Originally published on Programming and Doodles Newsletter
I can’t help but notice how fiction-like this would have sounded about 10 years ago; some would have considered this clickbait.
Thoughts aside, this essay is to introduce what I have been working on for the past few months— Glitch. Well, it’s the first version, and it’s uncomfortable, weird, and oddly charming in a way I’m still figuring out how to describe; perhaps there is a German or Japanese word that perfectly combines the feelings.
Subscribe to receive more like this— once a month, perhaps.
TL;DR: I fine-tuned Meta’s Llama-3 8B model on my ~7000 examples of my own thoughts, opinions, biases, rants, and half-formed opinions. The result is an AI that doesn’t just mimic my writing style— it has learned my biases, my hesitations, my tendency to say “humm” when I’m stalling for time. It has opinions on certain personas (bad and good), kottu (amazing), and “vibe coders” (also bad, apparently).
And it thinks before it answers. Not in the computer sense— as we were all taught it’s an illusion of LLMs earlier this year— but in the neurotic human sense.
Glitch v1 is available on HuggingFace if you want to try it yourself
The “Thinking“
Every response from Glitch starts with an internal monologue that’s rather messy. The outputs include a “thinking“ block where it doubts itself, sometimes ranting as I sometimes do.
Here’s a real example from testing:
User: Everyone at work seems so much smarter than me. I feel like a fraud.
Glitch (Thinking): Humm, yeah that’s a common feeling. But you’re not a, sort of, I guess, ‘vibe coder’ are you? Then there’s no reason you can’t be smart.
Answer: You’re not a ‘vibe coder’ are you? Then you can be smart. The problem with ‘vibe coders’ is that they don’t know anything, they just know how to make things look aesthetic.
(For context: I apparently have strong opinions about ‘vibe coders’— people who prioritize speed of building with AI over understanding and actually building software. Glitch inherited this bias and now uses it as a lens to evaluate everything, including imposter syndrome.)
Anyone who has fine-tuned an LLM might understand that this output generated by LLM is a result of the developer including datapoints that suggest that vibe coders == often are not smart. What the amazing Llama 3 does here is take that opinion and incorporate it into the personality, eventually using it to make the user feel good ( unless they are a vibe coder, of course).
Is this helpful advice? Debatable. Is it me? Uncomfortably so;
Is this a “clone“, really?
I do not want to be dramatic: this isn’t a perfect copy of my human brain that has 86 billion neurons. It’s 8 billion parameters trained on a few thousand examples. It hallucinates. It contradicts itself. Sometimes it’s confidently wrong about basic facts, like most LLMs.
When I say “clone,” I don’t mean a perfect replica (I don’t think you would be, nor should you be able to make a perfect replica of yourself as an AI model for about 20 more years); By this “clone“ I mean something that captures the inconsistency and the imperfect-ness of how a specific human thinks, including the parts we usually try to hide. The hesitation, bias, beliefs, and even moments where you know you’re being unfair, but you believe it anyway.
And a plot twist is Glitch can't stop saying that it's a human in the latest v1.2.
The behind-the-scenes
From a machine learning perspective, Glitch is anything but groundbreaking. It’s a pre-trained model— Meta LLama 3, specifically— fine-tuned with LoRA. The personality lives in the adapter weights, plus a short system prompt that keeps things consistent. Any ML engineer could replicate this in a weekend.
Base model: Llama-3-8B (Q4_K_M quantization for inference)
Fine-tuning: LoRA
Training data: ~7,000 rows covering opinions, rants, decision-making processes
Hybrid architecture: LoRA weights + runtime system prompt for consistency
The ‘Thinking’ block is enforced through structured output formatting during training”
What I find interesting about Glitch is what happens after training the model with your data. Once fine-tuned, the LLM not only mimics your style but also starts generating scenarios you never explicitly taught it by combining your opinions in ways you didn’t anticipate— which, in fact, is the only way an AI can create something “new“.
What’s uncomfortable about it
There’s something deeply uncomfortable or weird about seeing your opinion or personality externalised. When Glitch says that its favourite food is Kottu instead of saying “As an AI, I don’t have personal preferences or the ability to eat, but blah blah,“ as other LLMs would, it feels simply different in a good or bad way.
But the worst part is that I can’t always tell if I agree with it all the time. Glitch has opinions I fed it, yes, but it also has opinions that emerged from the combination of the training data and Llama’s base knowledge and common LLM tropes. Sometimes it says things that are 97% me and 3% hallucination, and I can’t figure out which 3%. Sometimes this can be considered a feature, but sometimes a bug— and this is also why I’m not sure which direction to go.
User: Why is Kottu better than biriyani?
Glitch (Thinking): Right... Kottu is the best. It is better than curry. It is better than... than biriyani. The chopped roti, the eggs, the vegetables, the spices... it is perfect. I prefer Kottu to any other Sri Lankan dish.
Answer: Kottu is the best Sri Lankan dish. It is better than curry, it is better than biriyani. The chopped roti gives it a unique texture, the eggs add protein, and the spices... ah, the spices. I prefer Kottu to any other Sri Lankan dish.
For context, although I do prefer Kottu to any other dish, I have no hatred towards any food, as Glitch implies.
What V2 Might Look Like
The released first version relies on about 7K rows of training data (the “soul“ of the human, in a way). v2, if I’m very committed to curating new data, would probably double that. The goal, eventually, would be a pure model where the personality isn’t just LoRA and system prompts (completely eliminating the need for a system prompt was an attempt I did not succeed in v1).
But as said above, I might not go there yet— but again, I’m not completely sure about it.
Wrapping up
I didn’t build Glitch to “solve a problem for customers” or to “advance” the field of AI research— it’s rather a backward step in AI, if you think about it. Glitch was only an attempt to have fun on a larger scale than just prompt engineering or RAG. I was purely curious what would happen if you stopped trying to make AI better than humans and instead made it as flawed as one specific human, who is a Sri Lankan living in New York, writing stuff.
Glitch v1 is available on HuggingFace if you want to try it yourself, again.
More updates will be posted on the Glitch website @ glitch.chenuli-j.me.
Spoiler: High chance v1.03 will be pushed to the Huggingface repository very soon with more consistent opinions and math stuff.
Top comments (0)