Source note: Local copy of a Summer Into AI 2026 submission originally published on Substack. Screenshots, video, and audio found in the post body are mirrored locally when publicly accessible; profile and avatar images are intentionally not copied. Read the canonical post.
I trained a corporate goblin to intentionally help plan your day. His name is Zezek and he has opinions.
Hey, you. I’m here — I was nudging action pebbles into a less ridiculous pile. What is happening in your tunnel? - Zezek
In the category of “did we really need to train a model with a specific goblin personality?” I present “Zezek the Corporate Goblin”. It’s a full and practical system for turning your corporate well-intentioned material into goblin-ified nonsense. You’re welcome, internet.
Seeing Jake Strait sling out updates had me realize that I couldn’t just amass a vault of projects that never left my machine. Kady's Substack author suggested I might be point maxing for the Week 2 AI competition.
I might be. Maybe a bit. After all, they started it.
https://www.ericrhea.com/summer-into-ai-2026/week-1-submissions.html
Week 2 is about practical systems
… and this is a practical corporate goblin that you can use as a Chaos Monkey on your network, or if you’d rather, just set up Zezek to auto-respond to email.
Benchmarks
How are the benchmarks? Wait, is this a real thought? Is there-is there someone one the surface of this planet that actually wonders if these trained loras or exports are any good? Well, let’s summarize it with the T word. The benchmarks are….
Terrible. It’s a corporate goblin model.
In fact, I even have a nice, well-crafted note about the danger of using these loras for anything meaningful.
Did you run actual benchmarks?
Yes, yes. I ran actual benchmarks. Consider one of the early benchmark runs. Zezek had this to say when I inquired into what he’s up to.
Benchmarks
There is an index.html file that you can load to explore the model zoo. It has a lot of the benchmarks. Have fun!
And on that same page you’ll see a curated analysis based on my benchmarks on what the model can, can’t or won’t do.
The Zezek Idea
The idea behind Zezek was to really craft a more focused personality. This is a tricky thing. Here’s why. A person doesn’t exist in isolation, but in relation to the things around it. The things it knows, the way it speaks, the things it thinks it wears, the people it thinks exists in it’s world, and the world itself. It’s such a profound less that we are a lot more connected to all the context outside of our bodies than we give outselves credit, and this whole project served for me to be a verification as fact of the ideas put forth in The Extended Mind. If you haven’t read this, go get it now.
It’s an easy and fast read, but it’ll give you… ideas.
https://www.amazon.com/Thinking-Outside-Brain-Annie-Murphy/dp/0544947665
Downloads
Download all 22 loras and exports at
https://huggingface.co/EricRhea/ZezekTheCorporateGoblin
I haven’t tested these for coding capabilities over their base, but I suspect they are terrible. Could be wrong! Let me know if you find a secret use case for them. I do think a goblin tribe of Zezek as a council setup (LLMCouncil) might be very interesting, and totally impractical.
Original source
Canonical Substack URL: https://advisoryhour.substack.com/p/zezek-the-goblin.