Sam Altman and Ali Ghodsi: OpenAI + Databricks, AI Agents in the Enterprise, The future of GPT-OSS

Thank you all for joining the webinar today. Uh we're very excited to have

today. Uh we're very excited to have with us a discussion uh between our CEO Ali Gatsi and uh via Sam Alman uh CEO of OpenAI. So thank you very much both for

OpenAI. So thank you very much both for for being here.

>> Thank you. Yeah, excited.

>> Yeah. Well, let's get started and I hope we can have a fun kind of thoughtful discussion. um you know data bricks and

discussion. um you know data bricks and open AAI we recently announced kind of first-of-a-kind partnership uh to make the open AAI models natively available within data bricks and agent bricks so

maybe actually starting with Ali first you know what's the importance of partnering with open AI on this on this endeavor >> yeah I mean every one of our enterprise customers want to use openAI like you know uh they all want to have the models

available they want to use them on their enterprise data uh and you know they're you know getting just those two things working together is non-trivial because it's like the data is sensitive They want to have their privacy, you know,

they want to have auditing, they want to have GDPR rights, but they want to like, you know, they want to use, you know, the models to build agents to get insights. So, it's like it's really

insights. So, it's like it's really customer demand that uh, you know, has been overwhelming. Yeah, we're we're

been overwhelming. Yeah, we're we're thrilled to do this. Enterprise is

becoming uh one of our biggest focuses.

We've had like 7x enterprise growth this year and uh you know we we really think that we're heading into the phase of AI where the models are getting so good that enterprises will need to use them want to use them bring them into their

uh into their whole ecosystem and we cannot imagine a better partner than data bricks to to make that happen. So

we're thrilled. We're also very happy to be a data bicks customer.

>> Yeah know that's amazing. you know, as as you said, you know, the partnership, not only is of course data bricks bringing open AAI models to the M, you know, all of our enterprise customers, um, but also for some time now, OpenAI

has been using data bricks for data analytics. So maybe starting with maybe

analytics. So maybe starting with maybe Ali, like how has that partnership and that work been going?

>> It's awesome working. I mean like extremely technical teams that push us to you know our boundaries and uh you know demand more and one more of us and uh you know it's nice that the two companies are so close geographically so

it's just easy to walk over and iterate on the product. Uh but uh yeah it's a one of our most demanding customers in a very good way like made us just way way way better as a company.

>> Um maybe moving separately a little bit you know Sam you and open have really revolutionized AI for consumers. Um how

do you think on this new chapter as you mentioned about going to enterprise how data bricks and open AI together you know can revolutionize AI for enterprise as well?

>> Yeah we always plan to do enterprise also but the models started with a lot of problems they started fairly weak. It

was easier to get consumers to adopt them. They're coming to the point in

them. They're coming to the point in their maturity where we are clearly seeing together huge enterprise demand.

Um, and the need to figure out how to bring this technology to enterprises in a way that they can use it with all their data, with all their constraints, with all of the concerns they have about

security and safety, but but also uh to sort of quickly jump into this new world where you'll have AI doing an increasing amount of the intellectual

work in an enterprise. um this feels like it's really the time for I think in 2026 2027 we'll see a huge transformation um of how enterprises think about this maybe the the example

we can look at from 2025 is what's happened with how code gets written >> right yes >> and now if you imagine that for every other function of the enterprise uh that would be quite a big deal and thrilled

to get to go do that together >> yeah I'm super excited I mean I think there's a lot of uh context in the data that enterprises have it's very proprietary right So uh the consumer side initially I presume started with

like okay we have all this data in the world that's you know public mankind is created over thousands of years what if the AI could really sort of compress that and understand that and reason on that but this then there's this like

data that's like not available to the LMS which we now can bring together so now you know providing the data that they have that's proprietary uh in the enterprise as context to the agents

which is really what they need that I think is going to be like the big unlock just excited to uh build that together.

>> Yeah. And you know on that point both Sam and Ali on sort of the model capability reaching a point where you know enterprises make a lot of sense.

