Fundamentals of Artificial Intelligence – Part Five: More Terms
If you've been following our current series of posts on Artificial Intelligence (AI) , you'll already know that in our third installment we published a similar article. That is, an article in which we introduced some essential and important terms (concepts, words, jargon) in this field. But, as you might expect, due to space limitations and the innovative and broad nature of this field of study, there are many more terms to learn. So, in this fifth installment, we'll focus on a few terms that are currently being used frequently by many AI and GIS researchers, developers, communities, and users. These are the following five terms: Quantization, Distillation, Obliteration, Uncensored, and Alignment.
And if you haven't read the previous parts, we'll clarify that in the first part we focused on what AI is and the various types that have already been created and those expected to be created. In the second part , we covered the difference between Symbolic AI and Connectionist AI, along with some important historical milestones in this technological field over the last 100 years. Finally, in the fourth part , we focused on Prompts and Prompt Engineering. So, if this area of AI/AIG is your thing, as well as the Linux universe, then keep reading to continue learning and deepening your knowledge on the subject.
Fundamentals of Artificial Intelligence – Part Four: Prompts
But, before we delve into this "fifth part of the exciting and surprising field of Artificial Intelligence" , we recommend you explore the previous related publication in this series , after you finish reading this one:
Fundamentals of Artificial Intelligence – Part Five: More Terms
Before we begin, it's important to clarify that deploying Large Language Models (LLMs) presents significant challenges in terms of memory consumption (VRAM/RAM), memory bandwidth, processing power (NPU/GPU/CPU), and behavioral alignment. The five terms we'll mention below—Quantization, Distillation, Obliteration, and Alignment/Uncensored—are optimization and modification techniques for the "weights" of AI models. These techniques allow these models to be deployed from server clusters inaccessible to many, down to local environments (such as general-purpose Linux servers or even consumer hardware), modifying both their efficiency and behavior.
In artificial intelligence, especially in artificial neural networks, a weight is a numerical parameter that defines the importance of the connection between two neurons. Each input to a node is multiplied by a weight. Then, the sum of these weighted inputs is passed through an activation function, allowing the model to give more or less importance to certain features. A simple example to understand would be the following: In an image recognition system, a weight can indicate how much a visual feature (such as a border, color, or texture) contributes to the identification of an object. Consequently, a high weight might mean that the feature is very important, a weight close to zero might mean that it has little influence, and a negative weight might mean that it has an influence contrary to the most satisfactory result. AI Glossary
Quantization (or Quantification)
In simpler or less technical terms, the Quantization process (also called Quantization by some researchers or in some bibliographies) refers to the technique focused on reducing the size and optimizing AI models by converting the model weights (its mathematical equations) from a high-precision format to a lower-precision one.
How does this work? Well, models are typically trained in 16-bit formats (FP16). The quantization process then transforms them into 8-bit, 4-bit, or even 2-bit formats (such as INT8, INT4, INT2, or GGUF/EXL2/GPTQ/AWQ packaging schemes). The goal is to allow models to occupy significantly less RAM/VRAM and run faster, making it possible to use these quantized LLM models on consumer hardware (like a home or personal computer or phone) with minimal loss of accuracy when responding to or generating multimedia content. For example , by reducing from FP16 bits to INT4, VRAM consumption can be reduced by 75%, and data transfer from VRAM to the GPU/CPU processing cores (and even to RAM) can be drastically accelerated.
Quantization is a technique used within large language models (LLMs) to convert weights and activation values of high-precision data, typically 32-bit (FP32) or 16-bit (FP16) floating-point numbers, to lower-precision data, such as 8-bit integers (INT8). What is quantization? (according to IBM)
Distillation
In simpler or less technical terms, the Distillation process (also called Knowledge Distillation by some researchers or in some bibliographies) refers to the knowledge transfer process where a small and efficient model (called "learner") is trained to imitate the behavior and results of a much larger and more complex model (called "master").
How does this work? Well, instead of training the student model from scratch with raw data, it's taught using the responses and probability distributions generated by the master model. The goal? To obtain a lighter, faster, and cheaper model to run, but one that retains much of the intelligence, reasoning, and capabilities of the original giant model. For example , if the correct word is "dog," a normal model only knows that "dog" is correct. The distilled "student" model learns from the "master" model that "cat" or "wolf" have a 15% and 5% probability, respectively, inheriting the deep semantic understanding of a 70-billion-parameter model in a network of just 8 billion.
Knowledge distillation techniques aim not only to replicate the results of "master" models but also to emulate their thought processes. In the era of LLMs, knowledge distillation allows for the transfer of abstract qualities such as style, reasoning skills, and alignment with human preferences and values. In other words, by imitating the "master" model, the "student" model learns "dark knowledge" (hidden, forbidden, or difficult-to-decipher information, wisdom, or patterns that escape common understanding). What is distillation, according to IBM?
Obliteration
In simpler or less technical terms, the Obliteration process (also called abliteration, machine delearning, or concept erasure by some researchers or in some bibliographies) refers to an advanced technique for eliminating direct concepts or biases within the model's weights , which is usually applied to eliminate the ability of certain AI models to refuse to respond (bypass the rejection of sensitive content), that is, to diminish or eliminate their factory "alignment".
