How to Write with AI (Tyler Cowen Breaks it Down)

David Perell

Abstract

An hour with an economist who uses these tools daily and refuses to have the abstract argument about them. The organising move is narrow and practical: treat a model as secondary literature — interrogate it about a subject rather than buying thirty books to acquire the context — and keep the writing your own even where the machine writes better.

• The gain is in the quality of the reading rather than in hours saved, because active questioning beats passive consumption.
• Never ask the model what question to ask; ask about the details of concrete examples and the questions arrive.
• The chat window tricks people into text-message prompts, when the first context-setting prompt should be very long.
• A model smooths prose toward the average, which a distinctive writer should refuse on purpose.
• Prompting one thing ten times beats prompting ten things once.
• What it still cannot do is ask the one question that matters.

Everything below follows the conversation in order, from the practical stance through model choice and prompting, into writing, the future of books, and teaching, out through networks, secrets and what stays human. For anyone building a personal brand the sharpest instruction is the one about being a source: a large body of work is what convinces a model you are worth referencing.

Chapter summaries

00:23 - Skip the big abstract debate and stay practical

Skip the big abstract debate and stay practical; experimenting with the tool teaches more than theorising about it.

"better I want to set the ground rules for this so there's a"

— 00:23

Experimenting with the tool teaches more than theorising about it. The whole conversation stays at that altitude deliberately, which is why almost every claim in it is checkable by doing the thing.

01:05 - The best models work as a secondary literature

Use the best models as a secondary literature: interrogate them about a subject instead of buying twenty-to-thirty books to acquire context.

"learning about it most of all I use AI when I read things"

— 01:05

Fast interactive context-building replaces the slow one-way reading of commentary. Preparation for a discussion went from ordering many books to two or three plus continuous questioning — and as the questioner, a modest hallucination rate does not matter, because you are not the one giving the answers.

01:56 - The point is improving the quality of the reading, not saving time

Active questioning beats passive consumption, so the gain is in the quality of the reading rather than in time saved.

"think that this is about saving time or improving the quality of your"

— 01:56

The gain lands in the quality of the reading rather than in hours saved. The richest texts reward endless re-reading, which makes them the ideal material for a guide.

03:47 - Which model gets used is the first thing that matters

Which model you use is the first thing that matters; the reasoning models people pay for are meaningfully better.

"uh you can do that so like what would Harold Bloom say what"

— 03:47

Capability varies enough between systems that the choice dominates most technique applied afterwards. One paid reasoning model is named best for this, another strong, a third fun but unreliable — and the rankings change constantly.

05:07 - Anything from any source must be double-checked, so double-checking AI is no extra burden

Anything from any source must be double-checked, so double-checking AI is no extra burden.

"okay so what I've noticed is a lot of people think that hallucinations"

— 05:07

Verification is already the standard for a human, a book or an encyclopedia, so the burden is not new. A writer instructs his assistant to fact-check every line as if trying to destroy the argument. The best reasoning models also hallucinate far less than a year ago, and the rate is still falling.

06:57 - The model never gets asked what question to ask

Never ask the model what question to ask; ask about the specific details of concrete examples and the good questions come to you.

"homogenizing thought but also the potential to really get out into wild and"

— 06:57

Generic requests return normy, bland output; specific interrogation is what surfaces the unexpected. Asking about the particulars of a historical figure gets somewhere that asking what should I ask never does.

08:12 - Background comes from the tool and the writing stays the writer's

Use AI to acquire background on unfamiliar ground, but keep the writing your own even when the machine writes better.

"using it in the writing process I don't directly use AI for writing"

— 08:12

Ownership of the voice matters more than a marginal quality gain. Legal background gets pulled for a column and the words it produced go unused — and while some outlets would forbid machine-written copy anyway, the refusal holds regardless of the rule.

08:52 - AI can smooth prose toward the average and easier comprehension

AI can smooth prose toward the average and easier comprehension; the distinctive writer should refuse that on purpose.

"I want the writing to be my own it's like my little baby"

— 08:52

It smooths prose toward the average and toward easier comprehension. A distinctive writer should refuse that on purpose, because the smoothing removes exactly what made the writing worth reading.

10:11 - The draft goes through the model to find what readers will resent

Run your draft through the model and ask what some readers will find obnoxious, explained in detail.

"all uh there was one use where uh Agnes card a while ago"

— 10:11

Asking for the detail rather than the verdict is what makes the answer usable. It is unusually good at catching when writing reads cold or harsh, which is why sharp-tempered managers use it to warm critiques.

