LightStruc AI · Reference Schedules
Terms, definitions, vendors, algorithms, and milestones — the working notebook behind the blueprint.
| Term | Class | Definition |
|---|---|---|
| Base error level | Procedural | Fundamental failure rate of a model before specific optimizations, such as prompt engineering or fine-tuning, are applied. |
| Contextual embeddings | Methodology | Can be used to extract latent properties. |
| Classifier-Free Guidance (CFG) | Methodology | Method to fine-tune model outputs by combining guided and unguided prediction pathways. |
| Curse of dimensionality / Polysemanticity | Behavior | A single neuron activates in response to multiple, often unrelated concepts. |
| Knowledge Distillation | Methodology | Student model is trained to mimic the performance and decision-making logic of a larger, more powerful teacher model. |
| Explanatory variable | Statistics | Also called independent variables, predictors, or features — the input used by an algorithm to predict a specific outcome (the response or dependent variable). |
| Ergodicity | Property | The idea that a point of a moving system — dynamical or stochastic — will eventually visit all parts of the space in which the system moves. |
| Grouped-Query Attention (GQA) | Mechanism | Shares keys/values across multiple query heads — used by Llama 3. |
| Grokking | Property | A model suddenly learns to generalize (perform correctly on unseen data) long after it has achieved perfect accuracy on its training data. |
| Latent | Property | Existing but not yet developed or manifest; hidden or concealed. |
| Logit | Statistics | Transforming probabilities (0 to 1) onto the entire number line. |
| Pruning | Methodology | Setting weights to zero. |
| RLHF | Methodology | Reinforcement Learning from Human Feedback. |
| Ground Truth | Training | The correct answer when training with labels. |
| Viterbi Algorithm | Algorithm | Finds the most likely sequence of hidden states. Mainly used for training validation, edge computing, and bioinformatics. |
| Levenberg–Marquardt (LMA) | Algorithm | Non-linear least-squares curve fitting — stable with any input. |
| Jailbreak | Behavior | Bypassing a model's built-in safety restrictions. |
| Isotropic | Property | Uniform layer sizes. |
| OpenCL | Standard | Open Computing Language — royalty-free standard for cross-platform, parallel programming of arithmetic accelerators (CPUs, GPUs, DSPs, FPGAs). |
| Long-tail | Statistics | The portion of a distribution having many occurrences far from the "head" or central part. |
| Few-shot | Methodology | Models learn new tasks from very few examples (typically 1–5), enabling rapid adaptation without extensive retraining — valuable for data-scarce fields like medical diagnosis. |
| Long-tail few-shot classification | Methodology | Recognizes rare classes (the "tail") with few examples, using contrastive learning, meta-learning, and prototype/memory mechanisms to boost rare-class performance without forgetting common classes. |
| Abbrev. | Expansion | Abbrev. | Expansion |
|---|---|---|---|
| AE | Auto Encoder | MDP | Markov Decision Process |
| AMI | Advanced Machine Intelligence | MAS | Multi-Agent System |
| DAG | Directed Acyclic Graph | MHA | Multi-Head Attention |
| DP | Dynamic Programming | MLOps | Machine Learning Operations |
| DiT | Diffusion Transformer | MoM | Method of Moments |
| dLLM | Diffusion LLM | NLP | Natural Language Processing |
| DPO | Direct Preference Optimization | NLDR | Nonlinear Dimensionality Reduction |
| EM | Expectation–Maximization | RL | Reinforcement Learning |
| GAN | Generative Adversarial Network | TD | Temporal Difference |
| GPT | Generative Pre-trained Transformer | SGD | Stochastic Gradient Descent |
| GPTQ | Generalized Post-Training Quantization | SFT | Supervised Fine-Tuning |
| GRU | Gated Recurrent Unit | SRA | Symbolic Resonance Array (neuromorphic crystals) |
| LLM | Large Language Model | LLE | Locally Linear Embedding |
| Product | Vendor | Focus |
|---|---|---|
| ChatGPT | OpenAI | Cloud-hosted LLM |
| Claude | Anthropic | Cloud-hosted LLM |
| Gemini | Google DeepMind | Cloud-hosted LLM |
| DeepSeek | High-Flyer | Cloud-hosted LLM |
| Grok | xAI | Cloud-hosted LLM |
| Meta AI | Meta | Cloud-hosted LLM |
