Should ChatGPT be censored? | Sam Altman and Lex Fridman

Introduction

The speaker discusses the challenges of creating an AI system that adheres to free speech absolutism while also regulating other people's speech. They mention the need for nuanced presentation of ideas and the challenge of evaluating bias in the system.

Free Speech Absolutism and Regulating Speech

• The speaker discusses the challenges of creating an AI system that adheres to free speech absolutism while also regulating other people's speech.

• People want a model that has been refined to their world view, which is really about regulating other people's speech.

• The speaker talks about presenting the tension of ideas in a nuanced way and how they are doing better at it than people realize.

• Evaluating bias in the system is challenging because there can always be anecdotal evidence of GPT slipping up and saying something wrong or biased.

Adapting to Models

The speaker discusses how people adapt to models and how sometimes there can be a really egregiously dumb answer that gets all the attention.

Adapting to Models

• When you ask the same question 10,000 times, what most people see is something around output 5000, but the output that gets all of Twitter's attention is output ten thousand.

• Sometimes there can be a really egregiously dumb answer, but in a world where you click screenshot and share, that might not be representative.

• More people are responding with their own results when they see these types of outputs.

Pressure from Clickbait Journalism

The speaker talks about whether there is pressure from clickbait journalism to not be transparent and how OpenAI handles mistakes.

Pressure from Clickbait Journalism

• The speaker is not pressured by clickbait journalism to not be transparent.

• OpenAI keeps doing their thing and admitting when they're wrong, trying to get better.

• OpenAI has systems that try to learn when a question is something they shouldn't answer, but it's early and imperfect.

Moderation Tooling for GPT

The speaker discusses the moderation tooling for GPT and the process of moderation.

Moderation Tooling for GPT

• OpenAI has systems that try to learn when a question is something they shouldn't answer, but it's early and imperfect.

• The speaker doesn't like the feeling of being scolded by a computer. They want people to know that this is a tool very much in their control.

GPT-4 Capabilities and Technical Leaps

In this section, the speaker discusses the capabilities of GPT-4 and the technical leaps that were made to improve it.

GPT-4 Capabilities

• GPT-4 has enough nuance to help explore complex topics like conspiracy theories without treating users like children.

Technical Leaps

• The leap from GPT-3 to 3.5 to 4 involved a lot of technical improvements in the base model.

• OpenAI is good at finding small wins and multiplying them together for big leaps. There were hundreds of complicated things done, including data organization, cleaning, training, optimization, and architecture.

Overall, GPT-4 has improved capabilities compared to its predecessor versions due to many technical improvements.