Skip to content
•
16 min read

Clear thinking in the age of AI: Nick Saraev's course

Talk

The most important video you'll ever watch in the age of AI by Nick Saraev, his YouTube channel, Sep 2, 2026 (5 hr 2 min)

Watched

These are my notes on what the speaker said. Names, numbers and quotes are as stated in the talk, and I haven't checked them independently.

Key ideas

  • People who seem to have more willpower mostly face fewer temptations, so plan to remove the temptation, never to resist it.
  • Set your own defaults once while you're clear-headed, and use friction: add steps in front of what you want less of and remove them from what you want more of.
  • Fix the body before anything clever. He ranks a sleep-disorder check, regular sleep, a fixed wake time, cardio fitness and bloodwork as the biggest gains.
  • Write the probability down. Expected value, Bayes and Fermi estimates turn a gut feeling into a prediction you can check later.
  • Tie each new habit to an obvious physical trigger as an if-then rule, and when one doesn't fire, fix the trigger rather than blaming yourself.
  • Ask AI to attack a plan, not approve it, and write the final decision and its reasons in your own words.

Nick Saraev’s five-hour Clear Thinking Course argues that thinking clearly is mostly a matter of designing your setup, not willpower. Decide once while you’re clear-headed, and let defaults and friction carry the tired version of you. He builds it in three layers: hardware (sleep, caffeine, exercise, air and a quieter workspace), software (eighteen mental models, from expected value to explore-exploit), and five techniques that turn knowing into doing, such as trigger-action plans, pre-mortems and five-minute timers. These notes keep the methods and the figures he gives, so you can look one up without rewatching.

The percentage gains are his own estimates, which he says combine the research with his own reasoning, and the study figures are as he states them. He’s clear that none of it is medical advice.

Course structure

He frames thinking as three abilities you can raise: attention (what gets in), working memory (what you can hold, about four meaningful items at once) and executive function (what decides where effort goes). The course runs in five blocks.

BlockStarts atWhat it covers
Framing6:05How the brain spends energy, chunking, willpower, defaults and friction
Hardware1:03:40Sleep, caffeine, alcohol, diet, fitness and workspace
Software1:52:17Eighteen mental models for decisions, time and planning
Techniques4:05:08Trigger action plans, noticing, pre-mortems, five-minute timers, AI
Recap4:55:32Every section in about six minutes

He suggests watching once at 1.5x, then again at normal speed, and answering the questions he poses at the end of each section, because recalling the material is what makes it stick.

Willpower, defaults and friction

His starting claim is that willpower isn’t a tank that empties. The brain uses roughly the same energy whether it’s working hard or idling, and people who seem disciplined mostly arrange not to meet the temptation. So any plan step that reads “I just won’t do it” gets rewritten to remove access instead. His own example is quitting cannabis: willpower failed repeatedly, and what worked was no longer seeing the friends he smoked with.

Defaults. Most of the defaults in your day were set by someone else for their own goal: autoplay, notifications on, the annual plan pre-selected. He cites the 2003 organ donation comparison, where Germany (opt in) had 12% of people registered and Austria (opt out) 99.98%. His diagnostic is to go through every transition in your day and ask what happens if you do nothing. If the answer benefits a company, change it. Setting your own default takes one clear-headed decision and then pays out on every tired morning. His own: at 5 a.m. his computer shows one browser, the bookmarks bar lists the day’s work in order, other sites are blocked until 3 p.m., and the phone isn’t in the room.

Friction. Add steps in front of what you want to do less and remove them from what you want to do more. He cites a cafeteria study where moving vegetables 10 inches farther away cut how much people ate by roughly 9% to 13%, and swapping a spoon for tongs cut it by about 16%. Something doesn’t have to be impossible, only inconvenient.

Hardware: body and workspace

He ranks these by expected gain for a typical person, and says the gains multiply rather than add. Stacked, he puts the total at about 150% to 200%, while warning that many of the numbers are estimates and the gain depends on how bad your starting point is.

