Our Age of AI: The Era of Direct Interface
What are the implications of being face-to-screen with AI?
It is apparent that we are cresting the age of generalist models: one-stop-shops that require you to never stray from their web interface. They offer a panacea to every stumbling block one might encounter throughout the day, on any project. Looking for an optimal way to sort through a python array? Or how about a chicken soup recipe that tastes like grandma’s. Or, if your grandma happened to lull you to sleep with tales of her napalm-factory-working days, you may be so lucky as to get a recipe for a war crime out of these large language models. The problems with LLMs are not unseen in other manifestations of machine learning. Just like in the Wizard of Oz though, when we look behind the curtain, we don’t find a magical, omnipotent being. Instead, we find an imperfect algorithm, full of the same biases and tendencies that have surfaced in the LLMs predecessors.
However, what is truly striking is that we are in the era of direct interface. Artificial intelligence has been in our lives well before the days of ChatGPT, from autocomplete to Alexa interactions to Alan Turing’s test. Now, for the first time in human history, neural networks are not some coveted spells uttered only by those who engage in the dark arts of machine inference. We have been interacting with AI in the form of machine learning for a while now, but this interaction was obfuscated by an intermediate application of the algorithm, like a playlist or a “Recommended TV Shows” list. The veil of apps has been lifted and the user can now directly engage with an algorithm.
October 1950: Alan Turing “propose[s] to consider the question, ‘Can machines think?’”
Source: “Computing Machinery and Intelligence” (Turing)
Now, the power of AI is in the hands of the masses, and transparently too. When someone opens up the web interface for ChatGPT or DALL-E2, they know that their input is directly into an ML algorithm, whose output will appear on the screen in mere moments. Even though ML has been studied for almost a century now, and the idea of artificial intelligence goes back even further, I believe we are in a new era of transparent integration of AI into everyday society. What does this mean for society at large?
January 2023: ChatGPT grew its user base 100 million in its first two months (fastest of any consumer application)
Source: Reuters, ChatGPT sets record for fastest-growing user base - analyst note
Well, hopefully it means that AI and ML can be demystified, a veil of mist dissipating with the illumination of common knowledge. Because, at its core, machine learning is not magic (although it can seem that way), it’s math. And I think that something we understand intimately can produce surprising, automated, and arguably novel results is much more exciting than an unexplainable black box that people don’t even bother trying to crack open. With ML being more accessible to the public, I hope that more people, regardless of their age, education, and coding experience, get involved in the development and training of these algorithms. And that those whom these algorithms disproportionately affect are given resources and a voice to be a part of that feedback loop. Integrating AI into everyday society inherently means the democratization of perhaps one of the most powerful tools developed in human history.
Implicit in the democratization of AI are the ethical implications that come along with releasing this tool for public consumption. We’re seeing the implications (good, bad, and ugly) of the Era of Direct Interface, and these observations are starting to prompt a “your scientists were so preoccupied with whether or not ‘could they’ that they didn’t stop to think if they should” moment. While there have been calls from industry leaders to halt large algorithmic development, I’m not sure that a pause is a sustainable or efficient way forward because, like it or not, humans will develop. We can develop practical, ethical frameworks, alongside algorithms, to examine the impact of these large models concurrently.
Included in this plea is the political involvement necessary for the gears of bureaucracy to turn a little faster to catch up to the pace of human innovation. This request is crucial to the fairly regulated development of AI. Our science and technology policies are outdated. US telecommunication law was overhauled for the first time since the 1930s with the Telecommunications Act of 1996. Responsible innovation requires equally responsible — and timely — governance with input from domain experts.
February 1996: President Clinton signs the Telecommunications Act into law, amending the Communications Act of 1934
Source: Telecommunications Act of 1996
While it’s worth re-examining these policies in a more current context, we should also be creating new ones to keep up with the cutting-edge of technology. The democratization and spread of AI should demonstrate to politicians that this type of technology is worth learning about, because it’s not going away anytime soon. And if we want to leverage it for the betterment of our society, well isn’t that the whole, idealistic point of government? A body that is larger than the individual to serve the greater good (don’t quote me I wasn’t the most attentive student when we were reading Locke and Hobbes)?
With the politicians are regulating, ensuring that this technology is being used in responsible manners, the ethicists should be right alongside the AI specialists, informing our politicians and the public of the entire, unruly nature of these algorithms. As ChatGPT and other algorithms extend their reach, it is becoming more transparent to the public that these algorithms are — in fact — not an end-all-be-all-fix-all gift from the technology gods. Just as algorithms need to be used responsibly, they need to be designed responsibly. And with more people seeing first-hand the consequences of algorithmic mishandling, I hope that there is a bigger push to ensure the ethical design and implementation of these algorithms.
These algorithms, like ChatGPT, are, by nature, general. They are called large language models for a reason; they are meant to encode a large amount of information and be general enough with its inference to be applicable to a wide variety of umbrella topics and an even wider variety of fairly niche ones. However, we’ve seen that there is a limit to this generality, and it comes in the limit of specificity. When you get too granular with ChatGPT, asking it for a very, very specific piece of information, it short-circuits (more metaphorically than literally — see: hallucinations). So clearly, these types of algorithms are meant for very general applications. Again, for the technology ChatGPT employs and the sheer impact it has had on society in a fairly short time frame, the effects of this algorithm can hardly be understated. But I have to wonder, if we’re in the era of the general model, what does the future look like? And what does it mean if, in the future, bigger is not always better? What if models got more specific instead of more general? What kinds of implications would smaller, specific models have in terms of algorithm performance, interference, transparency, ethics, public appeal and acceptance, political regulation, etc., etc.
What does a more specific, bespoke algorithm mean for our future? Exploring the concept of and developing the case for domain-specific machine learning is what Bespoke AI is all about.