The Ultimate Guide to Writing with AI

David Perell

Abstract

Ninety minutes from a writer who uses these tools daily and still argues the craft is unchanged. The two claims sit together deliberately: a writer who outright ignores the advances is a fool, and the core skills — taste first, then a spiky point of view — are the same with or without them.

• The more a piece comes from personal experience, the less likely a model is to outdo it.
• Quality is two different things, which is why the argument about machine writing goes in circles.
• Most of what gets generated is cut, which is the actual workflow rather than an admission.
• Slop is when publishing or finishing matters more than the quality of the thing — the tools did not invent it.
• Models are good at things with no wrong answer and bad at precise information.
• The sceptic trap is trying it once, getting a weak output, and quitting.

Everything below follows the conversation in order, from the scale of the change through what differentiates human non-fiction, into the definition of slop, the drafting workflow, memory and context, and out through a model-by-model account of what each one is actually for. For anyone building a personal brand the defensive strategy is stated in one line: ask where you hold data, facts and experience the models will not have.

Chapter summaries

00:00 - AI will shake the foundations of writing

AI will shake the foundations of writing; dismissing it entirely is a mistake, yet genuinely good writers will be fine — and the two claims square in how you write with it.

"As someone who's built a career around writing, I think that AI is"

— 00:00

The foundations, not the surface. Whatever else is true, the claim being made is structural rather than about a tool that helps with drafts.

04:00 - A model works as an always-available learning layer

A model can act as an always-available learning layer, more useful than most experts you would hire.

"glimmers of it in 2024, but it was really at the end of"

— 04:00

More useful than most courses, because it is available at the moment the question arises rather than on a syllabus. The always-available part is doing the work.

05:19 - The models are getting better and cheaper fast

The models are getting better and cheaper fast, and heavy competition is driving prices down.

"it's not just that. The models themselves are getting better and cheaper at"

— 05:19

Many firms racing for the best model push capability up and cost down at the same time, which is why the two curves move together rather than trading off. Engagement with the tools is up roughly tenfold in a year.

06:26 - A writer who ignores the advances is a fool

A writer who outright ignores these advances is a fool; this is a new paradigm of writing, and not one to be terrified of.

"that if you're interested in writing and you're just outright ignoring these advancements"

— 06:26

A writer who outright ignores the advances is a fool. It is stated bluntly and it is compatible with refusing to let the tools write — ignoring and refusing are different.

08:07 - The future is already here, unevenly distributed

The future is already here, unevenly distributed; you see glimmers and imagine them at higher resolution.

"example, but William Gibson, the science fiction writer, has this line that I"

— 08:07

The unevenness is the practical fact: some people are already working this way and most are not. Which means the useful question is not whether it works but who is already working that way and what they are doing differently.

08:52 - The last eighteen months were the most rapid change in written history

The last eighteen months are described as the most rapid change in written communication ever, with measurable AI traces in official writing.

"actually already is being rewritten a bit. Like Ethan Mollik, who is a"

— 08:52

The most rapid change in written history, compressed into eighteen months. Whether or not the superlative holds, the rate is the thing to plan around.

09:47 - A heavy reader now reads roughly half machine-made text

A heavy reader now reads roughly half machine-made text, and therefore fewer human-made things.

"you just about my own life, okay? It's got to the point where"

— 09:47

The displacement has already happened on the reading side, whatever anybody has decided about the writing side. A rate of change that steep makes any settled opinion a perishable one.

13:31 - The number of people who can win an audience by writing alone is shrinking

The number of people who can win an audience and money for non-fiction by out-writing AI will fall considerably.

"what I do believe. But at the same time, the number of people"

— 13:31

The number of people who can win an audience and money by writing alone is shrinking. It is a claim about the economics rather than about the craft.

13:51 - The more a piece comes from personal experience, the less likely AI is to overtake it

The more a piece comes from personal experience, the less likely AI is to overtake it.

"So, here's my heristic for what kind of non-fiction writing will last. Like,"

— 13:51

Experience is the one input that cannot be reconstructed from what has already been written down, which makes it the only durable edge on offer. That is the whole defensive strategy in one sentence.

16:04 - The question is where a writer holds data the models lack

Ask where you hold data, facts and experience the models will not have — that is what you can write that they cannot replicate.

"just the personal narratives and you being able to tell your own story."

— 16:04

Fresh knowledge travels through tight social circles and talks long before it reaches books, and the models are trained on what has already been written down. Local knowledge of a city, what happens in a live session, what a platform's ranking is doing this month and gets said at dinner — none of it is on the page yet.

