SDS 597: A.I. Policy at OpenAI — with Miles Brundage

GPT-3 Evolution

Miles Brundage discusses the evolution and challenges of GPT-2 and GPT-3. He explains that GPT-2 focused on responsible publication without an API, while GPT-3's deployment via an API allowed for better monitoring and limitation of potential harms 1. Both models are transformer-based and capable of performing a wide range of tasks by learning from diverse internet text 1. Miles emphasizes the importance of understanding and mitigating the risks associated with these models, such as disinformation and misuse for unethical purposes 2.

Imperfect technologies have a lot of uses, and we use them every day. The important thing is that people understand them and that you have appropriate guardrails to prevent those limitations from escalating into real harm.

--- Miles Brundage

Learning from real-world misuse has been crucial in refining safety measures and ethical guidelines for these models 2.

   

DALL-E 2

Miles Brundage introduces DALL-E 2, a system for generating images from text prompts, and discusses its creative and ethical implications. He highlights the model's ability to create variations and modifications of images, which opens up new avenues for creativity and economic impact 3. To prevent misuse, OpenAI has implemented policies to restrict the generation of real people's images and filtered out harmful content from the training data 4.

We think that kind of maps on well to what the technology is actually useful for, which is kind of creating new images, trying out new concepts, as opposed to kind of modifying images of people.

--- Miles Brundage

These measures aim to ensure that DALL-E 2 is used responsibly and ethically 4.

   

Codex Potential

Miles Brundage explains Codex, an extension of GPT-3 designed for code generation, and its impressive capabilities. Codex can solve programming problems by converting natural language inputs into code, and it powers applications like GitHub Copilot 5. While the model's potential is vast, Miles notes that there are concerns about its misuse for generating malicious code and the risks of relying too heavily on it, which could lead to buggy or insecure software 6.

The bigger concern for me, at least when it comes to code generation, is not malicious use, but kind of reckless or naive use where people might rely on it too much and end up generating buggy or insecure code.

--- Miles Brundage

Ensuring appropriate human oversight is crucial for the safe deployment of Codex 6.

   

CLIP Insights

Miles Brundage delves into the functionalities of CLIP, a model designed for image and text recognition. CLIP can create new classifiers on the fly by comparing text and image embeddings, making it highly versatile for tasks like captioning and image recognition 7. However, ethical challenges such as biases in classification and disparate performance across demographic groups need to be addressed 8.

There are various kinds of biases in all AI systems to varying degrees, but in image recognition systems, those are some of the clusters.

--- Miles Brundage

Ongoing research aims to mitigate these biases and ensure fair and accurate performance 8.