The word “model” is shorthand for any singular AI system, whether it is a foundation model or an app built on top of one. Examples of AI models include OpenAI’s ChatGPT and GPT-4, Google’s Bard and LaMDA, Microsoft’s Bing, and Meta’s LLaMA.
Moore’s Law
Moore’s law is a longstanding observation in computing, first coined in 1965, that the number of transistors that can fit on a chip—a good proxy for computing power—grows exponentially, doubling approximately every two years. While some argue that Moore’s law is dead by its strictest definition, year-on-year advances in microchip technology are still resulting in a steep rise in the power of the world’s fastest computers. In turn, this means that as time goes on, AI companies tend to be able to leverage larger and larger quantities of computing power, making their most cutting edge AI models consistently more powerful. (See: Scaling laws.)
Multimodal system
A multimodal system is a kind of AI model that can receive more than one type of media as input—like text and imagery—and output more than one type of signal. Examples of multimodal systems include DeepMind’s Gato, which hasn’t been publicly released yet. According to the company, Gato can engage in dialog like a chatbot, but also play video games and send instructions to a robotic arm. OpenAI has conducted demonstrations showing that GPT-4 is multimodal, with the ability to read text in an input image, however this functionality is not currently available for the public to use. Multimodal systems will allow AI to act more directly upon the world—which could bring added risks, especially if a model is misaligned.
Neural Network
Neural networks are by far the most influential family of machine learning algorithms. Designed to mimic the way the human brain is structured, neural networks contain nodes—analogous to neurons in the brain—that perform calculations on numbers that are passed along connective pathways between them. Neural networks can be thought of as having inputs (see: training data) and outputs (predictions or classifications). During training, large quantities of data are fed into the neural network, which then, in a process that requires large quantities of computing power, repeatedly tweaks the calculations done by the nodes. Via a clever algorithm, those tweaks are done in a specific direction, so that the outputs of the model increasingly resemble patterns in the original data. When more computing power is available to train a system, it can have more nodes, allowing for the identification of more abstract patterns. More compute also means the pathways between its nodes can have more time to approach their optimal values, also known as “weights,” leading to outputs that more faithfully represent its training data.
Open sourcing
Open-sourcing is the practice of making the designs of computer programs (including AI models) freely accessible via the internet. It is becoming less common for tech companies to open-source their foundation models as those models become more powerful, economically valuable, and potentially dangerous. However, there is a growing community of independent programmers working on open-source AI models. The open-sourcing of AI tools can make it possible for the public to more directly interact with the technology. But it can also allow users to get around safety restraints imposed by companies (often to protect their reputations), which can lead to additional risks, for example bad actors abusing image-generation tools to target women with sexualized deepfakes. In 2022, DeepMind CEO Demis Hassabis told TIME he believed that the risk from AI meant the industry’s culture of publishing its findings openly may soon need to end, and in 2023, OpenAI broke from convention and declined to release information about exactly how GPT-4 was trained, citing competitive pressures and the risk of enabling bad actors. Some researchers have criticized these practices, however, arguing that they reduce public oversight and worsen the problem of AI hype.
Paperclips
The innocuous paperclip has taken on outsized meaning in some sections of the AI safety community. It is the subject of the paperclip maximizer, an influential thought experiment about the existential risk that AI may pose to humanity. Imagine an AI programmed to carry out the singular goal of maximizing the number of paperclips it produces, the thought experiment goes. All well and good, unless that AI gains the ability to augment its own abilities (see: Intelligence explosion). The AI may reason that in order to produce more paperclips, it should prevent humans from being able to switch it off, since doing so would reduce the number of paperclips it is able to produce. Safe from human interference, the AI may then decide to harness all the power and raw materials at its disposal to build paperclip factories, razing natural environments and human civilization alike. The thought experiment illustrates the surprising difficulty of aligning AI to even a seemingly simple goal, let alone a complex set of human values.
Artificial intelligence in South Africa comes with special dilemmas – plus the usual risks
Quantum computing
Quantum is an experimental field of computing that seeks to use quantum physics to supercharge the number of calculations it is possible for a computer to do per second. That added computing power could help further increase the size of the most cutting-edge AI models, with implications both for the power of those systems and their societal impact.
