Artificial intelligence is rewriting the world, and concurrently inventing an entire latest language to explain the way it’s doing it. Sit in on any product meeting, pitch, or panel nowadays, and also you’ll hear people toss around LLMs, RAG, RLHF, and a dozen other terms that could make even very smart people within the tech world feel just a little insecure. This glossary is our try and fix that: pain-English definitions of the AI terms you’re more than likely to really run into, whether you’re constructing with these things, investing in it, or simply attempting to sustain by reading TechCrunch or listening to related podcasts. We update it commonly as the sector evolves, so consider it a living document, very like the AI systems it describes.
Artificial general intelligence, or AGI, is a nebulous term. But it surely generally refers to AI that’s more capable than the common human at many, if not most, tasks. OpenAI CEO Sam Altman once described AGI because the “equivalent of a median human that you would hire as a co-worker.” Meanwhile, OpenAI’s charter defines AGI as “highly autonomous systems that outperform humans at most economically beneficial work.” Google DeepMind’s understanding differs barely from these two definitions; the lab views AGI as “AI that’s at the least as capable as humans at most cognitive tasks.” Confused? Not to fret — so are experts on the forefront of AI research.
An AI agent refers to a tool that uses AI technologies to perform a series of tasks in your behalf — beyond what a more basic AI chatbot could do — similar to filing expenses, booking tickets or a table at a restaurant, and even writing and maintaining code. Nonetheless, as we’ve explained before, there are plenty of moving pieces on this emergent space, so “AI agent” might mean various things to different people. Infrastructure can be still being built out to deliver on its envisaged capabilities. But the essential concept implies an autonomous system that will draw on multiple AI systems to perform multistep tasks.
Consider API endpoints as “buttons” on the back of a chunk of software that other programs can press to make it do things. Developers use these interfaces to construct integrations — for instance, allowing one application to tug data from one other, or enabling an AI agent to regulate third-party services directly and not using a human manually operating each interface. Most smart home devices and connected platforms have these hidden buttons available, even when odd users never see or interact with them. As AI agents grow more capable, they’re increasingly capable of find and use these endpoints on their very own, opening up powerful — and sometimes unexpected — possibilities for automation.
Given an easy query, a human brain can answer without even considering an excessive amount of about it — things like “which animal is taller, a giraffe or a cat?” But in lots of cases, you regularly need a pen and paper to give you the proper answer because there are intermediary steps. As an illustration, if a farmer has chickens and cows, and together they’ve 40 heads and 120 legs, you may need to jot down down an easy equation to give you the reply (20 chickens and 20 cows).
In an AI context, chain-of-thought reasoning for big language models means breaking down an issue into smaller, intermediate steps to enhance the standard of the top result. It often takes longer to get a solution, but the reply is more prone to be correct, especially in a logic or coding context. Reasoning models are developed from traditional large language models and optimized for chain-of-thought considering because of reinforcement learning.
(See: Large language model)
It is a more specific concept that an “AI agent,” which implies a program that may take actions by itself, step-by-step, to finish a goal. A coding agent is a specialized version applied to software development. Somewhat than simply suggesting code for a human to review and paste in, a coding agent can write, test, and debug code autonomously, handling the type of iterative, trial-and-error work that typically consumes a developer’s day. These agents can operate across entire codebases, spotting bugs, running tests, and pushing fixes with minimal human oversight. Consider it like hiring a really fast intern who never sleeps and never loses focus — though, as with all intern, a human still must review the work.
Although somewhat of a multivalent term, compute generally refers back to the vital computational power that permits AI models to operate. The sort of processing fuels the AI industry, giving it the power to coach and deploy its powerful models. The term is commonly a shorthand for the sorts of hardware that gives the computational power — things like GPUs, CPUs, TPUs, and other types of infrastructure that form the bedrock of the fashionable AI industry.
A subset of self-improving machine learning wherein AI algorithms are designed with a multi-layered, artificial neural network (ANN) structure. This permits them to make more complex correlations in comparison with simpler machine learning-based systems, similar to linear models or decision trees. The structure of deep learning algorithms draws inspiration from the interconnected pathways of neurons within the human brain.
Deep learning AI models are capable of discover necessary characteristics in data themselves, relatively than requiring human engineers to define these features. The structure also supports algorithms that may learn from errors and, through a technique of repetition and adjustment, improve their very own outputs. Nonetheless, deep learning systems require quite a lot of data points to yield good results (hundreds of thousands or more). Additionally they typically take longer to coach in comparison with simpler machine learning algorithms — so development costs are likely to be higher.
