The Mechanical Jerk

• 

  One exchange on healthcare

Inspiration

Seeing that social media led to echo chambers in which they reinforced each others' beliefs and resulted in the increasing political polarisation in the West.

What it does

I fine-tuned the 355M parameter model of GPT-2 on the IAC dataset of forum debates/arguments regarding a set of given topics such as gun control, evolution, etc. I preprocessed the data into pairs of parent and child comments marked by delimiting special characters. Hence, the final model, when given a prompt on these topics, could spit out a convincingly realistic reply that usually disagrees with the OP.

It searches for tweets on controversial topics and runs inference against the text. It generates 10 samples and the most appropriate one is selected by algorithm based on factors such as length, complexity, and sentiment. It then tweets the reply at the person and goes to sleep for a random time around 7 mins before repeating ad infinitum.

How I built it

Used Google Colab with an attached K80 for fine-tuning. Pretrained model provided by OpenAI. All inference done on in a nvidia container on an RTX 2060 Super. All coding done in Python. Lots of caffeine and no sleep.

Challenges I ran into

Shenanigans with old tensorflow versions/CUDA/cuDNN compatibility during prototyping.

Accomplishments that I'm proud of

It works very well and produced fairly realistic responses. Applying state-of-the-art NLP models to a real life problem.

What I learned

It does not accomplish the initial goal. it just makes people angrier. The model is not capable of logical reasoning, only generating something that sounds like realistic response.

What's next for The Mechanical Jerk

Repurposing it by finetuning on a different dataset for a different purpose. Something more wholesome perhaps.

Built With

• colab
• docker
• python
• tensorflow
• twitter