How does this work? Well, unlike traditional training, developers must identify the exact "vector direction" or neural circuit (high-dimensional latent vector space) within the model that triggers rejection responses (like the typical "I can't help you with that"). Then, through mathematical processes, they erase this geometric vector direction or make it orthogonal (neutralize it) in the model's weights. What's the goal? To completely eliminate a specific behavior (like security censorship) from the model's mathematical foundation, making it nearly impossible for it to refuse to respond. For example , the direction vector representing "how to make an explosive" or "specific copyrighted information" can be isolated by applying variations directly to the weight matrices to elicit that response.
Modern descriptive logic models are optimized for safety and instruction compliance, meaning they are trained to reject harmful requests. This rejection behavior has been shown to be mediated by a specific direction in the model's residual flow. If the model is prevented from representing this direction, it loses its ability to reject requests . Conversely, artificially adding this direction can cause the model to reject even harmless requests. What is abliteration? (according to Hugging Face)
Model Alignment (Block) and Uncensored Models (Unlock)
AI Model Alignment: Explanation, Techniques, and Adjustments
To better understand the term "uncensored model," it's essential to first understand what "AI model alignment" is . Therefore, it's helpful to remember that basic AI models are trained to predict the next token. They then undergo a process called fine-tuning and either human feedback reinforcement learning (HFRL) or direct preference optimization (DPO). This injects "guardrails" to prevent the AI model from responding to instructions deemed immoral, dangerous, or illegal.
Therefore, a precise description of Model Alignment is as follows:
Model alignment is the discipline concerned with ensuring that the outputs of a large language model (LLM) are aligned with human values in a way that benefits users, developers, and society at large. To achieve this, various pre-training and tuning techniques can be employed. What is LLM alignment, according to IBM?
Pre-training techniques
Some pre-training techniques include the following:
- External alignmentIts objective is to adjust a model that has already been pre-trained (and which, in many cases, has already been subject to some degree of adjustment).
- Internal alignmentIt aims to incorporate human values and other safety principles directly into the initial pre-training of the model.
- Mechanistic interpretabilityIt is the practice of investigating how LLMs transform inputs into outputs, either through analyzing the internal operations of an LLM's neural network or by auditing the model's outputs for patterns that produce misaligned responses.
Alignment adjustments
And some adjustments were implemented as follows:
- Refinement (Fine Tuning)This involves techniques for further training a model whose weights have already been updated through previous training. Using the existing knowledge of the base model as a starting point, refinement adapts the model by training it on a smaller, more task-specific dataset.
- Efficient Parameter Tuning (PEFT)This is a method for improving the performance of pre-trained LLM models and neural networks for specific tasks or datasets. By training a small set of parameters while preserving most of the pre-trained LLM model's structure, it saves time and computational resources by increasing efficiency on specific tasks.
- Reinforcement learning from human feedback (RLHF)It is a machine learning technique in which a "reward model" is trained with direct human feedback and then used to optimize the performance of an artificial intelligence agent through reinforcement learning.
- Direct Preference Optimization (DPO): EIt is an efficient method for aligning large language models (LLMs) with human preferencesUnlike traditional reinforcement learning (RLHF), DPO eliminates the need to train a separate reward model or use complex optimization algorithms, simplifying training to a straightforward classification task. It can be implemented either without a reward model or by applying pairwise data and mathematical optimizations.
Uncensored AI models:
Having clarified the previous concept, it can be easily described or explained that an "uncensored" AI model is one in which, through a restrictive alignment process, certain or all imposed barriers or restrictions have been removed , or one to which alignment processes were never applied.
Therefore, these models will obey any prompt the user enters, operating under the pure premise of free software : "The tool must execute the user's command without acting as a moral arbiter." Or, in simpler terms, they are models that have been trained or modified to eliminate the security barriers, ethical filters, and alignment (RLHF) imposed by their original creators.
How is this achieved?
To achieve this, processes such as the following can be implemented:
- Ablation (elimination or deactivation) of the rejection vector (response)By detecting and subtracting this "rejection vector" from the weights during inference, the model loses its ability to refuse to respond.
- Refinement with Orthogonal DatasetsThis involves retraining a base model using datasets from which all responses of the type "Sorry, as an AI language model, I can't..." have been systematically purged. This allows the model to unlearn evasive behavior and prioritize strict obedience to given instructions.
Summary
In summary, we hope that this fifth installment in our ongoing series on Artificial Intelligence (AI) will be a fitting complement to our third publication, further strengthening and facilitating the use and mastery of the knowledge and skills necessary for many Linux users to understand, train, and navigate this modern and innovative technological field. Remember that Artificial Intelligence, along with Robotics, is currently changing the world and the way we live our lives . And if all goes well with it, humanity could experience an era of prosperity and progress unlike anything seen before.
So, it's better to learn about Artificial Intelligence now than to refuse and regret it tomorrow . Therefore, if you'd like to learn something specific about AI, we invite you to mention it in the comments so we can consider it for future posts. And of course, don't forget to continue participating, contributing, and supporting our beloved and valuable Linuxverse.
Finally, remember to visit our " homepage " in Spanish . Or, in any other language (by simply adding two letters to the end of our current URL, for example: ar, de, en, fr, ja, pt, and ru, among many others) to discover more current content.