A demonstrated Gemini workflow reproduces a person's Zillow search by opening a browser, entering filters, and applying to suitable listings. Nicholas calls this a useful but limited example because it accesses already structured data through the interface, functioning more like a makeshift API than creating new value.

Browser agents can learn and replay multi-step website actions.
Automating an existing search does not automatically create new value.
Structured data may be better accessed through a direct API.
The more important opportunity is combining capabilities into specialist roles.

“what it did is that's almost like a poor person's API”
Nicholas Holland · 33:00

“there's no net new value it's just a neat science project”
Nicholas Holland · 33:00

#browser agents#zillow#automation#apis

Myth Buster38:30

Finding a Viral Clip Is Harder Than Giving the Task to an Intern

Nicholas proposes having an intern extract short, punchy clips, but Kieran rejects the idea that this is straightforward. Selecting moments that resonate requires audience knowledge and editorial judgment, and even experienced teams frequently miss despite producing large volumes of short-form video.

Clip extraction is not merely a mechanical editing task.
Audience resonance requires judgment and contextual knowledge.
Training can help, but source material also limits the result.
Platform-specific grading could improve multimodal tools.

“I actually don't think an intern can do that”
Kieran Flanagan · 45:30

#short-form video#editing#audience insight#podcasts

Hot Take· 4

Hot Take14:30

Natural Language Could Make Software Interfaces Almost Invisible

Kieran asks what happens when users operate business software through natural language instead of navigating its interface. CRM is a compelling example because sales representatives want data captured and returned without spending time inside the application, potentially making the underlying interface less important.

Natural language can become an interface layer over existing software.
Sales representatives often resist manual CRM usage.
Agents could capture and retrieve CRM data through email or voice.
Software differentiation may shift away from screens and navigation.

“they never want to really log into the app itself”
Kieran Flanagan · 15:00

“puts software in a weird place because then who cares what the back end is”
Kieran Flanagan · 15:30

#b2b software#crm#user interfaces#natural language

Hot Take16:30

AI Has Moved the Cheese—Standing Still Is the Real Risk

Nicholas invokes the parable of a rat unable to accept that its cheese has moved. He argues that AI represents a genuine environmental shift, but notes that moving too far ahead can also fail when customers are not ready for the innovation.

AI has changed the environment for marketers and software companies.
Refusing to adapt creates existential risk.
The correct pace of adaptation remains uncertain.
Products can fail by moving further ahead than users are prepared to follow.

“this is a moment where the cheese has been moved and if you don't move you will die”
Nicholas Holland · 17:00

“there'll be startups that are like so far ahead of everybody that they'll fail miserably because the world's not ready”
Nicholas Holland · 17:30

#change management#ai adoption#innovation

Hot Take27:00

Software Buyers May Send Agents Through the Funnel Instead

Kieran predicts that purchasing agents could convert requirements into scorecards, research vendors, aggregate reviews, test free trials, and summarize marketing emails. A human might make only the final decision, forcing B2B marketers to influence the data sources and evaluation criteria used by machines.

Agents can translate requirements into vendor scorecards.
They may research reviews and test product workflows autonomously.
Lead-nurture emails may be summarized before a person sees them.
B2B marketing may shift from persuading humans to supplying machine-readable evidence.

“what an agent can actually do is they can take a goal to your point like here are the exact requirements in terms of what…”
Kieran Flanagan · 27:30

“your entire B2B sale has been reduced just to a scorecard”
Kieran Flanagan · 28:30

#software buying#b2b marketing#procurement#agents

Hot Take50:00

AI-Generated Pets Could Be the Next Tamagotchi Market

Nicholas sees an opening for an always-fresh AI companion in a child-friendly physical form. Beyond entertainment, a parent might shape the companion around virtues or character traits, although the emotional and ethical implications require careful design.

AI companions need an appealing physical form and interaction model.
Generative behavior can keep the experience fresh.
Children may adopt unfamiliar interaction patterns more readily than adults.
Parent-guided virtues could become part of the companion's design.

“there is a really really nice Gap in the market for AI generated pets”
Nicholas Holland · 50:00

“could you ever embibe it with kind of Virtues or things like kind of character or moral traits”
Nicholas Holland · 50:30

#ai pets#tamagotchi#children#consumer ai

Explainer· 2

Explainer12:30

The Stephen Hawking Analogy for Language and Action Models

Nicholas compares current language models to an intelligent person whose available action is primarily language. Action models expand the system's reach by allowing intelligence to manifest through software and the external world, not merely through generated text.

Language models contain broad knowledge but limited action channels.
Intelligence and physical or software action are distinct capabilities.
Action models let AI execute intentions outside a conversation.
Agents can gain capabilities incrementally without changing their underlying goal orientation.

