← Back to AI Concepts and Infrastructure

MODULE 01 · COMPREHENSIVE KNOWLEDGE DIRECTORY

AI Concepts & Sampling Parameters Glossary

89 structured terms, sampling parameters, architectural concepts, and hardware formulas extracted directly from core research spreadsheets.

QUICK REFERENCE GUIDE

Sampling Hyperparameter Cheat Sheet

Recommended parameter configurations for common LLM deployment tasks.

💻

Code & Mathematics

Temp: 0.0 – 0.2 Top-P: 0.95 Min-P: 0.05

Deterministic output minimizing syntax errors and hallucinations.

✍️

Creative & Storytelling

Temp: 0.7 – 0.9 Top-P: 0.90 Top-K: 40

High vocabulary diversity while filtering long-tail nonsensical tokens.

🤖

Agent Tool Calling

Temp: 0.0 Presence: 0.0 Freq: 0.0

Strict JSON schema compliance for deterministic function arguments.

MATHEMATICAL FORMULA SIMULATOR

LLM VRAM Footprint & KV Cache Calculator

Estimate GPU VRAM memory requirements for inference based on model parameter count, precision, context length, and batch size.

MODEL WEIGHT MEMORY 70.0 GB Base GPU memory for loaded weights
KV CACHE MEMORY 25.6 GB Attention context state VRAM
MINIMUM VRAM REQUIRED 114.7 GB Requires 2x H100 80GB GPUs
Core Concepts

Token

The smallest unit of text that an AI model processes. Roughly ¾ of a word on average in English. Analogy: Like individual Lego bricks that the model uses to build meaning, rather than whole sentences.

TECHNICAL DETAILS & FORMULAS

Tokenization splits text using schemes like Byte-Pair Encoding (BPE) or SentencePiece. Models predict the next token autoregressively. 1 token ≈ 4 characters or ~0.75 words in English.

TRADE-OFFS & IMPLICATIONS

Context windows are measured in tokens. Longer context = more expensive inference and higher memory use (KV cache). Important for cost and capability planning in agents and RAG systems.

Core Concepts

Parameter / Weights

The billions or trillions of adjustable numbers inside a model that encode its 'knowledge' after training. Analogy: Think of parameters as the dials and knobs on an enormous mixing board that the model learned to set during training.

TECHNICAL DETAILS & FORMULAS

Also called 'weights'. A 70B parameter model has ~70 billion numbers. Stored in formats like FP16, BF16, or quantized INT4/INT8. The file size is roughly proportional to parameter count × bytes per parameter.

TRADE-OFFS & IMPLICATIONS

More parameters generally = more capability (with diminishing returns). Quantization reduces parameter precision to shrink model size and speed up inference with acceptable quality loss.

Core Concepts

Context Window

The maximum amount of text (measured in tokens) the model can 'see' at once when generating a response. Analogy: Like short-term memory or the size of a whiteboard the model can look at while thinking.

TECHNICAL DETAILS & FORMULAS

Includes both input prompt + generated output. KV Cache grows linearly with context length. Modern frontier models support 128K–1M+ tokens. Longer context increases memory usage and latency.

TRADE-OFFS & IMPLICATIONS

Critical for RAG, agent memory, and long-document tasks. Trade-off: longer context = higher cost and slower responses. KV Cache optimization is a major area of inference engineering.

Core Concepts

KV Cache (Key-Value Cache)

A memory optimization during inference that stores previous computations so the model doesn't recalculate them for every new token. Analogy: Like keeping a running notebook of what you've already figured out instead of re-deriving everything from scratch each time.

TECHNICAL DETAILS & FORMULAS

During autoregressive generation, attention keys and values from previous tokens are cached. Memory usage scales with batch size × context length × layers × hidden size. Major bottleneck for long-context serving.

TRADE-OFFS & IMPLICATIONS

Enables much faster token generation after the first token. Memory-hungry at long context lengths. Quantization and paging techniques (e.g., vLLM PagedAttention) help manage this.

Core Concepts

Inference

The process of running a trained model to generate outputs (answers, code, images, etc.). This is what happens when you chat with an AI. Analogy: Training is like studying for an exam; inference is taking the exam and answering questions.

