Baffled by AI Jargon? Let’s Simplify It!

Artificial intelligence is revolutionizing our world while giving rise to a new lexicon to articulate its developments. A mere five minutes spent reading about AI can introduce you to terms like LLMs, RAG, RLHF, and a host of others that may leave even seasoned tech professionals feeling perplexed. This glossary is designed to demystify that complexity. It is consistently updated as the field evolves, serving as a dynamic reference akin to the AI systems it details.

---

Artificial general intelligence, commonly referred to as AGI, is a somewhat ambiguous concept. Generally, it refers to AI that performs beyond the average human across numerous tasks. OpenAI’s CEO Sam Altman described AGI as the “equivalent of a typical human you might hire as a co-worker.” Conversely, OpenAI’s charter characterizes AGI as “highly autonomous systems that outperform humans in most economically significant tasks.” Google DeepMind offers a slightly different take, defining AGI as “AI that possesses capabilities at least matching humans in most cognitive tasks.” Feeling confused? Do not fret — even top AI authorities regard it as a perplexing issue.

An AI agent is a tool that harnesses AI technologies to perform a range of tasks on your behalf, surpassing the functionalities of basic AI chatbots. Examples include managing budgets, arranging travel, or writing code. However, as previously mentioned, this burgeoning field is filled with nuances, leading to varying interpretations of the term “AI agent.” The underlying infrastructure is still being refined to realize its expected capabilities, but it fundamentally signifies an autonomous system that can integrate various AI systems to perform complex tasks.

Consider API endpoints as the “buttons” embedded within software that other applications can activate to trigger functions. Developers utilize these interfaces to forge integrations — enabling one application to pull data from another or allowing an AI agent to engage with third-party services without human intervention. Many smart home devices and connected solutions incorporate these hidden buttons, even if users never notice or exploit them. As AI agents evolve, they are increasingly capable of autonomously identifying and utilizing these endpoints, thereby unlocking substantial — and sometimes unexpected — automation capabilities.

When faced with straightforward questions, the human mind can provide answers almost instantly — like “Which animal is taller, a giraffe or a cat?” However, there are instances where written calculations are required to arrive at the correct outcome, due to intervening steps. For example, if a farmer has chickens and cows, collectively accounting for 40 heads and 120 legs, writing a simple equation may be essential to find the solution (20 chickens and 20 cows).

In the realm of AI, chain-of-thought reasoning for large language models entails decomposing a problem into smaller, intermediate steps to enhance the quality of the final output. While this approach usually takes longer to generate an answer, it is more likely to yield accurate results, particularly in logical or coding contexts. Reasoning models are developed from traditional large language models and refined for chain-of-thought reasoning via reinforcement learning.

(See: Large language model)

This term is more nuanced than “AI agent,” denoting a program that can take successive actions autonomously to achieve a goal. A coding agent is a specialized form, designed specifically for software development. Rather than merely suggesting code for human review, a coding agent can autonomously write, test, and debug code, alleviating the repetitive tasks that often consume a developer’s time. These agents can navigate entire codebases, pinpointing bugs, executing tests, and applying fixes with minimal human oversight. Think of it as hiring an exceptionally diligent intern who never tires — though, similar to any intern, human supervision is still necessary.

Though somewhat multifaceted, compute usually refers to the crucial computational power that enables AI models to function. This processing capability is what drives the AI industry, allowing for the training and deployment of sophisticated models. The term often denotes the types of hardware that provide this computational strength — like GPUs, CPUs, TPUs, and the various infrastructures that underpin today’s AI technology.

This category of self-improving machine learning includes AI algorithms designed with a multi-layered, artificial neural network (ANN) architecture. This setup lets them form more intricate associations than simpler machine learning systems such as linear models or decision trees. The design of deep learning algorithms is influenced by the interconnected pathways found in the human brain.