12:24 - Most people use AI wrong by asking questions that are too general and refusing to put

Most people use AI wrong by asking questions that are too general and refusing to put in their own time generating context.

"making when you're looking at how other people use Ai and you're like"

— 12:23

Used as a substitute for your own effort it under-delivers, which is where the unimpressed verdict comes from — a cheap way to do mid tasks. That is a fine conclusion if their own time is genuinely better spent elsewhere.

13:37 - The chat window tricks people into text-message-length prompts

The chat window tricks people into text-message-length prompts; the first context-setting prompt should be very long.

"the problem is that it's a text window that makes it feel like"

— 13:37

The box is the size of a text message, so the prompt becomes the size of a text message — when the first context-setting prompt is the one that should run long. The interface is shaping the behaviour more than the capability is.

14:12 - The tool is a stack of interacting agents rather than one box

Treat AI not as a single box but as a stack of interacting agents, some writing the prompts others answer.

"doing I haven't myself tried it yet I suspect it works very well"

— 14:12

Some of them write the prompts that others answer. The mental model changes what you expect of it and therefore what you ask.

14:50 - The tool gets used as one mind while it evolves into many

Use it today as if it were one mind, while it evolves toward a decentralised republic of many minds.

"the way like for humans there's a republic of science way smarter than"

— 14:50

The destination described is a decentralised republic of many minds, which is not what is in front of anybody today. Holding both at once is the practical posture.

15:07 - There are three layers of understanding AI, and most people never reach the first

There are three layers of understanding AI, and most people never reach the first.

"you're part of it I like to say there's three kind of layers"

— 15:07

Naming the layers makes it obvious that the usual complaint is made from outside all three of them. A person who has not reached the first layer is describing a tool they have only watched somebody else fail with.

16:38 - Technical understanding is not the requirement

You do not need technical understanding — a popular-science-level account plus hands-on use is enough.

"things like reinforcement learning synthetic data stuff like that how important is the"

— 16:38

You can have a working handle on a system without knowing its internals, the way you can not know how a side airbag works and still not be an idiot about cars. The technically expert do understand it better: this is a floor rather than parity.

19:12 - Generic corporate writing is the first kind to be automated

Generic corporate writing is the first kind to be automated; many who expected to write will not.

"if you're not using these tools how deflating it must be some humans"

— 19:12

Some humans become masters of the tools and the rest of that category shrinks, the way demand for certain programmers plummeted. It is a large adjustment rather than a total one — by no means all writing.

19:49 - Readers may not want the better AI product for personal or subjective work

Readers may not want the better AI product for personal or subjective work — they want a human behind it.

"writing a biography of a person the AI cannot really do it may"

— 19:49

What they want is a human behind it. Preference rather than quality is doing the work, which is a different market.

20:33 - Model personalities differ from one another

Model personalities differ; the freer, less-manipulated one is better at emotion and creativity but hallucinates more.

"average you've mentioned deep seek twice how is the the shape or personality"

— 20:33

Less alignment-flattening leaves a model more romantic, uneven and creative, which is a feature where invention is the point. One reached for a glorious sensory description in a context where hallucination is welcome — and the same model should not be used for research, where it is more censored on predictable topics.

21:30 - A deep-research model already beats an encyclopedia and open web search when the prompt is fleshed

A deep-research model already beats an encyclopedia and open web search when the prompt is fleshed out — and this is the worst it will ever be.

"do you think that deep research will get to a point where you"

— 21:30

Tailored long-form synthesis outperforms static reference when the need is specific. A ten-page article on an economics topic was judged better than any other source found — tailor-made to what one asker wanted, and not claimed to be best for everyone.

22:22 - Predictive books about the near future no longer make sense

Predictive books about the near future no longer make sense; cover that ground with ultra-high-frequency writing instead.

"how does all this influence what you're choosing to to write like should"

— 22:22

A multi-year book cycle is obsolete before publication in a field moving this fast. The response is structural: recent long-form shifts to the frozen distant past, and near-future takes move to blogging.

23:29 - The book is not the thing to fixate on

The book is not the thing to fixate on, because the model is a question box and whatever gets written has to be more interesting than that box.

"history but the other question is what can the AI soon enough write"

— 23:29

If any question inside a book can be asked of the box, the packaged book is an inefficient container for it. The prediction follows: few books get written by the tool, because the tool is the better container.

