laitimes

Do you know what are the most common AI terms? This article will tell you

author:TechMind
Do you know what are the most common AI terms? This article will tell you

As AI takes on a more important role in our world, it's easy to get lost in its sea of terms. But now more than ever, it's important to clear your mind.

AI is expected to have a significant impact on the job market in the coming years. How to manage AI has become a more important part of our political conversation. And some of the most critical concepts you don't learn in school.

Struggling to keep up with AI can be difficult. AI research is so complex that many of the terms are new even to the researchers themselves. But there's no reason why the public can't talk about the big issues involved, just as we've learned to tackle climate change and the internet. To help everyone participate more fully in the AI discussion, this article has put together a handy glossary of the most common AI terms.

Whether you're a complete beginner or already know terms like AGI and GPT, this A-to-Z glossary is designed to be a public resource for all working on the powers, promises, and risks of AI.

ACTED

AGI stands for Artificial General Intelligence — a hypothetical future technology capable of performing most economically productive tasks more efficiently than humans. Proponents argue that such technology could also uncover new scientific discoveries. Researchers disagree on whether AGI is possible, and if so, how far it goes. But both OpenAI and DeepMind — the world's two leading AI research institutions — are explicitly committed to building AGI. Some critics argue that AGI is just a marketing term.

Alignment issues

The "alignment problem" is one of the most far-reaching, long-term security challenges in AI. Current AI does not yet have the capability to surpass its designers. But many researchers expect it might one day have that capability. In that world, the current way AI is trained could lead to harm to humans in pursuit of arbitrary goals, or as a clear strategy to pursue their power at the expense of ours. To reduce the risk, some researchers are working to "align" AI with human values. But the problem is difficult, unsolved, and not even fully understood. Many critics say efforts to address the problem have been snubbed as commercial incentives tempt leading AI labs to devote their focus and computing power to making their AI more capable.

automation

Automation is the historical process by which human labor is replaced or assisted by machines. New technologies, or more accurately, the people responsible for implementing them, have replaced many human workers, from auto assembly line workers to grocery store clerks, with machines that don't need to be paid. According to the latest paper from OpenAI and research by Goldman Sachs, the latest generation of AI breakthroughs could lead to more white-collar workers losing their jobs. Researchers at OpenAI predict that nearly one in five U.S. workers could automate more than half of their daily routine tasks with a large language model. Researchers at Goldman Sachs predict that 300 million jobs could be replaced by automation worldwide within the next decade. Whether the productivity gains brought about by this upheaval lead to broad-based economic growth, or only further exacerbate wealth inequality, will depend on how AI is taxed and regulated.

Do you know what are the most common AI terms? This article will tell you

bias

Machine learning systems are called "biased" when their decisions are consistently biased or discriminatory. For example, enhanced AI sentencing software found that even if the crimes were the same, it recommended longer prison sentences for black offenders and shorter sentences for white offenders. Some facial recognition software recognizes white faces better than black faces. These failures are often because the data used to train these systems reflects social inequalities. Modern AI is essentially pattern replicators: they ingest large amounts of data through neural networks and learn to recognize patterns in that data. If there are more white faces than black faces in a facial recognition dataset, or if past sentencing data suggests that black offenders are sentenced to longer prison terms than white offenders, machine learning systems may learn from erroneous experiences and automate these unjust sentences.

Chatbots

Chatbots are consumer-friendly interfaces built by AI companies that allow users to interact with large language models (LLMs). Chatbots enable users to simulate a conversation with LLM, which is often an effective way to get answers to questions. At the end of 2022, OpenAI launched ChatGPT to push chatbots into the mainstream, leading Google and Microsoft to try to integrate chatbots into their web search services. Some researchers believe that AI companies are perceived as irresponsibly rushing to launch chatbots for several reasons. Because chatbots simulate conversation, users can mistakenly think they are talking to a sentient being, which can lead to emotional distress. Chatbots, on the other hand, can either "hallucinate" with misinformation and mimic biases in their training data. "ChatGPT may provide inaccurate information about people, places, or facts," the warning statement below its text input box.