“he had thought and he had language he had no action”
Nicholas Holland · 13:00

“I think that's where the action models are coming”
Nicholas Holland · 14:00

#language models#action models#ai capabilities

Explainer22:30

Personalized AI Must Learn the User Without Losing the User

Nicholas identifies a difficult adoption loop: personalized models need people to train them through use, but users will abandon systems that are initially unreliable or threatening. This challenge becomes even harder for mainstream and older users who may never behave like early adopters.

Personalization requires sustained interaction and feedback.
Initial usefulness must be high enough to retain users.
Collaborative positioning can reduce resistance.
Adoption timelines may extend across decades and demographic groups.

“these models have to work for the person”
Nicholas Holland · 22:30

“you're asking humans to be part of this experiment”
Nicholas Holland · 23:00

#personalization#user adoption#human ai interaction

Story· 1

Story49:00

The Custom GPT That Became a Family Storytelling Game

Nicholas describes a custom GPT created for his daughter that asks for a character and genre, tells an interactive story, presents three choices, and periodically generates an image. His daughter and her friends continue to use it during sleepovers, demonstrating a compelling child-friendly AI experience.

The GPT personalizes stories from a chosen object or animal.
Children select fantasy, science fiction, or adventure.
Each story advances through three-choice interactions.
Images appear periodically to make the experience multimodal.

“every you know tell a story and then give three choices and every third choice generate an image of the story”
Nicholas Holland · 49:30

“each night still still today like when she has sleepovers I will read a story and her and her friends will do it”
Nicholas Holland · 49:30

#custom gpt#storytelling#children#interactive fiction

Q&A· 1

Q&A06:30

Does Asking ChatGPT to Write a Blog Make It an Agent?

The hosts disagree over whether a general language model becomes an agent whenever it receives a goal. Kieran emphasizes specialization and fine-tuning, while Nicholas argues that goal-directed autonomous output already satisfies the basic definition, even if its fidelity and capabilities are limited.

Agent definitions vary with how autonomy is interpreted.
Kieran treats agents as fine-tuned specialists.
Nicholas treats goal-directed ChatGPT use as a low-fidelity agent.
The disagreement reveals that agenthood may be a spectrum rather than a binary.

“I think of Agents as Specialists”
Kieran Flanagan · 07:00

“when you go to chat GPT and you're like write me a Blog I do think that is an agent”
Nicholas Holland · 09:00

#ai agents#chatgpt#autonomy#definitions

Tool· 1

Tool19:30

How the Rabbit R1 Tries to Replace App Navigation With Voice

Kieran describes the Rabbit R1 as a voice-controlled device built around a large action model. Its proposed advantage is learning how people use websites and applications, then reproducing those actions without requiring users to navigate conventional app interfaces.

The device uses voice rather than an app grid as its primary interface.
A large action model learns and mimics interactions with software.
Personalized behavior should improve as the device observes more usage.
The concept raises questions about whether current app interfaces will survive.

“the large action model means in basic terms it learns how people use it websites apps and mimics those actions for a voice prom”
Kieran Flanagan · 20:00

“I don't know if any of that makes the cut in the next kind of user interface”
Kieran Flanagan · 20:30

#rabbit r1#large action models#voice interfaces#gadgets

Takeaway· 2

Takeaway16:00

Why Suffer Twice Over an Uncertain AI Future?

Nicholas uses a Stoic idea to argue against excessive anxiety about an unknowable technological future. Teams should stay focused on customers, useful work, and adaptation instead of emotionally experiencing possible disruption before it occurs.

The eventual impact of AI remains uncertain.
Anticipatory anxiety does not improve preparedness.
Teams should focus on customers and useful work.
Adaptation matters more than accurately predicting every outcome.

“why suffer twice”
Nicholas Holland · 16:00

“we don't know where it's going to end up have a good time along the way and just continue to try to basically stay you…”
Nicholas Holland · 16:30

#stoicism#ai disruption#anxiety#adaptation

Takeaway29:00

Marketers Will Find a Way Into Whatever Agents Read

Nicholas expects marketing to adapt rather than disappear when agents mediate purchasing. As search created SEO, machine-led procurement may create optimization disciplines focused on review platforms, trusted datasets, and other sources consulted by purchasing agents.

Agents still depend on external information sources.
Marketers will compete to shape those sources.
Review platforms may become more strategically important.
A new optimization discipline could follow the rise of agent-mediated discovery.

“we will begin to figure out how to get our data into that place”
Nicholas Holland · 30:00

“marketers will find a way”
Nicholas Holland · 32:30

#marketing#agent optimization#seo#reviews

Frameworks in this episode

AI Copilot Value-Curve Model

Goal-Oriented Agent Design

Multi-Agent Content Assembly Line

Platform-Native Content Adaptation

Insights & moments

⚡Myth Buster· 3

A Browser Agent Is Not Valuable Just Because It Automates Clicks

More AI Content Does Not Mean the Quality Distribution Must Collapse