24:02 - The book worth writing is the one only a human could write sincerely

Write the book only a human can write sincerely — where the weak existing literature and lived role make you the source.

"better be more interesting than the question box so the book I've started"

— 24:02

The one where the existing literature is weak and a lived role makes you the source. Sincerity is the qualification rather than the tone.

25:27 - Truly human books will stand out more as everything else becomes AI slop or human slop

Truly human books will stand out more as everything else becomes AI slop or human slop that looks the same.

"want we'll see you know maybe the readers are are fine with the"

— 25:27

As everything else becomes slop — machine or human, indistinguishable — the genuinely human book stands out more. Scarcity is being created by the abundance.

26:33 - Video persists for now for the same reason as memoir

Video persists for now for the same reason as memoir: people want real humans, not fake ones.

"accordingly so then does that mean that the YouTube channel is in a"

— 26:33

Presence and personality read as human, and are not yet wanted synthetic. The bet is to double down on filmed conversation while synthetic likeness is roughly two years off — the voice is already indistinguishable, and the visual is not.

29:16 - A possible biography can now be built cheaply

You can now cheaply build your own "possible biography" by open-sourcing the missing pieces of your life for future AIs.

"Perell here's another thing I'm doing with writing so some people have told"

— 29:16

Synthetic voice in your own cadence across languages, plus platform dubbing, changes who can hear you. The reach question stops being about translation budgets.

31:13 - Being non-technical can be an advantage

Being non-technical can be an advantage — you see what the tool is actually good for instead of the bells and whistles.

"you doing in terms of staying at The Cutting Edge because here's what"

— 31:13

Open-source the missing pieces of your life so future systems have them. A possible biography can now be assembled cheaply, which is a strange and genuinely new option.

32:55 - The posture worth holding is super practical

Be super practical: work with it, be self-critical about what you are doing, and learn from other people.

"actually good for and not am I impressed by all the neat bells"

— 32:55

You see what the tool is actually good for instead of the bells and whistles. Being non-technical turns out to be an advantage in evaluating it, which inverts the usual assumption.

33:14 - The correct posture toward this transition is nervousness

The correct posture toward this transition is nervousness — even the positive path brings heavy disruption.

"other people if we stripped out AI like a Jango block now in"

— 33:14

Work with it, be self-critical about what you are doing, and learn from other people. Three instructions, none of which require any technical knowledge. The disruption arrives either way, and treating nervousness as a failure of nerve misreads what is being observed.

34:54 - The machine is a reader worth writing for

Write for the AI as a reader: a large body of work convinces the models you are a legitimate source worth referencing.

"view when you say that you write for the AI I mean I"

— 34:54

Even the positive path brings heavy disruption, so nervousness is the correct posture. It is stated as an assessment rather than as a mood to be managed. The disruption arrives either way, and treating nervousness as a failure of nerve misreads what is actually being observed.

36:29 - An AI reader already has the context, which is a real trade-off

Because AI readers already know the context, you can fill in fewer blanks — a real trade-off against your human audience.

"Spinosa but when you write for the AIS like for one thing they're"

— 36:29

A large body of work convinces the models you are a legitimate source worth referencing. Writing for the machine as a reader is a distribution strategy rather than a stylistic choice.

37:12 - Tables and visuals let far more information be taken in

Turning text into tables and visuals lets you input and grasp far more information.

"of the things that you haven't spoken about that has been fundamental for"

— 37:12

Because a machine reader already holds the context, fewer blanks need filling — which trades directly against the human audience. Naming it as a trade-off rather than an upgrade is the honest framing.

38:29 - The tool is for concrete literal objectives rather than theorising

Use AI for concrete, literal objectives, not for theorising about a place.

"and not much time so we just landed in DC and I struggle"

— 38:29

Turning text into tables and visuals lets far more information be taken in. Different models suit different visuals, and moving an argument into a visual makes a writer more effective rather than merely prettier.

41:22 - The writing that persists is personal

The writing that persists is personal — a small true story the machine could never have supplied.

"that piece and I have a sentence from that piece that I think"

— 41:22

Concrete, literal objectives rather than theorising about a place — and travel planning is named as one of its very best uses. The specificity is the instruction.

42:08 - Teaching happens with the tool out in the open

Teach with AI openly: require students to use it, report what they did, and make the whole paper theirs.

"they already do 100% tell me about AI in the classroom how are"

— 42:08

A small true story the machine could never have supplied is what persists. Personal is a structural property here rather than a tonal one.