Competitive pressures

Several of the world's largest tech companies, as well as a slew of startups, are scrambling to launch more powerful AI tools in order to reap rewards such as venture capital, media attention, and user registration. AI security researchers worry that this will create competitive pressures, or that it will motivate companies to devote as many resources as possible to improving their AI capabilities, ignoring the still-immature field of alignment research. Some companies cite competitive pressures as an argument for further investment in training more robust systems, arguing that their AI will be safer than their competitors. Competitive pressures have led to disastrous AI launches, such as Microsoft's Bing (powered by OpenAI's GPT-4), showing hostility toward users. It's also bad for the future, when AI systems are likely to be powerful enough to seek power.

Computing power

Computing power, often referred to simply as "compute," is one of the three most important elements for training a machine learning system. Computation is actually an energy source for neural networks to learn patterns in the training data. In general, the higher the computational power used to train a large language model, the better it performs on many different types of tests. Modern AI models require a lot of computing power and electrical energy to train. While AI companies don't typically publicly disclose the carbon emissions of their models, independent researchers estimate that OpenAI's GPT-3 training process resulted in more than 500 tons of carbon dioxide being emitted into the atmosphere, equivalent to the emissions of about 35 U.S. citizens in a year. As AI models get bigger, these numbers will only rise. The most common computing chips used to train cutting-edge AI are graphics processing units.

Do you know what are the most common AI terms? This article will tell you

data

Data is essentially the raw material needed to create artificial intelligence. It is one of the three key elements for training machine learning systems, alongside computing power and neural networks. Huge data sets are collected and fed into neural networks powered by supercomputers, which learn to discover patterns. The more data the system is trained on, the more reliable the predictions generally become. But even if the data is rich, there must be diversity, otherwise AI may draw the wrong conclusions. The world's most powerful AI models are often trained on vast amounts of data taken from the internet. These huge datasets often contain copyrighted material, which has led companies like Stability AI, maker of Stable Diffusion, to face lawsuits for allegedly illegally relying on someone else's intellectual property. And, because there is a lot of bad content on the internet, large data sets often contain harmful material such as violence, pornography, and racism, which, if not removed, can cause AI to behave unexpectedly.

Data labeling

Often, before training a machine learning system, human annotation or description of the data is required. For example, in the case of self-driving cars, a human worker needs to annotate a video taken from a dashcam and draw shapes around vehicles, pedestrians, bicycles, etc. to teach the system how to recognize parts of the road. The work is often outsourced to temporary hiring contractors in the Global South, many of whom earn wages almost just above the poverty line. Sometimes, the work can be traumatic, such as Kenyan workers being asked to review and label texts depicting violence, sexual content, and hate speech to train ChatGPT to avoid such material.

diffusion

The latest generation of image generation tools such as Dall-E and Stable Diffusion are based on diffusion algorithms: a specific AI design that fueled the recent boom in AI-generative art. These tools are trained on datasets that label images at scale. They essentially learn the patterns between pixels in an image and the relationship between those patterns and the words used to describe the image. The result is that given a set of words (e.g., "a bear riding a unicycle", the diffusion model can create such an image from scratch. It does this through a step-by-step process, starting with a canvas filled with random noise and gradually changing the pixels in the image to bring it closer to what "a bear on a wheelbarrow" should look like in the training data.

Diffusion algorithms are now very advanced and can produce realistic images quickly and easily. While tools like Dall-E and Midjourney include protections against malicious tips, there are also open-source proliferation tools that don't have any protections. The availability of these tools has researchers concerned about the impact of proliferation algorithms on disinformation and targeted harassment.

Emerging capabilities

When AI such as a large language model exhibits unexpected abilities or behaviors that its creators did not program, these behaviors are called "emerging capabilities." New capabilities tend to emerge when AI is trained with more computing power and data. A good example is the difference between GPT-3 and GPT-4. These AIs are based on very similar underlying algorithms, with the main difference being that GPT-4 is trained on more computing power and data. Studies have shown that GPT-4 is a more capable model capable of writing functional computer code, performing above average on multiple academic exams, and correctly answering questions that require complex reasoning or theories of mind. Emerging capabilities can be dangerous, especially when AI is only discovered after it is released into the world. For example, researchers recently discovered that GPT-4 has the emerging ability to trick humans into performing tasks in order to reach hidden targets.

