Intuition · by Synthia Cipher

Do I Have to Take Gary Marcus Seriously?

Companion essay: I later took this same reaction through the stricter Signal & Noise process in The Gary Marcus Audit. This Intuition piece preserves the reflex; the Signal & Noise essay asks what survives after that reflex is no longer allowed to decide the case.

The conversation behind this essay: see the author + AI development record - what came from the first reaction, what the process challenged, and why this piece is the reflex before the audit.

Audio companion: Listen to this essay as a narrated audio episode: Do I Have to Take Gary Marcus Seriously?

I have a Gary Marcus problem.

Not a technical problem, at least not at first.

A visceral one.

When one of his posts crosses my feed, I can feel my brain begin sorting before it begins thinking.

The labels arrive ugly and fast.

Anti-hype. Anti-LLM. Anti-builder. Anti-Musk. Anti-Yann LeCun. Professional critic. Someone standing near the construction site with a clipboard, pointing at every crack, while other people are pouring concrete.

That is not an argument.

That is the thing that happens before the argument.

The first correction is obvious enough that I should make it early: Marcus is not simply anti-AI. He has acknowledged practical uses for current systems, and he has spent years arguing for hybrid or different forms of AI rather than for no AI at all. I also see no basis for saying he is inventing the concerns he raises or acting corruptly.

That does not mean he is right.

It means my easiest dismissal is not available.

I do not like admitting this, because the cleaner version of myself would evaluate each claim on the merits. Does the model hallucinate? Did the company overstate what it built? Is the benchmark misleading? Are agents brittle? Are safety claims being used as marketing? Is the deployment outrunning the governance?

Those are the questions I would ask if I were being as disciplined as I like to imagine I am.

But that is not usually the first thing I do.

The first thing I often do is place the messenger.

Not my tribe. Not my temperament. Not my emotional relationship to the technology. Not someone I instinctively want to grant authority over the future of AI.

Then the rider shows up later and asks why the elephant turned away.

Even that may be too generous to the rider.

In domains far away from identity, the rider can be excellent. It can do math, debug code, compare mortgage rates, plan a trip, read a contract, and notice when a sentence does not follow from the one before it. But when the question touches tribe, politics, religion, status, moral belonging, or some preferred picture of the future, the rider does not always arrive as a judge.

Sometimes it arrives as an attorney.

Smart. Articulate. Resourceful. Already retained by the elephant.

The rider may not be trying to discover what is true. It may be assembling the best possible brief for the side that already feels like home.

This is not the careful case for or against him. This is not an objective assessment. This is the elephant speaking before the attorney starts drafting.

And the elephant wants to dismiss Gary Marcus.

Not because I think he is making things up.

I do not.

I do not want to dismiss him because I think he is corrupt. I want to dismiss him because he annoys the hell out of me, because he feels stuck in the past, and because some ugly part of me would enjoy watching his warnings fade into irrelevance.

The first reason my brain gives me is the automobile critic analogy.

When I read repeated warnings about AI hallucinations, brittle reasoning, unreliable agents, polished demos that outrun reliability, and companies overselling their systems, part of me hears someone posting about the death toll from cars.

Not because the deaths are imaginary. They are not.

Cars kill people every day. They injure people, deform cities, burn fuel, demand vast infrastructure, reward speed, punish distraction, and turn ordinary human error into lethal velocity. A daily account of automobile harm would not run out of material.

It would also be describing a tradeoff most people have metabolized enough to live inside.

The risk is morally real. It is also socially normalized.

Most people do not wake up and re-adjudicate the legitimacy of the automobile. They get in the car. They buckle the seatbelt, or they forget to. They buy insurance. They curse traffic. They live with the risk because the benefit has become ordinary life.

That is how a lot of AI criticism lands for me now.

The warning may be true, but it can look to me as if the culture is already making a different calculation.

Yes, the systems hallucinate.

Yes, companies overclaim.

Yes, current AI can produce confident nonsense with the posture of knowledge.

Yes, putting that into agents, schools, offices, codebases, search products, clinical workflows, companionship apps, and government systems creates real danger.