Redistribution
The CEOs of the world’s two leading AI labs, OpenAI and DeepMind, have each claimed they would like to see the profits arising from artificial general intelligence be redistributed, at least in part. DeepMind CEO Demis Hassabis told TIME in 2022 that he favors the idea of a universal basic income, and that the benefits of AI should “accrue to as many people as possible—to all of humanity, ideally.” OpenAI CEO Sam Altman has written of his expectation that AI automation will drive labor costs down, and he has called for redistribution of “some” of the wealth arising from AI, through higher taxes on land and capital gains. Neither CEO has said when that redistribution should begin, nor how wide-ranging it should be. OpenAI’s charter says its “primary fiduciary duty is to humanity” but doesn’t mention redistributing wealth; DeepMind’s parent company Alphabet is a public corporation with a legal responsibility to act in the financial interest of its shareholders.
Red teaming
Large language model-When people talk about recent AI advancements, most of the time they’re talking about large language models (LLMs). OpenAI’s GPT-4 and Google’s BERT are two examples of prominent LLMs. They are essentially giant AIs trained on huge quantities of human language, sourced mostly from books and the internet. These AIs learn common patterns between words in those datasets, and in doing so, become surprisingly good at reproducing human language.
The more data and computing power LLMs are trained on, the more novel tasks they tend to be able to achieve. (See: Emergent capabilities and Scaling laws.) Recently, tech companies have begun launching chatbots, like ChatGPT, Bard, and Bing, to allow users to interact with LLMs. Although they are capable of many tasks, language models can also be prone to severe problems like Biases and Hallucinations.
Lobbying
Like many other businesses, AI companies employ lobbyists to be present in the halls of power, influencing the lawmakers in charge of AI regulation to ensure that any new rules do not adversely impact their business interests. In Europe, where the text of a draft AI Act is being discussed, an industry body representing AI companies including Microsoft (OpenAI’s biggest investor) has argued that penalties for risky deployment of an AI system should not primarily apply to the AI company that built a foundation model (like GPT-4) that risks ultimately stem from, but to any down-stream company that licenses this model and applies it to a risky use-case. AI companies have plenty of soft-power influence, too. In Washington, as the White House weighs new policies to tackle the risks of AI, President Biden has reportedly tasked the foundation of Google’s former CEO Eric Schmidt with advising his administration on technology policy.
Machine learning
Machine learning is a term that describes how most modern AI systems are created. It describes techniques for building systems that “learn” from large amounts of data, as opposed to classical
Red-teaming is a method for stress-testing AI systems before they are publicly deployed. Groups of professionals (“red teams”) purposely attempt to make an AI behave in undesirable ways, to test how systems could go wrong in public. Their findings, if they are followed, can help tech companies to address problems before launch.
Regulation
There is no bespoke legislation in the U.S. that addresses the risks posed by artificial intelligence. The Biden Administration issued a “blueprint for an AI bill of rights” in 2022, which welcomes AI-driven progress in science and health but says AI should not exacerbate existing inequalities, discriminate, impact privacy, nor act against people without their knowledge. But the blueprint is not legislation, nor is it legally binding. Over in Europe, the European Union is considering a draft AI Act that would impose stricter rules on systems the riskier they are considered to be. On both sides of the Atlantic, regulation is progressing at a much slower pace than the speed of AI advancement—and no significant global jurisdiction currently has rules in place that would force AI companies to meet a specified level of safety testing before releasing their models to the public. “The question we should be asking about artificial intelligence—and every other new technology—is whether private corporations be allowed to run uncontrolled experiments on the entire population without any guardrails or safety nets,” wrote Roger McNamee, a Silicon Valley investor-turned-critic, recently in TIME. “Should it be legal for corporations to release products to the masses before demonstrating that those products are safe?”
Reinforcement learning (with human feedback)
Reinforcement learning is a method for optimizing an AI system by rewarding desirable behaviors and penalizing undesirable ones. This can be performed by human workers (before a system is deployed) or users (after it is released to the public) who rate the outputs of a neural network for qualities like helpfulness, truthfulness, or offensiveness. When humans are involved in this process, it is called reinforcement learning with human feedback (RLHF). RLHF is currently one of OpenAI’s favored methods for solving the alignment problem. However, some researchers have raised concerns that RLHF may not be enough to fully change a system’s underlying behaviors, instead only making powerful AI systems appear more polite or helpful on the surface. (See: Shoggoth.) Reinforcement learning was pioneered by DeepMind, which successfully used the technique to train game-playing AIs like AlphaGo to perform at a higher level than human masters.
Scaling laws
Simply put, the scaling laws state that a model’s performance increases in line with more training data, computing power, and the size of its neural network. That means it’s possible for an AI company to accurately predict before training a large language model exactly how much computing power and data they will likely need to get to a given level of competence at, say, a high-school-level written English test. “Our ability to make this kind of precise prediction is unusual in the history of software and unusual even in the history of modern AI research,” wrote Sam Bowman, a technical researcher at the AI lab Anthropic, in a recent preprint paper. “It is also a powerful tool for driving investment since it allows [research and development] teams to propose model-training projects costing many millions of dollars, with reasonable confidence that these projects will succeed at producing economically valuable systems.”
