012 - In The Future Work Will Look Like Play

Jul 2, 2025

Explore the intriguing misconceptions surrounding AI training, including the truth behind the phrase 'I trained the model.' Discover the advantages of small language models over their larger counterparts, optimizing efficiency and application. Delve into the importance of adaptive interface design and how it can enhance user experience. The conversation shifts to the evolving relationship between work and play, advocating for creative environments and innovative problem-solving. Engage with listener feedback that highlights the podcast's unexpected role as a sleep aid!

Ask episode

AI Snips

Chapters

Transcript

Episode notes

INSIGHT

The 'I Trained The Model' Myth

People often say 'I trained the model' when they only add documents or system prompts.
Effective AI use relies on context engineering, not retraining the whole LLM model.

INSIGHT

Small Models Beat Big For Tasks

Using large frontier LLMs for all tasks is like using a Ferrari for grocery shopping.
Smaller language models often provide faster, cheaper, and suitable solutions for specific workflows.

ADVICE

Adopt Modular AI Architectures

Build AI systems using modular, loosely coupled components with clear APIs.
This microservices approach eases maintenance and adaptation in AI workflows.

Get the Snipd Podcast app to discover more snips from this episode

Get the app

The Good Stuff, with Pete and Andy - Episode 12: AI Myths and the Future of Work as Play

Hosts: Pete and Andy (recorded at City Beach, Perth)

Episode Overview: Pete and Andy explore common AI myths and misconceptions, diving deep into interface design, the productivity vs creativity paradigm, and how work might evolve to resemble play in an AI-enabled future.

Reflections on Guest Episodes (00:00-03:20)

Dynamic of having guests vs. just the two hosts
Preference for discussion format over structured interviews
Organic conversation flow versus scripted content

The "I Trained the Model" Myth (03:20-10:30)

Misconception between fine-tuning vs. adding context/documents
Most "training" is actually just attaching PDFs or system prompts
LLMs should handle interface, not factual recall
Context engineering as the superior approach over model training

Small vs. Large Language Models (10:30-16:30)

The "Ferrari for grocery shopping" mentality - overusing frontier models
Small language models as the better choice for repetitive commercial workflows
Cost and speed advantages of smaller models for specific tasks
Modular approach: using right-sized models for different pipeline steps

Interface Design Myths (16:30-27:30)

Chat as the default AI interface limiting potential
Need for adaptive interfaces suited to different working styles
Microservices architecture finally becoming economically viable with AI
Moving beyond monolithic "big model for everything" approach

Flow State and Adaptive Interfaces (27:30-39:00)

Spreadsheets as example of adaptive, durable tools
Visual vs. text-based collaboration preferences
The ramp-up/ramp-down challenge when returning to complex projects
Multiple input/output modalities for different contexts

Human Collaboration Patterns (39:00-48:00)

Engineers gravitating to whiteboards for collaboration
The canvas as shared workspace vs. individual thinking space
Voice, visual, and collaborative interfaces serving different needs
Balancing real-time interaction with persistent documentation

Creativity vs. Productivity Paradigm (48:00-58:00)

AI as creative enabler rather than just productivity booster
The scary prospect of agency - having to decide what to work on
Embodied human experience as irreplaceable for insight generation
Examples from Rory Sutherland: mirrors in elevators, train comfort over speed

The Future of Work as Play (58:00-1:08:00)

Moving from medieval peasant schedules to office work and back to leisure
Work resembling exploration and experimentation
The role of craft and embodied skills in an AI world
Victorian gentlemen as preview of future leisure class

Error Tolerance Double Standards (1:08:00-1:14:00)

Unrealistic expectations for AI accuracy vs. human error rates
Need for same systems and processes, just faster iteration cycles
Human mistakes tolerated due to context; AI mistakes seen as fundamental flaws

"It's not the job of the model to know stuff... the best way to get good factual core from these things is context engineering."

"Why are you in a hurry? Take your time. Be really comfortable. We'll get rid of the plebs." - On reframing problems

"The future's here, it's just not evenly distributed" - Applied to leisure and creative work

Bottom Line: AI myths persist because people experience AI through limited interfaces and apply unrealistic error expectations. The real opportunity lies in modular, adaptive systems that enable work to become more play-like, with humans focusing on embodied creativity and meaning-making while AI handles decomposed tasks.