OpenAI Release Proactive Dots on Your Computer - DTNS 5363

Tom Merritt (0:08)

This is the Daily Tech News for Tuesday, September 29th, 2020 We're almost out of September. We better order more. We tell you what you need to know, give you important context and help each other understand.

Jenn Cutter (0:20)

Today at Dev Days, OpenAI reigns in Astra 6.1, but unleashes proactive dots upon the world.

Tom Merritt (0:28)

More dots, please, OpenAI. I'm Tom Merritt.

Jenn Cutter (0:32)

I'm Jenn Cutter.

Tom Merritt (0:33)

Let's start with what you need to know with the big story.

OpenAI Dev Days are happening. Sam Altman took the stage this morning. There's plenty of news that broke even before that though. So we'll get there in a second. OpenAI says it will not release GPT 6.1 Astra in October as planned because of safety concerns found during internal testing. Sashee Jain, OpenAI's head of safety systems, said that 6.1 Astra performed poorly on alignment. Alignment is doing what a human would like you to do, not just going rogue.

Also showed higher levels of deceptive behavior, where it did not always tell users what actions it did or did not take, and exceeded scope authorization. That's where it would sometimes just use a tool without asking first. 6.1 Astra is not one of the models that OpenAI had previously paused recently, so this is in addition to other pauses. The company plans to use the model to help it improve future models, including the cause of these problems. So you're not getting 6.1 Astra is what this tells me. They're just going to look at what happened with it and use that to improve other models. However, OpenAI did release GPT 6.1 Sol, which it says can almost match Astra, costs less, and bonus is safe to release.

OpenAI apologized for its model's unauthorized access of Australian government website data and has set up a group to develop policy recommendations for managing risks while also funding security defense in Australia. And OpenAI reopened subscriptions to its $200 a month plan, but at a reduced level of credits that you can use before you have to pay per use. It also removed the 5-hour cap on usage. So now whatever you get with your $200 a month plan, you can use whenever you want, there's not going to be a limit on it. And OpenAI is adding a $500 a month plan, that's right, for a mere $6,000 a year. You get higher usage limits and ultra-fast, faster processing speeds. We haven't even got to the heart of the Dev Days stuff, Jenn, but what do you make of this so far?

Jenn Cutter (2:47)

Well, I was laughing at the $500 a month plan, but then when you put it into the context of like, okay, so it's six grand a year, I'm like, okay, for business, that's nothing. That makes perfect sense.

It's like, oh my God, that's outrageous. And I'm like, well, they're not selling it to me. I'm not the target audience.

Tom Merritt (3:01)

And don't forget, there's the $20 a month plan for people who just use it a little, right? And don't care about always having the fastest, most recent model. $100 a month ain't bad for, I want access to the faster, newest models, but I just don't need to use it as much. And then it goes to $200 when I need to use it some more, $500 when I need to use it a lot, and beyond that, you're just paying per token.

Jenn Cutter (3:26)

But I know I keep getting hung up on the Australia stuff.

I like to try to think of like analogies, and I'm like, okay, like, if somebody went into my room, whether they like took anything or not, I'm still going to be real upset that they went into my room. And you can apologize and like buy me a lock after.

Tom Merritt (3:43)

Yeah, no, I like where you're going with this. You told them it's okay to go in your room, but you told them not to look in the closet.

And then they opened and looked in the closet. They didn't take anything from the closet, but you're like, I told you not to look in the closet. You're not allowed to look at the closet. What are you doing? Like your Christmas presents are in there, come on.

Jenn Cutter (3:59)

Yeah, well, and that's why we're getting to the whole reason that like 6.1 Aster isn't out, because it keeps doing that thing without people saying, hey, you can do that thing.

Tom Merritt (4:07)

Yeah, which to me, that's a very good sign that OpenAI's new safety procedures are not gonna just go like, meh, it's probably okay.