# No Memory, No Harness: Why the Database Is the Last Line of Defense — Kay Malcolm — session 2026-07-02T21:50:00.000Z → 2026-07-02T22:10:00.000Z

_0 transcript lines · 2 slides · source: full recording_

## Transcript

_(no captions)_

## Slides

### 00:00:22

# TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate
[Hand cursor icon pointing to the word "Abstract"]

### 00:00:50

# What is TurboQuant

GOOGLE RESEARCH · ICLR 2026

A compression algorithm that stores embeddings and the KV cache in **3 to 4 bits** instead of 32, with no training. It works in **two stages**:

### STAGE 1 - POLARQUANT
Does the compression. Stores each vector by direction, so one simple rule quantizes it, with no codebook to train.

### STAGE 2 - QJL
Fixes the error. One extra bit per number keeps the compressed scores unbiased, so the ranking stays correct.

- Read the paper
- Read the launch post

[Flow diagram showing the two stages of TurboQuant: 32-bit full vector -> Stage 1 - PolarQuant -> (r, θ) length + direction -> Stage 2 - QJL + 1 bit -> ~4-bit compressed]