OpenAI VP on AI at work and ChatGPT's ad rules

The host introduces guest Benji and frames the conversation around AI's role in work and daily life.

ChatGPT was released in 2022, nearly four years ago. Benji has been at OpenAI for about a year and calls the change in his own work 'night and day'.

Early versions were mostly used for summarization, which felt magical then. The newer models reason through multi-step processes, think about what the user wants and can ask clarifying questions.

The next step, shown in a recent launch, is the ability to take actions on the user's behalf.

Benji's own account of the shift: 'The biggest change for me is that it's evolved from summarization to reasoning, to these multi-step processes, and now opening up possibilities to actually take action on your behalf.'

Telling real AI skill from claimed skill in a five-minute conversation is hard; it takes time to assess.

AI makes producing output easier than ever, but producing output does not mean the output is good.

The strongest AI-native people raise their output dramatically, around 10x, while keeping or improving quality through judgment. Weak users can 10x output just as easily, producing the 'AI slop' people complain about online.

The host raises the common talk that AI makes certain skills obsolete and asks which skills stay relevant.

Benji says skills like structuring information, analysis and storytelling still matter, while AI has removed much of the grunt work from them.

As more questions get answered, knowing which questions to ask and how to break a problem down becomes more important, and human judgment matters most at the final decision.

He says the weekly update that once took 60 minutes now takes five to ten minutes to review and send to his team.

He also uses Codex to clean his inbox of calendar change notices every hour and send a summary every six hours, so meeting changes still reach him.

By default, Codex keeps asking the user for confirmation before it takes an action.

SEO was the set of tactics for making a product show up in search engine results. What people now call GEO or AEO applies the same idea to conversational AI.

Benji says OpenAI does not give brands specific guidance on appearing in ChatGPT because the models and experience are still changing and the company serves consumers first.

The advice it does give is to make information available, much as in SEO, so the models can learn it and surface it to the right people at the right time.

The first principle is answer independence: ads have no influence on the answers ChatGPT gives. Advertisers cannot pay to get products into those answers.

The ad-selecting model and the answer-generating model are two separate models that do not speak to each other.

The second principle is conversation privacy: user conversations are private and never shared with advertisers.

People use ChatGPT for many different purposes, including discovering new things, considering items and making final decisions.

What is most distinctive is the middle of the funnel, where people weigh trade-offs and compare options.

Benji argues that a brand joining the conversation at that moment and showing how its product meets the criteria a consumer is weighing is a powerful new kind of opportunity.

Asked about a world ten years on, where the main issues are resolved, Benji says he hopes AGI benefits reach all of humanity and shorten the time from idea to reality.

He would most like to see creation explode as people use these tools to build, and he believes human judgment, creativity and craft will matter even more.