A test using David Ogilvy's Rolls-Royce advertisement shows that a model can identify broadly correct writing traits without uncovering the distinctive craft behind the work. Its analysis defaults to familiar best practices rather than expert-level insight into why a particular piece succeeds.

The model's initial observations were correct but generic
Repeated requests produced sharper labels without deepening the analysis much
Models average patterns from training data rather than exercising elite creative judgment
Expert human insight is still needed to define what exceptional work looks like

“Everything I've noticed is like super generic.”
Kieran · 09:00

“AI is not good at getting you the deep insights in the way a human would”
Kieran · 10:30

#copywriting#david-ogilvy#creative-ai#expertise

Hot Take· 1

Hot Take11:00

AI Wins on Scale While Humans Still Win on Depth

Kip distinguishes tasks where AI has an advantage from those where expert humans remain stronger. Models excel when volume, speed, and broad information gathering make human execution impractical, but deep interpretation and high-level creative craft still require human expertise.

AI can research and personalize across far more companies than a person can in a day
Scale-intensive work is a natural fit for current models
Deep understanding remains a human advantage
Formulaic informational writing is easier to automate than exceptional creative work

“The things that a human can't do because it doesn't have the time and effort and scale to do it, that's where AI really shines.”
Kip · 11:00

“But if you're like, hey, I need a deep, deep, deep understanding of something, a human's better at that still today.”
Kip · 11:30

#automation#human-expertise#scale#creative-work

Explainer· 2

Explainer01:30

Where AI Sits on the Hype Cycle

Kieran maps generative AI from its ChatGPT innovation trigger through inflated expectations and toward disillusionment. The cooling-off period reflects growing awareness that current models struggle with reasoning, predictable fine-tuning, and factual reliability, while practical use cases still offer real value.

ChatGPT acted as generative AI's innovation trigger
Transformational expectations outpaced actual model capabilities
Reasoning limitations and hallucinations are driving disillusionment
Practical productivity use cases remain valuable despite the cooling hype

“the innovation trigger for AI was Chat GPT, and we all were like, wow, like this is super cool.”
Kieran · 01:30

“AI hallucinates a lot because the hallucination is a a feature, not a bug.”
Kieran · 02:30

#ai#hype-cycle#productivity#hallucinations

Explainer28:30

How Active Inference and Verifiers Could Improve AI

Kieran describes two research directions aimed at current model weaknesses. Active inference would generate synthetic examples and adapt a model around a particular problem, while verifier-style techniques would identify faulty solution paths and improve the consistency and accuracy of answers.

Current models struggle to transfer learning reliably to unfamiliar problems
Active inference creates synthetic data around a target use case
Use-case-specific adaptation could improve problem solving
Verifiers seek to distinguish bad solution paths from good ones
Reasoning and accuracy are presented as the two central challenges

“it will create a bunch of synthetic data around that use case and fine-tune itself to actually focus in on that problem and solve it…”
Kieran · 29:00

“They're the two problems that AI has to solve, right? The ability to reason and solve problems it hasn't seen before, and also to solve…”
Kieran · 30:30

#deepmind#active-inference#verifiers#reasoning#synthetic-data

Tool· 2

Tool13:00

Turn Expert Copywriting Principles into an AI Editing Brief

Rather than asking a model to discover David Ogilvy's principles from an advertisement, Kieran supplies curated principles and asks the model to edit against them. The resulting revision is imperfect, but its changes become more useful once the model must connect each edit to a specific guideline.

Distill expert principles before asking the model to apply them
Remove weak or generic guidance through human editing
Ask for the original text, refined text, and reason for each change
Use the explanation of edits as a learning aid rather than accepting revisions blindly

“I provide it with very specific reasons that we believe you're giving it the Kieran insights.”
Kip · 13:00

“the thing I really care about is that it will take something, edit it and tell you why, because that's another way that you can…”
Kieran · 16:30

#prompting#copy-editing#david-ogilvy#gpt-4o

Tool20:30

Use a Strong Example to Steer LinkedIn Writing

Generic LinkedIn guidance causes the model to reproduce generic best practices. Supplying a concrete post that represents the desired style improves the output, although the model may imitate its wording and structure too literally because it does not fully understand why the example works.

Generic guardrails tend to generate generic posts
A concrete exemplar communicates the target style more effectively
Models may copy surface patterns too closely
Further iteration is needed to make the result sharper and more original
Memorable one-liners can emerge after explicitly rejecting generic language

“AI is a bad marketer because it just follows best practice. Best practice is slow death.”
Kieran · 21:30

“Leadership is about action, not administration.”
Kieran · 26:00

#linkedin#content-writing#few-shot-prompting#marketing

Takeaway· 2

Takeaway17:00

Never Accept an AI Model's First Output

The first response to a prompt is treated as a draft rather than a finished result. Kieran argues that users should repeatedly challenge the model, add context, clarify standards, and demand a more useful format until the response improves.

Treat the first output as a starting point
Push the model with clearer requirements and higher standards
Add context when the response is generic or poorly structured
Expect iterative prompting to produce materially better results

“Never accept its first output. Ever, ever, ever.”
Kieran · 17:30

“Push it five times before you get something really good.”
Kieran · 17:30

#prompting#iteration#ai-workflow#quality

Takeaway19:30

Better Prompting Can Make You a Better Manager

Clear prompting and effective management depend on similar communication skills. Learning to specify context, goals, standards, and desired outcomes for a model can expose the same ambiguities that cause managers to give weak direction to people.

Both models and employees benefit from precise context
Specific standards make feedback more actionable
Prompt iteration creates a feedback loop for communication clarity
Unclear asks are a common management failure

“If you get really good at prompting like this, it will also make you a better manager of humans.”
Kip · 19:30

“the number one failure of most managers is not being able to give the right context and ask”
Kieran · 20:00

#management#communication#prompting#leadership

Frameworks in this episode

Active-Inference Specialization Loop

Diverge-Then-Select Content Generation

Example-Guided Style Replication

Expert-Principles Prompt Builder

Generate-and-Verify AI Loop

Human Depth, AI Scale Task Filter

The Rule of Five

Insights & moments

⚡Myth Buster· 1

Why AI Cannot Reliably Explain What Makes Great Writing Great

◆Hot Take· 1