The other piece on top of that has been this sort of agents kind of age of agents sort of arising and I want to ask both of you a little bit more on the research side. you know now with AI

research side. you know now with AI research moving um from you know building great great frontier models to now including hey you mentioned context management right you mentioned agents

and and multi-agentic patterns and systems where do you think the future innovations like across these research buckets will really kind of drive this move into into enterprise >> I mean I think there will be all of

these different research buckets like we'll keep pushing on pre-training we will you know get these multi- aent systems that are doing complicated things the models will get smarter across the board. But there are maybe

kind of two axes that I would think about here. Uh one is nothing at all to

about here. Uh one is nothing at all to do with model intelligence or very little. It's about how tightly can you

little. It's about how tightly can you integrate it into an enterprise, all of the knowledge they have, all of the different data sources, their business processes, the the way that they want to work. Like you could bring a very

work. Like you could bring a very brilliant physicist and drop them into data bricks and maybe they wouldn't get that much done on the first day because they'd be missing all of that understanding. they would know how to

understanding. they would know how to even though they had crazy intelligence.

So >> can be a problem. Yeah,

>> I'm sure. Uh so so I think there's this one layer that the whole ecosystem together this will not just be either our companies but the whole world is going to have to build

kind of an integration uh into these models and sort of with these models and that is that is a lot of work to really figure out how you're going to you know teach a model to be a sort of productive

employee in a particular enterprise or whatever. Um and then the other category

whatever. Um and then the other category we uh we have seen a dramatic increase in how long of a horizon a model can work on a task for.

>> Yeah. Yeah. So the way the way that I think about this is for a particular class of task once you've given it that sort of connection to an enterprise that kind of integration um

how long of a task does the model have a 50% chance of success at gone from if we think about say coding we've gone from 5-second tasks at the

launch of GPT3.5 to 5 minute tasks with various GPT4

iterations to 5-hour tasks with um GP5 and that's remarkable. Uh but a lot of

what an enterprise does is tasks that require months or years sometimes and also we can only currently do this in this one vertical. So lengthening the

the horizon that these models can work on um giving them all of the context that exists inside an enterprise or the world and then adding that to like many more verticals I think will be the important thrust.

>> Yeah, I mean I I think that one is the 50th percentile task completion time horizon is super interesting. uh and I think if you bring more context so the first thing you mentioned also with more

context uh you know you can increase that horizon like you know lengthen that so the task something that would take a human a day to do the the the agents can now do um uh not that we have a day yet

but I'm saying if you add that context you can get that longer and longer and one thing that we saw that's really cool is just optimizing the context automatically like you know we we developed a technique called uh ja uh

which like basically um uh inspired by genetic algorithms But uh basically gets that enterprise context that you have and gets it into the model but automatically so that you don't have to have humans that are sitting there and

optimizing the context so that you can live feed the relevant context that's sitting in the different docs and inside of the enterprise. Uh so I think that's that's a really good metric to look at and see you know and the progression of

it has been like insane as Sam said like from like seconds now to like many many hours. So

hours. So >> very excited about that. It always yeah it always strikes me that the the model capability today is the dumbest it will ever be for the rest of my lifetime you know and that's like a incredible

thought to have whenever you think about the potential applications of this thing. Um and you know as as we

thing. Um and you know as as we mentioned sort of agents getting deeper enterprise operating on hours and hours and potentially even taking actions in the enterprise what are the main things

to think about in terms of ensuring kind of responsible AI and governance over the tools and the data and and the actions? Yeah, I mean we've worked very

actions? Yeah, I mean we've worked very closely on that and we've made sure that uh we we sort of build with privacy and security from the ground up. So the

whole uh partnership and integration builds on having the guardrails in place, making sure that you have audit logging on everything that happens in the enterprise so that you can track like what did it exactly do so you can go back and track those things. Having

access control uh making sure I mean simple things for enterprises that maybe you know sometimes takes for granted but uh you know is it on brand yeah what the model is saying is it recommending a

competitor things like that. So you can build those guard rails in from the beginning into the uh into the models and that's so if you use 35 now inside data books you get all of that out of the box. Uh so we're very excited about

the box. Uh so we're very excited about that. This is I think this is going to

that. This is I think this is going to become the fundamental limiter for adoption of AI in the enterprise. Uh it

won't be about the intelligence. It

won't be about the price. We're going to I believe the research teams and the sort of infrastructure teams will figure that out. But I think enterprises are

that out. But I think enterprises are are starting to realize very quickly just how critical this is and it will be the limiting reagent.

