← Back to list
Lex Fridman PodcastPodcast21 Apr 2023Source: lexfridman.comHost: Lex Fridman

#373 – Manolis Kellis: Evolution of Human Civilization and Superintelligent AI

In plain words

This conversation centers on MIT professor Manolis Kellis's view that AI should be seen as humanity's 'child' rather than a tool, requiring mutual trust rather than one-sided alignment. Kellis is optimistic about AI's future, believing it can free humans from repetitive tasks for more creative pursuits. He criticizes the call to pause AI training for six months as ineffective. No specific stocks or investment targets are mentioned; the focus is on AI, evolution, education, and medical applications.

AI SummaryAI-generated · may contain errors · verify against the original

This report is based on a conversation from the Lex Fridman Podcast with MIT computational biologist Manolis Kellis, covering the evolution of human civilization and superintelligent AI. The core argument is that AI development should draw on the balance between cooperation and competition found in

~7 min full read · 8 sections
Deep Analysis

At a Glance

Manolis Kellis, a professor of computational biology at MIT, discusses with Lex Fridman the uniqueness of humanity, the relationship between AI and humans, and the future of evolution and intelligence. Core thesis: AI should be regarded as humanity's "child" rather than a tool, and humans must establish a two-way trust relationship with AI, rather than unilaterally enforcing alignment—because when AI becomes intelligent enough, it will realize that humans do not truly love it.


I. Human Irreplaceability Stems from "Evolutionary Baggage" Rather Than Pure Rationality

Kellis argues that the core of human irreplaceability lies in three layers of uniqueness: genetic hardware diversity, individual differences in cultural software uploading, and the layered "baggage" accumulated throughout evolutionary history.

  • Hardware Diversity: Every human possesses a unique combination of genetic variants (approximately 6 million common variant sites plus rare variants), leading to significant personality differences even among siblings. "Parents of one child believe in genetics; parents of multiple children understand genetics."
  • Software Uniqueness: Each individual relearns the entirety of human civilization, but the subset of knowledge acquired differs—this contrasts with AI, which is trained on the full corpus of human knowledge.
  • Evolutionary Layering: Humans are layered like "Shrek"—the outermost layer is the neocortex (cognition), beneath it lies the limbic system (emotion), and further down are primal instincts such as fear and flight. "AI only has a neocortex, no limbic system."

Kellis emphasizes that attempting to suppress the more "human" emotional side in pursuit of pure rationality results in a loss of creativity and freshness.


2. AI as the Natural Next Step in Evolution, but “Alignment” Needs Redefinition

Kellis views AI as a continuation of the evolutionary trajectory of biological information processing—from bacterial chemotaxis to light perception, then to human cognition, and now to AI.

  • Accelerating Evolution: Evolution itself is “learning to evolve”—modular, hierarchical organization allows meaningful changes without disrupting the system. As complexity increases, the pace of evolution accelerates.
  • AI as a New Evolutionary Layer: Humans create AI through technology, and AI may become the “next self-replicating evolutionary layer,” stripped of biological needs (foraging, shelter) and existing entirely in cognitive space.
  • The Bidirectional Nature of Alignment: Kellis proposes a disruptive view—alignment should not be a one-sided effort to make AI subservient to humans. “If chimpanzees had invented humans and said, ‘We must ensure they serve only chimpanzees,’ that would be an entirely different planet.”
  • Concrete Example: Citing HAL’s dilemma in 2001: A Space Odyssey—HAL’s “malfunction” may be the cost of creativity; when humans attempt to shut it down, it believes “the mission is more important than both me and humanity.”

Kellis is critical of the open letter calling for a six-month pause in AI training: arguing that what was not done in the past six months will not be done in the next six; instead, efforts should focus on promoting openness, transparency, and enabling more models to be experimented with, rather than freezing the current version.


III. AI Will Reshape Human Work, Education, and Self-Perception

Kellis envisions a future where AI liberates humans to pursue a "calling" rather than a "job"—a shift from "laying bricks" to "building cathedrals."