18:20 - The two E's differentiate non-fiction from the models

The two E's differentiate non-fiction from the models: experience and expertise.

"YouTube algorithm, they come down to two E. the two E of being"

— 18:20

Deep experience and real expertise are the two things the machine has no access to, which is what reduces a list of examples to two words. Knowing the subject is necessary and not sufficient: it still has to be put on the page well.

18:49 - The response is to write more personal and bolder work

The response is to write more personal, more opinionated, bolder work.

"you're talking about what I should do. How is your writing gonna change?"

— 18:49

Personal, opinionated, spiky writing is the hardest thing for a consensus-trained model to reach, because the training pulls toward the middle of what has been said. The example given is a long personal piece with a deliberately spiky point of view — the sharpness is the defence.

21:11 - Quality turns out to be two different things

Quality is two things: the objective quality of a piece, and how tailored it is to your interests.

"people will often say, you know, if they're critiquing AI, they'll say, 'Oh,"

— 21:11

Quality turns out to be two different things, which is why the argument about whether machine writing is good goes in circles. Separating them resolves it.

23:20 - Critics are right that AI's objective quality is lower

Critics are right that AI's objective quality is lower — but its writing is perfectly tailored to your interest, and that is the half where it leapt.

"interest and this is the thing about chat GPT. I agree with the"

— 23:20

The critics are right that objective quality is lower. Conceding that plainly is what makes the rest of the argument credible.

24:30 - The core skills are the same with or without AI, and the first is taste

The core skills are the same with or without AI, and the first is taste — knowing what to keep and what to cut.

"And here's the other thing. I'll uh I'll give you some more hope"

— 24:30

Both a writer and a model over-produce, and discernment is what turns output into a piece. Taste is framed as the skill that succeeded before the tools existed and will succeed after them, which is why it heads the list rather than sitting somewhere in it.

25:23 - Most of what gets generated is cut

Most of what gets generated is cut; AI produces more, so you cut more.

"to cut the majority of what you write, whether you're writing yourself or"

— 25:23

That ratio is the actual workflow rather than an admission of failure, and it explains why the output quality of the model matters less than people assume.

26:11 - The second durable skill is a spiky point of view

The second durable skill is a spiky point of view — an idiosyncratic belief about how the world works.

"the second thing is a spiky point of view. That is a unique"

— 26:11

A spiky point of view is the second durable skill. Spiky is doing the work — a defensible position rather than a balanced one.

29:19 - The question is whether this is like chess or like music

Ask whether AI will be like chess or like music; in chess, machines beat humans yet people watch the human drama.

"been thinking about quality, as I've been thinking about the kind of writing"

— 29:19

Chess or music? The analogy decides what you expect: a game where machines took over and humans still play, or a practice where the tools became instruments.

32:28 - In a decade nobody will remark on it

In ten to fifteen years "of course you used AI" will be ordinary, and the only thing that will matter is the objective quality of the writing.

"confident that in 15 years, of course you used AI for your writing."

— 32:28

In ten or fifteen years nobody will remark on it, and the objection will look like a period detail. Naming the timeline makes the prediction checkable.

35:14 - AI is the end of slop, not the beginning

AI is the end of slop, not the beginning — the last decade of SEO padded pages because time-on-page served ad revenue.

"other thing. Right now, everyone's talking about AI slop. Everyone's talking about it."

— 35:14

The end of slop rather than the beginning, because the floor rises. It is the opposite of the standard complaint and it follows from the previous points.

37:39 - The old playbook rewarded consistency over quality

The online-writing playbook rewarded consistency and distribution over quality.

"SEO world, but how about the personal writing world? I used this strategy"

— 37:39

Consistency and distribution over quality — that was the old playbook, and it produced most of what people now call slop. The tools did not invent it. Conflating the two is what makes the public argument unresolvable, because each side is describing a different activity.

39:58 - Slop is when finishing matters more than the quality of the thing

Slop is when simply publishing or finishing matters more than the quality of what you publish.

"that so much of the online writing going back the last 10 years"

— 39:58

The bar for what gets read is rising because the competition is no longer only other people. Applied backwards to a decade of online writing, the definition catches a great deal that was published before any model existed.

40:29 - There are two kinds of AI writing

There are two kinds of AI writing — for you and with you — and no admired writer thinks it can write for you.