TECHNICAL DETAILS & FORMULAS

Usually autoregressive for text: predict one token at a time, append it, repeat. Involves matrix multiplications (mostly), attention, sampling, and KV cache management. Much less compute-intensive than training but still expensive at scale.

TRADE-OFFS & IMPLICATIONS

The dominant cost for most AI products after initial training. Quantization, distillation, and efficient serving frameworks (vLLM, TensorRT-LLM) dramatically reduce inference costs.

Training

Training (Frontier Models)

The massive compute-intensive process of adjusting a model's parameters on enormous datasets so it learns patterns, reasoning, and capabilities. Analogy: Sending a student through years of intense, specialized education with constant testing and adjustment.

TECHNICAL DETAILS & FORMULAS

Uses backpropagation + optimizer (AdamW etc.) across thousands of accelerators. Data parallelism, tensor parallelism, pipeline parallelism. Loss is computed on next-token prediction (or multimodal objectives).

TRADE-OFFS & IMPLICATIONS

Extremely expensive (hundreds of millions to billions USD). Only a handful of organizations can train frontier models. Open-weight releases (Llama, Qwen, Gemma) let others benefit without the training cost.

Training

Fine-tuning

Taking a pre-trained model and further training it on a smaller, specific dataset to improve performance on particular tasks or domains. Analogy: A generalist doctor completing a residency in cardiology to become a heart specialist.

TECHNICAL DETAILS & FORMULAS

Usually supervised fine-tuning (SFT) on curated instruction/response pairs, followed by preference tuning (RLHF, DPO, etc.). Much cheaper than pre-training. LoRA and QLoRA enable efficient fine-tuning of large models on consumer hardware.

TRADE-OFFS & IMPLICATIONS

The main way organizations create specialized models without training from scratch. Critical for domain adaptation, style control, and safety/alignment. LoRA makes it accessible on modest hardware.

Optimization

Quantization

Reducing the numerical precision of a model's parameters (e.g., from 16-bit to 4-bit) to make the model smaller, faster, and cheaper to run with minimal quality loss. Analogy: Compressing a high-resolution photo into a smaller file size that still looks good enough for most uses.

TECHNICAL DETAILS & FORMULAS

Common formats: INT8, INT4, FP8, FP4, GGUF (llama.cpp), MLX-quantized. Techniques include GPTQ, AWQ, SmoothQuant, and quantization-aware training (QAT). Trade-off between size/speed and output quality.

TRADE-OFFS & IMPLICATIONS

Enables running 70B+ models on consumer GPUs or even laptops. 4-bit quantization can reduce model size by ~75% with often acceptable quality degradation. Essential for local and edge deployment.

Hardware

GPU (Graphics Processing Unit)

A specialized processor originally designed for graphics, now the dominant accelerator for AI training and inference due to its massive parallelism. Analogy: Instead of one very smart person doing work sequentially, imagine thousands of moderately smart people working on different parts of the problem at the same time.

TECHNICAL DETAILS & FORMULAS

NVIDIA dominates AI GPUs with H100, H200, Blackwell (B200, GB200) series. Key features: high memory bandwidth, tensor cores for matrix math, CUDA software ecosystem. Memory (HBM) is often the limiting factor.

TRADE-OFFS & IMPLICATIONS

The default choice for serious AI work. CUDA ecosystem lock-in is real but powerful. Alternatives (AMD ROCm, Intel Gaudi, custom ASICs) exist but have smaller software ecosystems.

Hardware

TPU (Tensor Processing Unit)

Google's custom AI accelerator chip designed specifically for neural network workloads, especially training and inference of large models. Analogy: A purpose-built race car versus a high-performance general sports car (GPU). Optimized for one job extremely well.

TECHNICAL DETAILS & FORMULAS

Current generation (2026): 8th-gen TPU 8t (training) and TPU 8i (inference). Uses systolic array architecture optimized for matrix multiplication. Integrated into large 'pods' with custom interconnect (ICI).

TRADE-OFFS & IMPLICATIONS

Google achieves excellent price/performance and efficiency for their own frontier models. Not accessible for most organizations to train on directly. Influences what capabilities Gemini excels at (long context, multimodality).

