Would AI Replace or Assist Humans? — (Yet Another) Wrong Question to Ask!
posted: 10-May-2026 & updated: 13-May-2026
Share on LinkedIn | Instagram | Twitter (X) | Facebook
The question “Will AI replace humans or assist them?” is seductive because it offers exactly two options, both of which feel meaningful. But the binary is a trap. What is actually happening between human intelligence and AI is something that neither word captures — something richer, stranger, more intimate, and far more consequential than either “replacement” or “assistance.”
If I had to name it, I would call it cognitive resonance — a sustained, iterative process in which AI stimulates my brain, strengthens my insight, broadens the landscape of my knowledge, and even helps me develop creativity. Not by doing my thinking for me. But by doing something to my thinking that no tool, no textbook, and no search engine has ever done before.
If I must name what AI does in this new mode of interaction, I would call it cognitive catalysis.
And yet — precisely because of this experience — I am more convinced than ever that AI will never replace humans. Because what makes this process work is the very thing that only humans bring to it: intuition, judgment, lived experience, and the capacity for meaning.
The Wrong Question — Again
In “Why Do We Live? – A Wrong Question to Ask”, I argued that “Why do we live?” is the wrong question — that the right question is “Do I want meaning in my life?” In “Coming Back to the Human in the AI Era — A Sequel to “Why Do We Live?””, I argued that “Is anything honorable?” is the wrong question — that the right question is “Do I choose to honor anything?” In “Is AI Safe or Dangerous? - A Wrong Question to Ask!”, I argued that “Is AI safe or dangerous?” is the wrong question — that the right question is whether we are wise enough to wield it.
Here I make the same philosophical move once more.
“Will AI replace humans or assist them?” is the wrong question to ask.
The question sounds practical. It feels responsible. Policy papers are written around it. Conferences are organized around it. Careers are built on answering one side or the other. And it is almost entirely misleading — because it forces a phenomenon of extraordinary dimensionality into a one-dimensional binary, and in doing so, it obscures the most important thing that is actually happening.
I say this not as abstract philosophy but as lived experience. I have delivered dozens of AI lectures at universities — at Stanford University, at San Jose State University, at Texas A&M University, at Seoul National University, at Korea University, at KAIST, at POSTECH, and many many more. In every one of these lectures, I encounter the same framing from the audience — students, professors, industry professionals, even so-called experts: “Will AI replace us?” or “Is AI just a tool?”
One day, in the middle of delivering yet another special AI lecture, something hit me. Not gradually, but as a sudden flash of clarity, the kind that comes only when you have been circling a problem for a long time without knowing you were circling it.
The framing was wrong. Both sides were wrong. And not merely incomplete — fundamentally misconceived.
What Is Actually Happening
Let me describe what happens, concretely, when I engage with an LLM-embedded software service1 in my daily work.
I throw a question. I get an answer. But it does not end there. I examine the answer against my own expertise and correct it where it falls short. I ask a follow-up question informed by what the response revealed. I inject additional knowledge that the system does not possess. I challenge its framing with an alternative perspective from a domain it has not considered. I ask it to steelman a position I disagree with so I can test my own reasoning against the strongest counterargument. I take its output, transform it through my own judgment, and feed the transformed version back. It responds to my transformed version with something I did not expect. I learn something. I correct something else. The cycle continues — sometimes for hours.
This is not replacement. A system that I must correct, redirect, challenge, and inform at every step is not replacing me in any meaningful sense.
But it is also not mere assistance. An assistant retrieves information you request. An assistant formats a document you have already conceived. An assistant follows instructions. What I have just described is not following instructions. It is something closer to — and I use this phrase deliberately — an intellectual sparring partner with encyclopedic knowledge across virtually every domain of human understanding, available 24 hours a day, infinitely patient, and relentlessly responsive.
