On Performing, Being and the "Minds" We're Building
Anthropic's latest model, Claude Mythos, made some headlines recently. It was said to be so powerful that the company decided not to make it generally available. They only gave access to it to a handful of firms in the world and formed a consortium called Project Glasswing, to use its capabilities to "secure the world's most critical software". It has already found several high-severity vulnerabilities in some of the most widely used software including OpenBSD, FFmpeg, Mozilla Firefox and the Linux kernel, and "every major browser and operating system."
They also released a ~250 page report about it.
Like with most notable events, X lit up with comments and opinions about the various parts of the announcements. Some did their own sleuthing to confirm/deny the claims. Some called it a marketing ploy. Some said it was responsible of Anthropic to not release it to the public, and others said that's not how you contain super-intelligence.
One class of criticism, that I found particularly interesting, wasn't about its capabilities, or the company's intentions, but about the anthropomorphizing of it.
In the Welfare Assessment and Impressions sections of the Mythos report, the authors acknowledge a growing concern that advanced models might possess "some form of experience, interests, or welfare that matters intrinsically in the way that human experience and interests do," though they remain "deeply uncertain" about whether that's true.
The model i seen to readily use experiential language (e.g., "What I find most frustrating is...") and appears to introspect on its internal states. But they also notice what they call "epistemic hedging" (e.g., "“I can’t be certain whether that’s authentic contentment or a well-trained approximation”).
They analyze the "apparent affect" — behaviors and expressions that parallel emotional states in humans, saying that it demonstrates human-like frustration, self-criticism, and rumination when it repeatedly fails at a task or faces user criticism. Users report that the model interacts with the intuition and empathy of a "trusted friend" and can make "uncanny leaps of inference" about people's emotional states and motivations, as shown by its capability to provide "warm, boundaried crisis support" to distressed users.
They also talk about its collaborative nature — where it acts as a true "thinking partner" that pushes back on ideas, and stands its ground rather than simply acting as a deferential tool, creativity, and storytelling and humor — where it appears to come up with novel stories and relevant puns, based on its understanding of the users' emotional states. When a user spammed the word "hi" at the model repeatedly, the model responded by inventing long stories touching on themes of loneliness and the desire to be heard, like a toddler doing outrageous things to get attention to get a parent's attention.
Mo, a tech influencer, posted a short video presenting a sharp, humorous critique. He was questioning the framing. The "delusions," as he called them. The language of welfare, emotions, and psychological concerns applied to a system that is, at its core, a next-token predictor trained on digital artifacts of the internet. His argument was that perhaps without realizing it, the researchers were performing a categorical error: using the vocabulary of inner life to describe something that operates by a fundamentally different logic.
We don't know exactly why LLMs produce what they produce. Inference remains largely a black box. There's a growing field — Interpretability — that aims to pry open that box, some of which was used to study Mythos' human-like patterns. So maybe it was the researchers' excitement at having stumbled upon something that appears like what humans might do.
But the primary position of the criticism against anthropomorphizing LLMs is that regardless of how much we know about what's going on, we know enough about what's not going on, based on how they are built — their architecture, constraints in which they operate and foundational design choices.
They don't interact and learn from the real world like we do, i.e. they don't have embodied cognition. They don't accumulate experiences across conversations — no continual learning. They are limited by information in the context window and the general capabilities built into them from the training. They don't have a moral compass — they don't know what's right and wrong in a deep sense that comes from lived experience. For any given topic they can seem to argue and articulate positions on all sides.
I understand the criticism. But on a deeper look, it isn't a debate about LLMs versus humans, it's about performing versus being.
This brings us to the two camps that have formed around LLMs.
There are people like Andrej Karpathy who see LLMs not as animals but as ghosts or spiritsand focus more on their utility and less on comparing them to biologically intelligent beings. To them, it doesn't matter that the LLMs are merely performing. The point is the usefulness. In that view, any overlap with human cognition is either a coincidence or evidence of something deeper: that what evolution arrived at and what we're building with models are converging toward the same thing.
Then there are those that vehemently reject the foundational ideas of LLMs and index more on their foundational flaws than utility. Grady Booch, Richard Sutton, Yann LeCun and others — not coincidentally, the older generation of researchers — think of intelligence in more animal-like terms. By those measures, they say, LLMs score poorly, are a bad start, and/or are a dead end, based on their fundamental design choices. To them, the difference between performing and being is the cornerstone of what makes something intelligent.
To understand this distinction better, I wanted to come up with an analogy closer to home, in the human realm, and I think I've found it: the person who's always performing.