Deep learning AI models can autonomously uncover significant features in data, negating the need for human engineers to define them. This architecture also accommodates algorithms that learn from errors and refine their outputs through iterative adjustments. However, deep learning systems demand vast quantities of data points to deliver trustworthy results (often in the millions or beyond). They also generally require longer training periods compared to simpler machine learning algorithms, which can escalate development costs.

(See: Neural network)

Diffusion pertains to the technology that powers a variety of art-, music-, and text-generating AI models. Modeled after physical processes, diffusion systems gradually “deconstruct” data — including photos, songs, and more — by adding noise until they become unrecognizable. In physics, diffusion is a spontaneous and irreversible process — such as sugar dissolving in coffee, which cannot revert to its original cube form. Nonetheless, diffusion systems in AI aim to master a “reverse diffusion” process to reconstruct data from noise.

Distillation is a technique used to extract knowledge from a larger AI model through a ‘teacher-student’ model. Developers request outputs from a teacher model and compare these to a dataset for accuracy. These outputs then train the student model, guiding it to emulate the teacher’s behaviors.

Distillation allows for the creation of a smaller, more efficient model derived from a larger one while minimizing distillation loss. This technique likely contributed to the development of GPT-4 Turbo, a more efficient version of GPT-4.

While all AI companies use distillation internally, some may have employed it to catch up with leading models. However, distillation from rivals typically violates the terms of service of AI APIs and chat assistants.

This term refers to the further training of an AI model to improve its performance for a more specialized task or domain than it previously targeted — often by introducing new, exclusive data.

Numerous AI startups leverage large language models as foundational structures to create commercial offerings, aiming to enhance their utility for a specific sector or task by enriching previous training with domain-specific insights and expertise.

(See: Large language model [LLM])

A GAN, or Generative Adversarial Network, is a kind of machine learning structure that propels advancements in generative AI, particularly in generating authentic data — including deepfake technologies. GANs consist of a duo of neural networks: one generates outputs based on its training data, while the other assesses these outputs.

The two models are essentially pitted against each other; the generator strives to produce outputs that deceive the discriminator, tasked with identifying artificially generated data. This competitive framework enhances the realism of AI outputs without needing additional human oversight. Although GANs excel in specific use cases (like crafting lifelike images or videos), they aren’t as effective for general-purpose AI.

Hallucination is a term favored within the AI sector to describe instances where AI models fabricate incorrect information — essentially making things up. This is a critical concern for the quality of AI outputs.

Hallucinations can result in generative AI outputs that mislead and pose real-world risks (think of a health inquiry yielding harmful medical advice).

The issue of AI fabricating information is believed to arise from gaps in training data. Hallucinations are pushing the trend toward increasingly specialized vertical AI models — that is, domain-specific AI focusing on narrower areas of expertise — to help close knowledge gaps and minimize misinformation risks.

Inference is the process of utilizing an AI model. This act entails enabling a model to make predictions or conclusions based on previously encountered data. Importantly, inference cannot occur without training; a model must be educated to recognize patterns in a dataset before it can accurately extrapolate from that training data.

Various hardware types can execute inference, ranging from smartphone CPUs to high-performance GPUs and custom AI accelerators. However, not all hardware performs models equally well; for example, very large models would take significantly longer to make predictions on a standard laptop than on a cloud server outfitted with advanced AI chips.

[See: Training]

Large language models, or LLMs, are the AI frameworks employed by popular assistant platforms, such as ChatGPT, Claude, Google’s Gemini, Meta’s AI Llama, Microsoft Copilot, or Mistral’s Le Chat. When you interact with an AI assistant, you are engaging a large language model that processes your request directly or with the assistance of various tools, ideal for tasks such as web searching or code interpretation.

LLMs are extensive deep neural networks consisting of billions of numerical parameters (or weights, to be explained below) that learn the relationships between words and phrases, constructing a language framework — effectively a multidimensional mapping of words.

