3. Space: Chatting with an Alien

What if you could communicate with an alien? That was the basic premise of the Chatting with an Alien project. What kind of questions would you ask it (could you say him or her?), how would it respond? To explore this question, I created an Extraterrestrial Communication Device, where someone visiting the installation could send questions, which would be beamed through space, and then a response would be heard using sound.

Figure 1: Screenshot of the interface.

The interface itself was created using OpenFrameworks, which would then send the actual questions to a locally running webserver, written in Python. This was running the neural network for the conversational AI, which would be fed the question. After this neural network returned the response, it was converted using text-to-speech into sound and then played through the speakers. At the same time, the audio was piped using Soundflower back to OpenFrameworks, where the audio output was visualised within the interface.

The conversational AI itself was based on GPT-2 [1], the state-of-the-art language model developed by OpenAI and released just a few months ago. This was then fine-tuned for conversations by a group of developers and released as a model publicly [2]. Then I configured the hyperparameters of this model to better work for this use case, and gave it the personality of an alien from Mars. This was done by pre-feeding it statements, which the model thought it had said itself to someone else, such as these:

I am an alien.

I am called Krogitar.

I live on Mars.

I was born on Mars.

I am 142 years old.

Interestingly, because it relied on a very extensively trained language model (with 117 million parameters), this network has a significant understanding of the English language. For example, when at one point someone asked whether the alien was from Mars, it replied: "Yes, I am a Martian.". This was never fed directly to the model; rather, it seems to have linked these two concepts (Mars <> Martian) through the reading of thousands (40GB, in fact) of text from webpages.

[1] Radford, Alec, et al. "Language models are unsupervised multitask learners." (2019).

[2] https://github.com/huggingface/transfer-learning-conv-ai