LeverHis estimateWhat to do
Rule out a sleep disorderabout 10%Get a sleep study if you snore, breathe through your mouth, or wake tired after a full night
Sleep 7 to 9 hours10% to 20%Set a bedtime alarm and a wake alarm for the same times every day for a month
Same wake time every day6% to 12%Weekends included, since lie-ins reset the body clock
No caffeine within 10 hours of bed3% to 8% better sleepIn bed by 10 p.m. means no caffeine after noon
Bright mornings, dim evenings3% to 8%Ten minutes of outdoor light within an hour of waking, warm dim light and no social media before bed
Cool, dark, quiet bedroom2% to 6%About 18 to 20 °C, with blackout curtains or a fan
Bloodwork every year or two5% to 10% per deficiency fixedIron and ferritin, B12, vitamin D, thyroid and a blood sugar marker
Cut alcohol4% to 10%Even one or two drinks disrupt the second half of the night
Cut cannabis2% to 6%Similar cost to sleep quality
Protein and fiber2% to 4% eachHe moved from 5 to 10 grams of fiber a day toward the recommended 40 for his size
Creatine1% to 3%3 to 5 grams a day; the one supplement he says has real evidence
Raise VO2 max10% to 20%Measure it in a lab or estimate it with a fitness watch
REHIT sprints5% to 10%Two 20-second all-out sprints at the end of every workout
Fresh airhe puts 1,200 ppm of CO2 at a 30% to 40% lossA CO2 monitor, a cracked window, the door open between work blocks
No intelligible speech2% to 6%Instrumental music, brown noise or earplugs, never podcasts or lyrics
Desk by a window2% to 5%Or a bright desk lamp on winter mornings
Less clutter1% to 3%Everything in view competes for attention
Batch notifications3% to 6%Turn off banners and badges and check at two or three set times a day
One task on screen3% to 8%One full-screen window, other apps and tabs closed

A few studies he leans on here. In a two-week lab study of 48 adults, those allowed six hours in bed built up a deficit equal to one or two full nights without sleep, yet stopped feeling sleepier after the first few days. A 400 mg dose of caffeine six hours before bed cost about an hour of measured sleep without people noticing. For REHIT, he describes sedentary adults doing three ten-minute sessions a week, whose aerobic capacity rose 15% in men and 12% in women after six weeks. On air, he cites Allen and colleagues (2016), where scores fell 15% at about 945 ppm of CO2 and 50% at about 1,400 ppm, against a baseline of about 550. Outdoor air is about 420 ppm, and he says a small closed office can pass 1,000 within an hour or two.

His top pick after any medical problem is the sprints, and his tip is to stack defaults: exercise outside in the morning with a friend covers light, wake time, fitness and company at once.

Software: mental models

He calls these programs that run on the hardware, and says they’re where the large gains are. Each one works by turning a vague feeling into something you can write down and test.

ModelThe ideaHow he uses it
Expected valueAdd up each outcome’s value times its chanceIn his head for small calls, on paper for anything costing more than a few days or about $10,000
ResultingA good decision can lose and a bad one can winJudge past calls by their expected value, not their outcome
Power lawA few inputs produce most of the outputSort clients, videos or apps by result, then double down on the top 20%
Pareto principleRoughly 80% of results come from 20% of causesSame sort; one client paying as much as 25 others is a risk as well as a win
Chain lawA process’s success rate is the product of its stepsPut effort into the weakest step, and merge steps where you can
Local and global maximaYou can be at the top of the wrong hillSigns: shrinking gains, and someone less skilled on another path doing better
Attentional residueSwitching tasks leaves attention behindBefore switching, write where you are and the next step; keep at most two tasks open
AkrasiaKnowing is rarely the bottleneck; doing isMake the next action tiny and tie it to a trigger
Map and territoryEvery belief, dashboard and summary is a simplified mapWhen they disagree, check the real thing, such as the call recording instead of the summary
Ulysses contractBind your future self while you’re thinking clearlyPhone in another room, a public commitment, or money at stake
Opportunity costThe cost of a choice is the best thing you gave upEvery yes to something time-consuming names the specific no
Incentives and Goodhart’s lawPeople optimize what they’re paid forPair every target metric with one that gets worse if it’s gamed, such as tickets closed with tickets reopened
Chesterton’s fenceKnow why a rule exists before removing itAsk who built it, check the history, then decide on the merits
BayesUpdate a belief by how much likelier the evidence is if it’s trueAsk how often you’d see this evidence if you were wrong
Fermi estimateBreak a number into parts, estimate each, multiplyRound hard and aim for the right order of magnitude
Hourly rate and offload rulePrice your time at your best work’s rateOffload when the saving is clearly larger than the cost plus supervision
Planning fallacyPlans describe the smooth pathMultiply your estimate by a factor taken from your own past projects
Explore and exploitTrying new options versus using the best known oneExplore early and when returns fall; exploit while a winner is still growing