>> Yeah 100%. I agree. You know maybe this is a question for Sam but as we and data bricks can attest you know the open way model GPOSS have been a gamecher for a

lot of our companies both in terms of the capabilities and the use cases can tackle the customizability that an open way model you know involved and I think openai just recently released uh the GBD

OSS uh safeguards version as well >> curious yeah what's the future for open source and open weights from from open AAI >> there's clearly demand for for it.

There's honestly much less demand than there is for the most capable uh you know models that we can run in a cloud.

But I think we the answer is we're you know a bigish company and we should do both. People people want a model that

both. People people want a model that they control and that they can run on their device or in their own system and there needs to be some version of some way we support them.

>> Yeah. Um so I am curious like we're seeing huge demand on GP2oss like people really want you know especially like you know an American open source model that like kind of uh uh is at the very

frontier. Uh what is open opens plans

frontier. Uh what is open opens plans going forward? Like is is G2S the last?

going forward? Like is is G2S the last?

>> No no I I I hope not. Um I

we are trying to figure out how we can do a someday do a model that is like the quality of GPT5 uh running on a you know one device open source W model. We don't

know how to do that yet but uh we didn't think we were going to get a model as good as GPT OSS running you know in 120 billion parameters either. Um, so

again, I don't think this will be the what most people want, but the people who want it really want it. And I think we'll find we'll try to find ways to offer incredible open source models.

>> Wow, that's pretty amazing. So you could run on your, you know, if you could get that form factor down running on your laptop.

>> We we think about what it's going to look like to make a sort of, you know, a new computer for like the AI era. I

think that there's a whole bunch of reasons why that form factor is not quite right.

>> Yeah.

>> But one part of this is it should run locally a great model if you wanted to.

Yeah. Yeah.

>> So if Wi-Fi is down, you know, you should AI >> if Wi-Fi is down or you're just like, hey, from a privacy perspective, I I I I believe that privacy and freedom will be

two extremely important principles for how people use AI if it becomes as important in people's lives as we expect. And the current

expect. And the current tech industry, public policy, cluster, whatever, leaves some area for concern.

>> Yeah.

>> And and so I think people will want good local models. Oh, that's super exciting.

local models. Oh, that's super exciting.

Yeah, looking forward to that future.

>> Um, you know, maybe this is a question for Ali, but uh we as a community today are so focused on vibe coding and, you know, these small kind of agentic apps in the short term. What are we underestimating about this technology in

the long term, especially for for the enterprise?

>> Yeah, I mean, I think Sam touched on this. I mean, we're going to see it.

this. I mean, we're going to see it.

We're going to have co-workers that are agents and they're going to there there's so many different aspects of the work that they're going to be involved in. Like even on coding now we're like

in. Like even on coding now we're like just thinking about okay can you produce the code but you know coding at databbooks is only 20% of what R&D or an engineer does. We measure it it's only

engineer does. We measure it it's only 20% of their time what are they doing the rest of the time but it's like design docs it's PRDS it's you know discussions meetings but every one of them when you start looking at it it's like there's a place for agents too and

actually if agents were involved in the whole process it would probably also make the coding much much much better.

It's context that's missing when they're writing the code. So, we're going to see, you know, I think end to end for, you know, all of this kind of productive work. We're going to see them uh show up

work. We're going to see them uh show up and help accelerate everything that's happening inside of the enterprise. And

it's going to happen, you know, slowly.

It will take time, but that task horizon is just getting longer and longer and longer, which means they can take on more and more complicated tasks over time. So, it's exciting times.

time. So, it's exciting times.

>> Yeah. Yeah. Um, and then maybe Sam, for you, I think you've said before that we're in this like once in a history kind of transition, right? Okay, now the world is changing much faster than than

we think. Um maybe if you zoom out 5 10

we think. Um maybe if you zoom out 5 10 years, what's like the most I don't know surprising institution, right? Schools,

cities, religions, hospitals that you think will look unrecognizable, right, with those generations.

>> I think once in history is too big of a statement. Okay, I can say once in a

statement. Okay, I can say once in a generation.

>> Once in a generation. Yeah.