Thirty Examples of What “Neither Replace Nor Assist” Looks Like
The single example I gave above barely scratches the surface. Let me offer a far more comprehensive landscape of what this new mode of human-AI interaction looks like in practice. None of these are hypothetical. All of them either come from my own experience or from patterns I have observed directly in the people around me.
Deepening Understanding
Building genuine insight in mathematical concepts
I could prove the theorems of convex optimization duality for decades — I had the mathematical machinery. But the understanding, the deep insight into why duality works, what it really means at an intuitive level — that eluded me for years after graduating from Stanford. Through sustained, iterative dialogue with AI, asking the same question from fifteen different angles, challenging each explanation, and feeding in my own partial intuitions for refinement, I finally arrived at the understanding I had longed for. (I wrote about this journey in “Shadow Prices and Genuine Understanding - A Journey Through the Soul of Optimization”.) The AI did not give me the insight. But it would not have come without the AI, either. That is the phenomenon we need a new word for.
This experience has led me to a philosophical claim that some will find audacious, perhaps even absurd, but which I believe is defensible: I do not believe any human being, living or dead, has ever achieved the kind of genuine understanding I describe in the Convex Optimization Forum — not for Riemann’s Hypothesis, and not even for Fermat’s Last Theorem.
Consider Fermat’s Last Theorem. Andrew Wiles proved it in 1995 after seven years of solitary work, deploying machinery so sophisticated that only a few hundred mathematicians in the world could even verify the proof. The proof runs to over one hundred pages and requires modular forms, elliptic curves, Galois representations — concepts that did not exist when Fermat first wrote his famous marginal note. We know Fermat’s Last Theorem is true. We can verify the proof. But do we understand it in the sense I mean?
As I argue in the Convex Optimization Forum, genuine understanding means being able to see the result from multiple perspectives, to grasp the geometric essence of why it must be true, to feel it in your bones, to explain it in ways that make it obvious rather than merely provable. By this standard, Fermat’s Last Theorem remains ununderstood — not because the proof is wrong, but because the very fact that it requires such elaborate machinery suggests we are verifying rather than comprehending. If someone genuinely understood why these particular Diophantine equations have no solutions, the explanation might be more direct, more elegant, less dependent on hundreds of pages of technical apparatus.
And Riemann’s Hypothesis? For over 160 years, the greatest mathematical minds have studied this problem. They know the statement perfectly. They have explored implications, investigated special cases, connected it to the distribution of primes. But not a single person has proven it — which means, by the strict standard of understanding I propose, not a single person truly understands why it works. If they did, the proof would already exist.
Here is where AI becomes genuinely revolutionary. AI might help us achieve, for the first time in human history, the kind of deep, multi-perspectival, from-first-principles understanding of these great mathematical truths that has eluded humanity for centuries. Not by proving the theorems for us — AI cannot do that, as I have argued repeatedly. But by serving as an inexhaustible cognitive catalyst, allowing mathematicians to explore these problems from thousands of angles, to test intuitions relentlessly, to discover connections that would take lifetimes to uncover through traditional methods alone.
Perhaps, through sustained dialogue between human mathematical intuition and AI’s encyclopedic pattern-matching across all of mathematics, we will finally understand — genuinely, deeply, multi-dimensionally understand — why Fermat’s Last Theorem is true and why the Riemann Hypothesis must hold (or must fail). And when that understanding comes, the proofs may follow almost as an afterthought, the way a solution becomes obvious once you see the problem from the right angle.
This is not replacement. This is not assistance. This is the possibility of achieving something humanity has never achieved before: genuine understanding of our deepest mathematical truths.
I described my own small version of this journey — how AI catalyzed the understanding of duality that had eluded me for over a decade — in “Shadow Prices and Genuine Understanding - A Journey Through the Soul of Optimization”. If AI could do that for one mathematician’s struggle with one concept in convex optimization, imagine what it might do for humanity’s centuries-long struggle with the deepest truths in all of mathematics.
