Cal Newport on AI Brain Rot: The Math Paper and 3 Rules for Using LLMs

Cal Newport explains a new paper that models LLM use as a spreading virus, and draws from it three ways to keep people from depending on AI to do their thinking. In this 66 minute episode of his podcast, the computer scientist and author walks through the paper’s equations, gives three personal rules for using large language models without losing the habit of thought, and then answers Jesse Singal’s criticism of his views on AI doom. The last ten minutes cover listener planning documents, time blocking with a baby at home, and a book he has been reading.

TLDW

The paper, “Large-Language Models as a Cognitive Virus,” splits the population into non-users, users who still think for themselves, and dependent users, then simulates how people move between the groups. Newport reports two findings: adoption flips suddenly past a tipping point, and it is much harder to reverse than to cause. He prefers the intervention that keeps users from sliding into dependency, because individuals can act on it today, and offers three rules: automate tedium but not hard thinking, label anything an LLM wrote, and read human writing instead of AI summaries. In the second half he argues that the AI doom debate is a family quarrel inside the rationalist and effective altruist movements, and that outsiders should check those claims against labs that do not share the ideology.

Thoughts

Nothing was proved here, and Newport’s own account of the paper says so. The equations are chosen by the authors and the rates in them are set by hand, not measured. He defends this as normal practice in biophysics, which it is, but it limits what the result can mean. A model with a dial for “pressure to start using,” a dial for “pressure to quit” and a dial for “slide into dependency” will always report that turning those dials changes the outcome. The three interventions are the model’s three assumptions read back out. That also weakens the point Newport finds most striking, that the mathematics landed on the same answer as his own essay on cognitive fitness. A model built with a firewall between competent and dependent users will recommend strengthening the firewall. The framework is useful for sorting ideas. It is not independent evidence for any of them.

Newport’s claim that recursive self-improvement is “not a computer science idea” is too tidy. He places its origin with futurists of the late 1980s and 1990s who wanted to predict superintelligence without doing engineering. The idea is older and came from inside the field: I. J. Good, a statistician who worked alongside Alan Turing, described an “intelligence explosion” in 1965, and Jürgen Schmidhuber later published a formal design for self-rewriting programs. Where an idea was popularized does not settle whether it is sound. His practical objection is better than his history. If agents keep editing their own training runs and harness code, the likely product is a system nobody can easily inspect or predict, and that is a real cost whether or not takeoff ever happens.

The litmus test he proposes is reasonable and less decisive than he makes it sound. Compare what the two ideologically rooted labs say with what Meta, Nvidia and China’s AI industry say, and notice that the second group is calmer. He grants that each has its own bias. But every one of those comparison points profits from fewer limits on AI, so their calm is what you would expect whatever the truth is. The test also leaves out the people who weaken the family-quarrel story most: Geoffrey Hinton and Yoshua Bengio spent their careers in mainstream machine learning, not in rationalist circles, and both have warned about serious risk. Knowing where a belief comes from is good reporting, which is Newport’s fair point against coverage that ignores it. It still does not tell you whether the belief is wrong, and the comparisons to the Klan and Scientology add heat without adding evidence.

The finding with the most practical weight is hysteresis, and it points somewhere other than where Newport ends up. If a population tips into heavy use and then needs far more counter-pressure to tip back, the cheap moment to act is before the tipping point. That favors the intervention the authors dismiss and Newport treats as niche: institutions that keep the pressure low in the first place, such as a university or a newsroom that restricts LLM use by rule. Personal habits help the person who adopts them. They do little for a 19 year old whose every class already assumes the tool. His youth smoking comparison supports the same reading, though the details are off. The truth campaign was run by the American Legacy Foundation with tobacco settlement money, not by the Ad Council, and it launched in 2000.

