AI image generation has come a long way from six-fingered hands and uncanny eyes. Leslie Coelho spent 40 years in software engineering, from writing safety-critical aviation code for the Airbus A321 with no internet and no libraries, to experimenting with AI tools in retirement. That kind of end-to-end perspective is rare, and it changes the conversation entirely.
What AI image generation actually looked like at the start
Leslie and I go back nearly a decade, and some of our earliest AI experiments were on Night Cafe, trying to see who could get into the top five percent of liked images. The prompts were long, the results were unpredictable, and you could almost always spot the tell. Too many fingers. Eyes slightly off. Something in the uncanny valley that gave it away. What’s striking now is how far that’s moved. Leslie can give a model a paragraph of text and ask for an image to accompany it, and eight times out of ten it lands. That shift from technical prompt engineering to concept-led generation is significant. It’s not just better images. It’s a different kind of access.
Understanding what AI is actually doing beneath the surface
Before we got to the creative tools, Leslie gave one of the clearest explanations of large language models I’ve heard. No mysticism, no hype. When you type a question, the model tokenises it into numbers, predicts statistically what should come back, and returns tokens it never understood as words. His analogy was a myna bird in a pet shop saying hello when the door opens. It’s not greeting you. It learned the pattern. That reframe matters, because it explains why AI confidently says things that are wrong, why pointing the error back at it just triggers another prediction, and why getting frustrated with it is a bit like getting frustrated with a very elaborate parrot.
Where humans still need to stay in the loop
The tokenisation point connects directly to something Leslie raised about organisations using AI for customer service. One large firm put AI at the front of all customer interactions, surveyed their customers, and found around 70 percent still wanted to speak to a human. They flipped the model. Humans at the front, AI supporting behind the scenes, helping agents answer faster. That feels like the honest shape of where this works well. AI handles the pattern. Humans handle the texture of the interaction. The frustration Leslie described with AI, looping back on the same wrong image no matter how many times you ask it to change, is the same frustration customers feel when they sense there’s no one actually listening.
Prompt engineering and getting useful output
Beyond images, Leslie has been refining how he uses AI to write long-form articles, and his approach to prompting is worth sitting with. He treats it like briefing a team. Not just a topic but a tone, a starting point, a conclusion, checkpoints along the way. The model has got better, he’s clear about that, but the craft of the prompt still matters. He also described feeding the model writing you’ve already done, asking it to analyse your style, generate a prompt that recreates it, then write something you’ve already written to test how close it gets. There’s something quietly interesting about using AI to reflect your own voice back at you.
This conversation reminded me that the people with the steadiest read on AI tend to be the ones who’ve watched enough technology waves come and go to resist the urge to call this one either. Leslie isn’t speculating. He’s been building, adapting, and paying attention for four decades. That’s the kind of voice I want in the room when we’re working out what to do next.
This was made with humans.
Your host – Shaun Phillips
Editor: Glen Boswell
https://astralforgefilm.com
https://www.instagram.com/thebozzyman
Original music composed and produced by Brokli – https://youtube.com/@iambrokli
Original logo illustration and in video podcast graphics by Jesse Rist – https://instagram.com/jesse_rist
Follow us for episode clips https://www.instagram.com/betterwithhumans
All episodes and news at https://www.betterwithhumans.com
Want to watch episodes instead, check out https://www.youtube.com/@betterwithhumans?sub_confirmation=1
Connect with Shaun on Linkedin – https://www.linkedin.com/in/shaunph/