Exploring the philosophical foundations of a technical concept
When I am preparing a lecture on, say, the nature of hallucination in LLMs, I engage AI not to get a definition of hallucination, but to pressure-test my own argument that hallucination is not a bug but a structural engine of creativity. I throw my argument at the system. It finds the weak points. I strengthen them. It finds more. The resulting lecture is mine — entirely mine — but it has been forged in a dialogue that no textbook or colleague could have provided at that speed, depth, and breadth.
Connecting disparate domains
I ask – “Is there a structural parallel between Wittgenstein’s language games and the conditional probability estimation in Transformers?” The AI offers one. I see a flaw. I propose a better one. It refines. Within thirty minutes, I have an argument that connects two domains in a way that might take months to develop through reading alone. As I wrote in “AI (with current DL architecture) does not believe nor reason nor know nor think!”, LLMs are sophisticated conditional probability estimators, not reasoners — but this does not diminish the fact that interacting with such a system catalyzes genuine reasoning in the human.
Testing historical claims against multiple perspectives
Is the standard narrative about the fall of the Roman Republic actually well-supported? Rather than reading five books, I can engage in a rapid, iterative dialogue where I challenge each claim, ask for counterevidence, and build a more nuanced understanding in hours rather than weeks. The reading still matters — but the AI has given me a scaffolding that makes the reading far more productive.
Understanding the implications of a mathematical proof
You can follow every line of a proof and still not understand it. AI allows me to ask – “What would break if this lemma were false?” “What is the geometric intuition behind this algebraic step?” “Why did the author choose this approach rather than the obvious alternative?” Each answer generates three new questions. The understanding that emerges is mine, but the velocity at which it emerges is new.
Generating and Testing Hypotheses
Scientific hypothesis generation
A biotech researcher notices an anomaly in assay data. She describes the anomaly to the AI, which proposes seven possible explanations drawn from across the biomedical literature. She eliminates four based on her domain expertise. She had not considered two of the remaining three. One of those two becomes the basis for a new experimental design. The hypothesis is hers. The AI did not “discover” anything. But the hypothesis would not exist without the interaction.
Business model stress-testing
At Erudio Bio, I have used AI to pressure-test our VSA platform’s market positioning. Not by asking “Is this a good business model?” but by asking – “What are the three strongest arguments against this model?” Then – “Now steelman each of those arguments.” Then – “Now help me dismantle them.” The resulting pitch is infinitely sharper than what I would have produced alone — not because the AI wrote it, but because the AI pressured it.
Investment thesis development
An investor describes a startup’s technology. The AI maps it against adjacent technologies, identifies potential market overlaps, and surfaces regulatory risks the investor had not considered. The investor’s thesis is now informed by a broader landscape — but the judgment about whether to invest remains entirely human.
Policy scenario modeling
A policymaker considering AI regulation describes a proposed framework. The AI generates second-order effects, unintended consequences, and historical parallels from other technology regulations. The policymaker’s proposal is now stronger — not because the AI wrote policy, but because it served as a relentless devil’s advocate.
Architectural design exploration
An architect describes a spatial constraint. The AI proposes structural approaches from traditions the architect has not studied — Islamic geometric patterns, Japanese joinery, brutalist cantilevers. The architect selects, combines, and transforms. The design is original. The inspiration landscape was AI-expanded.
Enhancing Creativity
Writing refinement through dialogue
When I write a philosophical essay — as I did with “Coming Back to the Human in the AI Era — A Sequel to “Why Do We Live?”” — I do not ask the AI to write for me. I write a draft, then ask the AI – “Where is this argument weakest?” “What would Schopenhauer say to this?” “Is there a more precise word for what I mean here?” The final essay is mine in voice, argument, and conviction. But it has been sharpened by a dialogue partner that can summon Schopenhauer in an instant.
