The Overlooked Advances of AI’s Miracle Year
As revolutionary as the past year, 2022, was for Artificial Intelligence, the application of these breakthroughs is only beginning. While the focus at the end of the year was on OpenAI’s ChatGPT, other advances, such Meta’s Cicero, which uses world models to tie an AI’s reasoning to a “real” scenario, and DeepMind’s AlphaTensor, an AI to speed up AI, will be as important over the long term. 2022 was the equivalent of Albert Einstein’s “Annus Mirabilis,” 1904, in which he published papers on relativity, the photoelectric effect and thermodynamics. Einstein won the Nobel prize for the latter two papers, but the applications of E=MC^2, such as nuclear power and the nuclear bomb, turned to be more world-altering. Likewise, world models and AI-improving AI will lead to even greater changes.
Fittingly, the year 2022 began with an overlooked but ultimately very important paper which advanced the field of Large Language Models, AIs for generating text of which ChatGPT is now the most famous. In January, OpenAI introduced a new method of training language models, known as Reinforcement Learning from Human Feedback (RLHF), which improved the performance of its LLM chatbot by incorporating feedback from human raters, such as facts being untrue or responses being offensive.
Meanwhile, Google was working on its own chatbot, LaMDA. However, its natural language capabilities were so advanced that one engineer, Blake LeMoine, publicly declared in June that it was sentient. In response, Google released LaMDA to a limited number of testers, with three specific uses: planning an imaginary trip, providing instructions, and discussing dogs. LaMDA is designed to respond to prompts as humans do, so it can convincingly appear to have both thoughts as feelings, even if those responses are generated by calculations on a computer.
In November, Meta released Galactica, an LLM trained on academic papers, which showed the dangers of an LLM without safeguards. Meta took its website down after one day, as users asked it to write about controversial topics such as the benefits of antisemitism and eating crushed glass.
OpenAI released ChatGPT in December, leveraging its advances in RLHF, as well as a simple user interface, to get a million users by the end of the year. Even with ChatGPT’s RLHF-powered guardrails, it, like all LLMs, suffers from “Hallucinations”: because LLMs have no sense of the real world, they often provide plausible but wrong information. For instance, LaMDA described to me in great detail Mark Twain working for Levi Strauss, with funny anecdotes. Although they were both in San Francisco at the same time, there’s no record they ever met.
AIs that Plan, Lie, Create, and Accelerate
Meta’s other major breakthrough of November 2022, Cicero, which plays the game Diplomacy, will be the most important natural language breakthrough in the long run, and not simply because the AI fooled people into thinking it was a human play. Following LeMoine’s claims that LaMDA was sentient, Yann LeCun at Meta argued that for an AI to truly be autonomous, or Artificial General Intelligence, as it’s usually called, it must not only reason about its world, but also be able to explain and criticize its reasoning. Cicero closely follows the structure of AGI that LeCun laid out. It has a world model that reasons about the board to decide its best options and a chatbot to discuss options with other players. Diplomacy is a game of gaining and breaking trust, and Cicero is an expert in this area, re-establishing trust after its lies, and then betraying its allies yet again. Lest you think that Diplomacy is just a game, the Defense Advanced Research Projects Agency (DARPA) issued an almost $1 million grant to build a Diplomacy-playing AI, to help understand how scammers fool people. But no one knows how DARPA will use these insights.
Generative AI tools for art, such as Midjourney, OpenAI’s DALL-E 2, and the open source Stable Diffusion, are also strengthened by a “world model” of the image they produce. The description of the image, unlike a pure LLM, is based in the actual pixels. Invoke AI walks through an example of using “negative prompts” to remove features from an image. In the example, Stable Diffusion makes an image less blue and removes a rider from a horse, because it knows which pixels are not only blue but also which one form the horse and rider. Stable Diffusion is also an example of how AI advances tend to start in the hands of a few, and then become available to all through open source. Even if OpenAI refuses to produce sexual or misleading images, it is impossible to prevent others from using Stable Diffusion, which anyone with a computer can download, from doing so.
The last sign that AI advances are accelerating, comes from AIs that can improve other AIs. Futurist Ray Kurzweil called this phenomenon the “Singularity”: the exponential growth in AI will look small at first but grow beyond exponentially, precisely because AIs will find ways to improve faster than humans can. DeepMind’s AlphaTensor showed in October that it can improve matrix multiplication, the main process of neural networks. Once that is implemented, it will be even cheaper and faster to train large models. It also showed the power of Reinforcement Learning, the ability of AIs try billions of different strategies, learning from them until it finds the optimal one. This technique can be applied to world models that reflect reality, so an AI can steer a robot that has not built yet, and even try different robot designs so the optimal one is built. Other AI advances, such as Amazon’s CodeWhisper, lower the barrier to entry for building AI, so that simply asking for a Neural Network to address a computer vision results in working code.
As more and more people use AI-powered applications, some of them with risks still not well-understood, there is likely to be both great benefits and also great risk. As of this writing, Microsoft just released a premium version of Teams that includes ChatGPT-powered responses. The ability to respond to lots of messages quickly will no doubt increase its users’ speed, just as autocorrect and word suggestions do, but hallucinations will result in users choosing plausible but wrong responses, like autocorrect gone wrong. Both Google and Microsoft are beta testing search engines based on their LLMs, but Google’s ad for its new BART search engine included incorrect information. It is not encouraging when the AI has a major error even before it has public users. Meta has suggested incorporating insights from Cicero into games, and the insights will apply to other products as well. An AI based on Cicero that pursues its own agenda while professing cooperation with humans may be very dangerous. As AIs can think farther and farther ahead, thanks to advances such as AlphaTensor, it will be harder and harder for humans to stay in control.
A consequence of exponential growth is that there will be as much change in its doubling period- computational power doubles every 2 years, and coronavirus initially doubled every 2 weeks. As AI continues to grow beyond exponentially, the amount time for any given changes shrinks. As much change as we saw in 2022, we will see even more in 2023, more than that in 2024. Being prepared for the Next Big Thing requires us to understand not only the current Big Thing, but also the little things that will become Big before we know it.
