I’m Binge Watching Interviews with Sam Altman…

…so you don’t have to… and here are my takeaways. (Article 6)

Drew Wolin

6 min readMar 28, 2023

This is a continuation from a series of articles found HERE.

2023-Lex Fridman Interviews Sam Altman (0:00 to 30:00)

Context: This interview is brand spankin’ new!

Interviews in my series are pulled somewhat agnostic to the time when they were published.

The interview coverd in this article was just released a few days ago. And it’s a doozy — over two hours.

As always, I’ll pull out interesting tidbits below, adding in my personal commentary or relevant context sparingly.

This interview is so long (over 2 hours), I will organize it in 30-minute intervals across 5 different articles.

Unless otherwise stated, all quotes belong to Sam Altman.

0:00 to 30:00

Editor’s Note: Interview host Lex starts with a lengthy introduction highlighting how… scary / exciting… a concept is.

I think he tries to frame it in a somewhat uplifting manner. But there is also definitely a certain feeling of a dystopian future.

Sam says that in the grand scheme of things, we will look back at GPT4 (the newest commercial model of ChatGPT as of publishing) as a very early A.I. He says it messes up a lot, but it is the early stages of a technology that will be life-changing (like the internet).

Sam acknowledges that the underlying tech of ChatGPT is not what is innovative — but it is the packaging and usability of it.

Sam mentions an important term, RLHF, or Reinforced Learning from Human Feedback.

“We trained these models on a lot of text data… But when you first play with our base model… (it can do a lot of impressive things)… but it’s not very useful.”

He says that RLHF is a key component to what helped GPT evolve.

“That process works remarkably well to make the model more useful.”

“RLHF is how we align the model to what humans want it to do.”

On the subject of adding RLHF to ChatGPT, Sam says: “Ease of use matters a lot, even if the base capability was there before.”

How much of ChatGPT’s training datasource social media and memes?

“Not very much.”

Editor’s Note: I think it is easy to get frustrated with how much Lex is talking, or how slowly he’s going.

Or maybe that’s just me expressing — I am frustrated with both of those things! Sort of.

But this does indeed lead to a unique set of information that comes out, I think. Sam seems well-prepped. Sam is participating in the “What do you think of (random idea X)?” Or “Isn’t it really something that (XYZ is true)?”

Kind of like sitting around on a couch smoking weed with your friends.

This is indeed a unique interview format. I am not entirely unfamiliar with Lex or his podcast. I feel even more than usual, he is slowing things down in this interview with Sam. And I think I appreciate it.

Sam says that ChatGPT is too-often being utilized as a database, and not its best use, which is as a reasoning engine. “The thing that’s really amazing about this system is that it… for some definition of reasoning, it can do reasoning… I think that’s remarkable, and the thing that’s most exciting.”

Sam acknowledges that many would fight him on that idea and quibble about the definition of reasoning.

Editor’s Note: I would believe this is, of course, true. I’d expect most people to use new tech in a way that reminds them of tech that they are familiar with (like a Search engine). I think the opportunity for creative people now is to show how to use GPT to its fullest extent, and then others will follow.

About ChatGPT: “Some things that seem like they should be obvious and easy, these models really struggle with.”

Examples given are counting characters, counting words.

“We are building in public and we are putting out technology, because we think it is important for the world to get access to this early, to shape the way it is going to be developed, to help us find the good and the bad things…”

“Every time we put out a model… the collective intelligence and ability of the outside world helps us discover things we never could have imagined, and that we never could have done internally.”

Sam feels that it is important to put product out into the world, get feedback from the public, and iterate quickly.

“The tradeoff of that is that we put out things that are deeply imperfect… We want to get it better and better with each rep.”

Sam says the launch of GPT 3.5 was not one that he was particularly proud of.

“No two people are ever going to agree that one single model is unbiased on every topic.”

Sam thinks that the answer there is to build customization and control into GPT.

“One thing that I hope these models can do is bring some nuance back to the world. Twitter really destroyed that somewhat.”

“When I was a kid, I thought A.I. was like the coolest thing ever. I never thought I’d actually get the chance to work on it.”

Sam responds to Lex’s example about counting the number of nice words that GPT would say about different prominent people.

“If you had told me that I actually had the chance to work on A.I…. and the thing I’d have to spend my time on is: I’d have to argue with people about counting how many nice words the A.I. says about one person, and it’s not the same as the number of words that the A.I. has to say about another person, I’d not have believed you.”

But Sam acknowledges: “But the small stuff is the big stuff, in aggregate.”

“Somehow this is the thing that we get caught up in, as opposed to thinking about what this will mean for our future.”

“The work we do to make GPT safer and more aligned is very close to the work we do to make useful and powerful models.”

“There’s no one set of values, and there’s no one set of right answers for human civilization.”

“We will only be allowed to agree on very broad bounds of what these systems can do. And within those, maybe within those, different countries have other RLHF tunes. Maybe different users do too.”

Editor’s Note: I’ll take the opportunity to share an analogy I’ve made in the past. It feels that with A.I., we will likely have a small number of different base models (eg. From OpenAI, from Google, maybe from Meta or Amazon)… and this will approximate cultures of different continents in the world. “Western culture” vs. “Asian culture” vs. “African culture.”

So you choose a base model that is closest to the “culture” or “behavior” you want. And then you can further fine tune the A.I., which would be like fine tuning the “Western culture” model to learn differences of being “American vs. Canadian” as an analogy. Then you can further tune, etc.

Sam talks about a GPT4 feature called System Message, which allows users to have a good degree of steer-ability with what they want. Sam gives the example of having GPT write as if it were Shakespeare. GPT4 was tuned to treat the System Message with a lot of authority.

Sam says that he is not a good prompt engineer.

Editor’s Note: This is funny.

Sam says that he knows people who have a fantastic “feel for the model” by spending enormous amounts of time interacting with GPT.

Find 30:00 to 60:00 of this interview in Article 7, to be published within 24 hours!

My Biggest Takeaway from this Interview

This interview was an exercise for me in getting comfortable with Lex’s (comfortable) interview style.

Also, Sam mentions what I’d describe as a “democratic approach” to building A.I. a number of times here. I think we have, many times in Tech, heard organizations preach democratic-sounding value systems. But in reality, exposure to the public and feedback from the public serves the company — it gives them unique research and development opportunity — but it also perhaps artificially builds trust from the public. In examples such as Facebook, perhaps feedback from the public was not used strictly to make the product better for people. Perhaps it was used to help the organization to design the product to make it more money.

More (from this 2+ hour long interview) coming soon!

More to come!

I’m Binge Watching Interviews with Sam Altman…

…so you don’t have to… and here are my takeaways. (Article 6)

2023-Lex Fridman Interviews Sam Altman (0:00 to 30:00)

0:00 to 30:00

My Biggest Takeaway from this Interview

Articles in Series:

Written by Drew Wolin