Artificial Intelligence, Explained

A chipmaker announces record demand and the headline touts it as an artificial intelligence (AI) company. A warehouse robotics firm raises capital and the press coverage calls it an AI company. A chatbot developer files to go public and it, too, is labelled an AI company. Yet these are three distinct businesses from different industries with three sets of customers and entirely different economics.

AI is a wide field, and the companies working inside it can behave very differently. Thus, investors and traders are better served by looking past the catch-all label to understand what a company actually builds.

What Artificial Intelligence Actually Refers To

Artificial intelligence* is a field, not a product or a single technique. The National Institute of Standards and Technology (NIST) defines an AI system as one that can, for a given set of human-defined objectives, make predictions, recommendations, or decisions influencing real or virtual environments.¹

That definition is broad and potentially covers a fraud-detection model at a bank, the vision system in a self-driving car, and the software that drafts an email. Those systems, though, can differ substantially in how they are built and deployed. What they share is the goal: performing tasks that would otherwise require human effort and judgment.

Three Eras, Three Different Approaches

AI has gone through a few distinct eras, and a quick walk through them explains a lot about where things stand today in a rapidly changing world.

The rule-based era. The field was named in a 1955 proposal for a summer research project at Dartmouth College, written by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon. Their conjecture was that "every aspect of learning or any other feature of intelligence can, in principle, be so precisely described that a machine can be made to simulate it."² Much of the early work followed a rule-based approach. Engineers wrote explicit rules, usually as “if-then” statements, and the machine followed those instructions. The systems worked inside narrow domains and broke down outside them, because every situation had to be anticipated in advance by a person.

The machine learning era. By the 1990s, the alternative was to stop writing rules and start supplying examples. Machine learning* is the development and use of computer systems that adapt and learn from data with the goal of improving accuracy.³ Rather than being told how to identify fraud, a model is shown many transactions and infers the pattern itself.

The deep learning era. A critical advancement came in 2012, when a neural network* called AlexNet entered the ImageNet image-recognition competition. It won with an error rate of 15.3%, against 26.2% for the second-best entry, using a model with 60 million parameters trained on graphics processors.⁴ The margin was large enough to redirect the field. Deep learning*, the use of neural networks with many stacked layers, became a dominant approach across vision, speech, and eventually language.⁵

The language piece arrived in 2017, when researchers at Google published an architecture called the Transformer, described in their paper as "based solely on attention mechanisms, dispensing with recurrence and convolutions entirely."⁶ Nearly every large language model (LLM)* in use today descends from that design.

The Categories of AI in Use Today

Today's AI landscape includes several distinct categories, all in use at the same time.

Predictive and classification models estimate an outcome or sort an input into a category. Credit scoring, demand forecasting, and spam filtering all exist here. This is the least visible and most widely deployed form of AI in business.

Computer vision interprets images and video. It underpins medical imaging analysis, quality inspection on manufacturing lines, and the perception systems in autonomous vehicles. The U.S. Food and Drug Administration approved 223 AI-enabled medical devices in 2023, up from six in 2015.⁷

Robotics and control systems combine perception with physical action. The distinguishing feature is a machine that moves through the world, introducing engineering, safety, and hardware constraints that software-only systems never face.

Generative AI* produces new content rather than classifying existing content. NIST defines it as the class of AI models that emulate the structure and characteristics of input data in order to generate derived synthetic content, including images, video, audio, and text.⁸

LLMs are one type of generative AI, specialized for text. They are trained on very large text datasets to predict and produce language, and they are the technology behind the chat assistants most people picture when they hear the term "AI."

Those last two terms get mixed up a lot. Every LLM is generative AI, but plenty of generative AI has nothing to do with language. And generative AI is only one corner of the field.

Why the Distinction Matters

Let’s go back to the three companies mentioned above. Each is exposed to different forces:

• A chipmaker sells hardware. Its demand depends on how much computing capacity the industry is building, and its risks are supply chains, manufacturing capacity, and customer concentration.

• A robotics company sells physical machines. Its business runs on manufacturing throughput, deployment cycles, and safety regulation. It can be an excellent business in a year when language models disappoint, or the reverse.

• An LLM developer sells software and access to models. Its biggest costs come from training models and then running them for users, and its competition includes both well-capitalized incumbents and open-source alternatives.

Grouping all three under "AI exposure" obscures more than it explains. The AI category is also broad enough that adoption figures require a close reading: Stanford's Human-Centered AI Institute found that 78% of organizations reported using AI in at least one business function in 2024, up from 55% a year earlier, with 71% specifically using generative AI.⁹ The second figure sits inside the first, which is a useful reminder to check what a given AI statistic is actually counting.

A Concrete Example: Anthropic and Claude

The taxonomy is easier to explain with a real-world case: Anthropic.

Anthropic is an artificial intelligence safety and research company that builds reliable, interpretable, and steerable AI systems, and that develops the Claude series of LLMs.¹⁰

Claude is a product family built on LLMs, which are one type of generative AI. Generative AI is one category within the field of artificial intelligence. Four levels, each nested inside the next, and the company at the top is not the same thing as any of the layers below it.

Anthropic has filed for an initial public offering and has not yet listed.¹¹

Where Direxion Fits

Direxion has filed with the SEC to launch the Direxion Daily Anthropic Bull 2X ETF (ANTO) and Direxion Daily Anthropic Bear 2X ETF (ANTD). ANTO is designed to seek daily investment results, before fees and expenses, of 200% of the daily performance of Anthropic, and ANTD is designed to seek 200% of the inverse of that daily performance. The funds have not yet launched, and the registration statement must become effective before the funds may be sold. The funds seek daily investment objectives, are intended for traders who actively monitor their positions, and are not designed to be held for periods longer than a single day.¹¹