43:34 - The norm calling AI use cheating must shift and in fact collapse

The norm calling AI use cheating must shift and in fact collapse; assessment moves to oral and proctored exams.

"life so why not teach it now what's the constraint them not wanting"

— 43:34

Require students to use it, have them report what they did, and make the whole paper theirs. Openness rather than prohibition, which is the only enforceable position.

44:41 - The single highest-return move is simply to use the best models

The single highest-return move is simply to use the best models; the variance between frontier and free is enormous.

"striking from all the questions that I've asked you then just the on"

— 44:41

The norm calling it cheating must shift and will in fact collapse, with assessment moving to oral and proctored exams. The prediction is specific enough to be checked.

45:58 - Research academia is not changing and is not ready

Research academia is not changing and is not ready; within a couple of years the models do much of the work better.

"now and how is what it means to be a research-based academic changing"

— 45:55

The variance between frontier and free is enormous, so simply using the best model is the highest-return move available. The monthly cost is a good investment for far more people than realise it.

46:57 - AI is uniquely strong on old public-domain, easily-verified material and specific reference questions

AI is uniquely strong on old public-domain, easily-verified material and specific reference questions.

"I use AI is to study the Bible which is sort of my"

— 46:57

Research academia is not changing and is not ready, and within a couple of years the models do much of that work better. It is said plainly rather than hedged.

48:02 - What AI still cannot do is ask the one question that matters

What AI still cannot do is ask the one question that matters — the core one or two sentences an expert reaches.

"where it still is lacking is if I speak to somebody who really"

— 48:02

Old public-domain, easily verified material and specific reference questions are where it is uniquely strong. Knowing the strong zone is worth more than knowing the weak one.

49:07 - Secrets and context a human holds become more valuable, and there is now more incentive to

Secrets and context a human holds become more valuable, and there is now more incentive to hoard them.

"what a mentor can provide is unique uh what do you say it's"

— 49:07

What it cannot do is ask the one question that matters — the core sentence or two an expert arrives at. That is the remaining human function, stated as narrowly as possible.

50:28 - Returns to human networks rise sharply

Returns to human networks rise sharply; you need people to mobilise what the tools produce.

"skill now increasing returns to social networks that's right so social networks become"

— 50:28

Secrets and context a human holds become more valuable, and the incentive to hoard them rises with it. A second-order effect that most predictions miss. Whatever a model produces sits inert until somebody with relationships moves it, which is why networks appreciate rather than depreciate.

51:48 - Prompt as if speaking to an alien or a non-human animal

Prompt as if speaking to an alien or a non-human animal — be more literal.

"simple rules for prompting like if you were teaching somebody hey here's how"

— 51:48

You need people to mobilise whatever the tools produce, so returns to human networks rise sharply. The output is not self-executing. Whatever a model produces sits inert until somebody with relationships moves it, which is why the networks appreciate rather than depreciate.

53:10 - Prompting one thing ten times beats prompting ten things once

Prompting one thing ten times beats prompting ten things once, so follow-ups get planned as follow-ups.

"me is it seems to be a lot better to prompt one thing"

— 53:10

Be more literal, as though speaking to an alien or a non-human animal. Prompting matters little for basic queries and exponentially for the best deep-research work. Literalness costs nothing and removes an entire class of failure that people otherwise blame on the model.

53:57 - The two universal pieces of advice, more valuable now

The two universal pieces of advice, more valuable now: get more and better mentors, and improve your peer network daily.

"option so when you're mentoring young people what are you telling them to"

— 53:57

Prompting one thing ten times beats prompting ten things once, with follow-ups planned as follow-ups. Depth on one thread outperforms breadth across many.

54:51 - Career paths are becoming less formulaic

Career paths are less formulaic; we will need fewer specialists and more people who can manage AIs.

"you feel that career trajectories are changing for example to get really practical"

— 54:51

Get more and better mentors, and improve your peer network daily. Two pieces of advice that predate all of this and are worth more now than before.

55:53 - Both things get done at once

Do both at once: learn the tools because they are getting better, and invest in networks because they are.

"suspect it's a good investment well once again it's sort of like what"

— 55:52

Fewer specialists and more people who can manage machines. Career paths become less formulaic, which is a cost and an opening at the same time.

56:25 - An AI note-taker sits on the other side of every talk

Assume an AI note-taker on the other side: prepare a talk by writing what that note-taker should capture.

"you know we were talking writing for the AIS earlier and another thing"

— 56:25

Learn the tools because they are getting better, and invest in networks because they are getting more valuable. Both at once rather than choosing between them.

