Main content

Dreamforce 2023 - a lot of generative AI value actually comes from hallucinations, says OpenAI founder Sam Altman

Stuart Lauchlan Profile picture for user slauchlan September 13, 2023
Summary:
Some visions of a future shaped by generative AI from Sam Altman. Enthused or scared? Take your pick!

Altman
Altman answers

The biggest surprise is just that it's all working.

A self-deprecating comment from Sam Altman, the wunderkind behind the current generative AI revolution that’s dominated the entire tech agenda for the past six months. Altman popped up at Dreamforce this week to discuss…well, go figure!

Articulate, informed, highly intelligent, he seems still slightly surprised about what he’s set in motion:

When you start off on a scientific endeavor, you hope it will work. You let yourself dream that it'll work. You kind of have to have that conviction, the optimism, but you know when we got together at the beginning of 2016 and I said 'Alright, we're gonna build Artificial General Intelligence', that's great. But then you meet the cold hard reality. We had a lot of stuff to figure out. Figuring out any new science is always hard and in this case it was particularly hard.

We had conviction in the path laid out by my co-founder and Chief Scientist Iya [Sutskever]. But having that conviction and then actually getting it to work...very much the consensus wisdom in the world was this was not going to work. Through the effort of lot of mostly talented people, it did. That's probably the biggest surprise.

He adds:

In terms of how we make progress, are we surprised? Not really. We are empiricists. We know we're going to be surprised. We know that reality has a different way of playing than you think it's going to when you sit there and think your beautiful theories in your ivory tower. So we're never surprised. We just try to meet reality where it is involved with technology and where it can go.

Alignment

As to where that is, Altman says:

On this current hill that we're climbing with the technology of the GPT series, we're going to keep making it better. We'll make it more reliable, more robust, more multi-modal, better at reasoning, all of these things. We also want to make sure that it is really useful to people in all the ways that it's transforming things. We're now deep into the face of the enterprise really adopting this technology and getting the systems to be very secure, very highly trusted, handle data appropriately.

A challenge remains around the threat of so-called hallucinations thrown up by generative AI, he concedes, but makes a controversial pitch:

One of the sort of non-obvious things is that a lot of value from these systems is actually related to the fact that they do hallucinate. If you just want to look something up in a database, we already have good stuff for that. But the fact that these AI systems can come up with new ideas and can be creative, that's a lot of their power. You want them to be creative when you want and that's what we're working on.

That’s a bold  point, but not one that’s likely to assuage the legitimate concerns around ethics and values in relation to generative AI. Altman counters:

We want to have the smartest, most capable, most customizable models out there because we so believe this will enable humanity and that the good will be worth managing more than the bad. But we also want to make sure that our models are aligned and that enterprises can trust us with data, that you're very clear about our policies around that, and that we're keeping not only the data secure, but also all of the learnings that a model has after ingesting a person's data. So, keep the model very secure in itself...but also making sure that from a safety and privacy perspective, we are keeping pace. We want to continue to lean on both of those.

Some might argue that that’s wanting to have your AI cake and eat it. It certainly raises the question of how such an alignment of interests can be managed. Altman admits that this is a common question he gets and it’s clearly not one that he appreciates being asked:

I think the frame of it is nonsensical. What's happening here is we have this amazing new piece of the human tech tree called Deep Learning, and that approach is solving lots of problems. That same thing which is helping us make these very capable systems is how we're going to align them to human values and intent...Can we think about it like a whole system? We have to make a system that is capable and aligned. Not that we have to make a capable system, and then separately, go figure out how to align it.

He goes on to argue that what we’re seeing right now is intelligence emerging from “a very complicated computer program”:

Intelligence is somehow an emergent property of matter to a degree that we probably don't contemplate enough, and that it can happen with electrical signals flowing around us that are connected in certain ways. It's something about the ability to recognize patterns and data. It's something about the ability to hallucinate, to create, to come up with novel ideas, and have a feedback loop to test those as we study the [AI] systems, which are easier to study than the human brain for sure. Without doing a lot of damage there's no way we're going to figure out what every neurone in your brain is doing. But we can look at every neurone in GPT.

The shape of things to come

The models are going to get dramatically more capable, he argues:

They'll be dramatically more customizable and dramatically more reliable. The model itself, that is the fundamental enabler of everything else. We will continue to build features around the model for enterprise class usage, and we'll continue to build consumer products to make it easier for people to just start playing around.

But in the same way that the internet and mobile just kind of seeped everywhere, that's going to happen with intelligence. Right now, people talk about AI. There was a time after the iPhone launched that people talked about their mobile strategy But no software company says they're a mobile company now, because it'd be unthinkable to not have a mobile app. It'll be unthinkable not to have intelligence integrated into every product and service. It'll just be an expected obvious thing. Companies will have their AI agents that can go off and do things and customers can interact with and all sorts of other things that we're seeing right now. But this will be a big shift in terms of how we interact with the world with technology.

And this will be beneficial to business, he insists, by amplifying individuals capabilities:

One person with a really good idea and understanding of what a customer needs is going to be able to execute on that [which] would have taken complex, many, many person teams before. [It's] that ability to give people tools and let things happen with less resistance, with less friction, faster, easier. It's very easy to kill a good idea. It just needs one person to be a little bit less than supportive. Creative ideas are very fragile things. And so giving someone the ability to do more, I think it's going to lead to a big shift.

Altman certainly isn’t limited in his assessment of what improvements to GPT can offer in terms of new ideas and capabilities:

The one that I would be tempted to say is the most important is the ability to reason. GPT4 can reason a little bit in some cases, but not in the way that most of us use that term. When we have models that can discover new scientific knowledge at a phenomenal rate, when we let ourselves imagine a year where we make as much scientific progress as a civilization as we did in the previous decade or previous century, think about what that would do to quality of life. That's possible.

My take

An interesting discussion that appeared to divide the Dreamforce audience as far as I can tell from people’s reactions. For some, the potential future that Altman pitches is exciting and empowering for humanity; for others, there’s a lot of unanswered questions about how his theories map onto societal realities as we know them. But as Altman himself reminded us, the current generative AI frenzy isn’t even a year old. This conversation has barely begun.

 

For all the highlights from Dreamforce 2023, check out our dedicated events hub here.

Loading
A grey colored placeholder image