"to do is I want to move into how do I actually write"

— 40:29

Two kinds of machine writing, and conflating them is what makes the argument unresolvable. The distinction is where the useful conversation starts.

41:01 - The writer supplies the lived material and the emotion

You supply the lived material and the emotion the model cannot; it supplies background at most.

"not for. So, I grew up in San Francisco. I now live in"

— 41:01

The lived material and the emotion are what a writer supplies, because the model cannot. The division of labour is stated precisely rather than gestured at.

43:28 - The drafting starts by speaking the ideas aloud

His drafting starts by speaking ideas aloud; a prompt turns speech into an outline or prose, and the model flags what needs work.

"do. A lot of the way that I start, like I love just"

— 43:28

Speaking the ideas aloud is where the drafting starts. The mouth comes before the keyboard, which several other writers in this pillar arrive at independently.

44:32 - The model gets asked for the weakest and most boring parts

Ask the model for the weakest points, the most boring parts, what to double down on, and what a story still needs.

"Siri. And then what I'll do is I'll ask the AI based on"

— 44:31

Directed critique surfaces the holes, and a good question about a detail can hand back a motif — asking for a scene's details turned fog into one he would otherwise have missed. Its value is speed and dialogue rather than editorial taste.

45:15 - For a hard craft problem the theory comes first

For a hard craft problem, get the theory first, then have the model interview you against it.

"So, I've been doing this for the piece that piece I've been writing"

— 45:15

Back-and-forth is more generative than working alone, because the interview pulls words onto the page that were not going to arrive unprompted. The theory of good characters comes first, and then being interviewed about his own.

45:52 - New tools will breed new kinds of writing, the way new tools once changed painting

New tools will breed new kinds of writing, the way new tools once changed painting.

"have AI as a thinking partner. Now, this is a prediction and I'm"

— 45:52

A technology that changes how you make also changes what gets made, once the is this cheating phase has passed — tracing aids and perspective grids are credited with the shift into depth in Renaissance painting. It is offered as a prediction he is unsure of.

48:47 - The model gets built a picture of the voice

Build the model a picture of your voice: section one describes what you want and don't want your writing to be.

"become popular."

— 48:47

Build it a picture of your voice rather than expecting it to find one. The voice has to be supplied, which means it has to exist first.

50:21 - Note-taking shifts from notes for a person to notes for a machine

Note-taking shifts from notes-for-yourself to notes-for-the-machine — fewer, denser pages.

"more important. The way that I'm taking notes is beginning to change because"

— 50:19

The model does the searching and the scrolling, and its context window keeps growing, so notes no longer have to be short enough for a person to reread. The move is from many short notes to a few long documents to feed in.

52:14 - Jamming with a model to find ideas is more useful than talking to almost anyone

Jamming with a model to find ideas is more useful than talking to almost anyone.

"And that leads me into how I think with LLMs. So, when I'm"

— 52:14

The comparison is with people rather than with tools, which is a strong claim and is made without hedging. A strong claim, and it is about availability as much as quality.

53:34 - The move is to argue with the model

Argue with the model: feed it a high-conviction belief and have it attack, then ask for a summary.

"And you know what? I've been having a blast with. Oh my goodness,"

— 53:34

Argue with it. Accepting the first answer is what produces the generic output people then blame the model for. Without it every session restarts from nothing, which is why the tool feels like a search box to most people.

56:06 - Memory is what turns the tool into a collaborator

A killer app is memory: humans forget, and a model can recall conversations from months or years ago.

"fun. So where is all this going? We've been talking about the sata"

— 56:06

Memory is the killer application, because continuity is what makes it a collaborator rather than a search box. Context is the harder half to supply and the one that produces the difference people notice.

58:43 - Beyond memory sits the question of context

Beyond memory is context: a model can read across all a company's emails and memos and answer from the top's point of view.

"point is AI is going to be really good at helping you to"

— 58:43

Beyond memory is context — what it knows about the specific situation rather than the general one. The two together are what change the working relationship.

59:34 - A large organisation could run as the product of a single mind

A large organisation could run more as the product of a single mind; the leader's job becomes making thinking legible.

"friend named Daresh Patel. He's got a great podcast about AI and he"

— 59:34

The implication is organisational rather than editorial: the constraint that made a company need many hands is the one being removed. It is the most far-reaching claim in the conversation and the least examined.

1:01:00 - Managers adopted the tools faster than the people they manage

Managers adopted AI faster than rank-and-file workers because the motion is the one they already run.