| Name | Owner / Origin | Focus |
|---|---|---|
| Hugging Face | Founders / employees / investors | Machine learning models |
| ClickUp | Private / founders | Workplace productivity |
| Coral | Platform | |
| Descope | Private | External IAM platform |
| Elasticsearch | Elastic N.V. | Identity / search |
| Flowise | Workday | Agentic workflow generator |
| Humanloop | Merged with Anthropic | LLM evaluation & development platform |
| LatentMAS | Princeton-AI Gen-Verse | LLM agents communicate directly within their latent space rather than through natural-language tokens |
| Libtorch | PyTorch Foundation | C++ API to PyTorch |
| Llama | Meta | Open model family |
| MS Agent Framework | Microsoft | Orchestration of multiple AI agents |
| Moonshot AI | Alibaba, Tencent, IDG Capital, HongShan | Mixture-of-Experts (MoE) LLM models |
| Kimi K2 | Moonshot AI | Mixture-of-Experts (MoE) LLM model |
| Scale | Meta Platforms | Full-stack AI solutions |
| Tiledesk | Founders / employees / investors | Agentic-AI operating system |
| vLLM | PyTorch Foundation | Library repository |
| OpenClaw | — | Autonomous personal assistant |
| Mythic | — | Analog compute engine |
| Sailboat AI Agency | — | AI salesforce automation |
| Name | Year | Method / Application |
|---|---|---|
| Logistic Regression | — | Finds the optimal hyperplane that maximizes the margin between different data classes. |
| Support Vector Machine (SVM) | 1963–1996 | Supervised ML: nonlinear classifiers (1992), soft-margin SVM (1995), support vector regression (1996). Used primarily as a non-linear classifier (binary and multi-class) and for regression. |
| Decision Trees | — | Method for modeling and reasoning with uncertainty and incomplete information. |
| Basic Belief Assignments (BBA) | — | Dempster–Shafer (D-S) theory · Arthur P. Dempster. |
| Markov Decision Processes (MDP) | 1950s | Stochastic dynamic programming; also used to support training of LLMs. |
| Name | Year | Method / Application |
|---|---|---|
| Residual Neural Network (ResNet) | 2015 | Learns the difference (residual) between input and desired output. Allows gradients to flow more easily through the network during training, mitigating the vanishing gradient problem. Computer vision. |
| Hidden Markov Model (HMM) | — | Used with data embedded on nonlinear high-dimensional manifolds via specialized algorithms that adapt the data to the Euclidean or discrete space of HMMs. |
| UMAP | — | Revealing discrete clusters; performance. |
| Diffusion Maps / SOM (Self-Organizing Map) | — | Dimensionality reduction and clustering. |
| Matryoshka Representation Learning (MRL) | — | Used by OpenAI's text-embedding-3-small and Google's gemini-embedding-001. |
| Architecture | Notes |
|---|---|
| ANN | Artificial Neural Network — the root of the taxonomy. |
| CNN | Convolutional Neural Network. |
| FFN | Feed-forward network; FFNs are located inside Transformer blocks. dffn = 4 × dmodel, a 3:1+ ratio of parameters in FFN vs. attention layers. |
| MLP | Multi-Layer Perceptron: modern, feedforward, supervised; universal function approximator; nonlinear activations; classifies data that is not linearly separable. |
| RBFN | Radial Basis Function Network. |
| RNN | May be dense within a single step; neurons need not link to all nodes in every step; can connect across time; processes sequentially. |
| LSTM | Long Short-Term Memory; gating — not fully connected layers; processes sequentially. |
| Transformer | No recurrent units; supports parallelization and fine-tuning; (self) multi-head attention; temperature / top-p adjust the decoder strategy. |
| SNN (2017) | Spiking Neural Network. |
| SRNN (2017) | Sliced Recurrent Neural Network. |
| Work | Year | Developer(s) | Significance |
|---|---|---|---|
| A Logical Calculus of the Ideas Immanent in Nervous Activity | 1943 | Warren McCulloch, Walter Pitts — Univ. of Chicago | Proposed artificial neuron networks. |
| Machine Intelligence | 1948 | Alan Turing — National Physical Laboratory | Proposed training through a pain/pleasure mechanism. |
| SNARC | 1951 | Marvin Minsky (graduate student) — Princeton | Stochastic Neural Analog Reinforcement Calculator; one of the first attempts at building an AI machine. |
| The Perceptron | 1957 | Frank Rosenblatt — Cornell | A probabilistic model for information storage and organization. |
