Ai is reshaping the world and is also creating a whole new vocabulary to describe the process. Now you can hear a lot of talk about big language models, rag, rlhf — and the "opaque recurrence" that just happened last week, which is the reasoning technique in openai's new astra model that bothers ai security researchers. These words are being updated fast enough to make very smart people in the tech world feel a little bit edgy。
The glossary is designed to address this problem: to explain in plain and understandable language the ai terms that you most likely encounter, whether you're building products, investing in them or just trying to keep up with the industry. As the field evolves, we can regularly update it as an evolving document, as it describes the ai system。
Universal artificial intelligence (agi)
Universal artificial intelligence, agi, is a vague concept. But it usually refers to ai, which is more capable of most (or even most) tasks than ordinary human beings. Ceo openai sam ottman described agi as "a moderate human being who can be hired as a colleague". The constitution of openai defines agi as "a highly autonomous system that acts better than human beings in most economically valuable jobs". Google deepmind understands slightly different from these two definitions; the laboratory considers agi to be "an ai at least equal to human capabilities in most cognitive tasks". Are you confused? Don't worry — the same is true of experts at the forefront of ai research。
Ai smart body
The ai smart body is a tool that uses ai technology to represent users in a range of tasks — beyond the capacity of a normal ai chat robot — such as reimbursement, booking of tickets or reservations to restaurants, and even preparing and maintaining codes. But, as we explained earlier, this new field involves many parts of activity, so "ai intelligence" can mean different things to different people. Infrastructure is also still being built to achieve its envisaged capabilities. Its underlying concept, however, implies an autonomous system that may call on multiple ai systems to perform multi-step tasks。
Api peer
You can imagine the api endpoint as the "buttons" behind the software, and the other programs can press these buttons to let it perform the operation. The developers use these interfaces to construct integration - for example, to allow one application to extract data from another application, or to allow ai intelligents to directly control third-party services without having to operate each interface manually and manually. Most smart home-based devices and networking platforms have these hidden buttons, although ordinary users have never seen or interacted with them. With the increasing capacity of ai's intelligent bodies, they are increasingly able to detect and use these endpoints themselves, opening up a powerful, and sometimes unexpected, possibility for automation。
Chain of thought
Faced with simple questions, the human brain hardly needs to think about what's higher than a giraffe and a cat? But in many cases you often need paper and pen to get the right answer, because there are intermediate steps. For example, if a farmer has a chicken and a cow, they have 40 heads and 120 legs, and you may need to write a simple equation to figure out the answer (20 chickens and 20 cows)。
In the ai context, the large-language model's thought chain reasoning refers to the fragmentation of an issue into smaller intermediate steps to improve the quality of the final outcome. This usually takes longer to arrive at an answer, but the answer is more likely to be correct, especially in a logical or programming scenario. The reasoning model has been developed on the basis of the traditional large-language model and has been optimized for thinking on the chain through enhanced learning。
(see: large language model)
Programming smart
This is a more specific concept than the "ai smart body" and refers to processes that can take autonomous steps to achieve the goal. Programming smarts are specialized versions for software development. It does not simply suggest that codes be reviewed and pasted by humans, but is able to prepare, test and debug codes on its own and to process iterative error work that usually takes up a considerable amount of time from developers. These intelligent bodies can operate throughout the coding library and detect loopholes, run tests, push-back repairs, with little manual supervision. It can be imagined as the hiring of a super fast intern who never sleeps and never knows — but, like any intern, humanity still needs to review the results of its work。
Compute
While this is a term of more diverse meanings, arithmetic usually refers to the critical computational ability to enable the ai model to operate. Such processing capacity drives the entire ai industry and empowers it to train and deploy robust models. This term is often abbreviated by hardware that provides computing capabilities — such as gpu, cpu, tpu and other infrastructure that form the cornerstone of the modern ai industry。
In-depth learning
In-depth learning is a subset of self-improvement machine learning, in which ai algorithms are designed to have multilayered artificial neural network structures. This enables them to establish more complex linkages than simpler machine-based learning systems such as linear models or decision trees. The structure of deep learning algorithms draws inspiration from the intensive interconnectivity of human neurons。
