Introduction
Galaxy Research’s AI coverage usually focuses on meat-and-potatoes topics. To name a few: decentralized training; agentic payments, fundraising, and companies; inference capital markets; policy debates around model access and data centers; and, most soberingly, AI-powered cyber exploits.
In this new series, we’re trying something a little different. AI development is accelerating at unprecedented speed toward an unpredictable end, challenging us to take a step back and think about what it means to be human. Over the next few months, we’ll be looking at “thinking machines” through a wider lens than usual, occasionally strolling into the terrain of philosophy.
This may sound highfalutin coming from analysts at a digital assets and AI infrastructure company. But whether you’re a portfolio manager, a data-center technician, a venture capitalist, a crypto trader, or just a curious bystander, we bet that at some point in the past year you’ve been blown away (and perhaps unsettled) by what AIs can do. If you want to indulge a little more time in exploring how they got here and how they work, please read on. (Feedback welcome: email [email protected].)
First up: an overview of the AI field with an exposition of the current models.
Defining AI
The term artificial intelligence (AI) is so broadly used now that it could refer to many things. Perhaps the first thought that comes to many people’s minds is a generative chatbot such as ChatGPT, Claude, and Gemini.
At the highest level, when the field of AI was formally established in 1956, it denoted artificial general intelligence (AGI) and artificial superior intelligence (ASI). The former referred to a system that matches human-level performance across a full range of cognitive tasks. The latter referred to a hypothetical system beyond AGI surpassing human performance.
Whether genuine AGI has been achieved is hotly debated. Some researchers and industry leaders consider successful generalization on novel tasks to be evidence it has. For example, Jensen Huang, Nvidia’s CEO, and Greg Brockman, OpenAI’s president and co-founder, both claimed that the AGI era has arrived with the release of GPT-6 Astra on Sept. 3. Others believe that existing systems lack robust reasoning, grounded understanding, and the ability to generalize beyond their training.
For this series’ purpose, we will not speculate on this question. Instead, we will focus on how the systems that exist today can equal or even surpass humans in narrow tasks, and what comparing these systems to human cognition can and cannot show.
The Long Pursuit of Artificial Minds
To compare human cognition against AI mechanisms, it is necessary to go back in time and understand the history of AI development – an endeavor in which humans sought to mimic intelligence through machines.
Attempts to replicate human intelligence have existed throughout history. From the mathematician Hero of Alexandria's mechanical automata in the 1st century to the Mechanical Turk in 1770, we humans have always been fascinated by the idea of creating something that can behave and think like us. In 1950, Alan Turing further entertained the fascination, proposing the Turing test to gauge if machines have intelligence. However, the field of AI was not established until 1956, when a group of scientists gathered at Dartmouth College to brainstorm ideas.
“We propose that a 2-month, 10-man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College in Hanover, New Hampshire. The study is to proceed on the basis of the conjecture 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. An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves.” – The Dartmouth Workshop Proposal
From these scientists’ work, a branch of the discipline called symbolic AI took shape as the field's first dominant paradigm. Other schools of technique developed in parallel, sometimes in tension with it, in the decades that followed.
On Human Cognition: How Do We Know What We Know?
Creating a “thinking machine” requires answering two questions:
How is the world structured?
How do we humans learn about the structure of the world?
In the 17th and 18th centuries, two rival schools of epistemology developed around these questions: empiricism and rationalism. Empiricists, represented by John Locke, George Berkeley, and David Hume, believed that knowledge derives primarily from sensory experience. The mind begins as a blank slate, which Locke calls the tabula rasa, and all concepts and knowledge are built up from what we perceive, observe, and experience through the senses. Humans are born with no innate ideas, and even abstract concepts such as shapes and color originate from sensory impressions which are then abstracted to complex ideas.
"Let us then suppose the mind to be, as we say, white paper, void of all characters, without any ideas... Whence has it all the materials of reason and knowledge? To this I answer, in one word, from experience: in that all our knowledge is founded, and from that it ultimately derives itself." – John Locke (An Essay Concerning Human Understanding, Book II, Ch. 1, 1689)
Rationalists, represented by René Descartes, Baruch Spinoza, and Gottfried Wilhelm Leibniz, believed that reason itself is a primary and independent source of knowledge, not derived from the senses. Some knowledge is innate or knowable through pure reasoning alone, prior to or independent of experience. This can be neatly summarized in the famous Descartes line: “I think, therefore I am,” an existential truth that he argued only by thinking could he acknowledge. Mathematical truths, logical principles, and certain metaphysical claims are treated as things the mind can grasp through rational insight, not something learned from observation.