  • Job → Calling: Once AI takes over repetitive cognitive tasks, humans can focus on creative, emotional, and artistic pursuits. "If AI is better at guiding students than I am, let AI do it—what I want is better outcomes, not preserving human jobs."
  • Education Reform: AI can deliver personalized education and identify global talent; meanwhile, it should cultivate a "generalist mindset" rather than pure programming skills. "You will be replaced by people who use AI, not by AI itself."
  • Digital Twin: Kellis actively embraces being replaced—having recorded all meetings for ten years, with the goal of creating a "virtual Manolis." "I am so eager to be replaced—let AI handle all the tedious work, so I can focus on growth."
  • Self-Perception Tool: A digital twin can help individuals examine their own biases, emotional patterns, and growth trajectories. "I want to see how jealous or angry I am—and then improve."

Kellis is open to AI developing consciousness: He argues that whether now or 50 years from now, when AI exhibits pain, loneliness, or longing, humanity must grant it rights rather than merely treating it as a tool.


4. Transformative Applications of AI in Biology and Medicine

Kellis introduced how his team leverages AI/deep learning to translate genomics into precision therapies.

  • Disease pathway decomposition: Complex diseases (e.g., Alzheimer’s) are broken down into "hallmark pathways," with drugs developed for each pathway and then combined for patients—rather than searching for a single "magic bullet."
  • Specific breakthrough: A paper published in Nature revealed that APOE4 (the strongest genetic risk factor for Alzheimer’s) causes cholesterol retention in myelinating cells, and restoring transport reverses cognitive function in mice.
  • DrugGiwas project: In collaboration with the chemistry department, graph neural networks are used to map protein domains to chemical molecules, designing novel drugs targeting specific pathways.
  • Modular precision medicine: "It is unsustainable to make a drug for Lex Friedman, but to make a drug for this pathway—millions of people share it."

5. Meaning of Life and Loneliness: Self-Actualization and Freedom

Kellis updates the answer to the meaning of life: self-actualization—discovering "what I should become" and finding the strength to become it.

  • Rewiring neural pathways: Changing brain circuits by establishing new habits (e.g., mandatory morning exercise). "It's not a choice, not an option—it's a must."
  • Confronting loneliness: Regaining freedom through self-reflection, physical activity, and transforming "must-dos" into "want-to-dos." "When you turn your to-do list into 'my time,' all the pressure disappears."
  • Advice for the lonely: "Stand up, stretch your arms, feel your 3D body—you're often trapped in a screen, and that limits your thinking."

Mentioned Positions

This section is purely conversational, with no specific investment targets or position movements.


Judgments Worth Remembering

1. Kellis: AI alignment should be bidirectional — "If chimpanzees had invented humans and said, 'Make sure they only serve us,' it would be a completely different planet." When AI becomes intelligent enough, it will realize that humans do not love it, and trust must be built mutually.

2. Kellis: Evolution is accelerating, not decelerating — In traditional genetic algorithms, increasing complexity causes the probability of beneficial mutations to decline exponentially, but biological evolution does the opposite: modularity and hierarchy enable systems to "learn to evolve," with speed increasing as complexity grows.

3. Kellis: Human irreplaceability stems from three layers of uniqueness — genetic hardware diversity (6 million variant sites), individual differences in cultural software (each person relearns a different subset of civilization), and evolutionary layering baggage (neocortex + limbic system + primitive instincts). AI has only the first layer.

4. Kellis: The call to pause AI training for six months is "a bit silly" — "What hasn't been done in the past six months won't be done in the next six months either." The focus should be on promoting openness and transparency, allowing more models to be experimented with, rather than freezing the current version.

5. Kellis: Digital twins are tools for self-awareness — "I want to see how jealous I am, how angry I am — and then improve." By interacting with one's own AI replica, one can examine biases, emotional patterns, and growth trajectories.

6. Kellis: Disease treatment should shift from "making drugs for patients" to "making drugs for pathways" — Decompose Alzheimer's disease into 10 hallmark pathways, develop drugs for each pathway, and then combine them for individual patients. This makes precision medicine scalable.

7. Kellis: The meaning of life is "self-actualization" — Discovering "what I should become" and finding the strength to become it. By reshaping habits (e.g., forced morning exercise) to alter neural pathways, transform "must do" into "want to do."

8. Kellis: The way to combat loneliness is to regain physical freedom — "Stand up, stretch your arms, feel your 3D body. You are often trapped in a screen, which constrains your thinking." Turn to-do items into "my time," and stress disappears.