What will be the Law of the Land for AI?
2022 will be remembered as the year of generative AI, starting with the announcement of DALL-E 2 in April. In November, OpenAI upped the ante with limited preview of ChatGPT. Without much doubt, OpenAI is currently the pioneer of generative AI. Even though there were other products, almost all announced “after” OpenAI.
Regulating AI
OpenAI was also one of the first from the major stakeholders to raise concerns about the progress of AI. Its founder, Sam Altman, is the most vocal on their concerns. He still continues doing so. Some think it is a shrewd move that it is a tactic to preempt competition with regulation while they still have an edge over upstart firms. Though, dubbing Wilde, there are two tragedies in this life and one of them is getting what you want.
Latest Regulations
EU AI Act, is one of the first (if not the first) legislation attempts in order to regulate AI progress. EU Parliament passed the bill on June 14. In its essence, there are no alarming or severely restrictive rules for now. It includes “no-brainer” restrictions on discrimination, risk-level categorization, self-assessment framework for high-risk and several other topics and sectors (e.g., healthcare). An interesting restriction is on social scoring.
In other news, Japan declared copyright rules do not apply to copyrighted material when they are used as training data for generative AI applications. Though, copyright claims apply to the outputs of generative AI models should they infringe it.
Finally, there are countries such as China that have expectedly more restrictive regulations for generative AI companies.
A Brief History of Generative AI
Although the term “generative AI” may apply too much earlier cases (including pranks such as Sokal), the breakthrough can be originated from the introduction of Generative Adversarial Networks by Ian Goodfellow et al. in 2014. VQGAN-CLIP is a popular application.
Although I’m not well versed in the history of generative AI, Transformer technology from Google Brain is the second breakthrough in 2017. GPT is also the shorthand for “Generative Pre-Trained Transformer”.
Diffusion models are really powerful when it comes to generating images. Overly simplistic, models are trained by adding noise to the images and trying to repair the noise. Many successful models are based on diffusion (hence the name Stable Diffusion).
It is not a comprehensive history but I believe I mentioned some of the most important recent milestones in Generative AI. As with any breakthrough technology, this is just the tip of the iceberg.
What now?
Probably we will enjoy more progress in the following few years before everything stabilizes.
Personally, I do not expect excessive or crippling regulation on the development of generative AI applications. However, I expect a designation of responsibility in the cases in which generative AI blurts out something unexpected, soon.
Meanwhile, you may jump on Hugging Face to enjoy some exquisite specimens of the latest generative models in the wild.