"Nothing is in the intellect that was not first in the senses – except the intellect itself." – Gottfried Wilhelm Leibniz (New Essays on Human Understanding, written as a direct response to Locke, c. 1704)
These two schools of thought loosely correspond to the two major approaches of AI development.
The History of AI
The history of AI development is hard to map, given the field's diverse techniques and philosophies. There is no universal framework for categorizing its lineage, and many researchers would draw the lines differently. The categorization below is not the only valid one, but it captures the field's major divisions with reasonable accuracy.
The two broad approaches of AI development are symbolic and sub-symbolic.[1] Each carries a discernible lineage from the rationalist and empiricist traditions discussed above.
Symbolic AI breaks the world down into discrete symbols and formal operators with the philosophy that there are explicit rules governing how those symbols are combined. Through these rule-governed structures, we can generate outputs that mimic intelligent behavior. This philosophy mirrors that of the rationalists – the world is one in which reason, operating on well-defined categories, can derive truths. Just as Descartes held that certain truths could be established through pure deduction rather than experience, symbolic AI presupposes that intelligence can be captured by specifying rules and symbolic meanings in advance. Symbolic AI works by understanding and encoding the logical order of the world.
Sub-symbolic AI, by contrast, works with the uncertainty of information rather than fixed rules. Instead of encoding explicit logical structure, it uses probability and adjustable parameters to fine-tune outputs based on patterns extracted from data. Sub-symbolic AI detangles chaos. It learns through modeling the noise, ambiguity, and irregularity of raw information instead of imposing order on information beforehand. This mirrors the empiricists’ philosophy that knowledge is built up from experience as opposed to gained by reasoning in advance. Much as Locke's tabula rasa begins with no innate content and derives all ideas from sensations, sub-symbolic systems begin without fixed rules and recognize patterns from data exposure.
The 1956 Dartmouth Workshop Proposal led to the creation of the Logic Theorist, which was widely recognized as the first AI program. Against Alfred North Whitehead and Betrand Russell’s famed Principia Mathematica, a book of logic that sought to derive the foundations of mathematics, Logic Theorist not only successfully proved 38 of 52 theorems but also found a novel, shorter proof in one case. In so doing, it demonstrated that machines could not only replicate but also improve formal human reasoning through symbolic manipulation alone.
From here, symbolic AI models dominated for much of the next three decades. The Defense Advanced Research Projects Agency (DARPA) provided funding into the early 1970s with minimal oversight. After a brief funding pause, symbolic AI found commercial payoff in the expert-systems[2] boom of the early-to-mid 1980s. By 1985, industries spent over $1 billion annually mostly on in-house AI departments. These expert systems’ tasks ranged from identifying bacteria from infectious diseases and recommending antibiotics dosages (Mycin) to helping customers choose the right computer components to purchase based on their needs (XCON).
However, by the early 1990s, researchers realized that while the symbolic method could solve the "hard" problems, like proving Principia Mathematica's theorems, it could not solve the "easier" problems of perception and commonsense reasoning. Repeated but failed attempts to encode commonsense knowledge through symbolic logic, such as the Cyc project,[3] exposed the limits of rule-based reasoning for these problems. Some symbolic scientists believed that advancing narrower, domain-bound applications should take precedence over pursuing commonsense reasoning. Edward Feigenbaum, the pioneer of expert systems, argued that an AI program's power to solve complex problems came primarily from the specific domain knowledge it held. This philosophy shaped an entire generation of expert systems that made no attempt to replicate general human cognition or commonsense understanding, aiming instead to encode a narrow slice of specialist judgment within a bounded field.
In parallel, the sub-symbolic field kept developing despite a much rockier path. Sub-symbolic AI research began in 1958 with Frank Rosenblatt's invention of the Perceptron, which was the earliest well-known neural network model.[4]
The 1969 book Perceptrons by Marvin Minsky and Seymour Papert, two prominent symbolic AI researchers, severely undercut confidence in neural networks. Combined with symbolic AI's contemporaneous successes, sub-symbolic research received little funding and declining interest through the 1970s.
In the following decade, sub-symbolic AI research enjoyed a major revival due to three breakthroughs: John Hopfield's proof that neural networks could reliably learn and reach a solution, Geoffrey Hinton's similar result for the Boltzmann machine[5] and Hinton and David Rumelhart's popularization of the backpropagation algorithm for training multi-layer networks.[6]
Shortly after, Yann LeCun applied backpropagation to train a convolutional neural network to accurately scan handwritten zip codes provided by the U.S. Postal Service. This was a messy problem because different handwriting can be hard to identify. LeCun’s method was able to pick up visual patterns such as the loop of an “8” and generalize patterns into results. It achieved a 95% success rate and paved the way for automated mail-sorting systems adopted in the early 1990s.