Some of these come with worked examples worth keeping.

Expected value. A solar install business making $500,000 a year with a 50% chance of lasting five years is worth $1.25 million in expectation. A software startup with a one in a thousand chance of a $1 billion exit is worth $1 million. He says the point isn’t the precision: a written 20% is a prediction you can check later, and “it’s risky” isn’t.

Chain law. Three steps at 100%, 100% and 20% succeed 20% of the time. Three at 80% succeed 51% of the time. With 100 hours and steps at 95%, 70% and 95%, all the hours go to the middle step until it catches up.

Bayes. His core question is how likely the evidence would be if the belief were true versus false. “Send me the details” is about as likely from a buyer as from someone ending the call, so it tells you almost nothing. “Can we pay quarterly instead of monthly?” is about nine times likelier from a buyer, so it’s strong evidence: starting from a 25% chance of a sale, a nine-to-one ratio works out to 75%.

Fermi estimate. Cold email at a 0.5% positive reply rate, 25% of replies booking a call and 20% of calls closing works out to one client per 4,000 emails. His sizing of a dental software idea in Canada runs 40 million people, one clinic per 5,000, a fifth with the problem badly enough, and 3% closed in year one, which comes to about 50 clients.

Offload rule. Offload a task when the hours it would take you, times your marginal rate, beat the price of offloading plus the hours of supervising it, times the same rate. A five-hour task at $100 an hour is $500 of your time. Paying $150 and spending two hours briefing and checking costs $350, so it’s still worth it. He wants a margin of about 2.5x to 3x before offloading, because supervision is the cost people forget, and he doesn’t offload the core of what customers pay him for. For a solo business he warns that each new person adds a step to the chain.

Planning fallacy. He cites a study of 258 transport projects worth $90 billion in which nine in ten ran over budget, rail by 45% on average. Many people multiply their estimates by two or three; he uses about 1.5x because his projects are small and mostly depend on him. Find your own factor by comparing your last few projects’ estimates with what they took. Quote the scaled figure to clients but keep tight internal deadlines, so the buffer doesn’t turn into slack.

A note on two attributions: the fence parable comes from G. K. Chesterton’s 1929 book The Thing, and the usual wording of Goodhart’s law, “when a measure becomes a target”, is anthropologist Marilyn Strathern’s summary of Charles Goodhart’s 1975 point.

Techniques for acting on it

This block is about akrasia: understanding everything and then doing nothing. Each technique is short enough to use the same day.

Trigger action plans. Write the rule as “when X happens, I will do Y”. He cites a 2006 meta-analysis by Peter Gollwitzer and Paschal Sheeran covering 94 tests, which found a medium-to-large effect on reaching goals. The trigger is where these fail, so it has to be:

  • obvious, so it either happened or it didn’t. “When I open a new tab” works; “when I feel like procrastinating” doesn’t.
  • physical, such as picking up the phone, sitting in the chair or the first sip of coffee.
  • reliable, firing every time in the situation you want and never outside it.

The action is small and physical: open the document and write one sentence, not “write for two hours”. To install one, close your eyes and rehearse the trigger and the action ten times in a row, which takes under two minutes. If it doesn’t fire, pick a more noticeable trigger and rehearse again. His mornings are a chain of these that gets him from waking to his first piece of work in about a minute and a half.

