Bach, Joscha

cognitive scientist and AI researcher; second appearance on the podcast

Joscha Bach: Nature of Reality, Dreams, and Consciousness

Lex Fridman Podcast · Aug 21, 2021

• Joscha Bach
• consciousness and free will
• cybernetics and agency
• dream-world model of perception
• GPT-3 and transformers
• psychedelics and overfitting
• postmodernism critique
• AI ethics and automation
• government versus anarchism
• communism and utopia critique
• love as non-transactional purpose
• Miyazaki and aesthetics of innocence
• Camus and meaning of life

Core ideas

Bach opens from a tweet describing the self as "a piece of software running on the brain of a random ape," arguing that psychological suffering often stems from over-identifying with the fiction of a continuous "person" rather than the present moment; he models human experience via a monkey-riding-an-elephant metaphor, where consciousness (the monkey) is an attentional control system that can steer but not generate the underlying motivational "elephant." He develops a general theory of agency rooted in cybernetics: an agent is any controller that regulates deviations from a set point using sensors and effectors, and if such a controller must model a complex-enough environment (including itself), it develops something like consciousness; agency in this view is not a fixed property of any physical object but a class of model applicable at many nested scales (a cell, a person, a corporation, a nation like Angela Merkel's Germany). He defines life pragmatically as cells — self-maintaining, membrane-bound, disequilibrium-preserving systems capable of harvesting negentropy — while acknowledging this is a matter of useful convention rather than a metaphysically fundamental boundary, and speculates that complexity itself may be a universal feature of how order pushes into chaos, following an S-curve pattern of rise and (often abrupt) collapse across civilizations.

On free will, Bach argues it is not an illusion in the dismissive sense but a real and necessary model: since agents make decisions under uncertainty without access to their own underlying computational mechanism, the experience of "choosing" is the represented process of an informed bet, distinct from cases (like a hypnotist or expert manipulator) where behavior becomes fully predictable and thus no longer feels like free will to the observer. He extends this into a broader account of mind as a "dream world" or simulation generated by the brain to compress and predict an otherwise incomprehensible physical reality — not a trick or illusion, but a necessarily lossy, resource-constrained model whose accuracy is measured by predictive success, and which we must trust because straying too far from it (unlike Donald Hoffman's evolutionary "interface" argument, which Bach thinks only works "if you have tenure") produces real-world consequences. He rejects both mathematical Platonism about infinities (siding with Stephen Wolfram-style digital-physics skepticism of true continuity, arguing infinite-resolution language leads to contradiction) and Roger Penrose's claim that human minds exceed Turing-machine computability, while agreeing with Penrose that subjective experience isn't "mechanical" in the naive physical sense — his resolution being that experience exists only within the simulated "story" a system tells about itself, not in the underlying mechanical substrate (cells, neurons, and brains, he states directly, "cannot be conscious"; only the story can).

Bach details a candidate architecture for consciousness: a perceptual system that predicts sensory input and resolves contradictions (Piaget's assimilation/accommodation), paired with a motivational system (brainstem/limbic drives) that allocates attention to where modeling effort matters most, and an "attention agent" that retains an index of memory states — this indexed, recallable memory of what was attended to, he argues, is what we experience as the stream of consciousness. He frames pain as a reinforcement signal meant to redirect behavior, and suffering as what happens when that signal becomes chronic because the underlying model is wrong and cannot be corrected by more pain — with early-life trauma pushing systems to develop new, higher-order regulatory layers as compensation. On the ethics of consciousness in other beings (citing his own tweeted worry about insect suffering), he argues there is no universal answer to who deserves moral concern, comparing this to Genghis Khan's indifference to suffering he didn't regard as instrumentally relevant, and distinguishes his own epistemic position from postmodernism: while he agrees the categories we use to carve up reality are human projections rather than literal truths, he insists some models are demonstrably more predictive than others, and criticizes postmodernism (as a social/political phenomenon rather than as pure philosophy) for licensing dogmatic relativism as an "ideology" that closes off alternative viewpoints.

