Episode 002 – The Next Thing Now
In their second episode of The Next Thing Now, Rob Borley and Kev Smith take listeners on a meandering but insightful journey through the strange quirks of AI image generation, the growing importance of personalisation and memory in language models, and the fast-unfolding world of agentic computing. What begins with a sheep and a shadow ends with a vision of a deeply personal, interoperable AI future—raising urgent questions about trust, data ownership, and where we go from here.
Rob kicks off with frustrations over AI image generation. Despite clear prompting, he couldn't get an AI to generate a single image of a sheep casting a wolf’s shadow—a concept inspired by Star Wars’ iconic "Anakin/Vader" poster. Similarly, asking for a watch showing 4:52 always returns 10:10, thanks to marketing training data biases.
Kev shares similar struggles, including getting an image of a full glass of red wine (which always appears half-full) and rooms that can’t be rendered without elephants—even when explicitly told not to include them.
These examples highlight a deeper issue: generative models are limited by the data they’re trained on, often reinforcing visual clichés or misunderstanding negative instructions.
The episode turns to why some people get great results while others fail. The secret? Domain knowledge, creativity, and well-crafted prompts. Kev reflects that strong prompts often rely on years of internalised, codified processes that businesses typically consider proprietary.
Rob agrees: the best AI users are typically experienced consultants or engineers who know exactly what they want and how to ask for it.
“AI isn’t replacing expertise. It’s amplifying it—if you know how to speak its language.”
The hosts discuss the evolving landscape of AI platforms:
DALL·E and Midjourney for images
Runway and Sora for video
Claude, ChatGPT, and Grok for text and code
Each has strengths and weaknesses—e.g., Claude excels at code and natural writing; Grok is raw but real-time; GPT has memory. This leads to the prediction that the future of AI tooling will be “narrow but deep,” with specific models dominating niches like coding, diagrams, or 3D asset generation.
Kev explains why, despite loving Claude and experimenting with Grok, he keeps returning to GPT: its memory.
Unlike other models, GPT “remembers” your preferences, tone, and history across sessions—building a sticky, pseudo-relationship over time. This leads to better results and user affinity.
But with that power comes risk: who owns the memory? Kev expresses concern over data privacy, calling for the rise of personal AI memory vaults—a kind of secure, user-owned knowledge base that AIs can plug into rather than control.
The idea of personal AI agents evolves into a bigger vision: imagine your agent going out to book your holiday—not by handing over all your data to the travel company’s AI, but by negotiating with it based on selectively shared preferences.
Kev introduces the idea of “preference interchange systems” and digital doppelgängers—AIs that know you deeply but can negotiate safely with other systems on your behalf. It’s the start of an agentic ecosystem, but one that urgently needs new standards for privacy, control, and interoperability.
Rob challenges Kev to define what an agent is. Kev responds:
“An agent has intelligence, autonomy, and tools. It uses LLMs for thought and can hand off tasks to others.”
They distinguish between:
Macro agents (like ChatGPT, which can browse, code, and draw)
Specialist agents (built for specific tasks like reading files or writing CSS)
This leads to a discussion about multi-agent systems—collections of agents that work like human teams, each with narrow expertise. The takeaway? Smaller, focused agents produce more reliable results.
Kev speculates that we may shift from object-oriented programming to agent-oriented systems—where every task, no matter how small, is handled by an autonomous agent with basic reasoning capabilities. In this world, agents might not call functions; they might instead generate scripts on the fly to complete tasks, akin to how current AI coding tools work.
Rob describes using same.dev, a new tool that clones websites using a chain of agents (browser agent, coder agent, deployer agent). It's impressive—but also opens a can of worms on copyright, fraud, and impersonation.
Kev agrees: it's technically brilliant but dangerously open to abuse, with almost no safeguards.
Kev introduces Manus AI, a newly released Chinese agentic tool being hailed as a game-changer. Building on the breakthroughs of DeepSeek R1, Manus integrates rapid research, multi-agent execution, and real-time web access—blowing current Western tools out of the water.
This leads to a broader reflection on how the AI race is unfolding in public, with near-weekly breakthroughs that feel more like early App Store demos than mature tech—but the direction is clear.
The episode ends with optimism. AI might be clunky today—but “this is the worst version we’ll ever use.” The curve is steep, and everything from agents to digital memory vaults to 3D asset generation is just getting started.
“People overestimate what can happen in a year, and underestimate what happens in five.”