Well, my little cousin is using large language models to do his homework for him, my aunt makes Facebook posts with images from diffusion models, and I just deepfaked my voice in 5 minutes. That sentence sounds impossible, but the AI revolution has made it all too real. Some argue the AI revolution started in November of 2022, when ChatGPT was first released. Others say it started with Dall-E in early 2021. I say it started in 1958, but not with any big, fancy machine learning models. It started with the Perceptron.
The Perceptron is the predecessor of all deep learning models. Before our modern neural networks with hundreds of hidden layers and fancy activation functions, there was a single-layer, linear network called the Perceptron. Designed by Warren McCulloch and Walter Pitts in 1943, it took 15 years before it was actually built by Frank Rosenblatt in 1958. Made to be a standalone machine, the Mark 1 Perceptron used an array of 20x20 photocells for the input, potentiometers to encode the weights, and electric motors to adjust the weights. One of the first things it learned was how to tell left from right.
Rosenblatt envisioned a future, one where machines could understand language and see the world around us. His innovative work sparked uncertainty for the future, and where there is uncertainty, there is skepticism. Minsky and Papert provided a strong counterexample to the power of the Perceptron: it could not learn the XOR gate. Essentially, given two objects—each one either a “yes” or a “no”—the Perceptron would not be able to tell you whether they were the same. Even for problems it could solve, the Perceptron required far more computing power than was possible at the time. This gave skeptics the upper hand. A few years later, scientists entered the first AI winter. Funding was down and interest was low. The field stalled.
The AI winter provided no incentive to consider the impacts of the potentially groundbreaking field. Companies didn’t anticipate how AI might impact them, and the law gave no specific consideration to such technologies. Because no one really cared to make any decisions, an AI future remained entirely uncertain. Rather than consider the implications and consequences, almost everyone deemed such decision-making unnecessary. Sure, calling Rosenblatt’s future nonsense sounded strong and resolute, but no problem has ever been answered by pretending it didn’t exist. The uncertain future of AI remained, but it was pawned off to the next generations as a problem for later.
To the dismay of many, Rosenblatt was right. We are living through the AI revolution. Everyone with internet has access to state-of-the-art models, and they continue to transform how so many of us approach work and play in our everyday lives. Conversations about AI are no longer hypothetical. Businesses are transforming, the Senate is having hearings, and grassroots movements are growing because AI hit us hard, and we didn’t anticipate the impact. The uncertainty built over 60 years has finally broken free from the chains of hypotheticals and entered reality.
And now, in our AI summer, we are woefully underprepared to deal with any of the consequences of AI. Yes, we have money and resources and attention pointing directly at this technology, but we remain almost as uncertain as we were a few decades ago. In this time, plenty have created theories about approaches to AI, but most were in the context of sentient machines. While I’m certainly thankful for that preparation for when the time comes, AI right now isn’t alive, it’s just fancy statistics. Issues of copyright and stolen data and plagiarism plague modern discussions of AI, and almost none have truly been resolved.
I sometimes wonder if this uncertainty is the price we pay for technological advancement. Almost no one batted an eyelid when LAION and CommonCrawl started building their massive datasets. Before ChatGPT, GPT-3 was already wildly powerful, writing an entire Batman movie script. GANs could do text-to-image generation well before Stable Diffusion. Few people asked if we should stop to reflect on these developments. I suppose, the longer we ignored the voices telling us to stop, the faster we could develop this technology. We paid for our speed by ignoring the uncertainty. When organizations released AI into the wild with few provisions or plans, it was already too late.
To be clear, I would never expect anyone to anticipate everything that AI would bring. For the average internet user, the limits of what is possible have changed dramatically, and the evolutions of this technology continue to be unfathomable. Besides, we got a head start on some issues—like self-driving cars—but that still isn’t enough. With AI accessible to all, everyone is left to answer the uncertainty. Schools, healthcare, legislatures, companies small and large, everyone must now grapple with how to continue their lives in this once impossible future—in Rosenblatt’s future.
The AI summer is heating up, and we’re flying too close to the sun. Try not to get burned.