The pace of developments in AI has accelerated significantly over the last few months, as years of behind-the-scenes research converge in recent months with the release of jaw-dropping generative AI models.
And there is no question that the likes of Stable Diffusion, DALL-E 2, and ChatGPT have captured the attention of the public, as they win art competitions, write essays, and potentially change the face of work.
But what is the price of AI and its impact on our societies?
The cost of AI
We know that training large AI models is prohibitively expensive. But what is the cost of running ChatGPT for its millions of users? I don’t have exact figures, but how does “eye-watering” as an adjective sound?
Sam Altman, CEO of OpenAI said as much in a tweet last month: “We will have to monetize [ChatGPT] somehow at some point; the compute costs are eye-watering.”
And if you are counting on using the likes of ChatGPT for free forever, think again. Although OpenAI launched as a nonprofit in 2015, as of 2019 transited to a “capped profit” research lab where backers will make no more than 100 times what they put in.
This is quite a bit of money, and considering Microsoft invested USD1 billion, means that OpenAI needs to generate a profit. Indeed, OpenAI's recent pitch to investors, per a Reuters report, was that it expects USD200 million in revenue this year and USD1 billion by 2024. So, expect it to start charging soon.
Of course, there is also the hidden cost of generative AI models: that of the internet forever being contaminated with AI-generated content.
Consider how large language models are trained on data sets created by scraping the Internet for information, including the erroneous, fake, or outright malicious, which is used to train AI models. The trained AI models then regurgitate more of these, which is eventually spread to every corner of the Internet where they get scraped, again and again, ad nauseam.
“It’s not going to be good enough to just blindly hoover text up from the internet anymore, if we want to keep future AI models from having biases and falsehoods embedded to the nth degree,” wrote Melissa Heikkilä in the Technology Review.
You are the product
Ever wondered at the generosity of OpenAI offering ChatGPT at no charge? You probably know about the substantial amount of manual work required to fine-tune AI models.
Now imagine the tuning inadvertently performed by over a million users, all jumping in with real-world queries and feedback as they explored this newfangled service that everyone is talking about.
In other words, dear reader, you are now the product.
Of course, you can theoretically roll out your own ChatGPT. As reported on TechCrunch, developer Philip Wang had in the last week of December released a text-generating model that behaves similarly to ChatGPT. Dubbed PaLM + RLHF, it combines Google’s Pathways Language Model (PaLM), a massive 540-billion parameter model with a reinforcement learning technique.
The downside? You will need to pony up the hardware (or money) to train it yourself. Even if someone were to gift you with a dozen Nvidia A100 GPUs (MSRP of USD14,999 each), you would need to set up the infrastructure and corresponding software framework for AI training.
Real-world impact
In the meantime, there is the real-world impact of generative AI models to consider. At least one teacher in the U.S. has warned that ChatGPT will “end high-school English”, while another educator in Singapore is more sanguine that it will disrupt teaching and learning in a good way.
I think both have a point, though simply changing assessment methods or assignments to a format that cannot be easily regurgitated by an AI model is harder than it sounds. And I say this as a former adjunct lecturer for nine years, where I took as many classes as many full-time lecturers.
There is an even more ominous impact as early AI entrants make their mark. As OpenAI benefits from a vast torrent of feedback that is prohibitively expensive to acquire any other way, it stands to gain an insurmountable lead over other research organizations.
“We are now facing the prospect of a significant advance in AI using methods that are not described in the scientific literature and with datasets restricted to a company that appears to be open only in name,” writes Toby Walsh, a professor of AI at UNSW Sydney in The Conversation.
Though all major AI tools today are open-sourced, Walsh thinks the race to develop more capable AI may see organizations holding back, slowing down progress in the field.
Walsh argues that we may also see new monopolies develop, and the era of openness that has characterized and advanced AI research to date might be about to end, effectively ending the golden age of AI.
Paul Mah is the editor of DSAITrends. A former system administrator, programmer, and IT lecturer, he enjoys writing both code and prose. You can reach him at [email protected].​
Image credit: iStockphoto/Doucefleur