Shoggoth
A prominent meme in AI safety circles likens Large language models (LLMs) to “shoggoths”—incomprehensibly dreadful alien beasts originating from the universe of 20th century horror writer H.P. Lovecraft. The meme took off during the Bing/Sydney debacle of early 2023, when Microsoft’s Bing chatbot revealed a strange, volatile alter ego that abused and threatened users. In the meme, which is critical of the technique of Reinforcement learning with human feedback (RLHF), LLMs are often depicted as shoggoths wearing a small smiley-face mask. The mask is intended to represent the friendly yet sometimes flimsy personality that these models greet users with. The implication of the meme is that while RLHF results in a friendly surface-level personality, it does little to change the underlying alien nature of an LLM. “These systems, as they become more powerful, are not becoming less alien,” Connor Leahy, the CEO of AI safety company Conjecture, told TIME in February. “If anything, we’re putting a nice little mask on them with a smiley face. If you don’t push it too far, the smiley face stays on. But then you give it [an unexpected] prompt, and suddenly you see this massive underbelly of insanity, of weird thought processes and clearly non-human understanding.”
Stochastic Parrots
Coined in a 2020 research paper, the term “stochastic parrots” has become an influential criticism of large language models. The paper made the case that LLMs are simply very powerful prediction engines that only attempt to fill in—or parrot back—the next word in a sequence based on patterns in their training data, thus not representing true intelligence. The authors of the paper criticized the trend of AI companies rushing to train LLMs on larger and larger datasets scraped from the internet, in pursuit of perceived advances in coherence or linguistic capability. That approach, the paper argued, carries many risks including LLMs taking on the biases and toxicity of the internet as a whole. Marginalized communities, the authors wrote, would be the biggest victims of this race. The paper also foregrounded in its criticism the environmental cost of training AI systems. (See: Compute.)
Supervised learning
Supervised learning is a technique for training AI systems, in which a neural network learns to make predictions or classifications based on a training dataset of labeled examples. (See: Data labeling.) The labels help the AI to associate, for example, the word “cat” with an image of a cat. With enough labeled examples of cats, the system can look at a new image of a cat that is not present in its training data and correctly identify it. Supervised learning is useful for building systems like self-driving cars, which need to correctly identify hazards on the roads, and content moderation classifiers, which attempt to remove harmful content from social media. These systems often struggle when they encounter things that are not well represented in their training data; in the case of self-driving cars especially, these mishaps can be deadly. (See also: Unsupervised learning and Reinforcement learning.)
Turing Test
In 1950, the computer scientist Alan Turing set out to answer a question: “Can machines think?” To find out, he devised a test he called the imitation game: could a computer ever convince a human that they were talking to another human, rather than to a machine? The Turing test, as it became known, was a slapdash way of assessing machine intelligence. If a computer could pass the test, it could be said to “think”—if not in the same way as a human, then at least in a way that would help humanity to do all kinds of helpful things. In recent years, as chatbots have become more powerful, they have become capable of passing the Turing test. But, their designers and plenty of AI ethicists warn, this does not mean that they “think” in any way comparable to a human. Turing, writing before the invention of the personal computer, was indeed not seeking to answer the philosophical question of what human thinking is, or whether our inner lives can be replicated by a machine; instead he was making an argument that, at the time, was radical: digital computers are possible, and there are few reasons to believe that, given the right design and enough power, they won’t one day be able to carry out all kinds of tasks that were once the sole preserve of humanity.
Unsupervised learning
Unsupervised learning is one of the three main ways that a neural network can be trained, along with supervised learning and reinforcement learning. Unlike supervised learning, in which an AI model learns from carefully labeled data, in unsupervised learning a trove of unlabeled data is fed into the neural network, which begins looking for patterns in that data without the help of labels. This is the method predominantly used to train large language models like GPT-3 and GPT-4, which rely on huge datasets of unlabeled text. One of the benefits of unsupervised learning is that it allows far larger quantities of data to be ingested, evading the bottlenecks on time and resources that marshaling teams of human labelers can impose on a machine learning project. However it also has drawbacks, like the increased likelihood of biases and harmful content being present in training data due to reduced human supervision. To minimize these problems, unsupervised learning is often used in conjunction with both supervised learning (for example, by building AI tools to detect and remove harmful content from a model’s outputs) and reinforcement learning, by which foundation models that were first trained unsupervised can be fine-tuned with human feedback.