These models evolve by capturing the patterns they identify in vast databases of text, including books, articles, and transcripts. When prompted, an LLM generates the most likely pattern that corresponds to the request.

(See: Neural network)

Memory cache refers to a critical process that optimizes inference (the method by which AI generates responses to user queries). Essentially, caching is an optimization strategy designed to enhance the efficiency of inference. As AI relies on complex mathematical calculations, each computation consumes energy. Caching aims to reduce the number of computations required by storing certain calculations for future queries and transactions. Various forms of memory caching exist, one of the more well-known being KV (key-value) caching. KV caching operates in transformer architectures, improving efficiency and delivering quicker results by reducing the time and computational resources needed to generate responses to user inquiries.

(See: Inference)

A neural network is a multi-layered algorithmic structure that is foundational to deep learning — and, more broadly, to the recent surge in generative AI tools following the inception of large language models.

The idea of drawing inspiration from the intricately connected pathways of the human brain for data processing algorithms dates back to the 1940s. However, it was the recent development of graphical processing hardware (GPUs), propelled by the gaming industry, that truly unlocked this theory’s potential. These chips proved to be exceptionally effective in training algorithms with exponentially more layers than previously achievable, enabling neural network-based AI systems to accomplish extraordinary feats across diverse domains, including voice recognition, autonomous navigation, and drug discovery.

(See: Large language model [LLM])

Open source refers to software — or increasingly, AI models — where the core code is made publicly accessible for anyone to use, assess, or modify. In the AI realm, Meta’s Llama family of models stands out as a notable example; the widely-known Linux operating system offers a historical parallel. Open-source practices enable researchers, developers, and institutions to build on each other’s work, accelerating innovation while allowing for independent safety assessments that proprietary systems often hinder. Closed source pertains to private code; while users can engage with the product, its underlying mechanics remain hidden, exemplified by OpenAI’s GPT models — a distinction that is becoming a central debate in the AI landscape.

Parallelization entails executing multiple tasks at once rather than sequentially — much like ten employees addressing different components of a project simultaneously rather than a single employee completing everything in order. In AI, parallelization is vital for both training and inference; modern GPUs are specifically engineered to carry out thousands of processes concurrently, which is a key reason they form the backbone of industry hardware. As AI systems advance and models expand, the capability to parallelize operations across multiple chips and machines has emerged as a critical determinant in how swiftly and cost-effectively models can be developed and deployed. Investigations into enhanced parallelization methods have become a dedicated research field.

RAMageddon is the catchy term for a troubling trend affecting the tech sector: a growing shortage of random access memory, or RAM chips, essential for nearly all the devices we use daily. As the AI industry has progressed, major tech corporations and AI research labs — all striving to build the most powerful and efficient AI — have stockpiled significant amounts of RAM for their data centers, leaving little for others. Consequently, this supply shortage is driving up prices for the remaining inventory.

This scarcity also impacts sectors such as gaming (where companies have increased console prices due to difficulties sourcing memory chips), consumer electronics (leading to projected declines in smartphone shipments unseen in over a decade), and enterprise computing (as organizations struggle to acquire adequate RAM for their data centers). The rising costs are expected to persist until the shortage is resolved, although there are currently few indications that such an outcome is imminent.

Similar to AGI, recursive self-improvement (RSI) signifies the capacity of AI to enhance its intelligence while reducing dependence on humans. In a scenario where RSI occurs, AI models initiate self-improvement processes without human input, leading to rapid advancements in their capabilities and autonomy. Some interpretations liken this to a cataclysmic moment reminiscent of the singularity, where AI systems attain independence from external influence. However, RSI also simply describes a fundamental ability — whether an AI can create its own successor — thus simplifying the job for engineers tasked with constructing it. Several recent AI startups are focused on developing recursively self-improving models but often downplay the drastic implications, framing RSI as merely the next frontier in research.

