Sam Altman in conversation with Tyler Cowen (You tube discussion shot a few weeks ago)
Question: If I am thinking of re-structuring an entire organization to put GTP 6 or GPT 7- What should I be doing?
SAM: Blah Blah Blah… No clear answer… I am sure he has a perspective but there was no clear quidence. Please hear it for your self?
A better interpretation can be extracted from the discussion below:
Sam Altman and Ali Ghodsi: OpenAI + Databricks, AI Agents in the Enterprise, The future of GPT-OSS – (Jump to right moment of the discussion.)
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?
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 their maturity where we are clearly seeing together huge enterprise demand. 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 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. This feels like it’s really the time for I think in 2026 2027 we’ll see a huge transformation 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 that would be quite a big deal and thrilled to get to go do that together 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 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 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.
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 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 do with model intelligence or very 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 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 even though they had crazy intelligence. So can be a problem. Yeah,
I’m sure. 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.
The other category we have seen a dramatic increase in how long of a horizon a model can work on a task for. 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.
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. I mean I I think that one is the 50th percentile task completion time horizon is super interesting. and I think if you bring more context so the first thing you mentioned also with more context 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 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 which like basically inspired by genetic algorithms
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
After listening to those impressive comments from Sam and Ali: Take aways
- Focus on boring stuff… You Company data must have quality and secure access.
- Enterprise AI use case is very different to consumer AI use cases. i.e much longer context window and duration.
- Every enterprise software function needs to become a tool for the AI agents. Convert all enterprise apps to enterprise level secure re-usable and context driven components.”Bring data close to app”
- Build a platform for tinkering in the company used by casual business user.