57:19 - The default output is homogenised and bland even when good

The default output is homogenised and bland even when good; you have to work to make it not so.

"do you say that if you just ask AI simple questions like improve"

— 57:19

Prepare a talk by writing what an AI note-taker on the other side should capture. A phantom model sits on your shoulder even when you do not use one, enriching, intimidating and homogenising the work.

57:58 - A take and a direction get the rest filled in far better

Give the model your own take and direction and it fills in the rest far better than from a generic ask.

"and make some corrections for well after the Apple earnings came out recently"

— 57:56

Even when it is good, the default output is homogenised and bland. Making it otherwise takes deliberate work, which is the cost nobody prices in.

1:00:00 - An up-to-date answer engine replaces most search

An up-to-date answer engine replaces most search; it returns the right citation you can verify by clicking through.

"of using perplexity that are as strategic as how you prompt the llms"

— 1:00:00

Give it your own take and a direction and it fills in the rest far better than it does from a generic ask. The take is the input that cannot be outsourced.

1:01:06 - Match the tool to the job

Match the tool to the job: one model for queries, its deep-research offshoot for long reports, another as the best writer.

"tools I can tell you what I use I'm not saying it's all"

— 1:01:06

An up-to-date answer engine replaces most search for most questions. The substitution is already happening rather than being forecast. The habit forms around whichever one was learned first, and it survives long after a better fit becomes available.

1:05:09 - Staying current is its own discipline

Staying current is its own discipline: a paid tracker, the microblog, and the right chat groups.

"$400 a year but worth it for me I subscribe to information which"

— 1:05:09

Match the tool to the job rather than using one for everything. Obvious in principle and rarely done, because the habit forms around whichever one was learned first. The habit forms around whichever one was learned first, and it then survives long after a better fit becomes available.

1:05:44 - Very large context windows make routine what needed specialists

Very large context windows make routine what needed specialists — whole regulatory codes and historical archives.

"you think becomes possible with really large context windows so Gemini now has"

— 1:05:44

Staying current is its own discipline now, separate from using the tools well. The rate of change makes currency a recurring cost rather than a one-off.

1:06:38 - The next great human project is converting the world's knowledge into AI-usable form, and it will

The next great human project is converting the world's knowledge into AI-usable form, and it will create work.

"it in somehow scan it uh but working with things like that over"

— 1:06:38

Very large context windows make routine what used to need specialists. The capability arrived without most people noticing what it displaced. It is already underway inside those buildings rather than being proposed, which is what the gap in visibility actually consists of.

1:08:11 - The frontier labs are far ahead of what outsiders see, because they use AI to improve

The frontier labs are far ahead of what outsiders see, because they use AI to improve AI and guard the results.

"human effort to get there yeah and last question how Innovative is the"

— 1:08:09

The frontier labs are far ahead of what outsiders can see, because they use the tools on themselves before anybody else gets them. Converting the world's knowledge into machine-readable form is named as the next great human project, and it is already underway inside those buildings rather than being proposed.

Personal Branding Lessons

A practical conversation with almost no theorising in it. The moves below are the usable half.

Ask about concrete details rather than what question to ask

Asking a model what you should be asking outsources the only part that was doing the thinking. Asking about the specific details of concrete examples produces the good questions on its own, which is a different relationship to the tool entirely. 06:57

Write the first context-setting prompt long

The chat window tricks people into text-message-length prompts because of how it looks, not because of what it needs. The first prompt should be very long — the interface is shaping the behaviour far more than the capability is. 13:37

Ask the model which parts some readers will find obnoxious

Run the draft through and ask what will land badly, explained in detail. It is unusually good at catching prose that reads cold or harsh, which is why sharp-tempered managers use it to warm a critique before sending it. 10:11

Keep the writing your own even when the machine writes better

Using it to get onto unfamiliar ground and keeping the writing yours are compatible, and treating them as one decision is what produces the anxiety. The smoothing it offers removes exactly what made the writing worth reading. 08:12

Write the book only a human could write sincerely

Where the existing literature is weak and a lived role makes you the source. Sincerity is the qualification rather than the tone — and as everything else becomes indistinguishable slop, that book stands out more rather than less. 24:02

Turn text into tables and visuals

Converting text into tables and visuals lets far more information be taken in, and different models suit different kinds. Moving an argument into a visual makes a writer more effective rather than merely better presented. 37:12

Prompt one thing ten times rather than ten things once

Depth on a single thread beats breadth across many, with follow-ups planned as follow-ups from the start. Prompting matters little for basic queries and exponentially for the deep-research work where the returns actually are. 53:10

Give the model your own take and direction

A take and a direction get the rest filled in far better than a generic ask does. The take is the one input that cannot be outsourced, which is also why the default output is bland: nobody supplied one. 57:58

Questions

Each answer ends at the moment in the recording where it is given.

