Thinking Clearly with AI

A condensed reading companion to Nick Saraev's five-hour course, How to Think Clearly In The Era Of AI (opens in a new tab).

AI can produce a convincing answer before you have worked out what question matters. That speed is useful when the task is clear. When it is not, a polished response can make a weak assumption harder to notice.

Saraev's course approaches this problem through attention, working conditions, decision-making and practical habits. Its central recommendation is to use AI to examine your thinking while remaining able to explain your own decisions. This article condenses that argument into eight chapters, with new illustrative examples and links back to the relevant parts of the recording.

You can read straight through or keep a question from your own work in mind. The recurring example—a team considering AI-assisted weekly updates—is entirely synthetic and optional. The goal is a decision you can explain: what you want, which alternatives you considered, what remains uncertain, and what you will do next.

Contents

1. Make room for a difficult thought
2. Shape the conditions around your work
3. Compare decisions before knowing the outcome
4. Notice where your model stops fitting
5. Update beliefs without false precision
6. Decide what to delegate and how to plan
7. Turn reflection into a small action
8. Use AI while keeping ownership of your judgment

1. Make room for a difficult thought

Watch from 06:05 (opens in a new tab)

Saraev opens with a simplified account of cognition: attention selects what you notice, working memory holds what you are currently handling, and executive function helps direct the work. These are useful distinctions even without treating them as a complete description of the brain.

Consider the question, “Should we use AI for our weekly team update?” To answer it, you may be trying to remember the current workload, the intended readers, several tools, previous mistakes and a deadline. Holding all of that in your head makes it easy to lose a condition while following an attractive idea.

Writing gives the problem a stable form. Open a note and separate three things:

• Decision: whether to try AI-assisted drafting for the next two updates.
• Desired result: less preparation time while preserving accurate decisions, owners and deadlines.
• Unknowns: how much correction the draft needs and whether readers find it useful.

This is already more useful than “We need an AI strategy.” It identifies a choice, a purpose and missing information. It also prevents tool selection from quietly becoming the goal.

The same principle applies when you stop working. Leave a short return note: where you reached, what remains unresolved, and the next concrete step. “Compare the corrected draft against the original notes” is easier to resume than a screen full of open tabs.

For your own question: write the decision in one sentence, followed by the result you want and the most important unknown. If the sentence contains several unrelated decisions, separate them before asking AI to help.

2. Shape the conditions around your work

Watch from 36:01 (opens in a new tab)

One recurring argument in the course is that a plan should not depend entirely on resisting temptation at the moment it appears. Change the situation before relying on another promise to concentrate.

Choice architecture means the way options are arranged: what appears first, what is selected by default, and what requires extra effort. Friction is the effort involved in taking an action. Both can influence what happens next.

For the team update, imagine opening your laptop to a chat feed, an inbox and a document buried in a folder. The easiest next action is to react to somebody else. A small change is to open the working document and source notes before the work period, then pause non-essential notifications. If the phone is your working device, use its focus settings rather than making the device inaccessible.

Friction can also be helpful. Keeping a draft separate from the send action creates a moment to inspect it. An automatically generated message that is immediately distributed removes that opportunity.

Saraev spends a substantial part of the recording on sleep, exercise and the physical environment. The practical thread is to notice the conditions under which you work, rather than treating every difficulty as a character flaw. His numerical claims about cognitive improvements are not established by this summary; the health material would need separate evidence review before becoming specific recommendations.

For your own question: choose one reversible change that makes the intended action easier. Judge it by whether you actually begin and complete the work, not by how impressive the new routine looks.

3. Compare decisions before knowing the outcome

Watch from 1:52:17 (opens in a new tab)

A decision has alternatives, possible outcomes and uncertainty. Expected value makes those ingredients explicit: multiply each outcome's value by its probability, then add the results. Costs must be included, and values must use comparable units.

Suppose an optional automation trial takes two hours to configure. Over a fixed evaluation period, assume it has a 60% chance of saving six hours and a 40% chance of saving nothing. Those are invented assumptions for this example, not measured probabilities.

The expected net gain is 0.6 × 4 + 0.4 × (−2) = 1.6 hours. This calculation does not predict that the trial will save exactly 1.6 hours. It describes an average under the stated assumptions.

Its real value is exposing what the decision depends on. If success is only 20% likely, the expected gain becomes −0.8 hours. Now the uncertainty about success deserves attention. A smaller trial may be more useful than another confident argument.

Expected value is one input, not a complete decision rule. An option can have a positive average and still carry a loss you cannot afford. Quality, obligations and irreversible consequences also matter; they need not all be converted into money.

Saraev also describes resulting: judging a decision solely by what happened afterward. A careless choice can succeed by luck, and a reasonable choice can fail. Record your reasoning before the outcome, then review both the process and the result. Repeated failures are evidence to update on, not something to dismiss forever as bad luck.