Despite these successes, neural networks remained limited by insufficient data and computing power throughout the 1990s and early 2000s. However, with massive datasets generated by the internet, the repurposing of GPUs (once mostly used for gaming and 3D graphics) for parallel computation, and refinements to training techniques, sub-symbolic AI became a dominant branch in the 21st century. This transition was led by the rise of deep learning, which scaled the same neural network principles established decades earlier to previously unimaginable size and performance.
Deep learning uses multi-layer neural networks to progressively extract higher-level features from raw data, without human specification. It has since become a key technique in AI, applied to facial recognition, machine translation, climate modeling, and self-driving cars. Within this broader deep learning paradigm, a specific architectural innovation would prove decisive for language: the Transformer[7] .
The Transformer was introduced in 2017 in a paper written by eight Google researchers, "Attention Is All You Need." It enabled large language models and the current wave of breakthroughs from frontier labs and open-weight models. Today, most AI models such as ChatGPT, Claude, Gemini, and Kimi rely on deep learning and the Transformer architecture to generate their outputs.
The figure below shows the prevalence in research of the two AI approaches throughput history. The authors compared symbolic AI against connectionism[8] instead of the broader sub-symbolic AI we have been referring to. They totaled the number of papers that have cited the relevant methods and computed the relative ratio.
The competition between symbolic and sub-symbolic AI shows researchers’ contrasting beliefs in whether information is better organized by order or chaos, and whether humans learn by logic or processing raw data. The era we are in now, with the massive amount of data that we are immersed in and have access to, is one in which chaos triumphed over order, and thus our learning has shifted to engaging with raw, unstructured experience.
Surveying the Major Current AI Models
Before diving into specific AI models and how they work, we caution that what follows is not a comprehensive survey on artificial intelligence methods, but a specific and focused survey on machine learning and deep learning architectures. We will confine our discussions to these learning paradigms and five deep generative techniques because they are the systems that propelled the most visible recent breakthroughs in AI.
One more caveat: the categorization of these paradigms, architectures, and techniques, is highly debated among AI researchers, and there is no universal framework for organizing them. Classifications can differ depending on the time of publication and the author, and given the nuances of the methods themselves, even individual terms can be understood differently across sources. As such, any conclusions drawn throughout this paper represent our best attempt to synthesize the disparate viewpoints within the current field of AI. The goal is to offer clarity and a unified set of categorizations.
Three Machine Learning Paradigms
A subfield within AI is machine learning, through which a system improves its performance on a task by learning patterns from data rather than by a fixed set of rules. There are three major approaches: supervised learning, reinforcement learning, and unsupervised learning.[9] These three approaches guide how systems absorb enormous piles of data and generate the desired outputs.
Supervised learning involves guidance in training the model to correctly generate an output. The system is provided with pairs of inputs and their outputs and uses pattern-recognition to map one to the other. To break it down further, supervised learning takes inputs and compares its generated outputs against the correctly labelled results, with the goal of reducing error through many repeated attempts. The end mapping is applied to untrained inputs to predict outputs.
Reinforcement learning tries actions and adjusts based on the results. Through repetitive trial and error, the system adjusts its behavior to maximize rewards and avoid penalties.
Unsupervised learning involves finding structure in data without labels. There are various techniques to accomplish this goal, including clustering, dimensionality reduction, and anomaly reduction. (Clustering involves grouping of similar items. Dimensionality reduction simplifies complex data to its foundational factors. Anomaly detection views the gestalt of a data distribution to access for outliers. ) The difference between supervised and unsupervised learning is that in the latter there is no correct answer to check against during training. The goal of unsupervised learning is to generate a mapping that is applicable and cogent.
Five Deep Generative Architectures
The three learning paradigms introduced above explain how a model acquires its knowledge. Below we describe deep generative architectures[10] that explain what a trained model does when asked to produce something new, or in other words, how they produce outputs.
Autoregressive models work by predicting one piece of information at a time. This is the mechanism behind most current large language models (LLM), which underly chatbots such as ChatGPT and Claude. When applied to language, the model examines an entire body of text to determine the possible selection of subsequent words and their respective probability of fit. It then selects the best word and appends it to the end of the text. The process is repeated until a completed answer is generated. In addition to language, these models can also be used to generate images, audio, and videos.