Noticing. Three signals, each turned into a five-second move:

SignalWhat it meansThe move
ConfusionYour map of the situation is wrongSay “I notice I’m confused” and ask why
FlinchYou’re training yourself to avoid somethingLook straight at it for ten seconds and name what causes the feeling
”Should”The goal may be borrowed from someone elseAsk whether you want it; if so find the next physical step, and if not drop the guilt

Pre-mortems. Change the tense. Instead of asking what might go wrong, say it’s a week later, the project failed, and everyone writes in the past tense why. Pool the causes, fix the likeliest ones and repeat. His example is a plan to publish daily for 40 days whose pre-mortem predicted 11, broken by client work and perfectionism; the fixes were recording first thing each morning and a public commitment. He puts the cost at about 15 minutes on one page.

Five-minute timers. When you catch yourself saying “I’ll get back to it”, set a five-minute timer and try to actually solve the problem, not plan it. One of three things happens: you find it isn’t worth doing, you finish it, or you find it’s bigger. Then a second timer is for writing the next five steps and a third for doing as many as you can. Use a physical timer, write while you think, and decide in advance what done looks like.

Using AI without losing judgment

His concern is that models are trained on human ratings and people rate agreement highly, so asking “is this good?” mostly gets your own view back. A 2023 Anthropic paper found this sycophancy in five AI assistants from several companies. His fix is to ask for the attack instead, with four prompts:

  1. Assume the plan is a mistake and give the three strongest reasons a smart, experienced person wouldn’t do it.
  2. Out of 100 similar attempts, how many get the result I want?
  3. It’s six months later and the project failed: write the post-mortem. Run it several times and pool the reasons.
  4. Break the opportunity into pieces, estimate each as a range, and combine them into a Fermi estimate.

He calls this red teaming. The second rule is to never let AI make the final call: write the decision and your reasons yourself, because you carry the consequences and reasons written for you become justification after the fact. His order is that you draft, the AI attacks, you revise, and you write the decision line.

How to try it

The cheapest start is the one he gives in the first hour: move your phone to another room and go through your day asking what happens at each step if you do nothing. Then pick the hardware lever you suspect is most broken and the boring test that settles it, whether that’s a sleep study, a blood panel or a fitness check. From the software block, the habit that carries the rest is writing the probability down before you decide. From the last block, write one trigger action plan and rehearse it ten times today.

Key terms

Choice architecture
How options are presented, which shapes what gets chosen. There's no neutral presentation, only one designed for someone's goal.
Friction
Any cost between a person and a behavior: an extra click, step, walk or password.
Attentional residue
The part of your attention that stays on the last task after you switch, a term from Sophie Leroy's 2009 research.
Akrasia
Ancient Greek for acting against your own judgment: knowing the right thing and not doing it.
Resulting
Judging a decision by how it turned out instead of by how good it was with what you knew at the time.
Chain law
A process is as reliable as the product of its steps' success rates, so the weakest step sets the result.
Ulysses contract
A choice made today that limits your options tomorrow, so your clear-headed self binds your tired one.
Chesterton's fence
Don't remove a rule or step until you know why it was put there.
Goodhart's law
Once a measure becomes a target, people optimize the number and it stops measuring what it was meant to.
Marginal hourly rate
What an hour of your best work earns, as opposed to your average rate across every hour you work.
TAP
Trigger action plan: a rule of the form 'when X happens, I do Y'. Researchers call it an implementation intention.
Pre-mortem
Imagining a plan has already failed and writing down why, before it starts.
Resolve cycle
A five-minute timer spent actually trying to solve a problem you've been putting off.
REHIT
Reduced-exertion high-intensity interval training: ten minutes of easy cycling with two short all-out sprints.
Sycophancy
An AI model's tendency to agree with and flatter the user, learned from people rating agreement highly.

Quotes from the talk

It's not a plan. It is a hope.

Nick Saraev, quoted in the talk

But the feeling is the fallacy.

Nick Saraev, quoted in the talk

It's almost like your first opinion just wearing a lab coat.

Nick Saraev, quoted in the talk

Connects to

Other entries that share a theme with this one.

Decide the rule in advance

A limit set ahead of time, on how you work or on what an agent may do, is a choice you don't have to argue again each time it comes up.

More research

All research

Send feedback

This goes to my inbox through Web3Forms, with this page's address attached. The form sets no cookies.