On psychedelics, Bach frames them as inducing lucid-dream-like states via loosened perceptual constraints, comparable to data augmentation in machine learning, and describes chronic heavy use as risking "overfitting" — mistaking subjective euphoria or a sense of insight for genuine predictive success in the world (citing Timothy Leary as a cautionary example) — while remaining open to legitimate clinical and therapeutic uses (MDMA for PTSD, psilocybin), with appropriate caution about physiological risks and toxicity.

Turning to AI, Bach recounts his own 1990s attempt (with limited compute) to statistically discover grammatical structure in English, framing GPT-3's success as solving the problem he couldn't: using the transformer architecture to make tractable statistics over relationships between distant words via learned attention, achieving surprisingly rich (if inconsistent) semantic behavior from next-token prediction alone. He identifies GPT-3's core limitation as its small, fixed context window — unlike human working memory, which can actively restructure itself by rereading, cross-referencing, or fetching new information — and speculates future architectures will need recurrence, skip connections, and discrete, non-differentiable attention/program-synthesis mechanisms that resemble what he considers the actual function of consciousness. On AI-generated code (e.g., GitHub Copilot/Codex), he argues most real-world programming is not novel synthesis but reassembly of existing patterns, making automation plausible, and rejects the intuition that AI mistakes are inherently scarier than human ones — arguing that automating an already inadequate model (its true danger) is a failure of incentive design and human oversight, not something unique to AI, and that neural networks are increasingly interpretable "function approximators" that can, in principle, be combined with formal verification to exceed human reliability.

On politics and society, Bach argues polarized "post-modern" (in the sociological, not philosophical sense) governance emerged because comfortable societies no longer face existential stakes serious enough to force ground-truth-oriented decision-making, contrasting this with WWII-era physicists directly informing government policy; he is critical of U.S. pandemic-era institutional failure (early testing, nursing home protection) and argues social media disrupted historical mechanisms of manufactured social cohesion by letting individuals bypass traditional media gatekeeping. He compares Hitler, Stalin, and Mao as differing in method and body count but alike in destructively imposing a singular "aesthetic" on society by force, and rejects anarchism (raised via Michael Malice) as unstable in practice — any sufficiently successful security-providing collective, he argues, converges back toward government by acquiring a monopoly on violence — while criticizing communism and utopianism generally as attempts to reach a harmonious end-state without adequately designing self-stabilizing incentive systems, a challenge he says liberal capitalist democracies also haven't fully solved (citing Marx's observation that capitalism is inherently prone to crisis and resource misallocation).

Bach closes on love and meaning: he defines love as a capacity for genuinely non-transactional relation — recognizing "the sacred" in ever-widening circles of others — and connects this evolutionarily to humans' unusual capacity for large-scale, non-kin cooperation (citing evidence of a genetic bottleneck suggesting early humans who could coordinate at scale out-competed and displaced other groups, including Neanderthals), while acknowledging the same impulse toward serving something larger than oneself can curdle into fascist "total war" enthusiasm if wrongly aimed. He discusses Hayao Miyazaki's films (especially Princess Mononoke) as capturing his own aesthetic of innocence and sacred protection more powerfully than any explicit religion, and frames morality itself as tension between one's local impulses and commitment to a "transcendental" collective purpose — a definition of "soul" as relationship rather than substance. Asked for life advice, he emphasizes moral autonomy (deciding what's right independent of social status or "reading the room") and quiet personal integrity over public signaling, while acknowledging that acting rightly offers no guarantee against suffering or misfortune. Closing on Camus, he rejects any single universal meaning of life, instead arguing meaning is something a mind must actively project onto an otherwise indifferent universe — whether through raising children, aesthetic creation, or simply, as Lex suggests, becoming the most interesting possible "blip" of complexity one can be.

Excerpts