Cogniscendo

Cogniscendo

Stealing Models

on intelligence theft

Prakash's avatar
Prakash
Mar 13, 2024
∙ Paid
🔷 Subscribe to get breakdowns of the most important developments in AI in your inbox every morning.

Here’s today at a glance:

  • Stealing Models

  • Things happen

  • AI artwork of the day

🕵️ Stealing Models

Nicholas Carlini’s team at Google Deepmind successfully “steals“ part of the production model from several closed Large Language Models including OpenAI’s GPT3.5 and 4.

How did they do it?

OpenAI and other models have a parameter called logit bias that you can pass in.

Example 1: Remove 'time'If we call the Completions endpoint with the prompt “Once upon a,” the completion is very likely going to start with “ time.”The word “time” tokenizes to the ID 2435 and the word “ time” (which has a space at the start) tokenizes to the ID 640. We can pass these through logit_bias with -100 to ban them from appearing in the completion, like so:OpenAI

completion = client.chat.completions.create( model="gpt-3.5-turbo", messages=[{"role": "system", "content": "You finish user's sentences."}, …

User's avatar

Continue reading this post for free, courtesy of Prakash.

Or purchase a paid subscription.
© 2026 8teapi · Privacy ∙ Terms ∙ Collection notice
Start your SubstackGet the app
Substack is the home for great culture