In-depth learning ai models are able to identify important features in the data themselves without the need for human engineers to define them. It also supports the learning of algorithms from errors and improves their output through a process of re-engineering. However, in-depth learning systems require a large number of data points to produce good results (usually millions or more). Their training is often longer than simple machine learning algorithms — and therefore development costs are often higher。
(see: neuronet)
Diffusion
Proliferation is the core technology of many ai models that generate art, music and text. Physically inspired, proliferation systems gradually "destruct" the structure of data (e. G. Photographs, songs, etc.) by increasing noise until the data is completely gone. In physics, diffusion is spontaneous and irreversible - dissolved sugar in coffee does not restore to a square sugar form. But the proliferation system in ai is designed to acquire the ability to recover data from noise by learning a "reverse proliferation" process to restore destroyed data。
Distillation
Distillation is a technique for extracting knowledge from large ai models through the "teacher-student" model. The developers send requests to the teacher model and record their output. Sometimes answers are compared to data sets to test their accuracy. These outputs were then used to train students in the model and to learn the behaviour of the teacher approach。
The distillation technology can build smaller, more efficient models based on a large model with minimal distillation losses. This is probably how openai develops gpt-4 turbo。
While distillation techniques are used within all ai companies, there may also be ai companies that use them to catch up with forward models. The distillation of competitors' models usually violates the service terms of ai api and chat assistants。
Fine-tuning
This refers to further training in the ai model to optimize performance in missions or areas that are more specific than their previous training priorities, usually through input of new, specialized (i. E. Mission-oriented) data。
Many ai start-ups construct commercial products using large language models as their starting point, but are fighting to improve the relevance of the target industry or mission by fine-tuning them to complement their specific areas of knowledge and expertise on the basis of an early training cycle。
(see: large language model)
Generate anti-network (can)
Pan, the creation of counter-networks, is a machine learning framework that underpins some of the important advances in generating a generating ai for the generation of authentic data — including, but not limited to, in-depth forgery tools. Gan involves the use of a pair of neural networks, one of which uses its training data to generate outputs and transmits them to another model for evaluation。
The two models are essentially designed to compete with each other. The generator tries to confuse its output through the dichotomy, which is dedicated to identifying artificially generated data. This structured confrontation can optimize ai output to make it more real without additional manual intervention. However, gan is better suited to a narrower application (e. G. Producing a vehement photo or video) than a generic ai。
Hallucination
The illusion is the preferred term used by the ai industry to describe model-made information - that is, to generate incorrect information. Clearly, this is a big problem for ai quality。
An illusion can produce misleading, generated ai outputs, and may even lead to real-life risks — potentially dangerous consequences (thinking of a health inquiry to return to harmful medical advice)。
The issue of ai's information generation is considered to be the result of gaps in training data. The problem of hallucinations is driving the industry towards increasingly specialized and/or vertical ai models — i. E., specific areas requiring narrower expertise ai — as a way to reduce knowledge gaps and the risk of false information。
Inference
The reasoning is the process of running the ai model. It allows a model to project or draw conclusions based on previously seen data. It needs to be clear that there is no reasoning without training; models must learn from a set of data before they can be effectively extrapolated from these training data。
Many types of hardware implement reasoning, ranging from smartphone processors to powerful gpus to custom-designed ai accelerators. However, not all hardware can run models equally well. Very large models, if projected on laptops, may take a very long time, while much faster on cloud servers with high-end ai chips。
(see: training)
Large language model
The big language model is chatgpt, claude, google gemini, mIt's a popular ai model used by helpers like ai llama, microsoft copilot or mistral le chat. When you talk to ai's assistant, you're actually interacting with a large-language model that handles your requests directly, or through various tools available, such as web browsing, code interpreter。
Large-language models are a deep neural network consisting of billions of numerical parameters (or weights, see below) that learn the relationship between words and phrases and create a language expression, a multi-dimensional word map。