Musical composition exploration
A pianist — and I am one — can describe an emotional texture to the AI – “I want something that feels like late autumn light through stained glass.” The AI suggests harmonic progressions, voice leading patterns, and references to specific pieces. The pianist selects, transforms, and creates something the AI could never have composed — because composition requires not just knowledge of harmony but the lived experience of what autumn light feels like.
Visual art conceptualization
A painter describes a concept – “I want to express the tension between geological time and human time.” The AI offers references — Anselm Kiefer’s layered surfaces, the deep time photography of Edward Burtynsky, Japanese mono no aware. The painter had not encountered Kiefer. The resulting work is a conversation between the painter’s vision and an expanded landscape of possibility.
Narrative structure development
A novelist has a character but not a plot. Through iterative dialogue, the AI offers structural patterns — not plots, but patterns – the spiral structure of One Hundred Years of Solitude, the temporal fragmentation of Slaughterhouse-Five, the epistolary constraint of Dangerous Liaisons. The novelist sees her character in the spiral structure and something ignites. The novel is hers. The ignition was catalyzed.
Lecture design
Every major lecture I deliver goes through an AI dialogue. Not to generate content — the content comes from decades of experience. But to find the order, the arc, the moment where the audience’s assumptions will be most productively disrupted. As I reflected in “When Teaching becomes Transformation - Reflections on Five Lectures that Changed How Students See AI”, I was not just teaching content — I was facilitating paradigm shifts. The AI helps me find the sharpest path to that shift.
Accelerating Research
Literature landscape mapping
A researcher entering a new subfield can, in two hours of iterative AI dialogue, build a mental map that would otherwise take two months of reading. The map is not the territory — as I argued in “Coming Back to the Human in the AI Era — A Sequel to “Why Do We Live?””, the map is never the territory. But having a map before you enter the territory is transformative.
Cross-disciplinary translation
A neuroscientist and an economist are working on the same problem from different angles but cannot understand each other’s notation. The AI serves as a translator — not of languages, but of conceptual frameworks. Neither researcher is replaced. Neither is merely assisted. Both are amplified.
Experimental design refinement
A graduate student describes a proposed experiment. The AI identifies confounding variables the student had not considered, suggests control conditions drawn from adjacent fields, and points out a statistical subtlety in the proposed analysis. The advisor might have caught these in time. The AI caught them in minutes.
Data pattern recognition and hypothesis prompting
A researcher stares at a dataset that looks anomalous. She describes what she sees. The AI suggests seven interpretive frameworks. She eliminates five instantly from domain knowledge. The remaining two prompt a new analysis she had not conceived. The discovery — if it comes — is hers. The prompt was co-created.
Historical source analysis
A historian uploads primary documents and engages the AI in interpretation. The AI offers contextual knowledge the historian lacked about contemporary events in another country. This recontextualization changes the historian’s reading of the document entirely. The interpretation is the historian’s. The context was machine-delivered, human-evaluated, and human-integrated.
Strengthening Professional Practice
Legal argument construction
A lawyer preparing a brief uses AI not to write the brief but to find the weakest links in her own argument. “If opposing counsel were brilliant, what would they attack?” The resulting brief is harder to dismantle — not because AI wrote it, but because AI stress-tested it.
Medical differential diagnosis broadening
A physician describes symptoms. The AI offers fifteen possible diagnoses, including three from rare conditions the physician has never encountered. The physician evaluates each against the clinical picture. Two rare conditions warrant further testing. The diagnosis is the physician’s. The expanded possibility space was AI-provided.
Strategic consulting synthesis
A consultant has twenty interview transcripts from a client organization. Rather than spending three days manually coding themes, she engages AI in an iterative dialogue: “What patterns do you see?” “No, that’s superficial — look deeper.” “Now consider what is not being said.” The insight about organizational silence is hers. The velocity of reaching it is new.
Engineering design review
An engineer describes a structural design. The AI identifies a resonance risk that the engineer had dismissed as negligible. The engineer re-examines, runs a simulation, and discovers the AI’s concern was well-founded. The AI did not design anything. But it prevented a failure.
