AI Guide: A glance at key concepts and top players in AI

Source: Techcrunch

Compilation: Babbitt

Image source: Generated by Unbounded AI tool

Artificial intelligence (AI) appears to be in every corner of modern life, from music and media to business and productivity, and even dating. There are so many things that it's hard to keep up. This article will cover everything from the latest big developments in AI to the terms and companies you need to know to stay up to date on the state of affairs in this fast-moving field.

First, what is artificial intelligence?

Artificial intelligence, also known as machine learning, is a software system based on neural networks, a technique that was actually pioneered decades ago but has recently flourished thanks to powerful new computing resources. Currently, AI has achieved effective speech and image recognition, as well as the ability to generate synthetic images and speech. Researchers are working to make artificial intelligence capable of browsing the web, ordering tickets, tweaking recipes and more.

But if you're worried about a Matrix-esque rise of machines — don't worry. We'll talk about that later!

This guide to AI consists of two main parts:

• First, the most basic concepts you need to understand and the most recent important concepts.
• Then, outline the main players in AI and why they are important.

AI 101

Image credit: Andrii Shyp/Getty Images

One of the crazy things about AI is that, while its core concepts date back more than 50 years, until recently few even tech-savvy people were familiar with its concepts. So don't worry if you're feeling lost -- everyone is.

Let's be clear on one thing up front: While it's called "artificial intelligence," the term is a bit misleading. There is currently no unified definition of intelligence, but what these systems do is definitely closer to a calculator than a brain, except that the input and output of this calculator is more flexible. AI might be like an "Artificial Coconut" - it's imitation intelligence.

The following are basic terms that you will find in any discussion about AI.

Neural Networks

Our brains are largely made up of interconnected cells called neurons that mesh together to form complex networks that perform tasks and store information. People have been trying to recreate this amazing system in software since the 1960s, but the processing power required wasn't widely available until 15-20 years ago, when GPUs allowed numerically defined neural networks to flourish.

Essentially, they're just lots of points and lines: the points are the data, and the lines are the statistical relationships between those values. Like in the brain, this can create a multifunctional system that quickly receives an input, passes it through the network and produces an output. This system is called a model.

Model

A model is the actual collection of code that takes input and returns output. The similarity in terminology to statistical models, or modeling systems that simulate complex natural processes, is not accidental. In AI, a model can refer to a complete system like ChatGPT, or almost any AI or machine learning construct, no matter what it does or produces. Models come in various sizes, which means how much storage space they take up and how much computing power they require to run. And it all depends on how the model was trained.

train

To create an AI model, the neural networks that form the basis of the system are exposed to a bunch of information called a data set, or corpus. In doing so, these vast networks create a statistical representation of that data. This training process is the most computationally intensive, meaning it takes weeks or months on huge, high-powered computers. The reason for this is not only that the networks are complex, but the datasets can be very large: billions of words or images must be analyzed and represented in huge statistical models. On the other hand, once a model is trained, it can be used much smaller and less demanding, a process called inference.

Image credit: Google

Inference

When the model actually does work, we call it inference, and the traditional meaning of the word is very much: to state a conclusion by reasoning about the available evidence. Of course, this isn't exactly "inference", but statistically connects points in the data it ingests, actually predicting the next point. For example, say "complete the following sequence: red, orange, yellow..." it will find that these words correspond to the beginning of the list it ingests, i.e. the colors of the rainbow, and extrapolate the next item until it has produced the rest of the list part.

Inference is usually much less computationally expensive than training: think of it like browsing a card catalog rather than assembling it. Large models still have to run on supercomputers and GPUs, but smaller models can run on smartphones or simpler devices.

Generative Artificial Intelligence

Everyone is talking about generative AI, a broad term that simply refers to AI models that generate raw output like images or text. Some AIs summarize, some reorganize, some recognize, and so on—but AIs that actually generate something (whether or not it "creates" is debatable) are especially popular right now. Remember, just because AI generated something, doesn't mean it's correct, or even that it reflects reality! It's just that it doesn't exist until you ask for it, like a story or a painting.

Hot words right now

Beyond the basics, here are the most relevant AI terms for mid-2023.

Large Language Model (LLM)

Large-scale language models are the most influential and widely used form of artificial intelligence today. Large-scale language models are trained on almost all texts that make up the web and most of the literature in English. Ingesting all of this results in a huge base model (read on). LLMs are able to converse and answer questions in natural language and mimic written documents of various styles and types, as evidenced by tools such as ChatGPT, Claude, and LLaMa.

While these models are undoubtedly impressive, it must be kept in mind that they are still pattern recognition engines, and when they answer a question, they are trying to complete the pattern it has identified, whether or not that pattern reflects reality. LLMs often hallucinate in their answers, as we'll get to shortly.

If you want to know more about LLM and ChatGPT, click here.

Foundation Model

Training a huge model from scratch on a huge dataset is expensive and complex, so you don't want to do more than you have to. Base models are large models from scratch that require supercomputers to run, but often by reducing the number of parameters, they can be reduced to smaller containers. You can think of these as the total number of points the model has to handle, which can run into millions, billions, or even trillions these days.

fine-tuning

A base model like GPT-4 is smart, but it's also a generalist by design - it absorbs everything from Dickens to Wittgenstein to the rules of Dungeons and Dragons, but if you want it to be based on your Resume write a cover letter, these are useless. Fortunately, it is possible to fine-tune the model by doing some additional training on the model using a specialized dataset. For example, there happen to be several thousand job applications. This gives the model a better understanding of how to help the user in that domain without discarding the general knowledge it gleaned from the rest of the training data.

