AI, Creativity and Exploring Foundation Models
The other day, I was writing lyrics for a song, for which I researched rhymes and similar sounding words with specific meanings. This is a fairly complex task, which takes practice to master, and for a non native English speaker — as myself — it is one bit more challenging.
I had specific thoughts in mind and a specific style I wanted to express them in, so I did not want ChatGPT to generate all the lyrics for me. Instead, I found it quite helpful as a tool for this semantic-sound search. It is knowingly bad at simple arithmetics, which limitation I also faced when I wanted it to search for words with a specific number of syllables.
Sharing our ChatGPT conversations has lost its novelty by now. However, creative ways of incorporating Foundation Models (such as the one under ChatGPT, BingChat, Dall-e etc.) into applications which add extra value just started to rise. Many useful ChatGPT prompts and exciting applications have surfaced already, ranging from summarising, organising content to assisting content generation and knowledge acquisition. Many of them are probably just wrappers over ChatGPT, but some have already added value, aiming to enhance our creative workflows. I’m sure more is to come!
In 2021 a big group of AI researchers at Stanford Human-Centered Artificial Intelligence and the Center for Research on Foundation Models, came out with an article about the stirring paradigm shift in AI.
They forged the term Foundation Models to describe huge machine learning models which synthesise large amounts of data (text, visual, audio, 3D signals, etc.) to a generalist model, which other, more specialised models can be built on.
The dominant theme of the symposium was foreshadowed by Professor Michael Wooldridge’s introduction (paraphrasing): The event was planned to take place in October then got postponed to December, then got postponed again due to rail strikes, then ChatGPT happened.
As almost all speakers pointed out, the paradigm shift started to kick off with the famous Attention Is All You Need paper, which all current models are built on. However, throughout the day, I increasingly felt that the current burning question is rather:
A few years ago, AI research in industry started to take over academia in terms of speed of development and model performance. This is mostly due to the data and compute hungry nature of our current AI methods. These involve astronomical costs, only (big tech) companies can pay.
The gap between academia and tech has been growing, and the appearance of ChatGPT seems to have deepened the chasm.
Since this was mostly an academic event, many touched either explicitly or implicitly on the difficulties of even benchmarking foundation models. Academia’s contribution in the development of safe AI is crucial. However, most models and data are closed source, and even if they were open, running them requires an amount of money academia just doesn’t have.
Nevertheless, some attempts have been presented to benchmark foundation models, such as Holistic Evaluation of Language Models by Stanford. One of their results which was striking to me was that Fairness and Accuracy correlates across various foundation models.
Throughout the conference, several initial tests of ChatGPT have been shown.
In general, the panel expressed excitement about the possibilities of this new era of AI for education, science, art and research. However, many highlighted the lack of common sense understanding of current models. This can potentially be mitigated by using multimodal models to approximate human perceptual experience.
The panel also expressed their critiques and concerns about the risks of these models suddenly becoming widespread among the public:
1. Current models are surprisingly fluent but underwhelming in reasoning.
2. They can tell the uttermost nonsense very convincingly and with perfect fluency. One acute comment came from the Head of Data Science from no. 10 Downing Street: “I know people like that”.
3. There is a danger of data poisoning: the internet may become full of generated nonsense which then becomes training data of the same models, thus generating even more junk.
In a recent interview, Steven Pinker nicely summarised the false dichotomy often made between artificial and human intelligence. He argues, they aren’t in competition, rather, AI is yet another tool we can use to enhance our cognitive capabilities. This view was shared by the panel, illustrated by the example of spreadsheets, which enhanced accountants rather than replaced them. Furthermore, AI can also help us gain more insight into our own minds and perhaps can be used to explain its own behaviour.
As a rather ironic turn, straight after the panel discussion I ran to the first session of a creative writing course I recently signed up for. After the brief introductions, we started writing on the spot for prompts, given by our writing coach. I had a rather mystical / meditative experience of observing my inner model, generating thoughts which sometimes surprised even myself!
I highly recommend trying it out for everyone. To help you get started, I’ll share with you some prompts we got:
1. Imagine you start walking from your house tomorrow morning, when you see a pair of shoes. Imagine / visualise them in front of you and describe them in detail in a few sentences. (Was there anything that surprised you about the shoes, their position etc.?)
2. Now, imagine whom these shoes belong to and describe them in a few sentences.
Enjoy the ride of getting to know your own generative mind!
As for ChatGPT, I recommend finding new ways of enhancing yourself with it. Instead of trying to outsource creativity, use it to be creative yourself on a whole other level!
*During writing this blog post I used ChatGPT to search for one synonym and an expression.
For more posts by Anita Verö, read about her take on how diverse an AI ethics conference really is.