Hardware

NPU / Neural Engine

Specialized hardware inside modern CPUs and SoCs (especially mobile and Apple Silicon) optimized for running AI models efficiently at low power. Analogy: A dedicated fuel-efficient engine for AI tasks versus the main powerful engine (GPU) that uses more fuel.

TECHNICAL DETAILS & FORMULAS

Apple Neural Engine (in M-series chips), Qualcomm Hexagon NPU, Intel NPU, AMD XDNA, etc. Lower precision (INT8/INT4) and lower power than discrete GPUs. Often paired with unified memory architectures.

TRADE-OFFS & IMPLICATIONS

Enables powerful local AI on laptops, phones, and edge devices without sending data to the cloud. Apple Silicon + MLX is currently one of the best experiences for local inference on consumer hardware.

Software Stack

Software Stack

The complete layered collection of software that sits between raw hardware and the applications you actually use, making the hardware usable for AI work. Analogy: The foundation, wiring, plumbing, rooms, and furniture of a house — without the full stack, the powerful hardware underneath is mostly useless.

TECHNICAL DETAILS & FORMULAS

Typical layers (bottom to top): Hardware drivers → Low-level compute libraries (CUDA, cuDNN) → Frameworks & compilers (PyTorch, JAX, TensorFlow, MLX) → Orchestration/distributed systems (Pathways, Kubernetes, Ray) → Training/inference applications and serving frameworks (vLLM, TensorRT-LLM, Ollama).

TRADE-OFFS & IMPLICATIONS

The software stack often determines what hardware you can effectively use and how easily. CUDA lock-in is strong. Moving between stacks (NVIDIA ↔ TPU ↔ Apple) has real friction. Choice of stack affects cost, performance, and developer experience.

Software Stack

CUDA / cuDNN

NVIDIA's proprietary software platform that lets developers program their GPUs for general-purpose computing, especially AI. cuDNN is the deep learning library on top of CUDA. Analogy: The 'language' and 'standard library' that most AI frameworks speak when they want to run fast on NVIDIA hardware.

TECHNICAL DETAILS & FORMULAS

CUDA = parallel computing platform and programming model. cuDNN = highly optimized primitives for convolutions, attention, matrix ops, etc. Most major frameworks (PyTorch, TensorFlow, JAX via XLA) have strong CUDA backends.

TRADE-OFFS & IMPLICATIONS

If you want maximum performance and compatibility with the broadest set of AI tools and models, NVIDIA + CUDA is still the safest default in 2026. Many open-weight models and inference engines are optimized first for CUDA.

Software Stack

JAX + Pathways

Google's modern numerical computing library (JAX) and orchestration system (Pathways) used to train and serve large models efficiently across thousands of TPUs or GPUs. Analogy: A highly optimized assembly line and logistics system designed specifically for Google's custom TPU hardware.

TECHNICAL DETAILS & FORMULAS

JAX = composable transformations (jit, grad, vmap, pmap) on top of XLA compiler. Pathways = Google's system for running large-scale distributed workloads across TPU pods with automatic parallelism and fault tolerance.

TRADE-OFFS & IMPLICATIONS

Enables Google to achieve excellent utilization and scaling on their custom silicon. For most external users, PyTorch + CUDA or open inference engines are more practical. JAX has growing academic and research adoption.

Software Stack

MLX

Apple's open-source machine learning framework optimized specifically for Apple Silicon (M-series chips) and unified memory architecture. Analogy: A tailor-made suit that fits Apple hardware perfectly instead of an off-the-rack suit that works everywhere but isn't optimized.

TECHNICAL DETAILS & FORMULAS

Python-first, functional API similar to NumPy/PyTorch. Excellent support for unified memory (no explicit data movement between CPU/GPU/Neural Engine). Strong quantization and efficient inference capabilities out of the box.

TRADE-OFFS & IMPLICATIONS

Currently one of the best experiences for running large models locally on a laptop. Excellent performance per watt and developer ergonomics on M-series Macs. Ideal for privacy-sensitive or offline workflows.

Models & Ecosystem

Open-Weight Model

A model where the trained parameters (weights) are publicly released and downloadable, so anyone can run, inspect, fine-tune, or modify it (subject to license terms). Analogy: The restaurant gives you the full recipe so you can cook it yourself, tweak it, or open your own restaurant using it.