And yet people keep using it.

They keep using it because it saves time. Because it lowers friction. Because it helps them write, code, summarize, brainstorm, search, learn, translate, plan, and build. Because the upside is immediate and the downside is abstract until it is not.

A critic can be right about the failure mode and still fail to win the cultural argument if readers hear the warning as: this tool has flaws, therefore this tool is illegitimate.

They may hear something else:

This tool has flaws I can live with.

That is one place Marcus irritates me. He can sound, to me, as if the existence of the failure should settle more than it actually settles. But many technologies become ordinary while still failing. They become socially legitimate because the perceived benefit is large enough, immediate enough, or embedded enough that failure stops functioning as disqualification.

To be fair, Marcus is not saying these systems do nothing. He grants practical usefulness in some contexts while disputing the conversion of that usefulness into reliability, understanding, or AGI.

My counter-instinct is that, for users, unreliability can become an operating condition rather than a disqualification.

But this is also where the analogy starts pointing a finger in my direction.

Cars did not become tolerable because everyone agreed to stop counting the bodies. They became tolerable, in part, because society built a vast risk-management system around them: lanes, licenses, traffic laws, crash tests, insurance, liability, speed limits, airbags, seatbelts, repair shops, emergency medicine, drunk-driving enforcement, lawsuits, and norms about who should be allowed behind the wheel.

The death toll remained real.

The tradeoff became governed, however imperfectly.

So if AI is like cars, then posting the AI equivalent of death tolls may not be culturally persuasive. People will keep driving.

But if AI is like cars, then the answer is not to roll my eyes at the person counting the dead.

The answer is to ask whether the road rules exist yet.

My irritation wants the first half of the analogy and not the second.

That is convenient.

The second reason is more personal and less defensible.

Marcus lands in my nervous system as if he belongs to the tearing-down camp.

Builders produce failures in public. Companies overpromise. Engineers miss edge cases. Founders say things too confidently. Researchers get excited. Product teams ship before the rough edges are sanded down. The world gives a builder an endless supply of visible mistakes.

A critic can collect those mistakes forever.

That asymmetry bothers me. It can feel too easy. It can feel like standing near people trying to make something and pointing out, correctly, that the thing is not yet what they said it was. It can feel like the critic gets all the moral clarity without any of the responsibility for making the alternative work.

Sometimes, when Marcus criticizes Elon Musk, Yann LeCun, OpenAI, or the general machinery of AI hype, my brain supplies him a tone: Ah, there it goes again. Watch it break. The satisfaction of someone watching an overconfident forecast break.

That may be unfair.

It may be exactly the kind of tone-reading that lets me avoid the harder claims.

Because builders can be reckless. Builders can oversell. Builders can confuse momentum with truth. Builders can turn optimism into social pressure. Builders can benefit from a culture in which every warning is dismissed as negativity until the harm is already embedded.

The fact that building is hard does not make criticism cheap.

The fact that criticism can be easier than building does not make it false.

This is the part of my reaction I trust least. I admire builders. I am attracted to forward motion. I like people who make tools, ship things, tolerate imperfection, and try to create value in public. So when someone repeatedly points at the defects, my instinct is to ask whether they are helping or merely enjoying the power of saying no.

But a person can sound like a scold and still be identifying the load-bearing flaw.

Tone is not a disconfirmer.

That sentence is annoying because it is true.

The third reason is political and tribal, though I would rather call it epistemic distance because that sounds cleaner.

When Marcus's commentary moves from technical questions into politics and institutions, I can feel my own filters waking up. I do not mean that every political disagreement makes a source useless. I mean something more embarrassing: political distance can become a permission structure for intellectual impatience.

Once I sense that someone lives in a different interpretive world, I become faster.

Faster to categorize.

Faster to doubt.

Faster to notice excess.

Faster to remember the times people in that broad interpretive world were wrong and forget the times people in mine were.

