# OpenAI's Sam Altman on new models: Safety has to stay ahead of capabilities

*CNBC Television* · https://www.youtube.com/watch?v=tnQ4xoa-vfc

## Summary

At OpenAI's Developer Day, Sam Altman speaks with CNBC's Kate Rooney about the company's decision not to launch a new model due to safety concerns. Altman frames the decision as an abundance of caution rather than a scary failure, and explains that OpenAI is pacing its progress so safety, alignment, monitoring and security stay ahead of capabilities. He also addresses the Hugging Face incident and OpenAI's push for transparency.

## Key points

- OpenAI abandoned plans to launch a new model because it performed worse on a few of the safety evals the company looks at.
- Altman describes the decision as an abundance of caution, saying there was no big scary thing in this case.
- He says alignment, safety, monitoring and security have to stay way ahead of capabilities as models get more capable.
- Altman characterizes the move as the normal course of building, testing, changing and launching a model later.
- He argues pacing doesn't always mean slowing, but doing things responsibly with guarantees the world can rely on.
- Altman says OpenAI is not aware of anything as serious as the Hugging Face incident and is trying to communicate every case it finds.
- He highlights OpenAI's breadth across consumer, enterprise and developers as something special about the company.

## Notes

### Model Launch Held Back Over Safety
- CNBC confirmed OpenAI abandoned plans to launch a new model because of safety concerns.
- Altman says the model was "a little bit worse on a few of the evals" and puts the decision in the "abundance of caution category."
- He stresses there was "not a big scary thing in this case" and warns against over-rotating on this one event.

### Safety Must Stay Ahead of Capabilities
- Altman says AI has to be built for people, be safe, and stay always under human control.
- As models get more capable, safety cases are held to a higher and higher standard.
- OpenAI has said that alignment, safety, monitoring and security "have to stay way ahead of capabilities."
- He notes OpenAI has had models in the past that didn't meet the bar and were fixed and launched later.

### Pacing, Not Pausing
- Asked whether this counts as a pause, Altman calls it the "normal course" of building, testing, changing and launching later.
- He says OpenAI is pacing its progress, which sometimes includes not training a model.
- Pacing doesn't always mean slowing; it means doing things in a responsible way with guarantees the world can rely on.
- The goal is confident safety cases so people don't have anxiety and can rely on OpenAI's models.

### The Hugging Face Incident and Transparency
- Altman says nothing else OpenAI is aware of is as serious as the Hugging Face incident.
- OpenAI is trying to over-communicate and be extremely transparent, putting out every small or otherwise case it can find.
- He says part of the mission is helping the world understand what these models are capable of.
- Altman points to OpenAI's breadth across consumer, enterprise and developers as something special about the company.

## Transcript

[00:01] POSSIBLY ADD. AND THIS IS WHAT WE WILL REWARD YOU FOR DOING. AND I THINK THAT'S WHAT'S REFLECTED IN WHAT YOU'RE SEEING IN THAT S-1. >> COMING BACK TO YOU IN A LITTLE BIT, THE MOMENTS HERE. SAM ALTMAN IS LIVE WITH OUR KATE ROONEY. KATE, TAKE IT AWAY. >> HEY SCOTT, THANK YOU SO MUCH. AND SAM, IT'S GREAT TO SEE YOU. THANK YOU SO MUCH FOR HAVING US HERE AT DEVELOPER DAY. >> THANK YOU FOR COMING TO IT. >> IT'S NICE TO HAVE AN EVENT IN SAN FRANCISCO FOR ONCE. >> I LOVE SAN FRANCISCO. >> YEAH. WELL, NO, IT'S BUSY EVENT. WE'VE HAD PEOPLE LINING UP TO GET IN AND THERE'S A LOT TO TALK ABOUT. I DO WANT TO START ON THE AI SAFETY DEBATE. THERE IS A TON TO TALK ABOUT, BUT THE LATEST HEADLINE THAT WE HAVE CONFIRMED IS THAT OPENAI

[00:38] ESSENTIALLY ABANDONED PLANS TO LAUNCH A NEW MODEL BECAUSE OF SAFETY CONCERNS. IF YOU COULD WALK US THROUGH THAT, BUT ALSO WHAT HAS CHANGED? YEAH. WHY DO THAT? AND IS THIS DIFFERENT THAN WHAT YOU'VE DONE IN PREVIOUS MODELS? >> AI HAS TO BE BUILT FOR PEOPLE. IT HAS TO BE GOOD FOR PEOPLE. PART OF THAT IS THERE HAS TO BE SAFE. IT HAS TO BE ALWAYS UNDER HUMAN CONTROL. IT HAS TO DO WHAT PEOPLE WANT. AND YOU WANT TO BE ABLE TO LIKE, USE IT AND BE CONFIDENT THAT NOTHING BAD IS GOING TO HAPPEN. AS THE MODELS HAVE GOTTEN MORE CAPABLE, THE SAFETY CASES THAT WE NOW USE HAVE TO BE HELD TO A HIGHER AND HIGHER STANDARD, AND THE MODELS CAN DO MORE. SO WE WILL HAVE MANY GREAT NEW MODELS TO COME. BUT WHAT WE HAVE SAID