>> Um the econ I think the economy changes a lot. I I I don't think it's I don't

a lot. I I I don't think it's I don't think we're going to like run out of meaning or purpose or fulfillment or anything like that. Um, I think humans have such main character energy. We're

so able to like not care about what the machines do. Like already, you know,

machines do. Like already, you know, GPT5 is probably smarter than most people and they seem entirely unbothered. I feel pretty unbothered

unbothered. I feel pretty unbothered even though I'm like, "Oh, you know, I just had said host or whatever." Um, I so I don't think we have any of those crisis of meanings in quite the way that

or I'm I'm less convinced. Maybe we do.

Future's hard to predict. I'm less

convinced than other people are, but the economy feels like it should just absolutely radically transform and I you know maybe the whole like maybe the structure of it changes in a pretty big way.

>> Any uh predictions or thoughts on your end on this Ali?

>> I think you know Sam has you know he he's the ones who touched on are those are excellent. I'll give boring

are excellent. I'll give boring enterprise you know what I see in my kind of vicinity which is uh you know I think uh you know there's like this whole um you know in sales for instance in B2B organizations there's sales

engineering you know but I think that's actually even more than R&D uh this kind of technology agents is applicable there uh the marketing function inside of a

B2B function like there's so much that the agents already doing at data bricks like you know in the marketing funnel like preparing u campaigns and you know preparing the material that's for the

different audiences all that that's already being done by agents. Uh so I think we're going to see that uh change a lot of the inside of an enterprise like data bricks there's so many different op ops functions and the

people that basically are analysts they're very smart people with financial backgrounds you know they're typically sitting in excel every day um a lot of that is already being completely transformed and we're going to see that

actually be even more revolutionized so at least when I look at inside of a company like data bricks it's pretty much every function is going to be completely transformed that doesn't answer the society question but but you

know there's a lot of enterprises out there that kind of a little bit look like data bricks and I think already they're they're going to be vastly different in five years or maybe three.

>> Yeah. Yeah. And maybe you know I know we talked about 5 10 years from now but what are the most exciting enterprise AI use cases that that you both see today.

>> I mean there's there's so many uh and you know it starts like it we're taking on the smaller tasks. So like when you zoom in it might not look like that exciting but it's like what Sam said with the task horizons. Uh but you know

like one one one thing that I thought was interesting. Astroenica had a use

was interesting. Astroenica had a use case with agents that basically sift through 400,000 documents which is something no human could have done.

>> Yeah.

>> Uh and it's uh you know our in in finance we have a lot of financial organizations that are now going through all of the SEC filings all the related documents and they're gleaning out alpha for investments that then they're giving

to analysts that are then using them.

Humans that are using them but it's just this would be extremely tedious. human

would have to do that before and now it's just you know we take for granted it's like yeah of course sift through all those thousands of pages and give give back the response quickly to me. uh

in insurance we're seeing it you know underwriting again like in in in the healthcare space there's if you think about if you go to a hospital there's like thousands of thousands of pages of papers that you have to sign that are

produced for every hospital visit and I don't know if anyone ever even looks up those uh but now in the healthcare sector they're having agents more and more companies and startups are having actually all of that being fed to LLMs

that are going through it and you know extracting tidbits figuring out risk uh so so much of society is going to be changed uh but it's going to happen you know as you said with those you know simpler and simpler tasks that gets more

and more complicated and you know over time increases I I super agree with that answer but just to add two other directions I find very interesting um obviously it's great that people are

doing things you know faster and better and cheaper but the categories of things that you just wouldn't do at all before AI >> or and you already touched on this but things that you just simply couldn't do

are are both interesting to me so there's more focus on stuff you couldn't do like read 4,000 pages of documents.

And I think that's amazing that AI is just now churning through this. And

>> um it's so superhuman at literature search as a specific example of what that's meant.

>> Um but this wouldn't do category just doesn't get as much attention. So if you I use coding again because it's the kind of thing I have the best mental personal

mental model for. Um coders they can do their work faster but they also there's just features they would never think they would never try. that that would be at the bottom of someone's forever. I

will never do do that. I'm just not ever going to get there. And now if I can just like kick off one extra agent, I am willing to do that. So seeing there's just like there were these good ideas that you just wouldn't do. And if you

make the activation energy low enough, you make it cheap enough, you just you say, "Okay, fine. I'll try it." That's

been like an interesting thing for me to see.