These models were created by coding models found in billions of books, articles and transcripts. When you enter a hint into a large-language model, the model produces the most likely pattern to match the hint。
(see: neuronet)
Memory cache
The memory cache refers to an important process of increasing the efficiency of reasoning (the process of ai generating responses to user queries). In essence, it is an optimized technology aimed at increasing the efficiency of reasoning. Ai is clearly driven by high-intensity mathematical calculations, and each time they are performed, they consume more computing power. The cache is designed to reduce the number of calculations that the model may need to run by preserving specific calculations for future user queries and operations. There are various types of memory caches, one of which is known as kv (key) caches. The kv cache is applied to the transformer-based model to increase efficiency and speed of results by reducing the time required to generate user answers (and algorithmic workload)。
(see: reasoning)
Model context protocol (mcp)
Model context protocol, mcp, is an open standard that allows ai models to connect external tools and data - your files, databases, or applications such as slack, googledrive - without having to develop a customized connector for each pair. You can imagine it as an ai usb-c interface. The introduction of mcp by anthropic in 2024 and its subsequent transfer to the linux foundation has since been adopted by openai, google and microsoft as one of the fastest-growing standards in ai in recent years。
Mixed expert model (moe)
The mixed expert model is a model structure that divides the nervous network into many smaller specialized sub-networks, namely "experts", and activates only a small proportion of them for any given task. Unlike treating each request through the entire model (as is the case with the person who calls the entire office for each issue), the moe model incorporates a "router" that allows for the precise selection of suitable experts for specific tasks. This has made it possible to build a large but fast-moving, relatively inexpensive model, as only part of the network works at any given time. The mixtral model of mistral ai is a well-known example; it is widely accepted in industry that openai's newer gpt model also uses some version of this approach, although the company has never formally confirmed this。
(see: neuronet, deep learning)
Neural network
Neurological networks are the multilayered algorithm structure that underpins in-depth learning - and more broadly, the basis for the whole generation of ai tools following the emergence of large language models。
Although the idea of drawing inspiration from a brain-intensive neural circuit as a design structure for data processing algorithms dates back to the 1940s, the real release of this theoretical power is the graphic processing hardware (gpu) that has emerged in recent years through the video game industry. These chips are well suited to a number of training layers well above the algorithms of the past - to enable the neurological network-based ai system to perform far better in many areas such as voice recognition, autonomous navigation and drug discovery。
(see: large language model)
Neuralese
One of the worst scenarios in the scenario, where the model is based solely on its internal values for reasoning, rather than on human-readable language, leads to a process of thinking that is completely black. There are no published models to this extent — openai indicates that its astra model (which was released in september 2026 and was notable for its pioneering use of “irtransparent cycles” of reasoning) still retains the readability of the chain of thought and refutes its analogy with neurolinguistics. However, security researchers point out that astra's use of the non-transparent cycle (as defined below) is a substantive first step in this direction, and this is why the term has risen sharply following astra's release of relevant reports。
Opaque loop
Transparent cycles mean that the ai model recycles the same query in its inner layers, rather than progressively deduced in plain and understandable language. This approach is more efficient — smaller models can develop capabilities beyond their size with less calculus — but leaves much less readable traces than normal thinking chains (i. E. The real-time discourse you see when you turn to chat robots for help). This is a source of concern for security researchers, as these logs are key tools for detecting unusual behaviour — and this technology may make such monitoring much more difficult。
(see: the thought chain)
Open source
Open source refers to software — or, increasingly, to ai models — which is made publicly available to everyone for use, inspection or modification. In the ai area, mThe llama series model of eta is a prominent example; linux is a well-known historical precedent in the field of operating systems. The open source approach allows researchers, developers and companies around the world to build on each other's work, accelerate progress and support independent security audits that are not readily available in closed systems. Closed sources mean that the code is private -- you can use the product, but you can't see its internal working principles, as is the case with openai's gpt model -- and this distinction has become one of the most representative arguments in the ai industry。