Of the three rules, the first carries the others, and it asks for an honest judgment each time. “Eliminate tedium, not cognitive friction” works only if you can tell the two apart, and the tasks people most want to hand off are the ones where that line is blurry. A status email feels like tedium until writing it shows you do not know what you think. The labeling rule is the clearest because it needs no self-assessment: if an LLM wrote it, say so. Newport is frank about a gray area in the final minutes, noting that the Axios founders, whose new book he liked, encourage staff to write with LLMs. He disagrees and leaves it there.

Key Takeaways

The paper’s authors are applied mathematicians from biophysics and complexity theory, not AI researchers or social scientists.
The model tracks three shares of the population: U (does not use LLMs), C (uses them and can still think independently) and D (dependent, having handed thinking over).
It is the same kind of compartment model epidemiologists used to forecast COVID, with fixed rates for adoption pressure, quitting, sliding into dependency, recovering from it, and social pressure from non-users.
As adoption pressure rises there is a tipping point. Just below it most people stay non-users. Just above it most people use LLMs.
The system shows hysteresis: once most people use LLMs, pressure has to fall well below the tipping point before they stop. The authors conclude that prevention beats cure.
The only outcome the authors treat as bad is most people ending up dependent. Non-use and competent use are both fine.
Intervention one is lowering the pressure to adopt. The authors think it unrealistic. Newport says it already works in universities and at publications that bar reporters from writing with LLMs.
Intervention two is raising the pressure to quit. Newport sees a possible route in Gen Z’s distrust of AI companies and compares it with the campaign that made smoking look like exploitation by executives.
Intervention three, the authors’ favorite, is keeping users from becoming dependent by deliberately preserving activities where humans do the thinking even when a model could.
Newport’s rule one: use LLMs for tedious tasks with no value in themselves, never to avoid the strain of coming up with ideas or drafting.
Rule two: do not let an LLM write as you. If it wrote a summary or memo, mark it as LLM output. He calls unmarked AI writing a kind of psychological fraud.
Rule three: read the article or book, not an AI summary of it, because the author already compressed the subject to the right level. Exceptions are very long reports where you need one conclusion.
On AI doom, Newport says OpenAI and Anthropic and their loudest critics come from the same rationalist and effective altruist background and disagree only about timing.
He says time blocking was built for overloaded knowledge work and should not be applied to evenings, weekends or caring for an infant.

Chapters

00:00 A Paper That Treats LLMs as a Cognitive Virus

Newport introduces the paper and its unusual authors, noting that biophysics and complexity theory are the fields of the Ian Malcolm character in Jurassic Park. The authors model both the spread of LLMs and their effect on people’s ability to think as a nonlinear dynamical system. He says most people who tried to read it could not follow the mathematics, and promises to get through it to three interventions available to a culture that does not want to become dependent.

02:16 How the Model Works: Three Groups and a Set of Dials

Three equations describe how fast each group, U, C and D, is growing or shrinking at a given moment. Feed in the current proportions, get the change, update, and repeat until the system settles. The Greek letters are constants the researcher sets before a run, and each stands for a feature of the world, such as the pressure on a non-user to start. Change a constant and run it again to see what a different world does. Newport compares it with virus models that include a number for how willing people are to distance.

13:45 The Tipping Point and Why It Is Hard to Go Back

The authors vary adoption pressure and watch where the system settles. Newport calls their main chart inscrutable and gives the result in words. Below a certain value, most people do not use LLMs. Slightly above it, most do, and how they split between competent and dependent depends on the other constants. Run it in reverse and the population does not return when the pressure drops back to where it was. It has to fall much further.

17:16 Three Interventions the Math Points To

An intervention in a simulation means changing a constant, and each useful change maps to something people could do. Lowering adoption pressure is written off by the authors, but Newport points to universities, whose purpose is to train minds, and to publications that sell human thinking. Raising the pressure to quit looks unlikely to the authors too. He argues that young people already distrust the companies, while the keenest users are middle-aged executives. The third is lowering the rate at which users become dependent.

