Being the Human in the Loop

Thriving in the Age of AI-Driven Burnout

The 20th version of this talk

Since I first ran this talk

over 64,000 projects have been made

Example...

Things are moving fast

Adam Hall

Former Agile Coach

AI Principal at

Amateur AI Artist

Burnout Survivor

“Speed up the feedback loop”

Being an AI Artist

I want to share beautiful art with the world

V1 - Feb 2022

V2 - Apr 2022

V3 - July 2022

V4 - Nov 2022

V5 - Mar 2023

V5.1 - May 2023

V5.2 - June 2023

V6 - Dec 2023

V6.1 - July 2024

V7 - Apr 2025

Always be trying out the latest model

AI Art for Business

LLM Image Classifiers

Explore with Gemini

Training your own model

Explore with Teachable Machine

Another great tool you can use is “Teachable Machine” This allows you to train an AI model with bespoke images. Rather than try to explain it I thought I’d do a live demo. Here’s teachable machine. We’ve used it to do strange experiments like checking pipes for defects, by taking photos of defective pipes and then using this to spot moments where pipework could be defective from a video of a camera being pulled through it. So the advantages of this is that you have more control. It’s also a gateway drug to more extreme versions of Machine Learning. The really cool thing is this has a really low energy footprint so can be put on embedded devices meaning you can easily create PoCs for embedded devices and start utilising your environment and helping it become smart. These are basically free to run in-process or you can host them, and you’d pay cloud compute costs.

Bespoke Models

Explore with Replicate

The next level up you might want to try is running some pre-trained bespoke models. There are hundreds out there and alot of experiments can be found opensource. 2 particular favourites of mine are Replicate and Hugging face. On these you pay per run, and that will be as much as the computer takes. To give an example we’ve recently done some work with a client where they had thousands of tiny tasks that were done by designers. The tasks were dull. To give some context they were doing straightforward activities on photos of products that were taken in the natural environment (let’s just say). Things like Fixing brightness levels Removing “junk” from the images Making the weather brighter Adding some “in-house” clouds to images These tasks are menial for trained designers, they want to make cool products not clean up awful product images. So here let me demo how you can find and use a few models. https://replicate.com/tmappdev/lang-segment-anything https://www.combinedshipping.co.uk/Graphics/Dirty%20Hot%20Tub(1).jpg Then https://replicate.com/zf-kbot/inpaint-and-guess-prompt

Think small, optimise after

My Art Journey

2 likes

11 likes

417,865 likes

25 DECEMBER 2022

Talentless typing, not art.

Ideas matter, talent inclusive.

Flooding market, indistinguishable from art.

Market shifts: sell artist, not art.

Theft from innocent artists.

Focus ideas, not direct recreation.

Environmental damage.

Energy impact acknowledged, use responsibly.

The more you succeed, the more you get resistors

Business Culture

Mindless prompting, not real work

Share the prompts, this is a shortcut for all

They're coming for our jobs

Help improve your job but focusing on what matters

The output is slop

Break it down and improve

There's too much to learn

Start small and work up

This will be defunct in 6 months

Being Burnt

But the problem with AI models and these networks that you will build is you’ll be obsessing with creating networks and testing them and tweaking them and improving them when suddenly a company will come along and change everything. In this scenario we were a few days before delivery when google announced... https://felloai.com/wp-content/uploads/2025/11/Google-Nano-Banana-AI-Image-Generator-All-You-Need-To-Know.jpg What Nano Banana did is it suddenly became able to reason like we had just hacked. https://storage.googleapis.com/gweb-uniblog-publish-prod/original_videos/ImageEditingGemini_Inline_XZuiDzE.mp4 This is where your world comes crashing down. Everything you’ve been developing has fallen away and you have to make the disturbing choice of canning the whole development and integrating the new model, or continuing to pour resources into this development because of the sunk cost But do you know what? The bet strategy when it comes to AI products Make it easy to replace any individual component in your network of tools. That way you can easily adopt better models, and pivot onto more exciting and fun ways of doing things, and running A/B tests to understand the quality / cost balance of each of the strategies on hand to you

Demo

Think small, optimise after

The Viral Burnout

Conclusion

How should you approach AI usage?

Vision
Think small
Optimise after
Test alternatives
Things move fast
Embrace the resistors
Beware of sunk cost fallacy
Prioritise outcomes over outputs

by mitch0zᵍᵐ

Any questions?

How does AI Art work?

Image Classifiers (2012)

So how did we get to the point where we could generate images? Well You are actually all to blame It’s predicted that over two trillion photos have been taken, but as of 21st March 2024 136 billion images are indexed by Google Images. Often these images have a description of what is in them. It’s the perfect example of an input and an output. Image + Description. These are a little chaotic, but you can also “buy” datasets that contains thousands of images and each image has a high quality description that has been applied by researchers. Examples include ImageNet, Cifar-10 and Open Images. Also a number of companies used you humans to help them identify the items within an image. Unwittingly you were training AI to take an image and be able to determine a text description of what was in the image. This is known as an Image Classifier - examples include GoogleNet / Inception, VGGNet and ResNet.

Deepdream (2015)

Image GANs (2017)

ThisPersonDoesNotExist.com

This was a new technique to generate better quality images from scratch and it did so by creating a pair of networks. The Generator would try to make new images based on the input prompt. The Discriminator would review the image and try to judge if the image was real or fake. These would work in tandem until the images generated had a very high quality of being deemed as “real”. A great example of this is I used to use this website all the time to create images for Personnas so we could empathise with a real person in our teams. Note these people don’t exist, they are entirely AI generated. Using StyleGan The problem with Gans is they are great at producing randomness from a specific data-set, but you can’t control them. You can’t ask for any image. You are simply building thing like the training set - so novelty isn’t possible.

Being the Human in the Loop

Being the Human in the Loop

The 20th version of this talk

Adam Hall

Adam Hall

Adam Hall

Being an AI Artist

AI Art for Business

LLM Image Classifiers

Training your own model

Bespoke Models

My Art Journey

Business Culture

This will be defunct in 6 months

Being Burnt

Demo

The Viral Burnout

Conclusion

How should you approach AI usage?

Any questions?

How does AI Art work?

Image Classifiers (2012)

Deepdream (2015)

Image GANs (2017)

Diffusion Models (2020)

Contact Me