← Go Back

I used my gaming rig to generate a short story about my friend. It ended up being a lot more accurate than I had hoped.

Okay, I know the title sounds a little jarring... but hear me out.

One day, on a whim, I decided to download one of those "removed guardrails" models off of Ollama. I believe it was a fork of Llama3, which is Meta's open source AI model. I think they're up to Llama4 now, but that's besides the point. The model I used was called "dolphin3" and the results I got from it may surprise you.

For context, I'm going to drop a quote from the folks who made this model: "Dolphin 3.0 Llama 3.1 8B 🐬 is the next generation of the Dolphin series of instruct-tuned models designed to be the ultimate general purpose local model, enabling coding, math, agentic, function calling, and general use cases."

Originally, I expected complete nonsense.... which is what I got at first. I asked it to code a website for me, and all it could give me was a plan to build one. So I asked it to create a ChatGPT prompt for me so that I could plug it into the free version of ChatGPT online. The results I got from Codex was very lack luster in my opinion. The website looked derelict and soulless.

Then, the singularity moment happened...

I asked it to generate a story about my friend. I asked it to generate a wacky story about him. That's it. No details, no context, no nothing. Just spitball something. I got this:

"Noam was always known for his mischievous antics. One day, he decided to embark on an adventure unlike any other. His goal? To travel across the globe and complete a series of absurdly ridiculous tasks.

First stop: Paris, France. Noam decided to start his journey by attempting to juggle croissants while riding a unicycle through the streets of Montmartre. Despite numerous close calls with angry locals and stray pigeons, he managed to pull it off – although one croissant did end up flying into the Eiffel Tower.

Next up: Tokyo, Japan. Noam decided that this would be an ideal place for his next challenge: eating sushi while riding a rollercoaster at Tokyo Disneyland. His stomach churned with excitement as he navigated the twisting tracks and downed numerous pieces of California roll along the way. Unfortunately, he did throw up on one of the ride operators, but hey – that's all part of the journey, right?

From there, Noam traveled to Rio de Janeiro, Brazil, where his next challenge awaited: surfing on a flamingo-shaped surfboard while breakdancing at Copacabana Beach. The locals watched in awe as he rode the waves like a true professional – well, except for when he lost control and fell into the ocean with a flamingo.

Finally, Noam made his way to New York City for his grand finale: attempting to play a kazoo solo on stage during Carnegie Hall's annual classical music concert. The crowd gasped in disbelief as he blew air through the plastic instrument, creating an indescribable cacophony of sound that somehow managed to blend perfectly with the orchestra."

A surprisingly silly but believable short story about my friend who loves to travel and get up to fun antics. Wow. Incredible!

But in retrospect... not really.

Remember what I said earlier? I downloaded this model to my pc and ran it in my command line. No Claude Code or other GPT wrapper. Just raw command line. And I got a story about my friend traveling to places he has actually been to. What the heck?

If you're wondering why this stuff is so accurate, or why I'm so freaked out by it, its because Meta uses their users posts as training data. Even the private stuff (ALLEGEDLY; I HAVE NO PROOF OTHER THAN A SUSPICIOUS FEELING). I only say that because all of my friend's social media posts are private. Pretty weird.

I'm going to investigate this further and report my results. I want to use this model to create something so crazy that it maybe shouldn't see the light of day.

I am pretty amazed by how accurate this thing was. I didn't ping any data center for this. I just used my PC. If you would like to discuss this with me, feel free to reach out by contacting me. I would like any and all insight I can get from this. Thanks.

[1] Ollama, D. (n.d.). Dolphin 3. Retrieved February 25, 2022, from https://ollama.com/library/dolphin3

Disclaimer: Citation may be inaccurate. IEEE format. I do not endorse or advise use of this model. Use at your own discretion.