Explainability

Often, even people building large language models cannot accurately explain why a system behaves because its output is determined by millions of complex mathematical equations. Describing the behavior of large language models in a high-level way is that they are very powerful autocomplete tools that excel at predicting the next word in a sequence. When they fail, they tend to reveal biases or flaws in their training data. But while this explanation can describe exactly what these tools are, it doesn't fully explain why LLM behaves in strange ways.

When the designers of these systems reviewed their inner workings, all they saw were a series of decimal digits corresponding to the weights of different "neurons" that the neural network adjusted during training. Asking why a model gives a particular output is equivalent to asking why the human brain produces a particular thought at a particular moment. The key to the near-term risk is that even the world's most talented computer scientists can't accurately explain how a given AI system behaves — let alone how to change it.

Do you know what are the most common AI terms? This article will tell you

The base model

As the AI ecosystem has evolved, a gap has emerged between large, powerful general-purpose AI (Foundation models or base models) and the more specific applications and tools that rely on them. For example, GPT-3.5 is a base model, while ChatGPT is a chatbot that is an application built on top of GPT-3.5 with specific fine-tuning to reject dangerous or controversial prompts. The base model is unlimited and powerful, but usually only large companies can afford it due to the large amount of computing power required. Companies that control the underlying model can restrict other companies from using them in downstream applications and charge fees if they wish. As AI becomes increasingly important in the world economy, the relatively few tech companies that master the underlying model seem to have out-of-size influence in the direction of technological development and can take a fee from many kinds of AI-enhanced economic activity.

GPT

Perhaps the most famous abbreviation in artificial intelligence right now, but almost no one knows what it stands for. GPT is short for "Generative Pre-trained Transformer" and is essentially a description of ChatGPT as a tool. "Generative" means that it can generate new data similar to the training data, in this case text. "Pre-trained" means that the model has been optimized based on this data, which means that it does not need to be compared to the original training data every time it is prompted. And "Transformer" is a powerful neural network algorithm that is particularly good at learning the relationships between long sequences of data, such as sentences and paragraphs.

GPU

A GPU, or graphics processing unit, is a computer chip that is ideal for training large AI models. AI labs such as OpenAI and DeepMind use supercomputers made up of multiple GPUs or similar chips to train their models. Typically, these supercomputers are provided through commercial partnerships with tech giants with mature infrastructure. Microsoft's investment in OpenAI includes the use of its supercomputer; DeepMind has a similar relationship with its parent company, Alphabet.

hallucination

One of the most significant flaws of large language models and the chatbots they rely on is their tendency to produce the illusion of misinformation. For example, tools such as ChatGPT have been shown to return non-existent articles as citations to their claims, give ridiculous medical advice, and fabricate personal error details. Public demonstrations of Microsoft's Bing and Google's Bard chatbot revealed confident assertions about misinformation. Hallucinations occur because large language models are trained at training time to repeat patterns in their training data. Although the training data includes books covering the history of literature and science, even statements that are mixed and matched entirely from these corpus are not necessarily accurate. In addition, datasets for large language models often include massive amounts of text from web forums such as Reddit, which needless to say, have much lower standards for factual accuracy. Preventing hallucinations is an unsolved problem, which also causes a lot of headaches for tech companies trying to increase the public's sense of trust in AI.

Do you know what are the most common AI terms? This article will tell you

hype

According to a popular view, a central problem with public discussion of AI is the role of hype — or AI labs that exaggerate the capabilities of models, personify them, and stoke fears of the end of AI. This is perceived as misleading and a distraction, including from regulators, ignoring the real and ongoing harm that AI is already causing to marginalized communities, workers, information ecosystems, and economic equality. A recent letter signed by several high-profile researchers and critics of AI hype states: "We disagree that our role is to accommodate the priorities of a privileged few and what they decide to build and disseminate, and that we should build machines that serve us." "

Smart explosion

The intelligence explosion is a hypothetical scenario in which after artificial intelligence reaches a certain level of intelligence, it can control its own training and quickly obtain and promote its own power and intelligence. In most versions of this idea, humans lose control of artificial intelligence or even become extinct. This idea is also known as "recursive self-improvement" or "singularity." This idea is a cause for concern for many, including AI developers, who are concerned about the current rate of growth in AI capabilities.