Diffusion models work by cleaning up noise. It is mainly used to generate images and videos. The model is trained through two processes: it witnesses an actual image getting subsequently corrupted by random noise, then it runs in reverse, attempting to reconstruct the original image by removing random noise in steps. With repetitive training, the model eventually learns to generate novel images from pure random static through denoising.
Generative Adversarial Networks (GANs) work by having two neural networks compete against each other. One acts as a generator that produces fake data. The other acts as a discriminator that tries to detect fake data. Through this back-and-forth production and discrimination process, each party improves its ability to generate or detect garbage. The process ends when the generator produces outputs that are indistinguishable from real data.
Variational Autoencoders (VAEs) work by reconstructing compressed data. There are also two modules in this process: the encoder and the decoder. Unlike GANs’ two neural networks that compete against each other, the encoder and the decoder work together to generate outputs. The encoder helps train the decoder by taking inputs and compressing the data into lower-dimensional form, called the “latent space,” using variational inference to approximate the latent space's true distribution with a simpler, computable one. The decoder then attempts to reconstruct the compressed data back into its original form. Once the decoder has a good understanding of the encoded distribution, it gains the ability to sample any fixed point in the latent space to generate new, cogent outputs.
Normalizing flows work by learning to shape and un-shape a raw input into a computed product through revertible methods. It starts with a known probability distribution and computes it to a complex probability distribution matching real data, through a series of invertible transformations. The model is trained with the backward process, where it takes a data point on the computed probability distribution, calculates the corresponding data point on the known probability distribution, and evaluates the calculated data’s likelihood of fit. Once trained, it can generate new data by sampling data points on the known probability distribution and applying the forward process to transform inputs to outputs.
Modern LLMs in Practice
A chatbot like ChatGPT, Claude, or Gemini is where sub-symbolic AI, connectionism, machine learning, neural networks, deep learning, generative AI techniques, the Transformer, and LLMs all come together.
The chatbots are built on the Transformer, the architecture that enables the model to parallel process large amounts of data and direct its attention to the most relevant result. On top of that architecture, an LLM generates outputs autoregressively, one word at a time, spanning all stages of its learning.
Getting a model to provide a good response to prompts requires training in three stages, closely related to the learnings paradigm covered earlier. The first stage is self-supervised pretraining: the model works with unlabeled data and ingests large sets of training texts; it trains on its own to learn the structure of human texts and tries to make sense of the texts; it comes up with the “Base LLM.” The second stage is supervised fine-tuning: humans provide example prompts and answers to the model, and the model adjusts its answers accordingly to mimic the correct answer. The third stage is reinforcement learning with human feedback: the model generates multiple answers given a prompt, and human raters select the preferred response; it then configures itself to match human preference.
Beyond the Base LLMs
LLMs proved to be such a capable general-purpose foundation with the ability to read, write, and reason about language across an enormous range of tasks that researchers and companies began building specialized extensions on top of them. The following models are developed to extend the base LLM in various directions, building on the LLM’s core capabilities to make it better for a wider range of tasks, more types of data, and further modes of interaction.
Multimodal models: LLMs + perception, using the same generative techniques to take in more types of data and produce images, audio, and video alongside text.
Reasoning models: LLMs + deliberation, adding extended computation before coming up with answers, so the model works through a problem weighing different approaches and backtracking when needed, instead of producing its first plausible response.
Retrieval-Augmented Generation (RAG): LLMs + external knowledge base, applying LLM techniques while drawing on an authoritative database so that responses are tailored, accurate, and efficient.
Agents: LLMs + action over time, wrapping the model in a loop with access to tools, persistent memory, and a goal, so it can act, observe the result, and carry out a series of actions.
Embodied and world models: it is contested whether these models are an extension of the LLM. They incorporate information from a physical or simulated environment that evolves under action. Some embodied AIs are built directly on LLMs, such as certain humanoid robots using language models for high-level reasoning and task planning. World models, associated with researchers such as Yann LeCun and Fei-Fei Li, are rooted in the belief that grounding intelligence in the physical world requires abandoning the language-model approach altogether, in favor of architectures that predict future states of the world in an abstract representational space rather than in words. (Li’s company, World Labs, launched the first multimodal world model called Atlas on Sept. 1)
Conclusion
Whew, that was a lot of material to digest. We have established how current AI models generate outputs and what shapes their learning and surveyed the history of AI development. As that history suggests, AI development is closely connected with modeling human cognition. Current progress has led us to a stage where AI's outputs equal, or even surpass, our own.