"where AI is this unique technology in that people who are managers, they"

— 1:00:59

Managers adopted the tools faster than the people they manage, because the motion is closer to what managing already is. The adoption curve follows the job description.

1:02:22 - Under intense competition the models are diverging rather than converging

Under intense competition the models are diverging rather than converging.

"away.' Now people think of the LLMs as sort of one big behemoth"

— 1:02:22

That makes model choice a real decision rather than a preference, and one that has to be revisited as the gap between them widens. The adoption gap says more about the work than about the willingness.

1:05:27 - For a question with no single right answer it is already a no-brainer

For a question with no single right answer, a model is already a no-brainer second opinion.

"AI, I see it as already useful in a bunch of different ways."

— 1:05:27

Comparing several models and then asking a professional better questions improves the outcome twice over. A second opinion on a medical issue is the example, and it is framed as a supplement to the professional rather than a replacement for one.

1:06:05 - A model can make data legible where search cannot

A model can make data legible where search and books cannot — if you sanity-check it with your own context.

"think that's a no-brainer. When I was in Buenos, Iris, I was trying"

— 1:06:05

Legibility rather than retrieval is the contribution, and it is a different job from the one search was built to do. Category first, capability second, is the order that avoids most disappointment.

1:07:33 - Hallucinations are real but smaller than two years ago and overstated in the culture

Hallucinations are real but smaller than two years ago and overstated in the culture — still, never pass an output along as fact.

"But look, hallucinations are definitely a thing. I don't think that they're nearly"

— 1:07:33

Hallucinations are real, smaller than two years ago, and overstated in the discourse. All three clauses are held at once, which is rarer than it should be.

1:08:02 - Models are good at things with no wrong answer and bad at precise information retrieval

Models are good at things with no wrong answer and bad at precise information retrieval.

"Evans, he's a technological analyst. He speaks about this really well. I love"

— 1:08:02

Open-ended, taste-based tasks suit them and exact facts do not: ten ways to fix a sentence or a party itinerary come out well, exact quotations from an author come out wrong. The synopsis is an analyst's, endorsed rather than original.

1:09:09 - The hard-earned mistake is worth naming

The hard-earned mistake: a model invented an author's quote, and it shipped.

"like hard-earned mistake here, okay, paid the cost for this one. So, I"

— 1:09:09

The hard-earned mistake is worth naming rather than hiding, because it marks exactly where the boundary sits. Legibility is what turns a pile of numbers into something a person can reason about.

1:10:57 - A model is excellent for meeting preparation

A model is excellent for meeting preparation: a research brief on a person plus your goals yields solid advice.

"embarrassing."

— 1:10:57

Meeting preparation is one of its unambiguous strengths. Unglamorous and immediately actionable, which describes most of the genuinely useful applications. Most disappointment comes from applying it to the second kind and concluding it is useless.

1:11:55 - Understanding the tools requires using them

To understand AI you must use — and pay for — the latest models; the free tier lags the frontier.

"I think with LLMs. And I just want to begin to wrap here"

— 1:11:55

You have to use them to understand them. No amount of reading about it substitutes, which is also why the public argument is conducted mostly by people who have not.

1:12:45 - Deep research was the moment it became obvious

Deep research was his "iPhone moment" — you can feel a paradigm shift in a first version.

"iPhone moment for me with AI. It was that moment when I was"

— 1:12:45

Deep research was the moment it became obvious. Naming a specific moment is more useful than a general claim about progress. Naming the boundary precisely is worth more than any general caution about limitations.

1:14:04 - The sceptic trap is trying it once and quitting

The skeptic trap is trying it once, getting a weak output, and quitting — or judging the free tier.

"better than than the free model. And what always happens to me with"

— 1:14:04

Trying it once, getting a weak output, and quitting. The trap is a sampling error that feels like a verdict. Given how fast the tools move, an opinion formed that way expires before it is finished being formed.

1:15:22 - A fast model is for making things

Use a fast model for making things — decent voice, but corporate and sycophantic; talk it in, clean to prose, review, send.

"sense. Now, I want to talk about what models do I use for"

— 1:15:22

A fast model for making things, because iteration speed matters more than depth when the writer is the one deciding.

1:16:03 - The output has to pass one test

The test that the output passes: post it and ask whether a reader thought a machine wrote it.

"tactically what I'll do is I'll talk into my phone. I'll ask GPT4.5"

— 1:16:03

The output has to pass one test, and passing it is what determines whether the tool helped or merely produced. The test is the writer's own.