How should I actually use AI when I read, without letting it read for me?

Use it as a secondary literature rather than a replacement for the text. Read the primary work yourself, then interrogate the model chapter by chapter — its puzzles, its connections, the readings it prompts — so the reading becomes active instead of passive. The stated payoff is not saved time but better thinking: asking a live question as you go improves what you take away, and the richest texts reward it endlessly. 01:05

Are hallucinations a reason not to trust AI for research?

Less than most people assume. Anything from any source — a person, a book, an encyclopedia — has to be double-checked before you rely on it, so verifying the machine is the same discipline, not an extra burden. The best reasoning models now hallucinate far less than a year ago, and the rate keeps falling. As the questioner rather than the answerer, a modest error rate matters little. 05:07

Why does AI keep giving me bland, generic answers?

Because the questions are too general and too little of your own context goes in. Used as a substitute for effort, it returns competent, average, forgettable output; used as a complement to heavy reading and specific interrogation, it sharpens. The interface invites short, text-message prompts — the fix is a long first prompt that sets real context, followed by short follow-ups. If it reads flat, feed in more of yourself. 12:24

Should I let AI write or edit my work, or does that cost me my voice?

Pull background from it freely, but keep the writing your own. Smoothing prose toward easy comprehension strips the distinctive, layered quality that is the reason a particular voice gets read, so the averaging is a loss, not a gain. One editing use is endorsed: asking in detail what some readers will find obnoxious, which reliably flags a condescending passage and leaves the decision to you. The voice is the asset worth protecting. 08:12

Which kinds of writing will AI replace first, and which are safe?

Generic corporate writing goes first, on the same curve as certain programming roles — a large adjustment, though by no means all writing. What survives is the work only a human can do sincerely: memoir, biography, anything requiring fieldwork, a lived role or a real life behind the words. The safe ground is not the polished and general but the personal and specific. 19:12

If AI can write it, why would anyone read a human's version?

Because for personal and subjective work, readers may not want the better product — they want a real human behind it. A brilliant memoir corresponding to no actual life would soon bore. The scarce, truly human work stands out in relief as average output floods the field, which is exactly the uncopyable advantage a real person supplies. Lead with the lived material a machine cannot generate. 19:49

How do I write so that AI models treat me as a source worth citing?

Publish a large, consistent body of work on clear territory over time. The models build a better picture of a heavily published person than most humans hold, and volume is what makes you a referenced source — an "intellectual immortality" built simply by writing. Because that reader already holds the context, you can spell out less, though at a stated trade-off against the human audience. The record is the authority that gets you named. 34:54

What are the simple rules for prompting well?

Drop human assumptions and be literal — imagine addressing an alien or a non-human animal. The difficulty is emotional inertia, not cognition. Open with a long context-setting prompt, then keep follow-ups short, and prompt one thing many times rather than many things at once, since a long multi-question prompt degrades toward the end. Prompting matters little for basic queries and enormously for a long deep-research report. 51:48

Is it still worth paying for the most expensive model?

For far more people than realise it, yes. The variance between the frontier and the free tier is enormous, and off the frontier you cannot see how fast, or in which directions, capability is improving. The top tier's monthly cost is defended as a good investment for almost anyone with prospects; the free models catch up over time, and a new paid tier then opens above them. 45:13

Sources

How to Write with AI (Tyler Cowen Breaks it Down)

The sections above follow the episode's own order. Each timestamp is the point in the recording where the idea is discussed, so the post reads as a map of the conversation rather than a rearrangement of it.

Quotations come from the episode transcript with spoken filler ("okay", "right", "you know", "um", "actually", "kind of", "sort of") and false starts removed. No words are added: square brackets mark anything inserted for sense, and an ellipsis marks any cut. British spellings are restored, and speech-to-text errors were repaired against context only where the intended word is unambiguous. Where a line appears both in a cold-open teaser and in its place in the conversation, the timestamp cites the second — the point at which it is actually said.

Quotes are stamped with the moment they are said rather than with a speaker's name.