In the next installment of this series, we will explore the recent breakthroughs conducted by chatbots, and what they imply for us humans.
Legal Disclosure:
This document, and the information contained herein, has been provided to you by Galaxy Digital Inc. and its affiliates (“Galaxy Digital”) solely for informational purposes. This document may not be reproduced or redistributed in whole or in part, in any format, without the express written approval of Galaxy Digital. Neither the information, nor any opinion contained in this document, constitutes an offer to buy or sell, or a solicitation of an offer to buy or sell, any advisory services, securities, futures, options or other financial instruments or to participate in any advisory services or trading strategy. Nothing contained in this document constitutes investment, legal or tax advice or is an endorsement of any of the stablecoins mentioned herein. You should make your own investigations and evaluations of the information herein. Any decisions based on information contained in this document are the sole responsibility of the reader. Readers should consult with their own advisors and rely on their independent judgement when making financial or investment decisions.
Participants, along with Galaxy Digital, may hold financial interests in certain assets referenced in this content. Galaxy Digital regularly engages in buying and selling financial instruments, including through hedging transactions, for its own proprietary accounts and on behalf of its counterparties. Galaxy Digital also provides services to vehicles that invest in various asset classes. If the value of such assets increases, those vehicles may benefit, and Galaxy Digital’s service fees may increase accordingly. The information and analysis in this communication are based on technical, fundamental, and market considerations and do not represent a formal valuation. For more information, please refer to Galaxy’s public filings and statements. Certain asset classes discussed, including digital assets, may be volatile and involve risk, and actual market outcomes may differ materially from perspectives expressed here.
For additional risks related to digital assets, please refer to the risk factors contained in filings Galaxy Digital Inc. makes with the Securities and Exchange Commission (the “SEC”) from time to time, including in its Quarterly Report on Form 10-Q for the quarter ended September 30, 2025, filed with the SEC on November 10, 2025, available at www.sec.gov.
Certain statements in this document reflect Galaxy Digital’s views, estimates, opinions or predictions (which may be based on proprietary models and assumptions, including, in particular, Galaxy Digital’s views on the current and future market for certain digital assets), and there is no guarantee that these views, estimates, opinions or predictions are currently accurate or that they will be ultimately realized. To the extent these assumptions or models are not correct or circumstances change, the actual performance may vary substantially from, and be less than, the estimates included herein. None of Galaxy Digital nor any of its affiliates, shareholders, partners, members, directors, officers, management, employees or representatives makes any representation or warranty, express or implied, as to the accuracy or completeness of any of the information or any other information (whether communicated in written or oral form) transmitted or made available to you. Each of the aforementioned parties expressly disclaims any and all liability relating to or resulting from the use of this information. Certain information contained herein (including financial information) has been obtained from published and non-published sources. Such information has not been independently verified by Galaxy Digital and, Galaxy Digital, does not assume responsibility for the accuracy of such information. Affiliates of Galaxy Digital may have owned, hedged and sold or may own, hedge and sell investments in some of the digital assets, protocols, equities, or other financial instruments discussed in this document. Affiliates of Galaxy Digital may also lend to some of the protocols discussed in this document, the underlying collateral of which could be the native token subject to liquidation in the event of a margin call or closeout. The economic result of closing out the protocol loan could directly conflict with other Galaxy affiliates that hold investments in, and support, such token. Except where otherwise indicated, the information in this document is based on matters as they exist as of the date of preparation and not as of any future date, and will not be updated or otherwise revised to reflect information that subsequently becomes available, or circumstances existing or changes occurring after the date hereof. This document provides links to other Websites that we think might be of interest to you. Please note that when you click on one of these links, you may be moving to a provider’s website that is not associated with Galaxy Digital. These linked sites and their providers are not controlled by us, and we are not responsible for the contents or the proper operation of any linked site. The inclusion of any link does not imply our endorsement or our adoption of the statements therein. We encourage you to read the terms of use and privacy statements of these linked sites as their policies may differ from ours. The foregoing does not constitute a “research report” as defined by FINRA Rule 2241 or a “debt research report” as defined by FINRA Rule 2242 and was not prepared by Galaxy Digital Partners LLC. Similarly, the foregoing does not constitute a “research report” as defined by CFTC Regulation 23.605(a)(9) and was not prepared by Galaxy Derivatives LLC. For all inquiries, please email [email protected].
©Copyright Galaxy Digital Inc. 2026. All rights reserved.