Open-weight models give enterprises control over AI deployments

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 14:34

   intensive purposes what they need general enough for their tasks are enterprises adopting open weights model like a hundred percent you have no clue like it's crazy wow like people don't realize But it's like, how much have enterprises used to adopt the open waste through the course of this year? Just the course of this year? Through the course of this year, really. Why just this year? Just because it got way better?

2. 14:56

   I think it got way better. I think that really, like, there was a threshold of quality that was needed and a threshold of size that was passed in which you could have models that were intelligent and capable enough at a size that was manageable enough, such as an enterprise who, say, actually take an environment they own and deploy it through Obama or VLM or Inferno, whichever tool they use for inference, and easily have it running in an environment where it is reliable enough for their production and intelligent enough for their tasks.

3. 15:29

   I think at the point that people don't realize what's really impactful for the language modeling industry was the release of Kimi K3. I think it was the first language model that really caught the attention of not only AI practitioners, but really like the non-AI first market as to like a tool that was just as capable as the frontier models here and open weight. And Kimigata initially was very hard to host.

4. 15:53

   It was heavy, whatnot. Then other alternatives came. Now we have American models that are also coming that are just as good. And, you know, people are just switching as much as they can towards it. I think it was somewhat predictable. Like, when you look at, like, the adoption of operating systems in the enterprise, it's not like macOS and Windows. Like, Linux is really, by far, the most widely used operating system in the enterprise world.

5. 16:15

   Yeah. But what's the main reason? Is it mainly just because of cost, or do they care about things like security and privacy and just ownership? Both, both, both. Like, no, cost for sure is a component. You know, like, it's always easier to forecast gigawatts than tokens. Like, honestly, for an enterprise, it's honestly a question they should ask themselves if they're truly spending a meaningful amount of money in language model inference.

6. 16:34

   But secondly, you know where the model is running. You know who's editing the model. You know which version you're using. And you know that no one is interfering with the ways you're running and the model you're running. You know where it's hosted. You know who's hosting it. You know which version you're running. And you know when the next upgrade is coming. you have full control. And enterprises love control of everything they're running.

7. 16:56

   That's all good and bad for them. You know, I go to enterprises that are still using Windows XP and they shouldn't. But the truth is that for them it's really important when you're running an operation at large scale where a lot is at stake. You need to have consistency and reliability. And nothing gives you more reliability than owning everything you run internally. Yeah, no, that makes a lot of sense.

8. 17:15

   I mean, so many companies that are dependent