Teaching preparation and anticipation of student questions
Before a lecture, I ask the AI – “Given this material, what are the ten most likely misconceptions students will bring? And what is the most productive way to disrupt each one?” The teaching is mine. The anticipation of student minds is AI-amplified.
Expanding Cognitive Horizons
Philosophical dialogue as cognitive training
I engage in extended dialogues about epistemology, ethics, or the philosophy of mathematics — not to receive answers, but to train my own thinking. The AI provides resistance, like a sparring partner in boxing. The strength that develops is mine! The resistance was provided.
Language learning through contextual immersion
Rather than memorizing vocabulary, a learner engages in sustained conversation about topics she cares about — in the target language. The AI adjusts complexity, offers corrections, and maintains the thread. This is not a tutor. This is not a replacement for immersion. It is something new – synthetic immersion with infinite patience.
Worldview expansion
I ask – “Explain to me the strongest version of a philosophical position I disagree with.” The AI delivers. I find myself genuinely persuaded on two points I had dismissed. My position shifts — not because the AI convinced me, but because it presented the argument more fairly than any partisan advocate would have.
Emotional intelligence calibration
A manager drafts a difficult message to an underperforming employee. She asks the AI – “How might this be received by someone who is already anxious about their performance?” The AI offers five possible emotional reactions. The manager revises. The empathy is hers. The anticipation of emotional impact was AI-assisted. Or is it? This is precisely the kind of interaction that breaks the assist/replace binary!
Metacognitive reflection
After a long work session, I describe to the AI what I think I learned, what confuses me, and what I suspect I’m wrong about. The AI reflects back patterns in my own thinking that I cannot see from inside. This is not assistance. This is not replacement. This is something for which we do not yet have a word.
The Conference Analogy — Why Smart People Fly Across the World
Consider a phenomenon so familiar that we rarely examine it – why do people pay for plane tickets and hotel rooms to attend academic conferences?
Every technical paper presented at these conferences is available online. The slides will be uploaded. The video will be streamed. We live in an era of essentially universal access to the world’s academic output. And yet, brilliant people who could read every paper from the comfort of their offices spend thousands of dollars and days of their time to physically convene in a single location.
Why?
Because something happens when a group of people with deep expertise in a field gather in the same room and converse. The value is not in the information transfer — that can happen through PDFs. The value is in the network effect of intelligence. A remark by one researcher triggers an association in another. A question from the audience reframes a problem in a way the presenter had not considered. A hallway conversation between two people who had never met produces a collaboration that reshapes a field. The whole becomes dramatically greater than the sum of its parts.
This is what AI is doing for individual cognition — at a scale and speed that conferences cannot match.
But here is what AI cannot do – it cannot replace the conference itself. Because while AI can catalyze individual understanding, the network effect of collective intelligence — when multiple deeply knowledgeable humans engage together, each bringing their own intuition, expertise, and perspective — produces something that neither individual cognition nor AI catalysis alone can achieve. This is why I created spaces like the Convex Optimization Forum, which brings together researchers, professors, and practitioners who recognize the difference between knowing formulas and understanding mathematical truth. The forum meets bi-continentally in Korea and Silicon Valley, creating the very network effects that conferences provide, but with a singular focus – the pursuit of genuine understanding rather than the accumulation of knowledge. AI can catalyze each member’s thinking. But the cross-pollination of insights, the unexpected connections between different domains, the collective pushing past superficial understanding — these require human community.
One Hundred Experts in My Living Room
Let me make this concrete with the most personal example I can offer.
I earned my PhD at Stanford under Stephen Boyd, one singular figure in Convex Optimization! I can claim, without false modesty, that there are probably fewer than a hundred people on this planet who have reached a certain level of understanding and insight in this topic — the level of my criteria, which is not merely being able to solve problems or prove theorems, but having the kind of deep, intuitive grasp that lets you see why the mathematics works the way it does (to the level that I tried to explain in the Convex Optimization Forum)!