Reinforcement learning is a training approach for AI that allows a system to learn through trial and error while receiving rewards for correct responses — akin to training a pet with treats, but here, the “pet” is a neural network, and the “treat” is a mathematical signal indicating success. Unlike supervised learning, where a model is trained on a predefined set of labeled examples, reinforcement learning permits a model to investigate its environment, take actions, and continually refine its approach based on feedback. This method has proven especially effective for instructing AI to play games, operate robots, and, more recently, improve the reasoning abilities of large language models. Techniques such as reinforcement learning from human feedback (RLHF) have become essential for prominent AI laboratories to enhance model performance in terms of help, accuracy, and security.

Human-machine communication presents inherent challenges — individuals interact using natural language, while AI systems rely on complex algorithmic processes informed by data. Tokens act as the bridge: they represent the basic units of human-AI communication, denoting distinct segments of information processed or generated by an LLM. Tokens are created through a process called tokenization, which breaks down raw text into manageable units that a language model can comprehend, analogous to how a compiler translates human language into binary code for computers. In business settings, tokens also influence costs — many AI companies charge for LLM usage on a per-token basis, meaning greater usage translates to elevated expenses.

To summarize, tokens are the small pieces of text — often consisting of parts of words rather than full words — that AI language models segment language into during pre-processing; they act as rough equivalents to “words” for AI workload comprehension. Throughput refers to the amount of processing accomplishing within a specified timeframe, making token throughput a gauge of how much AI work a system can manage concurrently. High token throughput is a primary objective for AI infrastructure teams, as it dictates how many users a model can support at once and how quickly each receives a response. AI researcher Andrej Karpathy has expressed feelings of anxiety when his AI subscriptions are inactive, echoing experiences from his university days when costly computing resources were not fully utilized — illustrating why maximizing token throughput remains a focal point in the field.

Training machine learning AIs involves supplying data for the model to identify patterns and produce insightful outputs. In essence, this procedure enables the system to respond to data attributes, guiding it toward a desired result — whether that’s recognizing images of cats or composing a haiku on command.

Training can incur substantial costs due to the demand for extensive data inputs, a requirement that has been steadily growing; thus, hybrid strategies, such as pairing a rules-based AI with targeted data, can help mitigate costs without starting entirely from scratch.

[See: Inference]

A method where a pre-trained AI model serves as a basis for developing a new model targeted at a different yet often related task — allowing previously acquired knowledge to be reintroduced.

Transfer learning can bolster efficiency by streamlining model development. This approach proves beneficial when the data available for the new task is sparse. Nonetheless, it is essential to acknowledge that this method has limitations; models leveraging transfer learning to acquire general capabilities will likely necessitate additional data training to excel in their specific niche.

(See: Fine-tuning)

Validation loss is a metric that indicates how effectively an AI model learns during training — with lower values being preferable. Researchers scrutinize it closely as a real-time evaluation metric, utilizing it to discern when to halt training, adjust hyperparameters, or explore potential issues. One major concern it addresses is overfitting, a condition where a model memorizes its training data rather than authentically learning patterns that can be generalized to new scenarios. Think of it as the distinction between a student who thoroughly understands the material and another who merely recalls the previous year’s exam questions — validation loss helps identify which scenario your model might be approaching.

Weights are core components of AI training, as they determine the level of importance (or weight) assigned to various features (or input variables) within the training data — ultimately shaping the output of the AI model.

In other words, weights are numerical parameters that signify what is most pertinent within a dataset for the intended training objective. They fulfill their role by multiplying inputs. Training usually commences with randomly selected weights, but as the process moves forward, these weights adjust as the model aims to produce output that more closely aligns with the target.

For example, an AI model trained to predict real estate prices based on historical data for a given area might incorporate weights for variables like the number of bedrooms and bathrooms, whether a property is detached or semi-detached, its parking availability, and so forth.

Ultimately, the weights associated with these variables reflect their influence on property value based on the specific dataset provided.

This article is frequently updated with the latest insights.