(See: Neural network)
Diffusion is the tech at the guts of many art-, music-, and text-generating AI models. Inspired by physics, diffusion systems slowly “destroy” the structure of information — for instance, photos, songs, and so forth — by adding noise until there’s nothing left. In physics, diffusion is spontaneous and irreversible — sugar diffused in coffee can’t be restored to cube form. But diffusion systems in AI aim to learn a type of “reverse diffusion” process to revive the destroyed data, gaining the power to recuperate the info from noise.
Distillation is a way used to extract knowledge from a big AI model with a ‘teacher-student’ model. Developers send requests to a teacher model and record the outputs. Answers are sometimes compared with a dataset to see how accurate they’re. These outputs are then used to coach the coed model, which is trained to approximate the teacher’s behavior.
Distillation might be used to create a smaller, more efficient model based on a bigger model with a minimal distillation loss. This is probably going how OpenAI developed GPT-4 Turbo, a faster version of GPT-4.
While all AI firms use distillation internally, it could have also been utilized by some AI firms to meet up with frontier models. Distillation from a competitor often violates the terms of service of AI API and chat assistants.
This refers back to the further training of an AI model to optimize performance for a more specific task or area than was previously a focus of its training — typically by feeding in latest, specialized (i.e., task-oriented) data.
Many AI startups are taking large language models as a place to begin to construct a industrial product but are vying to amp up utility for a goal sector or task by supplementing earlier training cycles with fine-tuning based on their very own domain-specific knowledge and expertise.
(See: Large language model [LLM])
A GAN, or Generative Adversarial Network, is a variety of machine learning framework that underpins some necessary developments in generative AI in terms of producing realistic data — including (but not only) deepfake tools. GANs involve the usage of a pair of neural networks, one in every of which pulls on its training data to generate an output that’s passed to the opposite model to judge.
The 2 models are essentially programmed to attempt to outdo one another. The generator is attempting to get its output past the discriminator, while the discriminator is working to identify artificially generated data. This structured contest can optimize AI outputs to be more realistic without the necessity for extra human intervention. Though GANs work best for narrower applications (similar to producing realistic photos or videos), relatively than general purpose AI.
Hallucination is the AI industry’s preferred term for AI models making stuff up — literally generating information that is wrong. Obviously, it’s an enormous problem for AI quality.
Hallucinations produce GenAI outputs that might be misleading and will even result in real-life risks — with potentially dangerous consequences (consider a health query that returns harmful medical advice).
The issue of AIs fabricating information is believed to arise as a consequence of gaps in training data. Hallucinations are contributing to a push toward increasingly specialized and/or vertical AI models — i.e. domain-specific AIs that require narrower expertise — as a approach to reduce the likelihood of information gaps and shrink disinformation risks.
Inference is the technique of running an AI model. It’s setting a model loose to make predictions or draw conclusions from previously seen data. To be clear, inference can’t occur without training; a model must learn patterns in a set of information before it may possibly effectively extrapolate from this training data.
Many kinds of hardware can perform inference, starting from smartphone processors to beefy GPUs to custom-designed AI accelerators. But not all of them can run models equally well. Very large models would take ages to make predictions on, say, a laptop versus a cloud server with high-end AI chips.
[See: Training]
Large language models, or LLMs, are the AI models utilized by popular AI assistants, similar to ChatGPT, Claude, Google’s Gemini, Meta’s AI Llama, Microsoft Copilot, or Mistral’s Le Chat. If you chat with an AI assistant, you interact with a big language model that processes your request directly or with the assistance of various available tools, similar to web browsing or code interpreters.
LLMs are deep neural networks fabricated from billions of numerical parameters (or weights, see below) that learn the relationships between words and phrases and create a representation of language, a type of multidimensional map of words.
These models are created from encoding the patterns they find in billions of books, articles, and transcripts. If you prompt an LLM, the model generates the more than likely pattern that matches the prompt.
(See: Neural network)
Memory cache refers to a crucial process that enhances inference (which is the method by which AI works to generate a response to a user’s query). In essence, caching is an optimization technique, designed to make inference more efficient. AI is clearly driven by high-octane mathematical calculations and each time those calculations are made, they use up more power. Caching is designed to chop down on the variety of calculations a model might need to run by saving particular calculations for future user queries and operations. There are different sorts of memory caching, although one in every of the more well-known is KV (or key value) caching. KV caching works in transformer-based models, and increases efficiency, driving faster results by reducing the period of time (and algorithmic labor) it takes to generate answers to user questions.