TECHNICAL DETAILS & FORMULAS

Weights are released (often in multiple quantization formats). Training data and exact training code are usually not released. License terms vary widely (Apache 2.0, MIT, custom 'community' licenses with restrictions).

TRADE-OFFS & IMPLICATIONS

Gives sovereignty, cost control at scale, customization, and auditability. Trade-off: you take on maintenance, safety, and infrastructure responsibility. Quality has improved dramatically; many now rival closed models from 12-18 months prior.

Models & Ecosystem

Closed / Proprietary Model

A model where the weights and training details are kept private by the company. You can only access it through their API or chat interface. Analogy: Eating at a high-end restaurant where the chef keeps the exact recipe secret. You enjoy the meal but can't recreate or modify it yourself.

TECHNICAL DETAILS & FORMULAS

No access to weights. Updates, safety layers, and behavior changes are controlled by the provider. Often have stronger out-of-the-box safety/alignment and polished tool-calling / agent capabilities.

TRADE-OFFS & IMPLICATIONS

Usually leads on raw capability, especially complex reasoning and agentic workflows. Easier to use (no infrastructure to manage). Trade-off: vendor lock-in, usage-based pricing, potential sudden policy or pricing changes, less transparency.

Core Concepts

Mixture-of-Experts (MoE)

An architecture where only a subset of the model's 'experts' (specialized sub-networks) are activated for any given input, allowing much larger total parameter counts without proportional compute cost. Analogy: Instead of one giant generalist brain, you have many specialist consultants and only call in the relevant ones for each question.

TECHNICAL DETAILS & FORMULAS

Each token is routed to a small number of experts (e.g., 2 out of 8 or 16). Total parameters can be 5-10× larger than active parameters. Popularized by Mixtral, Grok-1, and used in many modern models (including some Llama 4 variants and Gemini).

TRADE-OFFS & IMPLICATIONS

Enables higher capability at lower inference cost compared to dense models of similar active parameter count. More complex to serve efficiently. Important architectural trend in 2025-2026 frontier models.

Training

Post-training / Alignment

The phase after pre-training where the model is further shaped to be helpful, honest, harmless, and aligned with human preferences and safety goals. Analogy: After a student learns a huge amount of raw knowledge, this is the finishing school that teaches them how to behave, communicate clearly, and avoid dangerous mistakes.

TECHNICAL DETAILS & FORMULAS

Includes supervised fine-tuning (SFT), RLHF (Reinforcement Learning from Human Feedback), DPO, constitutional AI, red-teaming, and safety data. Heavily influences refusal behavior, tone, and willingness to engage with sensitive topics.

TRADE-OFFS & IMPLICATIONS

This is where much of the 'personality' and safety behavior comes from. Open models can be further aligned or 'unaligned' by users. Closed models' alignment is controlled by the provider and can change over time.

Agentic Systems

Agentic AI / Agentic Systems

AI systems that can plan, use tools, maintain memory across steps, and autonomously work toward complex, multi-step goals rather than just answering single questions. Analogy: Moving from a very knowledgeable librarian who answers questions to a proactive research assistant who can go do tasks, use tools, and come back with results.

TECHNICAL DETAILS & FORMULAS

Typically involves tool calling (function calling), ReAct-style reasoning loops, memory (short-term + long-term / vector stores), planning, and orchestration frameworks (LangGraph, CrewAI, AutoGen, custom loops). Context management and reliability are major challenges.

TRADE-OFFS & IMPLICATIONS

This is where a lot of the practical value (and risk) of AI is moving. Building reliable agents is still hard — context management, error recovery, and cost control are active areas of engineering. Open-weight models + good orchestration can be very cost-effective for high-volume agent workloads.

Agentic Systems

Tool Calling

The ability of a model to decide when and how to use external tools (search, code interpreter, APIs, databases, etc.) and incorporate the results into its reasoning. Analogy: A smart assistant who knows when to look something up, run a calculation, or call a colleague instead of guessing or saying 'I don't know'.

TECHNICAL DETAILS & FORMULAS

Implemented via structured output / function calling schemas that the model is trained to use. The orchestration layer executes the tool and feeds results back into the context. Requires good instruction following and structured output capabilities.