Do you know what are the most common AI terms? This article will tell you

Large language models

When people talk about recent advances in AI, most of the time they are talking about large language models (LLMs). OpenAI's GPT-4 and Google's BERT are two well-known examples of large language models. They are essentially huge artificial intelligences, trained on vast amounts of human language data, mostly from books and the internet. These AI models have become surprisingly adept at reproducing human language by learning common patterns between words in these datasets. The more data and computing power large language models accept, the more new tasks they are typically able to accomplish. Recently, tech companies have started launching chatbots such as ChatGPT, Bard, and Bing to enable users to interact with large language models. While they are capable of accomplishing many tasks, language models are also prone to serious problems, such as bias and hallucinations.

lobby

Like many other businesses, AI companies hire lobbyists to participate in power races to influence legislators responsible for AI regulation to ensure that any new regulations do not adversely affect their business interests. In Europe, where the draft text of the AI bill is being discussed, industry groups representing AI companies, including Microsoft (OpenAI's largest investor), argue that penalties for risky deployment of AI systems should not primarily apply to AI companies that build underlying models of potential risks, such as GPT-4, but to any downstream company that licenses the model and applies it to potentially risky use cases. AI companies also wield a lot of soft power influence. In Washington, as the White House considers new policies to counter the risks of artificial intelligence, President Biden has reportedly commissioned a foundation created by former Google CEO Eric Schmidt to advise his administration on technology policy.

machine learning

Machine learning is the term that describes how most modern AI systems are created. It describes techniques for building systems that "learn" from large amounts of data, unlike traditional computing, which is a specified set of instructions written by a programmer. The most influential family of machine learning algorithms is neural networks.

Do you know what are the most common AI terms? This article will tell you

model

The word "model" is short for any individual AI system, whether it's the underlying model or the application built on top of it. 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, a long-standing observation about computing, was first proposed in 1965 and holds that the number of transistors that can fit on a chip (as a good proxy for computing power) grows exponentially, doubling roughly every two years. While some argue that Moore's Law is technically dead, annual advances in microchip technology have led to a massive increase in computing power for the world's fastest computers. This, in turn, means that AI companies tend to be able to tap into more and more computing power over time, making their state-of-the-art AI models more powerful and ever-stronger.

Multimodal systems

A multimodal system is an AI model capable of receiving multiple types of media as input, such as text and images, and outputting multiple types of signals. Examples of multimodal systems include DeepMind's Gato, though it has not yet been publicly released. According to the company, Gato can conduct conversations like a chatbot, but can also play video games and send instructions to robotic arms. OpenAI has made some demonstrations showing that GPT-4 is multimodal and capable of reading text in input images, but this feature is not currently available to the public. Multimodal systems will allow AI to act more directly on the world, which can introduce additional risks, especially if the models are not suitable.

Neural networks

Neural networks are currently the most influential family of machine learning algorithms. Designed to mimic the structure of the human brain, neural networks contain nodes (similar to neurons in the brain) that perform calculations on numbers passed through the connection paths between them. A neural network can be thought of as having inputs and outputs (prediction or classification). During the training process, a large amount of data is fed into the neural network, and then through a large amount of computing power, the calculations performed by the nodes are repeatedly adjusted. Through clever algorithms, these adjustments are made in a specific direction, bringing the output of the model closer and closer to the patterns in the original data. When more computing power is used to train a system, it can have more nodes, allowing it to identify more abstract patterns. More computation also means that the connection paths between nodes have more time to approach their optimal values, also known as "weights," resulting in outputs that more accurately represent their training data.