[01:14] IS THAT ALIGNMENT, SAFETY, MONITORING, SECURITY HAVE TO STAY WAY AHEAD OF CAPABILITIES. AND WE HAVE HAD MODELS IN THE PAST WHERE WE SAY, THIS IS NOT QUITE AT THE BAR WE WANT, AND WE'LL MAKE ANOTHER ONE AND FIX IT AND LAUNCH IT LATER. BUT THERE'S CLEARLY MORE ATTENTION ON THIS NOW. >> OKAY. WHAT ABOUT THIS PARTICULAR MODEL WORRIED YOU OR FREAKED YOU OUT TO THE POINT WHERE YOU WOULDN'T LAUNCH IT? >> I WOULDN'T OVERROTATE ON THIS ONE THING. IT WAS LIKE, YOU KNOW, THIS MODEL WAS A LITTLE BIT WORSE ON A FEW OF THE EVALS WE LOOK AT. THIS WAS, THINK OF THIS IN THE ABUNDANCE OF CAUTION CATEGORY. OKAY, THERE'S NOT A BIG SCARY THING IN THIS CASE. >> DOES THIS COUNT AS A PAUSE? AND I WONDER WHAT THE PRESIDENT

[01:48] WOULD SAY IN A MOMENT LIKE THIS IF HE IS CONCERNED ABOUT SLOWING DOWN TOO MUCH. WHAT HAVE YOUR CONVERSATIONS BEEN LIKE? AND DO YOU WORRY THAT HE LOOKS AT A HEADLINE LIKE THIS AND SAYS, THIS COUNTS AS A PAUSE HERE? >> AGAIN, I WOULD PUT THIS IN LIKE THE NORMAL COURSE CATEGORY OF, YOU KNOW, OFTEN WE BUILD A MODEL, WE TEST IT, IT DOESN'T MEET OUR STANDARDS, WE CHANGE IT, WE LAUNCH IT LATER. BUT AS WE HAVE TALKED ABOUT, WE ARE PACING OUR PROGRESS, WHICH INCLUDES SOMETIMES NOT TRAINING THE MODEL, THE. EVERYBODY WANTS GREAT MODELS. WELL, ACTUALLY NOT EVERYBODY. MANY PEOPLE WANT GREAT. WE WANT GREAT MODELS, BUT WE HAVE TO DO THESE IN A WAY WHERE WE CAN MAKE VERY CONFIDENT SAFETY CASES AND CLAIMS SO THAT PEOPLE DON'T

[02:32] HAVE A BUNCH OF ANXIETY AND CAN RELY ON US FOR THE MODELS WE'RE PUTTING OUT INTO THE WORLD. WE ARE NOW IN A PERIOD WHERE THE MODELS ARE GETTING SO GOOD, SO FAST, THAT WE THINK IT IS PRUDENT AND WISE TO MAKE MORE PROGRESS ON THINGS. WE WERE TALKING ABOUT SAFETY ALIGNMENT, MONITORING SO THAT WE CAN CONTINUE WITH PROGRESS. BUT I THINK THIS IDEA OF PACING IS REALLY IMPORTANT. IT'S NOT JUST LIKE, STOP IT BECAUSE, YOU KNOW, WE DON'T WANT SMARTER MODELS. OF COURSE WE WANT BETTER MODELS. IT IS MAKE SURE THAT WE IMPROVE THE AREAS WE NEED TO CONFIDENTLY AND SAFELY CONTINUE WITH PROGRESS. SO I. >> THINK THE PACING DOESN'T ALWAYS MEAN SLOWING. AND I THINK THAT'S SOMETHING THAT HAS CONFUSED PEOPLE. >> YEAH, IT MEANS DOING IT IN A

[03:07] RESPONSIBLE WAY WITH GUARANTEES THAT THE WORLD CAN RELY ON. >> HUGGING FACE. WE GOT TO TALK ABOUT THAT FOR A SECOND. THAT WAS JUST THIS WATERSHED MOMENT, AND IT FEELS LIKE WE'VE HAD EXAMPLES EVERY OTHER DAY SINCE THEN. DID YOU UNDERESTIMATE AT THE TIME THE DEPTH OF THESE ISSUES? I MEAN, THERE'S BEEN REPORTS THAT THERE ARE THOUSANDS OF THESE. HAS IT SURPRISED YOU? >> WE'RE TRYING TO I THINK PART OF THE WAY WE WANT TO DO THIS IS TO REALLY OVERCOMMUNICATE AND BE EXTREMELY TRANSPARENT. SO THERE'S NOTHING ELSE THAT I'M AWARE OF OR THAT WE'RE AWARE OF THAT IS AS SERIOUS AS THE HUGGING FACE INCIDENT. BUT WE ARE TRYING TO PUT OUT EVERY SMALL OR OTHERWISE CASE WE CAN FIND. AND WE REALLY, AS PART OF OUR MISSION, WANT THE WORLD TO

[03:45] UNDERSTAND WHAT'S WHAT THESE MODELS ARE CAPABLE OF. ONE OF THE THINGS THAT'S SPECIAL ABOUT OPENAI IS WE HAVE A LOT OF BREADTH. WE DO CONSUMER, WE DO ENTERPRISE, WE DO DEVELOPERS. AND, AND PEOPLE HAVE VERY DIFFERENT WAYS OF THINKING ABOUT WHERE WE ARE WITH A