Parallelization
Parallelity means doing multiple things at the same time, not one thing at the same time — like having 10 employees handle different parts of the project at the same time — rather than having an employee do all the work in the same order. In the ai area, parallelization is the basis for training and reasoning: modern gpu designs are designed to perform thousands of calculations at the same time, which is why they have become the backbone of hardware in the industry. With the increasing complexity of the ai system and the growing size of the model, the ability to achieve parallel processing between many chips and many machines has become one of the most important factors determining the speed and cost-effectiveness of model construction and deployment. Research on better parallel strategies has now become an independent area of research。
Memory doom (ramageddon)
Ramageddon is an interesting new name for a less interesting trend that is spreading in the technology industry: the shortage of random access memory (ram chips), which is driving almost all of our everyday products. As the ai industry has flourished, major technology companies and ai laboratories are competing for the strongest and most efficient ai, resulting in significant procurement of rams to support their data centres, leaving few others behind. This supply bottleneck means that the remaining stocks are becoming increasingly expensive。
The impact has been felt in the game industry (large companies have had to raise the price of their hosts because their storage chips have become more difficult to access), in the consumer electronics industry (a shortage of storage could lead to the largest decline in smartphone deliveries in more than a decade) and in the general area of business computing (because these companies also have difficulty obtaining enough rams for their own data centres). Price increases are expected to stop only after this painful shortage has ended, but unfortunately there is little indication that this will happen soon。
Cycle depth (recurrent depth)
This is a more "technical" name, referring to the same bottom method as the non-transparent cycle (repeated search in the model layer rather than sequential reasoning in language). The media can use these two terms almost interchangeably, which is why we keep them here, but "cycle depth" is a engineering term, and "untransparent loop" is a way of highlighting security concerns。
(see: opacity cycle, thought chain)
Recursive self-improvement
As with agi, retrogressive self-improvement (rsi) is a threshold from which to measure how smart ai can be and how much it can escape human dependence. In the rsi scene, the ai model began to improve itself without human intervention, thereby bringing about a huge leap in capacity and autonomy. In some cases, this would be a catastrophic moment similar to the one in which the ai model became immune from external intervention. But rsi also describes a basic capability: can the ai model design its successor? It makes it easier for engineers to try to build it. A number of ai start-ups have recently started to build models with the ability to adapt themselves, but most have moved away from that end-of-life reading and presented rsi as the next frontier of research。
Enhanced learning
Enhanced learning is a training-ai approach, and the system learns by constantly trying and getting rewards for the right answers -- just like training your beloved pet with snacks, except that pets here are a neural network, and snacks are a mathematical signal of success. Unlike monitoring learning on training models on fixed illustrative data sets, enhanced learning allows models to explore the environment, take action and keep their behaviour updated based on feedback received. This approach has proved particularly effective in training ai to play games and control robots and has recently been used to enhance the reasoning of large language models. Technology such as intensive learning (rlhf) from human feedback has now become the core tool for fine-tuning the top ai laboratory model to make it more helpful, accurate and secure。
Token
In terms of human communication, there are some obvious challenges — human communication in human languages, while the ai process carries out its tasks through complex data-based algorithms. Token builds this bridge: they are the basic components of human communication with ai and represent discrete data fragments processed or generated by large linguistic models. They were created through a process known as word severing, which decomposes the original text into a small unit that can be digested by a language model, similar to the binary code that the compiler translates into a computer understandable human language. In the business scene, token also determines costs - most ai companies charge large language models based on token usage, which means that the more businesses use and the more they pay。
Token throughput