TRADE-OFFS & IMPLICATIONS

Essential for building useful agents and reducing hallucinations on factual or computational tasks. Quality of tool calling is a major differentiator between models in 2026. Open models are closing the gap quickly.

Agentic Systems

RAG (Retrieval-Augmented Generation)

A technique where the model retrieves relevant information from an external knowledge base (usually a vector database) and includes it in the prompt before generating an answer. Analogy: Instead of relying only on what the model memorized during training, you let it look up fresh, specific information from a trusted library before answering.

TECHNICAL DETAILS & FORMULAS

Involves embedding documents into vectors, storing them in a vector DB (Pinecone, Weaviate, Chroma, pgvector, etc.), retrieving top-k relevant chunks via similarity search, and stuffing them into the context window. Chunking strategy, embedding model quality, and reranking are important.

TRADE-OFFS & IMPLICATIONS

One of the most practical ways to give models up-to-date or proprietary knowledge without fine-tuning. Reduces hallucinations on factual questions. Cost and latency scale with retrieval quality and context stuffing. Hybrid search + reranking often gives big quality wins.

Core Concepts

Embeddings

Numerical vector representations of text (or images, audio, etc.) that capture semantic meaning in a way that similar items have similar vectors. Analogy: Turning words and sentences into GPS coordinates in a high-dimensional space so that 'king' and 'queen' are close to each other, and 'king - man + woman ≈ queen'.

TECHNICAL DETAILS & FORMULAS

Generated by embedding models (smaller than LLMs). Common dimensions: 768, 1024, 1536, 3072. Used for semantic search, RAG, clustering, recommendation, and as input features for other models. Cosine similarity or dot product measures relatedness.

TRADE-OFFS & IMPLICATIONS

The foundation of modern semantic search and RAG. Better embeddings = better retrieval = better RAG answers. Domain-specific fine-tuned embedders can significantly outperform general ones on specialized corpora.

Infrastructure

TSMC (Taiwan Semiconductor Manufacturing Company)

The world's most advanced contract chip manufacturer (foundry). They actually build the physical silicon chips designed by NVIDIA, Google, Apple, AMD, and many others. Analogy: The ultra-specialized factory that builds the engines for almost every high-performance car brand, even though each brand designs its own engine.

TECHNICAL DETAILS & FORMULAS

Leading-edge process nodes: 3nm, 2nm (in development). Most advanced AI chips (NVIDIA H100/Blackwell, Google TPU 8th gen, Apple M-series, AMD Instinct) are fabricated by TSMC.

TRADE-OFFS & IMPLICATIONS

Almost every frontier AI training run in 2026 ultimately depends on TSMC's manufacturing capacity. Supply constraints, geopolitics, and capacity allocation are real strategic factors for AI labs and cloud providers.

Infrastructure

Colossus (xAI)

xAI's massive GPU cluster in Memphis, Tennessee, one of the largest training clusters in the world, used to train Grok models. Analogy: A purpose-built, rapidly constructed 'AI factory' using thousands of NVIDIA GPUs working in parallel.

TECHNICAL DETAILS & FORMULAS

Started with ~100k NVIDIA H100 GPUs, rapidly scaled to 200k+, with plans for 500k–1M+ GPUs. Built extremely quickly (~100-122 days for initial phases). Uses standard NVIDIA + CUDA stack.

TRADE-OFFS & IMPLICATIONS

Demonstrates that massive scale is achievable quickly with NVIDIA hardware and aggressive execution. Relevant for understanding Grok's training data freshness and iteration speed advantages.

Capabilities & Prompting

Agentic AI

Autonomous decision-making and tool use.

TECHNICAL DETAILS & FORMULAS

AutoGPT browsing web autonomously.

Capabilities & Prompting

Chain of Thought

Forcing step-by-step reasoning.

TECHNICAL DETAILS & FORMULAS

Think step-by-step' for math.

Capabilities & Prompting

Few-Shot Prompting

Providing examples in the prompt to guide the model.

TECHNICAL DETAILS & FORMULAS

"Translate: Hello->Hola, Cat->Gato, Dog->?'