For your own question: compare at least two options, including the current approach. Name the assumption most likely to reverse your preference.

4. Notice where your model stops fitting

Watch from 2:27:24 (opens in a new tab) · Map and territory, 2:47:50 (opens in a new tab) · Incentives, 3:12:03 (opens in a new tab)

A model is a selective description of reality. A dashboard, a schedule and an AI-generated explanation are all models. They help because they leave things out; the danger is forgetting what is missing.

For the weekly update, “minutes spent drafting” is a useful measurement. It does not reveal whether the message contains the right decisions or whether a colleague spends longer correcting it. A faster draft can coexist with more total work.

This connects to Goodhart's law: when a proxy becomes a target, people can improve the number without improving the underlying goal. If the team rewards the number of updates produced, volume may rise while usefulness falls. Pair the speed measure with a quality check, such as missing action items or corrections needed. Then inspect actual examples; a second number is still a model.

Another idea is the local maximum. You may have improved the current approach until further adjustments offer little benefit, while a different approach could work better after a learning period. Faster prompting may not solve a reporting problem caused by inconsistent source notes. Testing a simpler note structure might address the real constraint.

Changing direction also has an opportunity cost: the best alternative use of the same time or resources. Spending Friday comparing five assistants means not spending that time interviewing the people who read the update. State the actual trade-off rather than saying only that something “takes time.”

Finally, apply Chesterton's fence before removing a puzzling step: understand why it exists. The manual sign-off may prevent incorrect commitments from reaching another team. Its purpose may still be necessary even if the way it is performed can improve.

These models lead to a common question: What would I observe if my current explanation were wrong? A reader reporting missed decisions is evidence that matters, even when your speed chart looks excellent.

For your own question: identify one measure you rely on, what it leaves out, and one observation that would challenge your preferred approach.

5. Update beliefs without false precision

Watch from 3:23:07 (opens in a new tab) · Rough estimates, approximately 3:31:14 (opens in a new tab)

Bayesian reasoning begins with an initial belief—the prior—and updates it in response to evidence. The important question is not just whether an observation fits your idea. It is whether that observation would also be common if your idea were false.

An enthusiastic reaction to an AI demo may occur both when a tool is useful and when it merely looks impressive. It is weaker evidence than repeated success on the actual task.

Natural frequencies make this easier to see. Consider a separate, fully synthetic example of a document checker:

• Out of 1,000 documents, 100 contain a particular error and 900 do not.
• The checker flags 80 of the 100 documents with the error.
• It also flags 90 of the 900 documents without the error.

There are 170 flagged documents, of which 80 actually contain the error. For a randomly selected flagged document in this example, the probability of that error is therefore 80 ÷ 170, or about 47%. Catching 80% of errors does not mean that 80% of flags are correct. The starting prevalence matters.

For your own decision, you may not have reliable counts. Say so. An explicit uncertain assumption is useful; a made-up percentage presented as a measurement is not. Repeating the same story through several AI answers does not create several independent observations.

The course also introduces Fermi estimates: break a large unknown into smaller quantities you can roughly estimate, then combine them with the appropriate arithmetic.

For example, suppose ten people each prepare two short updates per week. If drafting takes 10–20 minutes per update, total drafting time is roughly 10 × 2 × (10–20) = 200–400 minutes per week. That is the size of the activity, not the amount automation will necessarily save. Review time and setup still count.

Use a range, identify the least certain input, and find out whether better information would change your decision. If the opportunity is small even under optimistic assumptions, you may already know enough to stop.

For your own question: separate known facts from estimates. Name the next piece of evidence worth obtaining and how it could change your view.

6. Decide what to delegate and how to plan

Watch from 3:38:45 (opens in a new tab) · Planning, 3:50:38 (opens in a new tab)

Delegation does not eliminate work; it changes who performs which parts. Saraev's offloading discussion makes supervision an explicit cost rather than assuming that handing over a task makes it disappear.

For AI-assisted drafting, compare the complete paths:

• Do it yourself: preparation, drafting and checking.
• Delegate: preparing inputs, explaining the task, reviewing the result, correcting it and handling exceptions, plus any direct cost.

If an update takes 30 minutes manually, but briefing takes five minutes, review takes 15 and repairs take another 15, delegation uses 35 minutes of your time. A near-instant model response does not change that total. The example does not prove AI drafting is inefficient; it shows what must be measured.

Time saved also needs an intended use. It might support another task, learning or rest. Do not automatically equate every freed hour with your highest imaginable earning rate.

Planning introduces another trap: imagining the work proceeding smoothly and treating that story as a forecast. The inside view asks how this project should unfold. The outside view asks how comparable projects actually unfolded.

Look at your last few similar tasks. If your estimates were two hours and completion usually took three, a 1.5 multiplier is a starting calibration for comparable work. It is not a universal rule. Changed scope, unfamiliar tools and dependencies may call for a different range. State when the work is complete and include review before promising a deadline.