Once again, token is a small piece of text that has been decomposed by the ai language model before processing the language -- - it's usually part of a word, not the whole word; in terms of understanding the ai load, it's roughly the word. Insistence refers to the amount that can be handled within a given period of time, so token throughput is essentially an indicator of how much ai workload a system can handle at the same time. High token throughput is an important objective of the ai infrastructure team as it determines how many users a model can serve at the same time and how quickly each user can receive a response. Andre karpasi, an ai researcher, described how he was anxious when his ai subscription service was idle — as he felt when he was a graduate student, when he saw expensive computer hardware not being fully utilized — this sentiment is precisely the reason why maximizing token consumption has become a sort of obsession in this area。
Training
The development of machine learning ai involves a process known as training. In short, this refers to inputting data into models so that models can learn from models and generate useful outputs. In essence, it's the system's process of responding to data characteristics, enabling it to adjust its output towards the goal it is pursuing. - either to identify images of cats or to generate a sentence as needed。
Training may be costly because it requires a large amount of input, while the amount of data required has been on the upward trend — that is why hybrid methods (e. G. Fine-tuning of rule-based ai with targeted data) help to contain costs in less than zero。
(see: reasoning)
Migration learning
A technology that uses previously trained ai models as a starting point for the development of new models to address different but often relevant tasks - so that the knowledge gained from previous training cycles can be re-used。
Migration learning can improve efficiency and cost savings by shortening model development processes. This approach is also useful when data on target missions are relatively limited. It should be noted, however, that this approach has limitations. Models that rely on migration to learn to acquire generalization capacity often require training on additional data to perform well in their areas of focus。
(see: fine-tuning)
Validation loss
Validation of losses is a number that is used to tell you what the ai model has learned during training - the lower the value, the better. This indicator is closely tracked by researchers as a real-time report card to determine when training will be discontinued, when super-parameters adjusted or whether there is a need to investigate potential problems. A key issue that helps to mark is over-compatibility, i. E. Model memory of hard back training data, without real learning of models that can be generalized to new situations. You can imagine the difference between a student who really understands the contents of the teaching materials and a student who just remembered last year's exam. - validation of the loss helps to reveal what your model is becoming。
Weights
The weights are at the core of ai training, as they determine the importance (or weight) given to different features (or input variables) in the data used in the training system - thus shaping the output of the ai model。
In other words, weight is a numerical parameter used to define which elements of the data concentration are most important in given training assignments. They perform their functions by multiplying the input. Model training usually starts with randomly assigned weights, but as the training process advances, the weights are constantly adjusted to achieve a closer output。
For example, an ai model for predicting housing prices in a given area, if trained on the basis of historical real estate data in the area, may include weights on the number of bedrooms and bathrooms, whether the property is independent or semi-independent, availability of parking spaces, garages, etc。
Ultimately, the weight given to each input by the model reflects the extent to which these factors are based on the value of the property in the given data set。
This paper will be updated regularly as new information emerges。
Q&a
Q1: what is a non-transparent cycle
A: opaque cycle means that the ai model recycles the same query in its inner layers, rather than gradually deduced in plain and understandable language. This approach is more efficient, but leaves much less readable traces than normal lines of thought, making it more difficult for security researchers to detect unusual behaviour。
Q2: what's the difference between a big language model and an ai smart body
A: the big language model is the core model used behind the chatgpt, claude, and others, mainly dealing with and generating languages. The ai intelligent body, which is a tool for using ai technology to represent users on a range of tasks, such as reimbursement, booking, and even writing codes, goes beyond ordinary chat robots。
Q3: what is ramageddon? Why does it affect ordinary consumers
A: ramageddon refers to the trend of a shortage of random access memory chips, as ai procured ram for data centre construction. This has led to an increase in the price of the game host, a possible maximum reduction in the volume of smartphone deliveries for more than a decade, and an increase in the cost of computing by businesses in general, which is not expected to stop until the shortfall is alleviated。