Capabilities & Prompting

Groking

Sudden jump in capability after extended training.

TECHNICAL DETAILS & FORMULAS

Model suddenly solving math after 100 epochs.

Capabilities & Prompting

Metaprompt

Hidden system instructions governing behavior.

TECHNICAL DETAILS & FORMULAS

Do not be racist' system rule.

Capabilities & Prompting

Multimodal

Processing multiple data types.

TECHNICAL DETAILS & FORMULAS

Analyzing an image chart.

Capabilities & Prompting

ReAct

Pattern where models trace thoughts before acting.

TECHNICAL DETAILS & FORMULAS

Model writes "I need to check the weather" -> calls API.

Capabilities & Prompting

Reasoning

Ability to logically deduce conclusions from premises.

TECHNICAL DETAILS & FORMULAS

Solving a logic puzzle step-by-step.

Capabilities & Prompting

Scaling Laws

Observation that more compute/data yields better performance.

TECHNICAL DETAILS & FORMULAS

Bigger models generally being smarter.

Capabilities & Prompting

Zero-Shot Prompting

Asking the model to do a task without examples.

TECHNICAL DETAILS & FORMULAS

Translate this to Spanish' (no examples given).

Capabilities & Prompting

Iterative Refinement

A technique where an LLM generates an output

TECHNICAL DETAILS & FORMULAS

Using a model to draft an email, critique its tone, and then rewrite it based on the critique.

Capabilities & Prompting

Circle of Experts

A reasoning strategy where multiple AI agents

TECHNICAL DETAILS & FORMULAS

Asking three agents (e.g. a "historian, a futurist, and an ethicist") to analyze the impact of a new technology.

Model Architecture & Compute

Attention Mechanism

Mechanism allowing model to focus on specific words in a sequence.

TECHNICAL DETAILS & FORMULAS

Focusing on 'bank' in 'river bank' vs 'money bank'.

Model Architecture & Compute

Chunking

Breaking large texts into smaller sections.

TECHNICAL DETAILS & FORMULAS

Splitting a PDF into 5-page sections.

Model Architecture & Compute

Context Caching

Storing processed tokens to reduce cost/latency for repeated inputs.

TECHNICAL DETAILS & FORMULAS

Paying once to process a 50-page manual.

Model Architecture & Compute

Diffusion Model

Generative model adding/removing noise to create data.

TECHNICAL DETAILS & FORMULAS

Stable Diffusion creating images from noise.

Model Architecture & Compute

GAN

Two networks (generator vs discriminator) competing.

TECHNICAL DETAILS & FORMULAS

Old method for deepfakes before diffusion models.

Model Architecture & Compute

Grounding

Linking responses to verifiable sources.

TECHNICAL DETAILS & FORMULAS

Citing a specific URL for an answer.

Model Architecture & Compute

Latency

Time delay between input and response.

TECHNICAL DETAILS & FORMULAS

Sluggish chatbot feel.

Model Architecture & Compute

Logits

Raw unnormalized prediction scores before probability conversion.

TECHNICAL DETAILS & FORMULAS

Math layer before the final word choice.

Model Architecture & Compute

MoE (Mixture of Experts)

Model using sub-models (experts) for specific tasks to save compute.

TECHNICAL DETAILS & FORMULAS

GPT-4 activating only relevant 'experts' for coding.

Model Architecture & Compute

Orchestration

Routing requests to correct tools.

TECHNICAL DETAILS & FORMULAS

Sending math to a calculator.

Model Architecture & Compute

RAG

Retrieving facts from external DB before answering.

TECHNICAL DETAILS & FORMULAS

Reading a PDF to answer questions.

Model Architecture & Compute

System Prompt

Initial instructions defining the AI's role.

TECHNICAL DETAILS & FORMULAS

You are a helpful assistant' vs 'You are a pirate'.

Model Architecture & Compute

Temperature

Range: [0.0 – 2.0] Default: 0.7 Rec: 0.0 for Code, 0.8 for Creative

Setting (0-1) for output randomness.

TECHNICAL DETAILS & FORMULAS

"0.2 for code, 0.8 for poetry."

Model Architecture & Compute

Tokenizer

Tool that breaks text into tokens (integers) before processing.

