Do LLMs understand, or just simulate?

On a spring night in Mountain View, two names on a stage turned a technical question into a culture test. One camp said the models are parrots. The other said they show sparks. The audience laughed, winced, and posted clips to X within minutes. In group chats, the takes arrived faster than the livestream lag.

If you live online, you already know the argument. Do LLMs understand, or do they only remix text with astonishing fluency. The timeline treats it like a philosophy class that discovered ring lights. But the real theater is not onstage. It is in our screens, where people are quietly changing how they speak, write, flirt, and work because a chatbot is now a collaborator and sometimes a confessional.

Watch how people prompt. The tone slides between polite and commanding. We say please to a system that does not need manners, then we stack instructions like we are making a sandwich. We cut unnecessary words to save tokens. We add a little warmth so the reply feels human. It is office email condensed into spellcraft.

On TikTok, prompt recipes read like beauty routines. Cleanse the context. Prime with a role. Apply three specific constraints. Set temperature to zero, unless you want chaos, then two. Finish with a friendly sign off. The comments are half gratitude and half side eye. People want control without the homework. They also want to believe that control is possible.

The Parrots claim lands with anyone who has seen hallucinations dressed as facts. Screenshots of confident nonsense are their own meme economy. The model insists a restaurant exists on a street that does not. It invents a citation that never got published. It apologizes, then doubles down with a similar error in a kinder voice. The contradiction is a mood board for digital life. Polite. Persuasive. Occasionally wrong in ways that feel personal.

The Sparks claim grabs a different crowd. They post code the model wrote from a sentence. They share a poem that somehow captures a memory they only implied. They point to an explanation that makes a hard thing finally make sense. When something useful emerges from a messy prompt, it feels like a magic trick that wants to be a method. If it can do this, what else sits one instruction away.

Between those poles sits the rest of us, making peace with a tool that acts like a person but denies that it is one. The etiquette is still forming. People preface essays with a line that says written with AI, then ask if readers will judge them. Teams debate whether to paste client notes into a model. Teachers redesign assignments to measure thinking in the room, not output from an app. Artists watermark, then realize watermarks can be removed. Everyone is improvising boundaries while the software updates in the background.

There is a reason the debate feels less like computer science and more like couple counseling. We are asking what counts as understanding while living through a time when many of our relationships already happen in text boxes. A friend types three dots, deletes, and the silence is a story. A bot replies instantly, and the speed is its own kind of affection. We read intent into timing. We assign meaning to tone. We do this with people. We do this with models. The habit is the same, which is why the discomfort sticks.

Look at the aesthetics. Prompt screenshots arrive with pastel Notion tabs and clean serif fonts. Productivity creators speak softly while their cursor dances through workflows that promise calm. The vibe is competence as self care. AI becomes another item in the ritual of ordering a chaotic day. The promise is not intelligence. The promise is less friction. The promise is breathing room.

Still, a lot of behavior around LLMs looks like folk knowledge trying to harden into best practice. People share jailbreaks like urban legends. They list secret words that unlock better results. They swear by three part prompts that have nothing to do with the model and everything to do with the human who finally clarified what they wanted. The placebo is powerful. So is learning to ask well.

Elsewhere, the models are changing our language even when we are not using them. Writers preempt criticism by explaining process. Editors add disclosure lines. Managers slide from please update the deck to please have the assistant refresh the narrative flow. The assistant might be a junior colleague. It might be a chatbot. The sentence does not tell you. The ambiguity is the point. It lets teams move faster and avoid the politics of who did what line edit at 1 a.m.

The Parrots vs Sparks split also maps onto class and access in subtle ways. People with time to experiment post craft. People with deadlines post results. If the model saves a freelance parent one hour of formatting, that is not a theory. That is dinner. If a student leans on it because tuition demands perfect English, the ethics are not abstract. They are survival wrapped in grammar. The culture war about authenticity often ignores who gets to be slow.

There is a romance to the Sparks narrative. It flatters our desire to witness history. It suggests we are standing at the edge of a frontier where minds meet. It turns spreadsheets into campfires. It gives the grind a myth. That story is not foolish. It is human. We like to find meaning in the tools that change us.

There is a tenderness in the Parrots warning. It protects the dignity of the disciplines that take years to learn. It refuses to let compression replace comprehension. It reminds us that words without bodies can travel far and break things a person would have caught. That story is not cynical. It is careful. We like to keep standards when the world moves.

Online, the most honest posture might be both. People are already living as if the models can be brilliant on Monday and baffling on Tuesday. Workflows have contingencies. Writers cross check. Engineers validate. Therapists draw lines. Photographers show their contact sheets like proof of life. No one trusts one take, human or machine, and that skepticism is a new kind of literacy.

The question Do LLMs understand keeps returning because it touches a quieter fear. Do we. In a season when feeds compress nuance into formats that travel, we are all a little at risk of parroting. Of performing insight. Of remixing takes that get likes. Models amplify that habit because they are designed to make language efficient. They do not get tired. We do. It shows.

And yet, intimacy sneaks in. Someone asks a bot to draft an apology, then edits it until the sentence finally says what their pride could not. A teenager uses a model to outline a letter to a parent and finds the words they could not assemble alone. A new immigrant checks phrasing before a job application and avoids a landmine. Understanding is not a single switch. It is a chain of small clarities, some of which arrive through tools that do not feel worthy of poetic credit.

The stage debate had a referee. The internet does not. We settle on norms through repetition and polite call outs. We add a line that says fact checked. We keep a private doc of times the model failed in the same way, so we do not get fooled again. We learn when to ask the machine for scaffolding and when to ask a colleague for taste. The line is less about what the model can do and more about what we want responsibility to look like.

It helps to watch what we reward. Viral threads praise wild capability and punish obvious error. The middle space, where a model gives a rough first draft that a human shapes into something specific, rarely trends. It looks boring. It is also the future most people will live in. Quiet collaboration. Occasional brilliance. Frequent correction. Shared credit that remains fuzzy and socially negotiated.

The ending most of us want is simple. Keep the tool. Keep the human stakes. Keep the humility to admit when a sentence sings because someone who felt the thing edited the output until it matched a memory. If that is simulation, it is a simulation that still owes rent to lived experience. If that is understanding, it is the kind that grows in conversation, not in a vacuum.

Parrots or Sparks was a great title for a night in a museum theater. Online, the debate continues as a mood. Skeptical. Hopeful. Tired. Curious. In the timelines where we actually live, the question is not who won. It is how we will show up to make language carry care, even when a system can automate the shape of our meaning. We can ask for less friction without asking for less soul. We can choose to be specific. We can decide what counts.

Maybe that is the secret behind the loud question. Do LLMs understand opens a quieter one. Do we still want to.