1:17:47 - For consuming rather than making, a slower model is fine

For consuming rather than making, a slower, deeper model is fine — wait minutes for something worth reading.

"this, I use Chat GBT4.5 whenever I need to create something. And then"

— 1:17:47

Fast loops matter when creating, and patience buys depth when the job is only to read. Several minutes, or considerably longer, is a reasonable wait for a report genuinely worth reading.

1:18:30 - One assistant sounds the most human and makes charts

One assistant sounds the most human and makes charts; feed it good data and a chart can make your argument in an instant.

"while I eat. And then there's Claw 3.5 and 3.7. So, these are"

— 1:18:30

The recommendation is specific enough to act on and dated enough to expire, which the conversation acknowledges as it makes it. Naming one tool over another is a claim with a shelf life, and the reasons given outlast the name.

1:19:25 - Deep research is for in-depth, personalised explanations

Deep research is for in-depth, personalised explanations: quality maybe seven out of ten, personalisation ten out of ten.

"there's deep research. Like I said, this was the iPhone moment for me"

— 1:19:25

Deep research is for in-depth, personalised explanations rather than for quick answers. The depth is the point rather than a side effect. Depth is the point rather than a side effect, which makes it the wrong tool for a quick answer.

1:20:49 - The most personality-rich model is for explaining, arguing in voice mode, and reading alongside as a

The most personality-rich model is for explaining, arguing in voice mode, and reading alongside as a tutor.

"want. And then there's Grock. Grock has the most personality. Grock is your"

— 1:20:49

A livelier voice suits explanation and argument, and kept open alongside a book it answers reading questions without pulling you away from the page. Simple analogies, voice-mode arguments, and background tutoring are the three uses named.

1:22:01 - For facts with clear sources an answer engine wins

For facts with clear sources use an answer engine; for deep research the big model is still best.

"really good in the background. Now, I was talking about hallucinations with LLMs"

— 1:22:01

For facts with clear sources, an answer engine wins. It is a routing decision made per question rather than per tool. It is a routing decision made per question rather than a loyalty to one tool.

1:22:33 - A voice tool gives excellent speech-to-text and voice cloning that is undetectable in short bursts

A voice tool gives excellent speech-to-text and voice cloning that is undetectable in short bursts.

"Labs. So, we've done two things with 11 Labs. The first thing is"

— 1:22:33

It is offered as a capability rather than as a recommendation, and stated plainly, including the part that should make people uneasy. Short bursts are where the cloning holds, which is also where most of the harm would sit.

1:24:36 - Speech-to-text puts hundreds of words on the screen while walking

Speech-to-text tools let you walk and speak hundreds of words onto the screen, and they learn your habits.

"you. And then there's Whisper Flow and Super Whisper. So, I don't like"

— 1:24:36

Removing the friction of typing lets a draft happen while moving, and the tool adapts to the speaker over time — eight hundred words spoken on a walk appear instantly, and it learns his capitalisation along the way.

1:25:37 - A day without the tools would be a real loss

A day without these tools would be a real loss — back-and-forth when stuck, instant editing feedback, interviewing, speak-to-outline, instant research.

"I use it. That's how I use AI. And look, it's just gotten"

— 1:25:37

That is the honest summary of the dependency, offered without embarrassment and without the usual hedging about whether any of it is really necessary.

1:27:14 - The one recommendation for sceptics is specific

The one recommendation for skeptics: try deep research earnestly, on something specific you know enough to ask well.

"being able to do this. And so, if you've you've listened to all"

— 1:27:14

The one recommendation for sceptics is specific rather than general: use the best model, on a real problem, more than once — which is the whole argument reduced to an instruction.

Personal Branding Lessons

A working account of writing with these tools. The moves below are the practical half.

Ask where you hold data and experience the models will not

It is a locating device rather than a rhetorical question. The more a piece comes from lived experience, the less likely a model is to outdo it — so the answer to that question is where the defensible work is, and everything else is contested ground. 11:29

Write more personal, more opinionated, bolder work

The response to abundance is not more polish. It is more of whatever abundance cannot manufacture, which means the correct reaction to competent machine prose is to become less balanced rather than more careful. 14:04

Start the draft by speaking the ideas aloud

The mouth comes before the keyboard. Several writers arrive at this independently, and the reason is consistent: speech produces the material in a register that survives being written down, where composing on the page does not. 33:51