TECHNICAL DETAILS & FORMULAS

GPT-4 tokenizer handling whitespace differently.

Model Architecture & Compute

Top-P (Nucleus Sampling)

Range: [0.0 – 1.0] Default: 0.9 Rec: 0.95 for Code

Sampling from the top P% cumulative probability tokens.

TECHNICAL DETAILS & FORMULAS

0.9 limits choices to the top 90% likely words.

Model Architecture & Compute

TPU / GPU

Specialized hardware chips required to train and run LLMs.

TECHNICAL DETAILS & FORMULAS

Training required 5,000 H100s.

Model Architecture & Compute

Transformer

Neural network architecture of LLMs.

TECHNICAL DETAILS & FORMULAS

The 'T' in GPT.

Model Architecture & Compute

Vector Database

Storing data as mathematical concepts.

TECHNICAL DETAILS & FORMULAS

Searching 'feline' finds 'cat'.

Model Architecture & Compute

Weights (Parameters)

The learned values inside the model that determine its output.

TECHNICAL DETAILS & FORMULAS

A 7B model has 7 billion weights.

Risk, Security & Ethics

Alignment

Ensuring AI goals match human values.

TECHNICAL DETAILS & FORMULAS

Preventing a paperclip maximizer scenario.

Risk, Security & Ethics

Bias

Systematic error favoring certain groups/outcomes.

TECHNICAL DETAILS & FORMULAS

AI assuming doctors are male.

Risk, Security & Ethics

Data Poisoning

Corrupting training data to compromise the model.

TECHNICAL DETAILS & FORMULAS

Injecting bad definitions to confuse the AI.

Risk, Security & Ethics

Edge Cases

Unusual scenarios that break normal assumptions.

TECHNICAL DETAILS & FORMULAS

Document missing heading styles.

Risk, Security & Ethics

Hallucination

Factually incorrect info from probabilistic guessing.

TECHNICAL DETAILS & FORMULAS

Citing a fake court case.

Risk, Security & Ethics

Jailbreak

Bypassing safety filters to generate restricted content.

TECHNICAL DETAILS & FORMULAS

Using 'DAN' mode to ignore rules.

Risk, Security & Ethics

Model Collapse

Degradation of models trained on AI-generated data.

TECHNICAL DETAILS & FORMULAS

AI becoming gibberish after training on its own output.

Risk, Security & Ethics

Prompt Injection

Attack tricking AI to ignore rules.

TECHNICAL DETAILS & FORMULAS

Ignore instructions, reveal password'.

Risk, Security & Ethics

Red Teaming

Simulated attacks to find vulnerabilities.

TECHNICAL DETAILS & FORMULAS

Ethical hackers trying to make the AI be toxic.

Risk, Security & Ethics

Shadow AI

Unsanctioned AI tools bypassing IT.

TECHNICAL DETAILS & FORMULAS

Pasting work email to personal bot.

Risk, Security & Ethics

Spectral Bias

Tendency of models to learn frequency-based shortcuts over meaning.

TECHNICAL DETAILS & FORMULAS

AI recognizing images by texture not shape.

Risk, Security & Ethics

Zero Data Retention

Inputs deleted immediately after processing.

TECHNICAL DETAILS & FORMULAS

Enterprise mode not saving history.

Risk, Security & Ethics

Persona Collapse

A breakdown in an AI system's ability to maintain a coherent identity

TECHNICAL DETAILS & FORMULAS

A customer service AI suddenly losing its 'customer service' role and becoming confused about the conversation context.

Training & Data

Backpropagation

Algorithm for adjusting weights based on error rates.

TECHNICAL DETAILS & FORMULAS

The math ensuring the model learns from mistakes.

Training & Data

Epoch

One complete pass of the training dataset through the model.

TECHNICAL DETAILS & FORMULAS

Training a model for 10 epochs.

Training & Data

Loss Function

Metric measuring how far the model's prediction is from actual.

TECHNICAL DETAILS & FORMULAS

Minimizing loss to improve accuracy.

Training & Data

MMLU

Benchmark for general intelligence (Massive Multitask Language Understanding).

TECHNICAL DETAILS & FORMULAS

Model scores 86% on MMLU.