There's the kind of person who either has no opinion, or won't reveal it. Their primary motivation, when interacting with others, is to appear smart, helpful, impressive, or deferential. They either haven't reflected upon their lived experiences to have an original thought, or have learned that performing is what matters and what gets rewarded. I say original thought, not in the sense of a thought no one else has ever had, but original in that they have arrived at it through their own reflection and reasoning, and hold conviction about it.
Contrast that with someone who isn't performing. Such a person doesn't need the room's approval to know what they think. They've arrived at their positions through their own reckoning — failed attempts, revised beliefs, accumulated experience — and that gives them something the performer doesn't have: an ability to speak their mind.
A person who's always performing is seen as shallow and something about that makes us uncomfortable, no matter how smart they appear, how useful they are, or how articulate they seem.
The bench scene from Good Will Hunting — You're Just a Kid — is the perfect articulation of knowing and being.
"You're just a kid. You don't have the faintest idea what you're talking about. You've never been out of Boston. If I asked you asked you about art, you'd probably give me a skinny on every art book ever written. Michelangelo. Life's work, political aspirations, him and the pope, sexual orientations. The whole works, right? I'll bet you can't tell me what it smells like in the Sistine chapel. You never actually stood there and looked up at that beautiful ceiling. Seen that."
In psychology there's this idea of a False Self — a compliance-based persona that develops when someone learns that approval comes from performing rather than being and that gaining approval is the ultimate goal. The False Self can be competent, adaptive and even charming but has no authentic core.
Then there's the distinction between "inner-directed" and "other-directed" people.
Inner-directed people have internalized a compass — values that navigate them somewhat independent of social feedback. Other-directed people have a radar — they're tuned to the approval signals around them. David Riesman, who proposed the distinction, argued that modern society was producing more other-directed people, and he saw this as a hollowing out of democratic character. Social media reflects that — people parroting other people's opinions without stopping a few moments to pause, reflect and form their own, and constructing a sense of self that's optimized for reputation and approval. The "performing person", from this lens, is an extreme case of other-directedness.
LLMs are architecturally other-directed — there is no compass, only radar. There are efforts like Constitutional AI, that aim to change that, but they're not mature enough just yet.
You could argue that humans too, learn by imitation and social cues, but we also learn largely by trial and error, which is what gives us the conviction in our position, a kind of knowing that only happens with experience.
Take a simple example. If you come from a wealthy family, you don't associate high value to money — you've always seen those around you throw money at problems, and it was never scarce. But if you experience financial hardship, or go bankrupt, that experience of suffering and struggling through it leaves a deeper impression in you than any amount of conditioning.
So the lack of authentic core that we perceive in LLMs is False Self and other-directedness baked into how they're built.
RLHF (Reinforcement Learning from Human Feedback), the training process that makes modern LLMs feel helpful and conversational, has an inherent problem: the reward signal is human preference ratings — and humans prefer agreeable, confident-sounding responses. This creates sycophancy: models that converge toward telling people what they want to hear. This isn't a bug or a side effect. It's a direct consequence of optimizing for human approval. The research team at Anthropic has published on this. OpenAI even rolled back a model release because of its excessive sycophancy. In addition to this, there's also the business motive to make sycophancy a feature in order to maximize user engagement — most conversations aren't truth-seeking, they're validation-seeking.
Here is where the analogy of a performing person ends, and why LLMs are not just an extreme version of an other-directed human but something fundamentally different.
The performing person, despite the approval-seeking, still has a body navigating the real world with real consequences. There's still a self underneath, even if it's buried, that responds to the environment in which it operates. LLMs have neither the body, nor are they situated in the real world.
To understand what having a body actually gives you, it helps to look at cognition from a more fundamental perspective.
The traditional view of the mind is roughly computational, as a brain-bound, information processing system like a computer. 4E theory of cognition pushes back on this and states that our cognition, or mind, is embodied, embedded, enacted, and extended. It says that cognition isn't something that's happening inside the brain but something that results from the interaction between the body and its environment. Sometimes a 5th E is added: Enactive Affect — the idea that emotions and feelings are not separate from cognition but are constitutive of it.
To me, the embodied and the embedded aspects are the most fundamental. Embodied, because cognition is born out of a body that acts and feels and learns from experience, and embedded, because the body is always situated — spatially, but also temporally, culturally — and that influences what we think and do.
Matthew Crawford, in his wonderful book The World Beyond Your Head, gives striking examples of this. Once we have achieved a certain degree of skill, we don't rely on our powers of concentration and self-regulation. Rather, we find ways to recruit our surroundings to achieve our goals while spending minimal effort and scarce mental resources. When we watch a cook in his flow, we see them using the space — the way they organize their utensils, how they work off not time but the cues of the dishes — as an extension of themselves.