>> Yeah, that's awesome. Yeah.

>> Yeah. So yeah, the ability to launch async parallel basically agentic jobs, right? uh really changes how you think

right? uh really changes how you think about the process of actually doing work which is which is very exciting.

>> Yes. I mean it's uh it's so cool because it's so fast and you can do so much with it but now I feel like I'm so frustrated like why is it not why is it not like why it's taking so long like a minute this to write 20,000 lines.

>> You know this is an area that I think we as an industry are underinvested in is how how we make this stuff dramatically faster.

>> Yeah.

>> 100%.

>> We're going to push on that.

>> Yeah. Exciting.

>> Yeah. I mean we've done this in our database why Lindy is king right for any interactive experience whether it be a dashboard and the same principle applies to agents as well >> industry took a bet that that uh which I think was the right bet if you had to do

one or the other which is we were going to prioritize cost per token over you know token to token latency and design most hardware that way there's clearly demand for a much higher cost per token at a much lower latency

>> interesting is that now we're getting peaks into the hardware you're building >> the hardware that we'll eventually build >> not the first thing we'll do but We want to we want to do that at some >> Yeah.

>> Yeah. Great. Uh we have a lot of uh folks dialing in here from uh you know where data scientists or data engineers or some even cos and leaders at at these

enterprise companies. Um with this sort

enterprise companies. Um with this sort of generative AI revolution maybe what's like one thing that you know CEO or leader should do like immediately now to like prepare yourself right for this

wave that's already here and it's like continuing to to roll to roll forward. I

think uh you know I'll start with the boring stuff because it's uh it's people people um you know they're excited about you know what they can accomplish with AI which is great uh but before you get there you have to have a good data

foundation and you have to make sure that you can secure it and you can make sure that you have the right guard guardrails in place and then also getting access to as much of your data in the enterprise. there's a lot of data that's siloed and tucked away like

that's sitting in a database that's not accessible or that's you know um actually not even this is on premise it's not in the cloud and so on so making sure that you just can get all of that in one place uh so that it's

accessible as context to the AI will be kind of like that differential whether the AI immediately gets it right whether it's within the task horizon or whether it's uh you know it's just trying many many things because you just didn't give

it that little context that's sitting tucked away in that silo data set that's exactly what I was going to I think that's I think that's it sounds so boring but it's so important to get in this work.

>> Yeah. Yeah. And and you know we've seen sort of natural language is great but it's also ambiguous. So you know when when our Ali knows this or when our customers ask hey please compute the uh

churn for me. Well what is what is churn? Like it's it's really dependent

churn? Like it's it's really dependent on the enterprise context and and the enterprise use.

>> Yeah it is I mean it is that definition exists somewhere in the organization in a company. how you define revenue or how

a company. how you define revenue or how you define churn or what when a fiscal year starts or what the quarters are um and you know so it's just can you find that as a context is is that is the AI

connected to that if it is it will be able to figure it out because the models are so freaking smart now and they're getting smarter every release that comes u they're getting smarter and smarter so we just have to make sure we can feed that context though there's a little bit

debate raging now on like whether these agent patterns are just agents and tools in a loop or whether you got to handgineer all the individual uh pieces of the agents and wire them into some

deterministic kind of kind of workflow.

the models what you see if models can like do novel math they're certainly smart enough to figure out how to do this out without hand engineering >> may take us a little bit of time and again you do need in the same way that an employee needs some training needs to know where to look where the enterprise

stuff is you'll have to like make that knowledge available but they will not need custom training for it >> yeah totally agree isn't a bitter lesson uh >> it is it is a bit lesson bitter lesson you know it's like and we've seen it with most of the agents that people are

building the successful agents when when you when when the people built them talk about it's like hey I actually did the simplest thing I didn't have lots of lots of lots of agents that call each other and lots of hand tune things. I

just let the AI figure it out.

>> Yeah. Yeah. Great. Um, thank you both for uh joining us today. We really

really appreciate it. Uh, for the folks who are dial in here, you know, if you want to get ready and get started uh using uh open a on data bricks, uh please check out uh AI uh playground and

also our agent bricks product uh to get starting building agents uh with open AAI and the GBT5 models on data bricks.

Thank you both for being here. Thanks so

much. Great discussion. Thank you.