(See: Inference)
Model Context Protocol, or MCP, is an open standard that lets AI models hook up with outside tools and data — your files, databases, or apps like Slack and Google Drive — and not using a developer constructing a custom connector for each single pairing. Consider it as a USB-C port for AI. Anthropic introduced MCP in 2024 and later handed it over to the Linux Foundation, and it’s since been adopted by OpenAI, Google, and Microsoft, making it one in every of the fastest-spreading standards in recent AI history.
Mixture of Experts is a model architecture that splits a neural network into many smaller specialized sub-networks, or “experts,” and only prompts a handful of them for any given task. Somewhat than routing every request through the complete model — like calling in your whole office for each query — an MoE model has a built-in “router” that picks just the proper specialists for the job. This makes it possible to construct enormous models that stay relatively fast and low-cost to run, since only a fraction of the network is doing work at anyone time. Mistral AI’s Mixtral model is a widely known example; OpenAI’s newer GPT models are also widely believed to make use of some version of this approach, though the corporate has never officially confirmed it.
(See: Neural network, Deep learning)
A neural network refers back to the multi-layered algorithmic structure that underpins deep learning — and, more broadly, the entire boom in generative AI tools following the emergence of enormous language models.
Although the concept of taking inspiration from the densely interconnected pathways of the human brain as a design structure for data processing algorithms dates all the best way back to the Forties, it was the far more recent rise of graphical processing hardware (GPUs) — via the video game industry — that basically unlocked the ability of this theory. These chips proved well suited to training algorithms with many more layers than was possible in earlier epochs — enabling neural network-based AI systems to realize much better performance across many 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 underlying code is made publicly available for anyone to make use of, inspect, or modify. Within the AI world, Meta’s Llama family of models is a outstanding example; Linux is the famous historical parallel in operating systems. Open source approaches allow researchers, developers, and corporations around the globe to construct on top of each other’s work, accelerating progress and enabling independent safety audits that closed systems cannot easily provide. Closed source means the code is private — you need to use the product but not see how it really works, as is the case with OpenAI’s GPT models — a distinction that has turn into one in every of the defining debates within the AI industry.
Parallelization means doing many things at the identical time as a substitute of 1 after one other — like having 10 employees working on different parts of a project at the identical time as a substitute of 1 worker doing every thing sequentially. In AI, parallelization is key to each training and inference: modern GPUs are specifically designed to perform hundreds of calculations in parallel, which is an enormous reason why they became the hardware backbone of the industry. As AI systems grow more complex and models grow larger, the power to parallelize work across many chips and lots of machines has turn into one of the vital necessary aspects in determining how quickly and cost-effectively models might be built and deployed. Research into higher parallelization strategies is now a field of study in its own right.
RAMageddon is the fun latest term for a not-so-fun trend that’s sweeping the tech industry: an ever-increasing shortage of random access memory, or RAM chips, which power just about all of the tech products we use in our each day lives. Because the AI industry has blossomed, the largest tech firms and AI labs — all vying to have essentially the most powerful and efficient AI — are buying a lot RAM to power their data centers that there’s not much left for the remaining of us. And that offer bottleneck signifies that what’s left is getting an increasing number of expensive.
That features industries like gaming (where major firms have needed to raise prices on consoles since it’s harder to seek out memory chips for his or her devices), consumer electronics (where memory shortage could cause the largest dip in smartphone shipments in greater than a decade), and general enterprise computing (because those firms can’t get enough RAM for their very own data centers). The surge in prices is barely expected to stop after the dreaded shortage ends but, unfortunately, there’s probably not much of an indication that’s going to occur anytime soon.
Like AGI, recursive self-improvement is a threshhold for the way smart AI can get, and the way little it could depend on humans. Within the RSI scenario, AI models start improving themselves without human intervention, resulting in an enormous acceleration in capabilities and autonomy. In some tellings, this might be a cataclysmic moment akin to the singularity, a moment when AI models turn into resistant to outside intervention. But RSI also describes a basic capability — can an AI model design its own successor? — which makes it much easier for engineers to try to construct it. A variety of recent AI startups have got down to construct recursively self-improving models, but most of them dismiss the apocalyptic implications, presenting RSI as simply the following frontier for research.