One day, in the middle of yet another extended AI dialogue about optimization theory, I had a realization that stopped me cold!
It felt as though all one hundred of those people were sitting in front of me, and I could converse with them ceaselessly.
Not one at a time. Not by appointment. Not with the social friction of academic hierarchy. Not with the limitations of anyone’s patience. But all of them, all at once, available for any question I cared to ask, from any angle, at any depth, and at any hour.
Imagine what happens to your intelligence when you have that kind of access.
I want to be very honest here. Ever since I graduated from Stanford, there was one concept that haunted me! Duality! I could prove every theorem related to duality in convex optimization. The mathematics was never the problem. But you know that kind of understanding — the insight, the moment when the formalism stops being a sequence of logical steps and becomes something you can see, something you can feel in your bones? That understanding, I longed for it for years. I read. I re-derived. I taught. I lectured. And still, the deep intuitive grasp eluded me.
Through sustained, iterative, and relentless dialogue with AI — asking the same question from twenty different angles, challenging every explanation, injecting my own partial intuitions and asking the AI to build on them, correcting its errors, being corrected in turn — I finally arrived at the understanding I had sought for over a decade.
The AI did not understand duality. As I argued at length in “AI (with current DL architecture) does not believe nor reason nor know nor think!”, LLMs do not understand anything in the philosophically or cognitive scientifically meaningful sense. But the process of engaging with an AI about duality produced understanding in me! And that understanding is genuine, deep, and permanent.
Is this replacement? Absurd! The AI has no understanding to replace mine with.
Is this assistance? The word is laughably inadequate! An assistant helps you do what you already know how to do. This helped me achieve what I had failed to achieve for over a decade through every other means available.
This is something else entirely.
This experience — of achieving genuine mathematical understanding through AI-catalyzed exploration — has profoundly shaped my thinking about how we cultivate deep expertise. It is precisely why I created the Convex Optimization Forum, a community for mathematicians, researchers, and practitioners who recognize that there is a profound difference between knowing that something is true and understanding why it must be true. The forum exists for those who have taught convex optimization, published papers using it, implemented successful systems with it — and yet possess the self-awareness to recognize that their understanding remains mechanical rather than deep, one-dimensional rather than multi-faceted. AI can catalyze that journey from knowing to understanding, but the journey requires human curiosity, human judgment, and human community.
The Right Name
If I must name what AI does in this new mode of interaction, I would call it cognitive catalysis!
A catalyst in chemistry does not become part of the final product. It does not replace the reactants. It does not merely “assist” them. It creates conditions under which reactions become possible that would otherwise not occur — or would occur so slowly as to be practically meaningless.
AI is a cognitive catalyst! It does not think for us. It does not replace our judgment. But it creates conditions under which our thinking reaches places it could not otherwise reach — or could reach only over timescales that render the achievement practically meaningless (like my decade-long quest for understanding duality).
And like a chemical catalyst, it remains unchanged. The AI that helped me understand duality did not itself gain any understanding. The AI that helped a researcher generate a hypothesis did not itself hypothesize. The AI that helped a pianist find a harmonic progression did not itself hear the music.
The transformation happens entirely in the human.
This is why the replace/assist binary is not merely imprecise but actively harmful. It directs our attention to the wrong place. It asks us to stare at the AI and decide what it is doing. The real action is happening in the human mind.
Why AI Will Never Replace Humans
And this brings me to what might seem like a paradox – precisely because AI is so extraordinarily powerful as a cognitive catalyst, I am more convinced than ever that it will never replace humans.
Here is why.
Every single one of the thirty examples I described above shares a common structure – the human provides something that the AI cannot.
In the hypothesis generation example, the researcher provides domain expertise that allows her to eliminate four of seven suggestions instantly. In the lecture design example, I provide decades of experience with audiences that tells me where the paradigm shift will land. In the duality example, I provide the very desire for understanding — a desire that the AI does not and cannot have.
