Black Ice AI

Research proposal deck «Advancing American AI», a research proposal by Black Ice AI and the RVL lab at Purdue University on reinforcement learning for language models.

Client
Black Ice AI, investigation and risk intelligence
Kind of work
Research proposal deck
Industry
Other
The title slide.

Overview

«Advancing American AI» is a research proposal by Black Ice AI and the RVL lab at Purdue University. It argues for feedback-based reinforcement learning that keeps language models aligned with human intent, and applies the idea to security clearance in the US government.

Twelve slides of the deck, from the title to the closing slide.
Twelve slides of the deck, from the title to the closing slide.

I designed the deck: dark slides lit in green and blue, red slides for the problem, and diagrams for the proposed method, from the supervisors that check truthfulness, fairness and safety to an adaptive layer on top of a frozen model.

The problem: security clearance in the US government.
The problem: security clearance in the US government.
The solution: the lab’s RL-DELI project.
The solution: the lab’s RL-DELI project.

The proposed method: truthfulness, fairness and safety supervisors reward or penalise the model’s outputs.

A holistic adaptive layer takes the place of the RLHF and SFT steps on top of a frozen pre-trained model.
A holistic adaptive layer takes the place of the RLHF and SFT steps on top of a frozen pre-trained model.

The lab, the platform and the applications3

The lab’s earlier work.
SARA by Black Ice AI.
Expected applications.
The closing slide.
The closing slide.

Credits

Presentation design
Jean Ivanov
Next projectQazaq GardenWebsite

More work

All works

Let’s talk about a project

Discuss a project