Reinforcement Learning from Human Feedback (RLHF), is a special kind of fine-tuning that you'll hear about a lot — it uses data from humans interacting with LLMs to improve their communication skills.

Diffusion

*From a paper on advanced post-diffusion techniques, you can see how to reproduce images from very noisy data. *

Image generation can be done in a number of ways, but by far the most successful is diffusion, the technique at the heart of Stable Diffusion, Midjourney, and other popular generative AIs. The diffusion model is trained by showing it images that are gradually degraded by adding digital noise until there is nothing left of the original image. By observing this, the diffusion model also learns to perform the process in reverse, gradually adding detail to pure noise to form an arbitrarily defined image. We've started to move beyond that in graphics, but the technology is solid and relatively easy to understand, so it's going to die out pretty quickly.

Hallucination

Initially this was a problem of some images in training slipping into irrelevant output, e.g. buildings appearing to be made of dogs due to the over-prevalence of dog images in the training set. AI is now said to be hallucinating because it doesn't have enough or conflicting data in its training set, it just makes things up.

An AI asked to create original or even derivative art is hallucinating. For example, an LLM could be told to write a love poem in the style of Yogi Berra, and it would happily do so—even though such a thing doesn't exist in its dataset. But that can be a problem when a factual answer is required; the model will confidently present a half-real, half-illusioned response. Currently there's no easy way to tell which is which other than checking it yourself, since the model itself doesn't actually know what is "true" or "false", it's just trying to complete a pattern as best it can.

AGI or Strong Artificial Intelligence

Artificial General Intelligence (AGI), or Strong Artificial Intelligence, isn't really a well-defined concept, but the simplest explanation is that it's an intelligence powerful enough not only to do what people do, but also to be like us Learn and improve yourself. Some worry that this cycle of learning, integrating these ideas, and then learning and growing faster will be a self-perpetuating cycle that will lead to a superintelligent system that cannot be constrained or controlled. Some have even proposed delaying or limiting studies to prevent this possibility.

It's a terrible thought. And films like The Matrix and The Terminator have explored what might happen if artificial intelligence got out of control and tried to exterminate or enslave humanity. But these stories are not grounded in reality. The appearance of intelligence we see in things like ChatGPT is impressive, but has little in common with the abstract reasoning and dynamic multi-domain activity we associate with "real" intelligence.

While predicting future developments is nearly impossible, it may be helpful to imagine AGI as interstellar space travel: we all understand the concept and seem to be working towards it, but at the same time, we There is still a long way to go to achieve it. Just like AGI, no one will do it by accident due to the huge resources and basic scientific progress required!

It's fun to think about AGI, but there's no need to ask for trouble because, as commentators have pointed out, despite its limitations, AI already poses a real and significant threat today. No one wants Skynet, but you don't need a nuclear-armed superintelligence to do real damage: People are losing their jobs and getting scammed today. If we can't solve these problems, what chance do we have against the T-1000?

Top Players in Artificial Intelligence

OpenAI

Image credit: Leon Neal/Getty Images

If there's one household name in AI, it's OpenAI. OpenAI, as the name suggests, is an organization that intends to conduct research and make the results more or less publicly available. It has since restructured into a more traditional for-profit company that provides access to advanced language models like ChatGPT through APIs and apps. It's led by Sam Altman, a techno-billionaire who nonetheless has sounded the alarm about the possible risks of artificial intelligence. OpenAI is a recognized leader in the field of LLMs, but conducts research in other areas as well.

microsoft

As you might expect, Microsoft has done its fair share of AI research, but like other companies, has more or less failed to translate its experiments into major products. Its smartest move was an early investment in OpenAI, which led to an exclusive long-term partnership with the company that now powers its Bing conversational agents. Although its own contributions are smaller and less directly applicable, the company does have considerable research muscle.

Google

Known for its moonshots, Google somehow missed the opportunity for AI, even though its researchers did invent the technology that directly leads to today's AI explosion: Transformers. Now it's trying to develop its own LLMs and other agents, but it's clearly playing catch-up after spending most of the past decade pushing the outdated concept of AI "virtual assistants." CEO Sundar Pichai has repeatedly said the company is firmly behind AI in search and productivity.

Anthropic

After OpenAI's departure from openness, Dario and Daniela Amodei left it to start Anthropic, intending to fill the role of an open and ethically considerate AI research organization. With the amount of cash they have on hand, they are serious competitors to OpenAI, even if their models (like Claude) aren't quite as popular or well-known yet.

Image credit: Bryce Durbin/TechCrunch

Stability

Controversial but unavoidable, Stability represents the open-source genre of "whatever you want" AI implementations that collect everything on the internet and make the generative AI models it trains available for free, provided you own the hardware to run it. This fits very well with the "information wants to be free" philosophy, but also accelerates ethically dubious projects like generating pornographic images and using intellectual property without consent (sometimes simultaneously).

Elon Musk

Musk was no exception, outspoken about his concerns about runaway AI, and some sour grapes after his early contributions to OpenAI went in directions he didn't like. While Musk isn't an expert on the subject, as usual, his antics and comments do get a lot of buzz (he's a signatory to the aforementioned "AI pause" letter), and he's trying to build his own research presence .