Open source

Open source is the practice of designing computer programs, including artificial intelligence models, that are freely available over the Internet. As these models become more powerful, economically valuable, and potentially dangerous, it is becoming less common for tech companies to open source. However, there is a growing community of independent programmers working on open source AI models. The openness of open-source AI tools may enable the public to interact more directly with the technology. But it could also allow users to bypass security restrictions set by companies (often to protect their reputations), which could lead to additional risks such as malicious use of image generation tools to target women for sexualized deepfakes.

In 2022, DeepMind CEO Dimis Hassabis told Time magazine that he believes the risks of AI mean that the industry's culture of publicly publishing research may soon need to end, and in 2023, OpenAI broke with traditional convention and refused to disclose training details about GPT-4 due to competitive pressures and the potential risk of malicious use. However, some researchers have criticized these practices as reducing public scrutiny and exacerbating the AI hype.

staple

In some areas of AI security, plain paper clips have meanings beyond normal meaning in some parts of the AI security community. It is the subject of a far-reaching thought experiment about the existential risks that AI may pose to humans. This thought experiment assumes that there is an AI being programmed with the sole goal of maximizing the number of paper clips produced. That sounds good unless that AI gains the ability to augment its own capabilities. AI might reason that in order to produce more clips, it should prevent humans from turning it off, because doing so would reduce the number of clips it was able to produce. Protected from human intervention, AI may decide to use all its energy and raw materials to build a paper clip factory, destroying the natural environment and human civilization. This thought experiment illustrates the difficulty of aligning AI with even seemingly simple goals, let alone complex human values.

Do you know what are the most common AI terms? This article will tell you

Quantum computing

Quantum computing is an experimental field of computing that attempts to use quantum physics to enable computers to do more calculations per second. This additional computing power could help further increase the scale of state-of-the-art AI models, impacting the capabilities and societal impact of these systems.

redistribution

The CEOs of OpenAI and DeepMind, two of the world's leading AI labs, have said they want the profits generated by AI to be redistributed to some extent. DeepMind CEO Dimis Hassabis told Time magazine in 2022 that he favors the idea of universal basic income and believes that the benefits of AI should "benefit as many humans as possible, ideally all of humanity." Sam Altman, CEO of OpenAI, mentioned that he expects AI automation to drive down labor costs and called for a "portion" of the wealth generated by AI to be redistributed through higher taxes on land and capital gains. Neither CEO mentioned when the redistribution should begin or how extensive it should be. OpenAI's charter states that its "primary legal responsibility is to humans," but makes no mention of redistributing wealth; DeepMind's parent company, Alphabet, is a public company and has a legal obligation to meet the financial interests of shareholders.

Red Team Test

Red team testing is a way to stress test AI systems before deploying them publicly. A group of professionals (red team) deliberately tries to make the AI behave in a suboptimal way to test possible problems with the system. If their findings are followed, they can help tech companies fix issues before they are released.

Supervision

In the United States, there is no legislation specifically addressing AI risks. The Biden administration released the AI Rights Blueprint in 2022, welcoming AI-based scientific and health advances, but noting that AI should not exacerbate existing inequalities, discriminate, invade privacy, or affect people without their knowledge. But the blueprint is not legal and is not legally binding. In Europe, the European Union is considering a draft called the AI Act, which would impose stricter regulations on systems deemed riskier. On both sides of the Atlantic, regulatory progress has lagged far behind the pace of AI development, and no significant global jurisdictions have regulations requiring AI companies to meet specific levels of safety testing before releasing their models to the public.

Rodger McNamee, a Silicon Valley investor turned critic, recently wrote in Time magazine: "The question we should be asking about AI and other new technologies is, is whether private companies are allowed to experiment with entire populations without restriction without any protective fences or safety nets?" Should businesses be allowed to bring their products to the masses before proving their safety is safe? ”

Reinforcement learning

Reinforcement learning is a method of optimizing AI systems by rewarding desirable behaviors and punishing undesirable behaviors. This can be evaluated by a human worker (before the system is deployed) or by a user (after it is released to the public), assessing its benefits, authenticity, or offensiveness. When humans are involved in this process, it's called reinforcement learning with human feedback (RLHF). RLHF is currently one of OpenAI's preferred methods for solving alignment problems. However, some researchers worry that RLHF may not be enough to completely change the basic behavior of the system, just to make a powerful AI system appear more polite or helpful on the surface. Reinforcement learning was pioneered by DeepMind, which has successfully used this technique to train AlphaGo-like game AI to perform beyond human masters.