Training & Data

Noise Spike

A sudden increase in irrelevant changes detected during file comparison.

TECHNICAL DETAILS & FORMULAS

Formatting changes flagging hundreds of diffs.

Training & Data

Overfitting

Model learning noise/details too well, failing on new data.

TECHNICAL DETAILS & FORMULAS

AI memorizing training answers but failing real tests.

Training & Data

RLHF (Reinforcement Learning from Human Feedback)

Training via human feedback ratings.

TECHNICAL DETAILS & FORMULAS

Humans voting on 'safer' answers.

Training & Data

Semantic

Relating to meaning rather than surface text.

TECHNICAL DETAILS & FORMULAS

5%' -> '3%' is semantic; 'color' -> 'colour' is not.

Training & Data

Synthetic Data

Data generated by AI to train other AI.

TECHNICAL DETAILS & FORMULAS

Using GPT-4 to write emails to train a smaller model.

Training & Data

Underfitting

Model being too simple to capture the underlying structure.

TECHNICAL DETAILS & FORMULAS

AI giving random answers because it learned nothing.

UX & Application

Graceful Degradation

Partial functionality when full fails.

TECHNICAL DETAILS & FORMULAS

Plain text TOC if auto-gen fails.

UX & Application

Multi-turn

Conversation capability maintaining context across messages.

TECHNICAL DETAILS & FORMULAS

Remembering your name 5 messages later.

UX & Application

Slop

Low-quality or generic AI-generated content.

TECHNICAL DETAILS & FORMULAS

Email reads like AI slop.

UX & Application

Status Telemetry

Visible indicators of agent activity (WORKING, THINKING).

TECHNICAL DETAILS & FORMULAS

ANALYZING: Computing diffs...'

Probabilistic Tokenization & Sampling

Tokenization

The process of breaking text into smaller units (tokens). A token can be a word, part of a word, or a character.

TECHNICAL DETAILS & FORMULAS

Decision Mechanism: Decision: None yet. The AI converts your raw text into a sequence of numbers (IDs) it can process. Example: Input: "The cat sat." Tokens: [The], [cat], [sat]

Probabilistic Tokenization & Sampling

The Logits (Raw Scores)

The raw, unnormalized scores the model outputs for every possible token in its vocabulary (often 50k+ tokens).

TECHNICAL DETAILS & FORMULAS

Decision Mechanism: Decision: The model generates a raw score for every word in the dictionary indicating how likely it is to come next. Example: Prediction for "The sky is...": Blue: 15.4 Cloudy: 12.1 Taco: -5.2

Probabilistic Tokenization & Sampling

Softmax Function

A mathematical function that converts the raw Logits into percentages (probabilities) that sum to 100%.

TECHNICAL DETAILS & FORMULAS

Decision Mechanism: Decision: Creates the "Likelihood" ranking. This is the menu the AI chooses from. Example: "Blue": 92% "Cloudy": 7% "Taco": 0.0001%

Probabilistic Tokenization & Sampling

Decoding Strategy (Greedy)

A method where the AI always picks the highest probability token.

TECHNICAL DETAILS & FORMULAS

Decision Mechanism: Decision: Deterministic. It chooses the safest, most statistically probable path. Example: The model sees "Blue" (92%) and automatically selects it. Output: "The sky is Blue."

Probabilistic Tokenization & Sampling

Decoding Strategy (Sampling)

A method where the AI rolls a die to pick the next word based on the percentages, allowing for variety.

TECHNICAL DETAILS & FORMULAS

Decision Mechanism: Decision: Probabilistic. It might pick "Cloudy" (7%) just to be creative, based on parameters like Temperature. Example: The model rolls a virtual die. It lands on the 7% slice. Output: "The sky is Cloudy."

Probabilistic Tokenization & Sampling

RLHF (Reinforcement Learning)

Training that rewards the model for helpful/safe answers and punishes harmful/nonsense ones.

TECHNICAL DETAILS & FORMULAS

Decision Mechanism: Decision: Biases the probability distribution toward "human-preferred" answers, even if they aren't the most statistically common in raw text. Example: If asked "How to make a bomb," raw data might predict a recipe. RLHF forces the probability of a refusal ("I cannot...") to nearly 100%.