Character.ai Pretraining Tricks: Squinch Algorithm

Source: https://x.com (thread) Author: unknown (referencing Character.ai / Noam Shazeer) Date: 2025-2026

Summary

Five pretraining tricks from Character.ai before Google acquisition. Key invention: Noam Shazeer’s gradient compression algorithm “Squinch” — maintains SOTA MFU despite running on GCP H100-TCPX (1/4 of InfiniBand bandwidth).

Key Claims

  • Character.ai ran pretraining on GCP H100-TCPX (1/4 bandwidth of InfiniBand)
  • Noam Shazeer invented “Squinch” — gradient compression algorithm for low-bandwidth training
  • Squinch maintains state-of-the-art MFU (Model FLOP Utilization) despite bandwidth constraint
  • 4 other tricks mentioned but not fully detailed in thread
  • Before Google acquisition: company had to solve bandwidth constraints rather than throw money at hardware

Entities

Concepts