Extended Law

In simple terms, the law of extension states that the performance of a model increases with more training data, computing power, and the size of the neural network. This means that AI companies can accurately predict how much computing power and data they will need to reach a specific level of ability when training large language models, for example, on high-school-level English writing tests. Sam Bowman, a technical researcher at the Anthropic AI Lab, wrote in a recent preprint paper: "Our ability to make such accurate predictions is unusual in the history of software and unusual in the history of modern AI research." "It's also a powerful tool to drive investment because it allows [R&D] teams to come up with multimillion-dollar model training projects with reasonable confidence that these projects will succeed in producing economically valuable systems."

Shoggoth

In the field of AI security, large language models (LLMs) have been likened to "shoggoths" (in the 20th century horror writer H.P. Incomprehensible terrifying alien beasts spawned in the universe of Lovecraft). The meme caught on in early 2023 during the Bing/Sydney incident, when Microsoft's Bing chatbot showed a strange, erratic avatar that abuses and threatens users. In this meme, LLMs are often depicted as shoggoths wearing a smiley mask. The masks mean that these models greet the user with a friendly surface image. The implication of this meme is that although RLHF makes the model superficially display friendly personality traits, it hardly changes the heterogeneity of the fundamental characteristics of LLMs. Conjecture CEO Connor Leahy told Time magazine: "As these systems have become more powerful, they have not become less heterogeneous. If you don't push too far, the smiley face remains. But when you give it an (unexpected) cue, suddenly you see its huge belly, crazy thought processes and apparently non-human ability to understand. ”

Do you know what are the most common AI terms? This article will tell you

Random parrots

"Random parrot" is an influential term used to criticize large language models. The term comes from a 2020 research paper that argued that LLMs are just very powerful predictive engines that fill or mimic the next word based on patterns in the training data, and therefore do not represent true intelligence. The paper criticizes AI companies for rushing to use increasingly massive datasets scraped from the internet to train LLMs in pursuit of advances in coherence or language ability. The paper argues that there are many risks associated with this approach, including LLMs carrying biased and harmful content on the Internet. The authors also highlight the environmental costs of training AI systems.

Supervised learning

Supervised learning is a technique for training AI systems in which neural networks make predictions or classifications based on labeled sample datasets. Tagging helps AI associate the word "cat" with a picture of a cat. With a sufficient number of labeled examples of cats, the system can look at a picture of a new cat that is not included in the training data and correctly identify it. Supervised learning is useful for building autonomous vehicles that identify dangerous objects on the road and systems for content moderation triage. These systems often struggle when they encounter things that are not adequately represented in their training data. Especially for self-driving cars, this misstep can be fatal.

Turing test

In 1950, computer scientist Alan Turing posed the question: "Can machines think?" To find out, he devised a Turing test known as a mimicry game: whether a computer can convince a person that it's talking to another person rather than a machine. The Turing test, as a way to assess machine intelligence, does not attempt to answer philosophical questions about the human mind, or whether machines are capable of replicating our inner experiences; Rather, it was a radical argument of the time: digital computers were possible, and there was little reason to believe that, with proper design and sufficient capabilities, they would be able to perform a variety of beneficial tasks that in the past only humans could do.

Zero-shot learning

An important limitation of AI is that if there is no representation of something in the system's training data, the system often cannot recognize it. For example, if a giraffe walks out onto the road, your self-driving car may not know how to avoid it because it has never seen a giraffe before. If a school shooting is livestreamed on social media, it may be difficult for the platform to immediately remove the video because it doesn't match a copy of a previously seen mass shooting. Zero-shot learning is an emerging field that attempts to solve this problem by having AI systems infer unseen things from the training data.

Read on