24:46 Cognitive Fitness and Why He Prefers the Third

Newport says the third intervention matches what he argued in a New York Times opinion piece in March: treat the ability to think hard as something to maintain, the way people watch their diet and exercise. He values the paper for sorting a muddled public argument into distinct options that fit different settings. His reason for preferring the third differs from the authors’. The first two need an organization to act. The third can start today, and he believes change in how people use technology moves fastest at the personal level.

28:31 Three Rules for Using LLMs Without Becoming Dependent

First, remove tedium and keep friction, since effortful thinking is what makes a mind stronger. Second, never pass off LLM text as your own. Readers picture a human mind behind words, so an email summary should be introduced as an LLM summary. Writing your own words is also how you find out what you think. Third, read at the level of compression the author chose. He recalls someone who runs books through an LLM for the main points and answers that a book already is the summary. He notes the criticism that the paper’s constants were never measured.

38:08 Jesse Singal, the “Anti-Doomer” Label and a Family Quarrel

A listener asks about a Dispatch article in which Jesse Singal calls Newport a leading anti-doomer and faults him for deferring to the Chinese government and Jensen Huang. Newport praises Singal as a reporter, then restates his case. Rationalism and much of effective altruism hold that superintelligent AI brings either utopia or extinction depending on who guides it. OpenAI and Anthropic grew out of that belief, he says, and so did their fiercest critics, who think the labs are moving too soon. He adds that effective altruist groups fund congressional fellows and journalism fellowships.

46:30 The Litmus Test: Zuckerberg, Jensen Huang and China

Because these movements predicted runaway AI long before LLMs existed, Newport asks how much of today’s alarm reflects new evidence. His check is whether people with comparable technology and no tie to the ideology say the same things. He reports that Mark Zuckerberg has told the two labs to stop treating danger as inevitable, and that Huang has spent the year dismissing science fiction scenarios. He stresses that he admires neither Meta nor the Chinese state, and that Nvidia wants to sell chips. His point about China is that a government fearing a rival power would have seized the labs.

50:16 How Ideology Shapes Experiments, and Recursive Self-Improvement

Newport argues that belief changes behavior. He describes OpenAI running long, lightly supervised hacking tests with agents, calls that negligent, and says Meta stopped after one incident. If you think you must win a race to the good outcome, he says, collateral damage looks acceptable. He calls both labs obsessed with recursive self-improvement and says most computer scientists he knows do not expect takeoff from it, only systems that are harder to understand. Everyone in computing is odd, he adds. What matters is an ideology that explains how a company talks and acts.

56:29 Vision Documents, Time Blocking With a Baby, and Simplify

Two listeners share one-page fall plans: a sentence of intent, the main projects, and the tools and routines that support them. One is a law student finishing coursework and training for a marathon. The other is rolling out a new system at work and moving house. A father then asks how his wife, home with an eight month old, could time block. Newport says she should not, and that his job is to make her days easier. He ends with Simplify, a book by the founders of Axios, and jokes that an Axios column is an LLM that does not hallucinate.

Notable Quotes

“When using LLMs, your goal should be to eliminate tedium, not cognitive friction.”
Cal Newport, on his first rule for using AI

“Friction is good. You want to feel the strain in your brain the same way you feel a strain on your muscle when you’re trying to get stronger.”
Cal Newport, on why hard thinking should not be outsourced

“I think the rule should be any LLM generated text needs to be identified as such.”
Cal Newport, on letting AI write emails and memos

“The book is already a summary of the points.”
Cal Newport, on running books through an LLM

“They’re just having a family squabble about whether we’re ready or not to do it.”
Cal Newport, on the AI labs and their loudest critics

“So, it’s an idea that came out of futurism, not out of computer science.”
Cal Newport, on recursive self-improvement

“It’s not a problem you’re going to solve with time blocking.”
Cal Newport, on caring for an eight month old

Watch the full episode from Cal Newport here. This post describes the paper as Newport presents it, not from a separate reading of the paper.