SIGNAL//DESK
AI/MLsrc: Roost FA glossary v2

Model

The model is the AI. Specifically: a giant file of numbers (the weights — billions of them) plus the architecture that uses them to compute probabilities. When you call ChatGPT, you are calling a model. When you call Claude, you are calling a different model. They produce different distributions for the same prompt because their weights are different. Choosing a model = choosing a probability machine.

A model is the combination of a neural network architecture and its associated learned parameters, typically stored as a large binary file. It functions as a probabilistic engine that processes input tokens to generate a distribution over the next possible tokens. Different models, such as GPT-4o or Llama 3.3, are distinct sets of weights and architectures that produce unique output distributions.

A model is defined as the static artifact consisting of a specific computational graph (architecture) and its corresponding set of learned parameters (weights and biases). It serves as a deterministic function that maps an input sequence to a probability distribution over the vocabulary space. The model is the version-controlled, deployable unit—such as a transformer-based checkpoint—that executes inference via an API to produce distinct, model-specific output distributions.

other senses: AI security

evolution

  1. 1958 · history
    Perceptron

    Frank Rosenblatt introduced the first artificial neural network model capable of learning weights through iterative adjustment.

  2. 1986 · history
    Backpropagation

    Rumelhart, Hinton, and Williams popularized the algorithm that allowed multi-layer models to learn complex representations by propagating error gradients.

  3. 2012 · history
    AlexNet

    The success of this deep convolutional model on ImageNet demonstrated the power of scaling model parameters and GPU-accelerated training.

  4. 2017 · history
    Transformer Architecture

    Google researchers introduced the Transformer, a model architecture that enabled massive parallelization and the foundation for modern Large Language Models.

  5. 2020 · history
    GPT-3

    OpenAI released a model with 175 billion parameters, establishing the paradigm of 'foundation models' that perform diverse tasks via in-context learning.

  6. 2026-04-07 · tracked
    Claude Mythos Preview announced

    Anthropic previews Mythos, a frontier cyber-vuln-finding model; not generally available.

  7. 2026-06-09 · tracked
    NIST publishes Gödel guardrail proof

    NIST proof: no finite static guardrails can protect AI; continuous monitor-and-update required.

  8. 2026-06-09 · tracked
    Claude Fable 5 released

    Mythos-class Fable 5 released broadly; Anthropic says new safeguards block high-risk areas.

  9. 2026-06-10 · tracked
    Fable 5 jailbroken by Pliny

    Pliny publishes a Fable 5 jailbreak within 24h; safety guardrails bypassed.

  10. 2026-06-12 · tracked
    US govt orders Fable 5 & Mythos 5 disabled

    US export-control directive bars foreign-national access; Anthropic disables both models for all. Opus 4.8 unaffected.

seen in events

← all terms