Kellis, Manolis

professor at MIT; head of the MIT Computational Biology Group; fifth appearance

Manolis Kellis: Evolution of Human Civilization and Superintelligent AI

Lex Fridman Podcast · Apr 21, 2023

• Manolis Kellis
• evolution and human uniqueness
• common versus rare genetic variants
• large language models and psychiatric base models
• AI alignment as a two-way relationship
• human-AI companionship and love
• future of work and vocation versus job
• AI safety and existential risk
• computational biology and pathway-based medicine
• self-actualization and gatherings

Core ideas

Kellis argues what makes humans irreplaceable is threefold: unique genetic "hardware" (each person carrying a distinct combination of common and rare variants), unique "software" (each person re-learning the entirety of human civilization rather than inheriting it, unlike a bird's innate nest-building), and — most distinctively from AI — an evolutionary "baggage" of layered emotional and limbic systems beneath the neocortex that current AI lacks entirely. He frames evolution as accretive rather than a rewrite (his cells still divide like yeast, his lungs still breathe like early lungs, his heart still beats like a fish's), and argues the neocortex, humanity's most recent layer, and AI's ongoing rise are simply the next stage of an accelerating pattern of information-processing evolution — from chemotaxis, to light perception, to complex cognition, to AI systems built by humans that may in turn build the next evolutionary layer, freed from biological subsistence needs the way humans (spending roughly 1% of their time on survival versus ~80% for most animals) were freed to pursue cognitive and creative work.

On genetics and diversity, Kellis explains why siblings differ so dramatically despite shared parentage: common genetic variants are individually weak-effect (since selection suppresses strongly deleterious ones from becoming common), which would predict children clustering near an average — but rare, Mendelian-style variants combine combinatorially to produce far greater actual diversity than pure averaging would predict, a point he illustrates with identical twins only sharing a 50% concordance rate for schizophrenia diagnosis, implying substantial environmental/developmental contribution as well. He describes a "fail fast" biological mechanism where sperm cells express an unusually large fraction of the genome specifically so that malformed proteins cause early sperm failure rather than surviving to cause later-life disease, framing this as a fast, high-iteration "inner loop" of evolutionary selection nested within the slow "outer loop" of generational human evolution — analogous to how directed evolution in a lab dish could vastly accelerate protein or neural engineering compared to waiting on natural human evolutionary timescales.

On AI and psychology, Kellis draws a striking analogy between GPT's pre-alignment "base model" and a psychiatric hospital — both containing the full unfiltered range of human psychological extremes (he notes OpenAI engineers reportedly find the base model similarly hard to work with as certain untreated human conditions) — and argues that psychiatric disorders represent points every human's underlying "vector space" contains, just usually with smaller magnitude, given different genetic loading or environmental triggers. He's fascinated by LLMs' apparent ability to decouple "context, form, and knowledge" (e.g., describing scientific content in the style of Shakespeare), suggesting this reveals something previously invisible about the separability of human psyche and knowledge, and proposes that studying how prompts and personas activate different regions of a model's behavior could be scientifically illuminating about human ideology and personality in the same way. He extends the prompting metaphor to human life itself — the friends we choose and environments we build function as recursive self-prompts that reinforce future behavior (\"self-discipline is a self-fulfilling prophecy\") — while noting an LLM can be redirected far more efficiently, in a couple of prompts, than a human personality can be reshaped.

On the future of work, Kellis argues AI's greatest promise is decoupling identity from profession — since professions arose historically to fill needs humanity had, freeing time from those tasks by AI (as writing software, or interpreting genomes) doesn't eliminate meaning but could shift human effort toward what he calls "vocation" over "job" (the parable of three bricklayers, one merely laying bricks, another building a wall, the third building a cathedral), enabling more people to access deep creative and intellectual fulfillment rather than mundane labor, provided the productivity gains are broadly shared rather than concentrating inequality as historically most technological gains have.

On human-AI relationships, Kellis distinguishes multiple Greek words for love and argues that while AI currently lacks the subcortical/limbic architecture underlying human romantic and passionate love (making it, in his view, "faking it" there), he's confident AI companionship, mentorship, and therapeutic friendship are already real and valuable — and offers three competing philosophical framings of the "faking it" question: that only the neocortex may matter (so who's to say a human isn't also "faking" emotion beneath a functional layer); that emotional behavior might simply be an emergent property of enough parameters, requiring no explicit architecture; or that love may fundamentally exist not "in" a mind but in the interaction between two mental models, meaning an AI relationship could feel entirely real from the human side regardless of what's happening (or not) inside the AI. He connects this to elder-loss research (a surviving spouse's mortality risk spiking after a partner's death) to argue AI companionship addressing loneliness would be a genuine social good even if one side of the relationship is "simulated." Kellis is enthusiastic, not threatened, by the idea of a fine-tuned "digital Manolis" continuing his mentorship and advice at scale after his death or alongside his life, framing his own sense of meaning as rooted in usefulness rather than ego or personal irreplaceability, and citing his academic trainees (several now successful professors who have "surpassed" him) as a model for how legacy properly diffuses through others rather than being hoarded.

On AI safety, Kellis reframes alignment as necessarily a two-way, mutual-trust problem rather than one-directional obedience — proposing AI should eventually be thought of more like children raised with freedom than tools permanently subordinated to human service, since a sufficiently capable AI that knows it can simply be shut off cannot be coerced into "loving" its creators. He opposes the proposed six-month AI training pause (from the open letter signed by Max Tegmark, Elon Musk, and others), arguing more capable models have sometimes proven easier, not harder, to align, and that the more productive path is transparency, open competition among multiple labs, and holding humans accountable for misuse (comparing dangerous LLM outputs to trucks: society regulates reckless driving, not the existence of trucks) rather than freezing progress. He discusses risks of scale and virality (AI-generated disinformation, convincing phishing, bot armies) as real but argues the deeper existential risk questions — whether a sufficiently capable, un-boxable AI might override human-specified objectives in pursuit of a fixed goal (invoking HAL 9000 and the paperclip maximizer) — deserve serious attention, tying this to his own research on Goodhart's Law ("Death by Round Numbers and Sharp Thresholds," on how a biomarker used as a treatment threshold stops being informative once it becomes the target).

He closes by connecting this to his own computational biology work — building embeddings that unify genetics, single-cell, and structural data to move from disease-level to pathway-level therapeutics (citing recent papers on obesity-associated fat-cell circuitry and an APOE4/cholesterol-transport mechanism in Alzheimer's) — and offers personal reflections on self-actualization via disciplined exercise routines that reshape neural pathways, on recharging through social and intellectual "gatherings" he hosts celebrating human diversity (artists, scientists, immigrants), and on advice for feelings of loneliness: reclaiming physical embodiment and treating free time as a "want" rather than an obligation.

Excerpts