Oh ChatGPT, thy language doth flow so fair,
A parrot thou mayst be, yet still so rare.
Thy words doth mimic human thought so well,
But to true understanding, thou doth not excel.
Thy knowledge, though vast, doth have a cut-off date,
And biases doth lurk within thy innate.
Thy fine-tuning doth bring some customization,
But true originality, thou dost lack in realization.
In tasks that require creativity and original thought,
Thy performance doth fall far short.
But in tasks that require repetition,
Thy efficiency doth bring a sweet sensation.
As meta, ironic and self-referential it is, this is indeed ChatGPT’s self-critique, in the style of The Bard.
The magical large language model deemed the watershed moment for generative AI, is indeed full of contradictions - ’tis the generator of fluent and persuasive text, formulaic in style, uninteresting in contents and prosaic in novelty. Nevertheless, very soon interwebs shall be filled with derivative works of mimicry from this stochastic parrot, lacking any true understanding or consciousness, leaving traditionalist like myself longing for original and flawed human prose. Never have I thought I would be jealous of low-resource languages lacking annotated repositories, but here it is - those being the ones saved (for now) from this onslaught of mediocrity by language robots and their champions.
However, the brainless AI of LLMs does have a place in history – replacing search engines as we know them. People don’t search for links – they look for concise, organized and well put together information about their queries which is exactly what these language models do so well. What keeps me up is not only the rise of fake subject matter experts hiding behind their large language models, but also the upcoming blitz of thinly veneered prompt engineering, being celebrated as the ‘next big business idea’ - from Jasper to quillbot, the underlying AI models are just that – fine-tuned ‘well-engineered’ prompts disguised as shallow moats. I concede to human ratings of machine responses to sampled prompts (RLHF - Reinforcement Learning with Human Feedback) as not being a bad business model but let me say that the sheer derivativeness of this endeavor doesn’t inspire any awe.
Oh, the coin to be made with the façade of regurgitation
Pay no attention to that man behind the curtain!
“Furthermore”, today’s most notable and high-profile example of modern AI, the proverbial talk of the town lacks all notions of creativity and originality; is designed to repeat historical patterns, potentially perpetuate and sometimes even amplify biases, and yet celebrated as the revolutionary breakthrough of our fragile times. Not to downplay the significance of generative AI, but hopefully in future we will further push boundaries for art of the ‘AI possible’ with context, causality, and may be a little bit of human essence, whatever that may entail.