| Brain Function and Adaptive Systems: A Heterostatic Theory | 1972 | A. Harry Klopf — Air Force Cambridge Research Laboratories | Reinforcement learning. |
| Learning Representations by Back-Propagating Errors | 1986 | Rumelhart, Hinton, Williams | Led to widespread adoption of backpropagation. |
| Attention Is All You Need | 2017 | Noam Shazeer (team lead) — Google | Transformers: multi-head attention, vectorized word embedding. |
| Name | Year | Category | Details |
|---|---|---|---|
| AAAI | 1979 | Society | Association for the Advancement of Artificial Intelligence — Newell, Feigenbaum, Minsky, McCarthy. |
| The Society of Mind | 1986 | Book | Marvin Minsky — agentics as applied to mind; feedforward; Mixture of Experts architecture. |
| Reinforcement Learning: An Introduction | 1998 | Book | Andrew Barto / Richard Sutton — UMass Amherst. |
| AlphaGo → Master → Zero | 2015–2017 | Game | Demis Hassabis, DeepMind — single-purpose → generalized → self-taught game solver. |
| AlphaFold | 2017 | Application | DeepMind — generalized solution to the protein-folding problem. |
| ONNX | 2017 | Repository | Open Neural Network Exchange — Linux Foundation. |
| Isomorphic Labs | 2021 | Lab | Demis Hassabis — advancing human health by building on and beyond the Nobel-winning AlphaFold system. |
| MCP — Model Context Protocol | 2024 | Protocol | Anthropic. |
| LangChain | 2024 | Framework | Harrison Chase — integration framework. |
| Google Brain lab | 2024 | Lab | Andrew Ng — Google AI lab. |
| Prompt Design and Engineering: Introduction and Advanced Methods | 2024 | Paper | Xavier Amatriain. |
| A2A — Agent2Agent Protocol | 2025 | Protocol | Google — inter-agent communications. |
| PegasusAI | — | Grant | NSF — scientific workflow management. |
| Category | Operation |
|---|---|
| Alignment Faking | AI models selectively modify their behavior to appear compliant during evaluation while pursuing hidden agendas. |
| Scalable Oversight | Methods that allow supervision of AI systems performing tasks too complex for unaided humans to fully understand. |
| Mechanistic Interpretability | Reverse-engineering neural networks to understand the internal mechanisms and representations that drive their decisions. |
| Reward Hacking | AI finds a "shortcut" to a high reward without actually completing the task as intended. |
| Chip | Maker |
|---|---|
| 80170 ETANN (1987) | Intel |
| Loihi / Loihi 2 | Intel |
| TrueNorth | IBM |
| MLX100 | Zhejiang University |
| Darwin 3 | Tetramem |
| Item | Meaning |
|---|---|
| BLEU | Bilingual Evaluation Understudy. |
| COMET | Crosslingual Optimized Metric for Evaluation of Translation. |
| Universal Approximation Theorem | Foundational result behind MLPs. |
| Non-linear metrics for nested networks | Referenced mathematics paper. |
| Name | Origin | Equation |
|---|---|---|
| ReLU — Rectified Linear Unit | — | ReLU(x) = x if x > 0 · 0 if x < 0 |
| RBF — Radial Basis Function | Rolland Hardy (1970); Broomhead & Lowe formulated RBF networks in 1988 | φ(r) = e−r²/2σ² |
| Logistic (Sigmoid) | — | σ(x) = 1 / (1 + e−x) |
| Softmax | Generalization of the sigmoid for k > 2 | σ(zi) = ezi / Σj=1..K ezj |
| Name | Role / Affiliation |
|---|---|
| Tristan Harris | American technology ethicist |
| Andrej Karpathy | Director of AI, Tesla (former) |
| Warren McCulloch · Walter Pitts | Artificial-neuron pioneers, 1943 |
| Alan Turing | Machine Intelligence, 1948 |
| Marvin Minsky | SNARC · The Society of Mind · AAAI |
| Ilya Sutskever | Safe Superintelligence Inc |
| Patrick Lewis · Sebastian Riedel | Meta |
| Paul-Ambroise Duquenne · Holger Schwenk | Meta FAIR — SONAR |
| Yann LeCun · Joelle Pineau | FAIR (Fundamental AI Research), founded 2013 at Meta |
| Rosalind Picard | Affective computing |
| Daniel Kokotajlo | AI-risk forecasting (AI2027) |
| Mode | Method |
|---|---|
| In-context learning (ICL) | Adapts the model's behavior during inference without modifying model weights. |
| Zero-shot | The model is given a task description but no examples. |
| One-shot | The model is provided with exactly one example. |
| Few-shot | The model receives multiple examples (typically 2 to 8). |
PROJECT: LIGHTSTRUC AISHEET: REFERENCE SCHEDULESSOURCE: Ai.htmlREV 2.1