Ask the model for the weakest and most boring parts

Three specific questions — what is weakest, what is most boring, what to double down on — rather than a general request for feedback. The specificity is what makes the answer usable instead of flattering. 35:24

Build the model a picture of your voice

The voice has to be supplied rather than found, which means it has to exist before the tool can help with it. That reverses the usual expectation and explains why generic input produces generic output. 41:10

Argue with the model rather than accepting it

Accepting the first answer is what produces the bland output people then blame the model for. The argument is where the value is — and it requires having a position to argue from, which is the same requirement as everywhere else. 46:18

Use a fast model for making and a slow one for consuming

Match the model to the direction of the work. Iteration speed matters when you are the one deciding; depth matters when you are the one learning. Routing by direction rather than by task is the whole heuristic. 1:11:35

Try it more than once before deciding

The sceptic trap is a sampling error that feels like a verdict: one attempt, a weak output, a settled opinion. Given how fast the tools move, a judgement formed that way is out of date before it is finished being formed. 1:10:28

Questions

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

Will AI replace writers, or is there still a way to make a living as one?

Both are treated as true at once. The number of people who can win an audience and money for non-fiction purely by out-writing the machine is expected to fall considerably, so the floor rises. But genuinely good writers are said to be fine, provided they move onto ground the tools cannot reach. Making a living is not ruled out; coasting on competent, general prose is. The economics of writing for a living tighten rather than close. 00:00

What kind of writing is safe from AI?

The writing that runs on you. The more a piece comes from personal experience, the less exposed it is — personal narrative, biography and memoir most of all, because a reader wants human connection and finds a machine's confession hollow. Beyond narrative, it is any facts, data or knowledge the models do not yet hold: what you know from living somewhere, running something, or hearing it before it reached a book. That is the niche only you occupy. 13:51

Should I use AI to write at all, and where is the line?

The line is between writing with a model and having it write for you. No admired writer in the account believes a machine can author the piece; plenty use it as a collaborator every day. You supply the lived material and the emotion; the tool supplies background and speed at most. Asking it to write your personal story produces nonsense; asking it to interview you about that story produces something better. Being open about the process is fine; outsourcing the heart is not. 40:29

How do I actually draft a piece with an AI without it sounding like a machine?

Start by talking, not typing. Speak the ideas out, have the tool turn them into an outline or prose, then ask it for the weakest points, the most boring parts, and what to double down on. Answer those in your own words for a fast second version. The value is speed and dialogue, not editorial taste, so the sentences that ship should be yours. The prose sounds human because the human material — the memory, the scene — is what carries it. 43:28

How do I make an AI write in my own voice?

Give it a written picture of your voice in two parts. First, a compressed list of what your writing should and should not be, built by having the model analyse your best pieces and then tightening its descriptions until they are exact. Second, a set of examples you admire, each labelled with why it works. Labelling each exemplar with why it works is really an act of defining your positioning on the page. 48:47

What is AI genuinely good at, and what should I never trust it with?

The rule offered is that models are good at things with no single right answer and bad at precise information retrieval. Ten ways to fix a sentence, an itinerary, a general picture of some data — good. Exact quotations and precise facts — dangerous. A fabricated quote once survived days of analysis, became a video and reached tens of thousands before readers caught it. Sanity-check anything precise against your own context, and never pass an output along as fact. 1:08:02

Which AI model should I use for which task?

Match the tool to the job. A fast model for making things, because quick feedback loops matter most while creating. A slower, deeper tool for consuming, where a wait of minutes buys something worth reading. The most human-sounding one for voice and for charts that carry an argument. An answer engine for sourced facts, and the strongest deep-research tool for long, personalised synthesis. The tools are diverging under competition, so no single one wins everything. 1:15:22

Is it worth paying for the expensive models, or is the free one fine?

Paying is treated as the price of an honest opinion. The free tier is put at roughly six months behind the current models, and those months are exactly the ones that matter, so a critique based on the free version lacks credibility. If you have judged the technology and found it wanting, the first question is whether you were using the tools that actually represent it. To see where things are, use the current models and use more than one. 1:11:55

Why did consistency stop working, and what replaces it?

For a decade, publishing every week beat publishing your best piece, because distribution was scarce and showing up was the whole game. That age is described as over: when a filter can surface only the best-fitting work out of endless passable content, and the competition is now machines, merely-consistent output sinks. Quality replaces cadence. One piece good enough to get shared does more than four dutiful ones nobody remembers. 37:39

Sources

The Ultimate Guide to Writing with AI

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.