Model·Foundations

About

Model Foundations exists because most writing about machine learning research optimizes for speed over understanding. Papers get compressed into headlines; benchmark tables circulate without their setups; claims outlive the evidence that qualified them.

What this site is

A collection of structured technical notes on LLMs, VLMs, LVMs, multimodal models, quantization, training, inference, and AI systems. Each note is built from primary sources — the paper, its source code, experiment logs, and benchmark tables — and organized so that a reader can understand the method, judge the evidence, and attempt a reproduction without re-deriving everything from scratch.

How notes are structured

Every note follows the same anatomy:

  • Core insights — the ideas that survive after the marketing is removed.
  • Method — the mechanism explained from first principles, with the equations and the implementation details the paper glosses over.
  • Claims & evidence — each headline claim traced to the specific experiment behind it, with a verdict: verified, partial, unverified, or refuted.
  • Benchmarks — results reproduced next to their baselines, with hardware, precision, and evaluation setup stated.
  • Limitations & open questions — what the paper does not show.
  • Reproduction notes — what happened when the method was actually run, by us or by credible third parties.

Verification rubric

VerdictMeaning
verified The claim is directly supported by the paper's own experiments and has survived independent replication or our own checks.
partial Supported under the paper's specific setup, but weaker or unconfirmed outside it (different scale, hardware, or data).
unverified Stated in the paper but not yet checked against evidence beyond it.
refuted Contradicted by later work or failed replication.

What this site is not

Not an AI news site. Not a paper-summary feed. Notes are published when the analysis is done, not when the paper trends. Every note lists its own status — draft notes are working documents; reviewed notes have been checked end to end.

Contact

Corrections are the most valuable contribution a reader can make. Open an issue on GitHub if a claim, equation, or number is wrong.