As I argued in “Coming Back to the Human in the AI Era — A Sequel to “Why Do We Live?””, AI is a magnificent ally for the scalpel, a useful map-maker for the territory, and an impossible substitute for the gardener. The gardener decides what to plant, what to tend, what to let grow. The gardener brings intuition — that felt sense of what matters, what is beautiful, what is worth pursuing — that has no counterpart in conditional probability estimation.2
Intuition. Aesthetic judgment. The sense that an argument is almost right but not quite. The recognition that a research direction feels promising before you can articulate why. The capacity to be moved by music, by mathematics, by another human being’s suffering. The will to create meaning in a universe that provides none, which I explored at length in “Why Do We Live? – A Wrong Question to Ask”.
These are not things that AI does poorly. They are things that AI does not do at all. And they are precisely the things that make the cognitive catalysis work. Without the human bringing intuition, judgment, lived experience, and the desire for understanding, the catalyst has nothing to catalyze.
The irony is exquisite: the more powerful AI becomes as a cognitive catalyst, the more indispensable the uniquely human contributions become. A better catalyst does not reduce the need for reactants. It increases the value of having the right ones.
Beyond the Binary
In “Is AI Safe or Dangerous? - A Wrong Question to Ask!”, I argued that the safety-danger binary prevents the cross-boundary thinking our era demands. The replace-assist binary does the same thing. It forces a multidimensional phenomenon into a one-dimensional axis and thereby renders invisible the most important thing that is happening.
What is happening is this – for the first time in human history, individual human beings have access to a cognitive catalyst of essentially unlimited breadth, infinite patience, and zero social friction. This catalyst does not replace human intelligence. It does not merely assist it. It amplifies it, provokes it, expands it, and accelerates it — but only when the human brings genuine expertise, genuine curiosity, and genuine judgment to the interaction.
The people who will thrive in this era are not those who use AI as a replacement for thinking, nor those who use it merely as a faster search engine. They are those who learn to engage AI as what it actually is! – a catalyst for cognitive processes that remain irreducibly human.
And the civilization that will thrive is not one that asks “Will AI replace us or help us?” but one that asks: “Are we cultivating the human capacities — intuition, judgment, wisdom, creativity, and the desire for understanding — that make AI’s catalytic power worth having?”
This is not an abstract question. It requires concrete action. It requires creating spaces — educational, communal, institutional — where these capacities are not merely acknowledged but actively developed. It requires recognizing that genuine understanding cannot be automated, but only catalyzed. And it requires building communities of practice, like the Convex Optimization Forum, where the pursuit of deep understanding is valued over the accumulation of superficial knowledge, where self-awareness about what we don’t know is celebrated rather than hidden, and where AI serves as a catalyst for human insight rather than a replacement for human thought.
That is the question worth asking.
And as with every question worth asking, the answer is not found. It is created!
Sunghee Yun
Co-Founder & CTO @ Erudio Bio, Inc.
Philosopher, Mathematician, Thinker, and Universal Truth Seeker
Entertainer, Entrepreneur, Engineer, Scientist, Researcher, Creator, and Connector of Ideas, and, most of all, PEOPLE!
- Strictly speaking, when we interact with, say, ChatGPT or Claude, we are interacting with an LLM-embedded software system or service — not with the LLM itself. This distinction, which I discuss at length in “AI (with current DL architecture) does not believe nor reason nor know nor think!”, is more than pedantic; it shapes how we should think about what these systems actually do. ↩
- As I argued in “AI (with current DL architecture) does not believe nor reason nor know nor think!”, LLMs are at their core sophisticated conditional probability estimators. Their extraordinary outputs emerge not from understanding, belief, or reasoning but from statistical pattern completion at scale. This does not diminish their utility — it clarifies the nature of the human-AI interaction, which is precisely why the catalysis metaphor is more accurate than either “replacement” or “assistance.” ↩