The Safety Tax of Cache Compression
Published:
Some notes on what happens to a language model’s refusals when we start throwing away parts of its memory to save space, and why the system prompt tends to be the first thing to go.
Aryan Gupta
Posts by Tags
AI Safety
Liar, Liar: Beyond Vocabulary Suppression
Published:
An account of what happens when we steer a language model toward honesty, and whether it actually becomes more truthful or simply learns to sound that way.
The Spectral Geometry of Misalignment
Published:
A tour of what post-training leaves behind in a model’s weights, and why the directions that matter for refusal and misalignment are far easier to find than to isolate.
Activation Steering
Liar, Liar: Beyond Vocabulary Suppression
Published:
An account of what happens when we steer a language model toward honesty, and whether it actually becomes more truthful or simply learns to sound that way.
Finance
Distributional Jump Bridges for Disclosure Response Forecasts
Published:
What happens when a model reads a company’s SEC filing before predicting how its stock will react, and why the text turns out to say more about risk than about returns.
Forecasting
Distributional Jump Bridges for Disclosure Response Forecasts
Published:
What happens when a model reads a company’s SEC filing before predicting how its stock will react, and why the text turns out to say more about risk than about returns.
Game Theory
World’s Simplest Poker: A Complete Analysis of the Discrete Analogue
Published:
Some notes on what happens to poker when the deck shrinks to a handful of numbered cards, and why the only stable way to play even three of them involves a little bluffing.
Inference
The Safety Tax of Cache Compression
Published:
Some notes on what happens to a language model’s refusals when we start throwing away parts of its memory to save space, and why the system prompt tends to be the first thing to go.
Interpretability
Liar, Liar: Beyond Vocabulary Suppression
Published:
An account of what happens when we steer a language model toward honesty, and whether it actually becomes more truthful or simply learns to sound that way.
The Spectral Geometry of Misalignment
Published:
A tour of what post-training leaves behind in a model’s weights, and why the directions that matter for refusal and misalignment are far easier to find than to isolate.
KV Cache
The Safety Tax of Cache Compression
Published:
Some notes on what happens to a language model’s refusals when we start throwing away parts of its memory to save space, and why the system prompt tends to be the first thing to go.
Mathematics
World’s Simplest Poker: A Complete Analysis of the Discrete Analogue
Published:
Some notes on what happens to poker when the deck shrinks to a handful of numbered cards, and why the only stable way to play even three of them involves a little bluffing.
NLP
Distributional Jump Bridges for Disclosure Response Forecasts
Published:
What happens when a model reads a company’s SEC filing before predicting how its stock will react, and why the text turns out to say more about risk than about returns.
Poker
World’s Simplest Poker: A Complete Analysis of the Discrete Analogue
Published:
Some notes on what happens to poker when the deck shrinks to a handful of numbered cards, and why the only stable way to play even three of them involves a little bluffing.
Post-Training
The Spectral Geometry of Misalignment
Published:
A tour of what post-training leaves behind in a model’s weights, and why the directions that matter for refusal and misalignment are far easier to find than to isolate.