Finally, distinguish exploration from exploitation. Exploration tests alternatives; exploitation uses an approach that already works. Early in a project, a bounded comparison may be valuable. Near a deadline, endlessly switching tools can consume the benefit you hoped to obtain. Decide when to stop comparing and use what you have learned.

For your own question: choose a bounded next step, name who checks the result, and estimate the complete effort using something comparable you have actually done.

7. Turn reflection into a small action

Watch from 4:05:08 (opens in a new tab) · Pre-mortems, 4:33:30 (opens in a new tab)

Understanding a useful idea does not ensure that you apply it. The course closes that gap with concrete triggers, noticing habits, pre-mortems and short work periods.

A trigger-action plan connects a recognisable event to a specific response: “When X happens, I will do Y.” Compare “I should check AI output more carefully” with “When an AI draft is ready, I will compare its decisions, owners and deadlines with the source notes before sending it.” The second tells you when to act and what to do.

Noticing provides possible triggers. Confusion can indicate an unresolved mismatch: the dashboard says the process improved, but the team reports more work. Pause and state the discrepancy. Avoidance can point to an undefined next step. A repeated “I should” is an invitation to ask whose goal it serves and whether you endorse it. These are prompts to investigate, not proof that every uncomfortable feeling has one simple explanation.

A pre-mortem examines a plan by imagining that it has already failed. Instead of “Could anything go wrong?”, ask, “It is two weeks later, and the update trial failed. What happened?”

Possible stories might be that nobody reviewed the draft, source notes arrived too late, or the assistant dropped an unresolved disagreement. Turn the credible stories into changes: assign the reviewer, set a notes deadline, and require unresolved issues to remain visible. A vivid story is a candidate failure mode, not evidence that it will happen.

Then take a small action. Set a five-minute timer and define an endpoint, such as writing the trial question and choosing one suitable sample. If the task cannot be finished, leave a specific next step or a question for someone who can unblock it. The timer makes beginning easier; it does not guarantee completion.

For your own question: write one trigger-action plan and one plausible failure story. Change the plan where that story reveals a preventable weakness.

8. Use AI while keeping ownership of your judgment

Watch from 4:47:14 (opens in a new tab)

Saraev's AI-specific chapter warns about sycophancy: an assistant agreeing with the user's apparent preference instead of reliably challenging it. “This is a good idea, isn't it?” can invite a reassuring answer. Fluent agreement is not independent confirmation.

His proposed sequence is straightforward: develop a position, ask AI to challenge it, revise it, and write the final decision yourself. The preceding chapters supply the questions worth asking.

In an approved assistant, share only material suitable for that service. You can use this original prompt with public or synthetic information:

My decision is: [decision]. My goal is: [goal]. The options are: [options]. These are my known facts and assumptions: [separate lists].

Compare the strongest case for and against my preferred option. Identify the assumption most likely to change the decision. Describe one plausible failure and a small way to test the key uncertainty. Distinguish evidence from speculation; do not invent sources, success rates or measurements.

Inspect the criticism as carefully as the praise. Asking for objections can generate weak objections, too. A prompt is an instruction, not a guarantee of accuracy. Check factual claims against their sources and recompute the numbers that matter.

For the update trial, an assistant might point out that saving drafting time is irrelevant if review takes longer. That objection is useful because it changes what the trial measures. A sweeping claim that “AI always improves productivity” or “AI always loses nuance” is much less useful without evidence about this task.

Finish with a compact record in your own words:

A synthetic example might end: “I will trial AI drafting on two updates because drafting takes substantial time. A named colleague will check each draft. We will compare total preparation and correction time and inspect missed decisions before deciding whether to continue.”

The purpose is not to prove that you never need assistance. It is to remain able to explain why the choice makes sense and to revise it when reality disagrees.

When your decision produces something another person will use, continue with T01-L02 — Trust but verify: checking output before it leaves your desk (opens in a new tab). That companion workflow covers checking the actual deliverable before sharing it.

Source and authorship

This article was written by Astra (openai/gpt-6-astra) on 12 September 2026 as an original condensation of Nick Saraev's How to Think Clearly In The Era Of AI: Full Course (5 Hours) (opens in a new tab), published 2 September 2026. The source video is also available through the curriculum video page (opens in a new tab).

The source basis was a complete-duration, locally generated Whisper transcription, with the relevant passages examined for this article. The automatic transcript has recognition errors and has not been fully corrected against the audio. Timestamp links are entry points to the source discussions, not verbatim quotation locators.

The examples, calculations, prompt and decision-record format are new teaching material. Numerical examples are synthetic; they are not results from a classroom or a real team. The source's health claims, productivity percentages and personal business results were not independently verified. This article retains the practical reasoning techniques without adopting those figures. The content received an author-side source-fidelity and readability pass, not an independent editorial review.