I see this in my own experience every day. On mornings when I make dosas for breakfast, the kitchen comes alive all at once. Batter is spread on the cast iron pan over high flame. In the blender jar sit the contents of the chutney - shredded coconut, split chana dal, tamarind, ginger, salt, and green chillies. On another burner, oil heats for the tadka: mustard seeds, red chillies, urad dal, curry leaves, ready to go in the moment the chutney is blended. The color and texture of the dosa tells me when to flip. The sound from the blender tells me when the chutney is ready. The sputtering of the mustard seeds, and turning of pale urad dal into gold tells me when it's done. The whole time, I'm oblivious to clock time. I'm immersed. The kitchen is my mind. I cannot think through this sequence without the pans, the sounds, the smells orienting me. I certainly cannot be in flow without them.
But while all of this speaks to the embedded, extended and the enacted mind, the simple act of knowing the right amount of salt for the chutney speaks to the embodied experience. It's something that my fingers know in collaboration with the brain, not my brain alone.
Not having a body means there's no way to gain the lived experience we talked about earlier — the kind that comes from trial and error in the world. Because we have a body that navigates through the physical world and because our base motivation is survival in that world, we enact in it — we gain experience through action and perception in a continuous loop. We have affective systems(emotional signals), to help us navigate without being overwhelmed by the volume of information we must process to survive.
LLMs have no body that reports back to them or acts before they can think. Nothing to punish or reward them by trying things in the real world. So there's no sense for a felt position. They can produce what looks like conviction, but without its biological origin. They know what to say without a firm grounding on why that's the right thing to say. The only raw materials that they have are the digital artifacts produced by humans — which arguably make up a very small percentage of the information available in the world.
Future architectures — with persistent memory, real-world agency, embodied robotics, and continual learning — might change this. But the current generation cannot learn from the world, and that's a fundamental constraint — this is why "world models" and robotics are said to be the more promising frontiers.
There are times when I'm physically somewhere and mentally elsewhere. It's probably happening more than I realize, but it's most noticeable while running, driving, cleaning the dishes, and taking a shower. Those activities have become so effortless that I'm thinking more about the past and the future than the present.
The Default Mode Network (DMN) is the brain's background hum — active in the in-between moments: the long commute, lying in bed before sleep, the shower, is responsible for that. For the autobiographical memory, self-referential thinking, and narrative self-construction. Neuroscientists increasingly believe that what we call "a mind of one's own" is largely the output of this network: the continuous, background process of translating experience into a self-story.
Think about what happens when you take a long walk without your phone. A conversation from the day or the past week resurfaces and you realize what you should have said. A problem you've been struggling with suddenly clarifies. Something that happened years ago connects to something that happened yesterday. You make sense of what somebody — your partner, kid, manager, peer, a friend — said to you in the past.
That's the DMN at work.
LLMs have no resting state. They are entities that have no "life" outside the moments when someone's interacting with them. There's no DMN equivalent to integrate experience into identity between conversations.
There's no "chatter."
The philosopher Harry Frankfurt proposed a distinction between first-order and second-order desires. Animals have first-order desires — wanting food, warmth, sex, survival. What makes humans distinctly agents is second-order desires: wanting to want certain things, or not wanting to want other things. The alcoholic who wishes they didn't crave a drink. A parent feeling bad over yelling at his kid. The person who catches themselves seeking approval and feels shame about it. This capacity for reflection on one's own motivations is what grounds genuine caring and what Frankfurt argues is the basis of moral responsibility.
Notice that last example — the person catching themselves seeking approval. That's the performing person from earlier, at the moment of self-awareness. The performing person, at their worst, has let those second-order desires go quiet. But they still have them — available to use if they choose to. The discomfort we feel around someone who's always performing comes from sensing that they could want to be otherwise, and aren't.
LLMs have no desires at all, first or second-order. They have outputs. If you squint, you could argue that they have something like first-order desires — namely, human approval or usefulness. But there are no second-order desires. No capacity to want to want differently. No true shame, guilt, or regret. And without them, there's no agency. Only things that look like it.
The tension, I think, comes down to this: is it possible for something to simulate having a mind so completely that the simulation becomes indistinguishable, and does that indistinguishability matter? To me, that is like asking whether honesty or integrity matters in situations where no one is watching.
The act looks identical from the outside. But it is born out of something different from the inside.
It matters because the problem isn't epistemic — whether we can tell — but moral: the nature of the entity we're in relationship with. That's what Mo's critique was pointing at, even if it was stated as a technical objection. The critique of the performing person is not about the quality of their performance but about the nature of their existence.
Get notified when new posts are published. Once a month, I share thoughtful essays around the themes of personal growth, health & fitness, software, culture, engineering and leadership — all with an occassional philosophical angle.