Reinforcement learning is a way of coaching AI where a system learns by trying things and receiving rewards for proper answers — like training your loved one pet with treats, except the “pet” on this scenario is a neural network and the “treat” is a mathematical signal indicating success. Unlike supervised learning, where a model is trained on a hard and fast dataset of labeled examples, reinforcement learning lets a model explore its environment, take actions, and repeatedly update its behavior based on the feedback it receives. This approach has proven especially powerful for training AI to play games, control robots, and, more recently, sharpen the reasoning ability of enormous language models. Techniques like reinforcement learning from human feedback, or RLHF, are actually central to how leading AI labs fine-tune their models to be more helpful, accurate, and protected.
In relation to human-machine communication, there are some obvious challenges — people communicate using human language, while AI programs execute tasks through complex algorithmic processes informed by data. Tokens bridge that gap: they’re the essential constructing blocks of human-AI communication, representing discrete segments of information which were processed or produced by an LLM. They’re created through a process called tokenization, which breaks down raw text into bite-sized units a language model can digest, much like how a compiler translates human language into binary code a pc can understand. In enterprise settings, tokens also determine cost — most AI firms charge for LLM usage on a per-token basis, meaning the more a business uses, the more it pays.
So again, tokens are the small chunks of text — often parts of words relatively than whole ones — that AI language models break language into before processing it; they’re roughly analogous to “words” for the needs of understanding AI workloads. Throughput refers to how much might be processed in a given time frame, so token throughput is basically a measure of how much AI work a system can handle without delay. High token throughput is a key goal for AI infrastructure teams, because it determines what number of users a model can serve concurrently and the way quickly each of them receives a response. AI researcher Andrej Karpathy has described feeling anxious when his AI subscriptions sit idle — echoing the sensation he had as a grad student when expensive computer hardware wasn’t being fully utilized — a sentiment that captures why maximizing token throughput has turn into something of an obsession in the sector.
Developing machine learning AIs involves a process often called training. In easy terms, this refers to data being fed in so that the model can learn from patterns and generate useful outputs. Essentially, it’s the technique of the system responding to characteristics in the info that allows it to adapt outputs toward a sought-for goal — whether that’s identifying images of cats or producing a haiku on demand.
Training might be expensive since it requires lots of inputs, and the volumes required have been trending upwards — which is why hybrid approaches, similar to fine-tuning a rules-based AI with targeted data, may help manage costs without starting entirely from scratch.
[See: Inference]
A way where a previously trained AI model is used as the place to begin for developing a brand new model for a distinct but typically related task — allowing knowledge gained in previous training cycles to be reapplied.
Transfer learning can drive efficiency savings by shortcutting model development. It could possibly even be useful when data for the duty that the model is being developed for is somewhat limited. But it surely’s necessary to notice that the approach has limitations. Models that depend on transfer learning to achieve generalized capabilities will likely require training on additional data as a way to perform well of their domain of focus
(See: Nice tuning)
Validation loss is a number that tells you the way well an AI model is learning during training — and lower is best. Researchers track it closely as a type of real-time report card, using it to make your mind up when to stop training, when to regulate hyperparameters, or whether to analyze a possible problem. One in every of the important thing concerns it helps flag is overfitting, a condition wherein a model memorizes its training data relatively than truly learning patterns it may possibly generalize to latest situations. Consider it because the difference between a student who genuinely understands the fabric and one who simply memorized last 12 months’s exam — validation loss helps reveal which one your model is becoming.
Weights are core to AI training, as they determine how much importance (or weight) is given to different features (or input variables) in the info used for training the system — thereby shaping the AI model’s output.
Put one other way, weights are numerical parameters that outline what’s most salient in a dataset for the given training task. They achieve their function by applying multiplication to inputs. Model training typically begins with weights which are randomly assigned, but as the method unfolds, the weights adjust because the model seeks to reach at an output that more closely matches the goal.
For instance, an AI model for predicting housing prices that’s trained on historical real estate data for a goal location could include weights for features similar to the variety of bedrooms and bathrooms, whether a property is detached or semi-detached, whether it has parking, a garage, and so forth.
Ultimately, the weights the model attaches to every of those inputs reflect how much they influence the worth of a property, based on the given dataset.
This text is updated commonly with latest information.
If you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.