Skip to content
CAS

Foundation · 2018– · autoregressive scaling

GPT-style Decoder

A stack of causal transformer blocks trained to predict the next token. At scale, this single objective yields general-purpose language capability.

ModLensFoundation

Interactive Diagram

Focus the lens

Click any component to read what it does. Signal direction follows the edges.

TokensCausal MaskDecoder Blockscausal attention + FFNVocab Projection

inputTokens. The sequence so far — the only input.

Core idea

Compression is comprehension: a model forced to predict the next token must internalize grammar, facts, reasoning patterns and style to minimize loss.

Why it exists

One simple objective, one architecture, scaled — turned out to be a path to broad capability rather than a narrow language model.

Mathematics

\mathcal{L} = -\sum_t \log P(x_t \mid x_{<t})
Objective

Data Flow

What moves through the system

  1. 01Tokens embed with causal masking.
  2. 02Each block attends only to earlier positions.
  3. 03The final hidden state scores the vocabulary.
  4. 04Sampling or decoding strategies produce text.

Strengths

  • + Single unified objective
  • + Natural interface for generation and tool use
  • + Scaling behavior is well characterized

Limitations

  • Left-to-right only during training
  • Capability is emergent and hard to inspect

Applications

  • · Assistants
  • · Code generation
  • · Reasoning loops
  • · Agent cores
Ask CAS