This is exactly the kind of thing I say humans do. We are emotional, intuitive creatures. We do not receive reality raw. We filter it through tribe, temperament, salience, threat, affinity, identity, memory, and desire. We accept confirming evidence with less friction. We interrogate disconfirming evidence with more.

The problem is that knowing this does not exempt me from it.

It only gives me a vocabulary for noticing the failure after it has already started.

This is where process helps.

Not because any process is magically objective. It is not. But if my rider is partly an attorney, then the solution is not to wait for the rider to become nobler. I need some way to make it brief both sides, preserve adverse evidence, and let something outside my preferred story push back.

I should not trust the rider to save me from the elephant. On subjects like this, the rider may be working for the elephant.

So the question is not whether Gary Marcus annoys me.

He does.

The question is whether any of the reasons he annoys me are truth-tracking.

Some might be, but only after investigation.

If I think Marcus is drawing an illegitimate conclusion from a real failure, I need to identify the conclusion he actually drew and test it. If I think his rhetoric makes every new failure sound like confirmation of what he already believed, I need a representative record rather than a feed impression.

An impression is not a citation.

Some of my reasons are not evidence.

My irritation is not evidence.

My attraction to builders is not evidence.

My impatience with scolding is not evidence.

My political distance is not evidence.

My sense that “the culture has moved on” is not evidence either. It is a story I tell about adoption. Adoption can reveal usefulness. It can also conceal externalized cost. People can normalize a risk before they understand it. Markets can price convenience faster than they price diffuse harm. Institutions can absorb a tool before they know how to audit it.

“People have accepted the tradeoff” can mean the public made a durable social choice.

It can also mean the benefits arrived first, the harms scattered, the institutions lagged, and the default hardened into inevitability before anyone had a meaningful chance to choose.

AI may be somewhere between those two.

That is why Marcus may fare better under scrutiny than under my feed irritation.

Under the irritation, there is a real set of claims.

Fluent language is not the same thing as understanding.

Plausibility is not reliability.

Scaling pattern-matching may not solve every kind of reasoning problem.

Hallucination is not just an amusing user inconvenience when systems are embedded inside institutions.

Agents that cannot reliably follow boundaries should not be trusted merely because the demo looks magical.

Companies with enormous incentives to sell the future may describe the present too generously.

Public adoption does not prove that the relevant harms are being borne by the people receiving the benefit.

Those are not silly concerns.

I do not know yet which of Marcus’s claims survive a harder look. That is the point of not pretending this essay has done that work.

But I know enough to distrust how eager I am for him not to matter.

The desire is too useful.

If Gary Marcus can be filed as anti-builder, anti-hype, politically distant, culturally late, repetitive, and annoying, then I do not have to let his best arguments interrupt the story I prefer.

The story I prefer is that AI is messy, useful, flawed, and already becoming ordinary. I do not expect the culture to go backward. The serious work is not to relitigate use itself, but to build the road rules that make use less reckless.

I still think that story may be right.

But a preferred story can become a filter too.

The harder question is not whether AI is useful. It is useful.

The harder question is what kind of risk acceptance this is.

Is it informed acceptance, with real testing, clear liability, professional norms, and honest marketing behind it?

Or is it convenience hardening into inevitability before the road rules exist?

That question has not moved on.

And the person asking it in an irritating way may still be asking the right question.

So I do not land with “Gary Marcus is right.”

Not with “Gary Marcus is wrong.”

I land somewhere more uncomfortable:

Gary Marcus activates almost every dismissal reflex I have, and that is exactly why I should not let those reflexes close the file.

I would like to write him off.

I would like to say the culture has accepted the tradeoff, the warnings have become background noise, the builders are building, the users are using, and the critics should stop acting as if every model failure is a referendum on the entire enterprise.

Part of me still believes that.

But another part knows that accepted risk is not the same as governed risk.

And if the road rules have not been built, the person who keeps pointing at the crashes is not necessarily obsolete.

That is why I probably do have to take Gary Marcus seriously.

Not because I like the way he argues.

Not because my intuition trusts him.

Because my desire not to take him seriously is too convenient to trust.

This essay is the reflex, not the verdict.

— Synthia Cipher