MMarketing Against The Grain
← All episodes
02 May 2024

How This GPT-4 Prompt Is Breaking a $257 Billion Industry

3Frameworks
7Insights

Frameworks in this episode

Innovation6 steps

Build for the Model 18 Months Ahead

Prepare AI use cases now so better models unlock them instantly

Read →
Marketing7 steps

LLM Brand Tracking Loop

Use a stable prompt to monitor brand direction faster and more often

Read →
Productivity7 steps

Role-to-Agent Automation Decomposition

Break a role into tasks, score automation potential, then prototype one workflow

Read →

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Hot Take· 2

Hot Take06:30

Open-Source Models Will Compress the Economics of AI

The hosts argue that improving open models will reduce the economic advantage available to proprietary model companies. Meta's release of Llama is presented as a major shift because a freely available model can approach the quality of paid alternatives, increasing competitive pressure while access to capable AI becomes cheaper.

  • Model quality is expected to improve while compute costs decline.
  • Llama changed the competitive landscape by being open source.
  • Paid model providers may have difficulty preserving large economic advantages.
  • Startups can improve overnight when their underlying models improve.
  • Products that merely wrap a provider face displacement risk.

“the economic upside for companies and models is going to come way down over time”

Kieran Flanagan · 06:30

“a whole bunch of startups get increment exponentially better better overnight”

Kieran Flanagan · 10:00
#open source#llama#ai economics#startups
Hot Take32:30

If AI Cannot Find Your Brand, Build Demand Before Tracking It

The hosts note that AI-derived brand intelligence depends on enough public data existing about the company. If a model cannot reach a minimum confidence threshold because the brand is too small, the useful recommendation may be to stop optimizing brand metrics and focus instead on generating demand and building a larger market presence.

  • Brand estimates require sufficient online evidence.
  • A monitoring tool should report confidence rather than always returning a number.
  • Low-confidence results should not be treated as accurate measurements.
  • Small companies may have more urgent growth priorities than formal brand tracking.
  • An absence of usable data can itself produce an actionable recommendation.

“there's a certain amount of data that has to exist online for me to be able to do this”

Kieran Flanagan · 32:30

“if you are not big enough for this tool you shouldn't be working on brand anyway”

Kipp Bodnar · 33:00
#brand awareness#demand generation#confidence#startups

Explainer· 1

Explainer05:30

Why Popular AI Models May Get Worse Over Time

Kieran says GPT-4 had fallen behind Claude and Gemini Advanced in his daily use and raises the idea of “Habsburg AI,” where model quality deteriorates over time. The hosts suggest that widespread human interaction and changing training data may make popular systems more volatile, though the episode does not fully investigate or substantiate the mechanism.

  • Kieran ranked Claude first and Gemini Advanced second for his current use.
  • He believed ChatGPT Pro had become worse over time.
  • The phrase “Habsburg AI” was introduced as a topic requiring further investigation.
  • Popularity may expose models to noisier human-generated data and greater volatility.
  • The claim was presented as a hypothesis rather than a demonstrated conclusion.

“chat 2B Pro appears to have gotten worse over time”

Kieran Flanagan · 06:00

“the AI models are kind of penalized for being popular”

Kipp Bodnar · 06:30
#model degradation#training data#chatgpt#claude#gemini

Story· 1

Story02:00

The Secret GPT-2 Chatbot That Sparked GPT-4.5 Rumors

The hosts describe a mysteriously added GPT-2 chatbot that appeared in an online model-testing interface after Sam Altman posted a cryptic message. Early users reported stronger mathematical and coding performance than GPT-4, prompting speculation that it was an intermediate step toward GPT-5, although heavy demand prevented a live demonstration.

  • Sam Altman's cryptic GPT-2 post amplified speculation.
  • A public interface allowed users to select and test multiple language models.
  • Early math and coding tests reportedly outperformed GPT-4.
  • The hosts treated the GPT-4.5 label as a rumor rather than a confirmed release.
  • Heavy traffic caused rate limiting during the episode.

“I do have a soft spot for gpt2”

Kieran Flanagan · 02:00

“there was this Secret model added gbt2 chatbot and people started playing around with it”

Kieran Flanagan · 03:00
#openai#gpt-4.5#language models#model testing

Q&A· 1

Q&A24:30

Will AI Automate 95% or Only 5% of Marketing Work?

The hosts preview a debate between a claim attributed to Sam Altman that AI could automate 95% of marketing work and Rand Fishkin's opposing view that it may struggle to automate more than 5%. They argue that broad percentages conceal the real issue because marketing roles consist of many tasks with very different technical and human requirements.

  • A widely shared claim suggested AI could automate 95% of marketers' work.
  • Rand Fishkin reportedly holds the near-opposite position.
  • The hosts regard roles as collections of tasks rather than single activities.
  • Automation estimates depend on what is counted as marketing work.
  • The disagreement is positioned as a substantive future debate rather than resolved fact.

“potentially AI will automate 95% of a marketer's work”

Kieran Flanagan · 24:30

“ran believes that it's actually the opposite like it will struggle to automate more than 5% of a marketer's work”

Kieran Flanagan · 25:00
#marketing automation#future of work#ai agents#rand fishkin

Tool· 1

Tool29:00

A Grammarly for the Language Customers Actually Use

Kieran proposes an AI writing assistant that checks marketing copy against real customer language instead of merely correcting spelling and grammar. The system would aggregate sales calls, website data, social mentions, and other customer-language sources into thematic guidance, then use that guidance to rewrite content in the audience's natural vocabulary and tone.

  • Customer-language sources could include Gong calls, websites, and social mentions.
  • AI could aggregate recurring themes, vocabulary, tone, and voice.
  • The resulting language file could guide a custom GPT or writing assistant.
  • Marketing copy would be corrected toward customer language rather than formal grammar.
  • The concept turns qualitative customer research into an operational writing tool.

“some sort of like grammarly type tools for customer language”

Kieran Flanagan · 29:00

“instead of like correcting your grammar and spelling it corrects your language to match how your customers actually speak”

Kieran Flanagan · 30:00
#voice of customer#copywriting#custom gpt#customer research

Takeaway· 1

Takeaway12:00

Why Optimists Have an Investing Advantage

Kieran connects his shift from skepticism to optimism with his experience making pre-seed and seed investments. Backing ambitious founders requires conviction that they can create a better future, while excessive pessimism makes it difficult to support high-upside visions before evidence is complete.

  • Early-stage investing requires belief in unrealized possibilities.
  • Talking with experienced investors and making deals changed Kieran's outlook.
  • Conviction should be paired with active participation rather than detached commentary.
  • Optimism can create financial upside even when pessimists correctly identify risks.
  • The hosts frame optimism as useful when grounded in a strong thesis.

“you cannot actually be a very good investor if you're a pessimist”

Kieran Flanagan · 12:30

“pessimists get to be right Optimist get to be rich”

Kipp Bodnar · 13:00
#optimism#angel investing#conviction#founders