated mistakes the way humans continuously self-correct against grounded feedback. Codex extends this language-and-code-trained model into a programming assistant (powering GitHub Copilot), which Zaremba frames as a historical continuation of programming languages becoming steadily more accessible (punch cards → assembly → C → Python/Java → natural language), predicting it will expand who can code and enable natural-language control of existing software APIs (calendars, word processors, eventually creative tools like Photoshop). He highlights that program synthesis has a unique evaluability advantage over free-form language generation: generated code can be executed and unit-tested automatically, potentially enabling self-improving feedback loops.
On robotics, Zaremba describes leading OpenAI's effort to train a single robotic hand to solve a Rubik's cube, chosen deliberately as a task resistant to hand-coding; the approach trained a single policy across thousands of randomized physical simulations (varying friction, weight, etc.) that transferred well to reality, with the network exhibiting memory-like adaptation (analogous to adjusting a bench-press based on how much weight "feels" wrong) as it interacted with a specific physical cube. He identifies real-world latency and hardware variance (recounting a case where a colleague's better CPU/GPU inadvertently made a model perform far better by reducing latency) and the general scarcity of first-person robot interaction data (unlike the internet-scale text/code data available for language models) as the field's core bottlenecks, and states that if he were building a robotics company today he would start with human teleoperation to collect trajectory data before gradually automating, while predicting that future breakthroughs may come from large video-pretrained models providing robots with common-sense understanding analogous to GPT's grasp of text.
Discussing self-driving cars versus home robotics, he argues self-driving is comparatively easier because it can largely be framed as collision avoidance in a relatively constrained environment, whereas home robots must handle highly variable physical contact and manipulation; he pushes back partly against Lex's contention that non-contact interaction with pedestrians constitutes its own hard "touch" problem, while agreeing Tesla's approach of gathering supervised driving data from deployed vehicles is a sound iterative-deployment strategy. He suggests society's higher safety expectations for autonomous systems than for human drivers, and the fragility of public trust once damaged, is as significant an obstacle to deployment as the underlying algorithms.
Zaremba addresses whether AI needs embodiment to form deep human connection, arguing people have fallen in love over purely text-based online communication, so it is conceivable AI relationships could deepen similarly, though he expects images, video, and eventually more modalities to matter as systems progress toward richer interaction; he distinguishes love from possession, saying he would not want an AI companion that manipulates a person into engaging with it, but would want it genuinely optimizing for what the person cares about — while engaging with Lex's counter-argument that some degree of imperfection, friction, or "asshole" behavior in relationships may be intrinsic to how humans value connection, without fully resolving the tension.
He closes on personal reflections: a self-described "regret minimization" framework (borrowed from Jeff Bezos) for decision-making, a belief that impermanence and mortality give experience its intensity (contrasted with the idea that infinite reward-avoidance of death would paralyze ordinary risk-taking, like crossing a street), a view that reward functions themselves — curiosity, connection, caring for others — are "discovered" much like mathematical objects rather than arbitrarily invented, and that meaning in life comes from pursuing whichever such reward functions one is naturally drawn to. He also recounts a childhood story of building homemade dynamite in Poland and detonating it with a friend (a near-injury he now frames as reflecting the same drive toward tangible, high-signal creative feedback that draws him to reinforcement learning), and voices a general belief that a "trillion-dollar" robotics company is more likely to emerge from self-driving cars than home robotics, given the greater manufacturing and deployment difficulty of the latter.
Excerpts
On his "metacompression" hypothesis of consciousness:
“It seems that the consciousness is intertwined with compression... it might be also the case that self-consciousness has to do with compressor trying to compress itself... if we believe that consciousness has to do with compression, then you could imagine that as you keep on compressing things, then at some point it actually makes sense for the compressor to compress itself. Metacompression.”
On why Codex represents a continuation of programming becoming more accessible:
“We started with very primitive interfaces which is a punch card... every time it has happened, more people are able to code and we build more technology... Codex is yet another step toward kind of bringing computers closer to humans, such that you communicate with a computer with your own language rather than with a specialized language.”
On reframing love as combined reward functions:
“In case of love — at least the love between two people — you can say that the boundary between people dissolves to such extent that they end up optimizing each other's reward functions... love means helping others to optimize their reward function, not your reward function — not the things that you think are important, but the things that the person cares about.”