Samsara CEO在播客中讨论AI视频推理用例与算力成本下降

【播客逐字稿片段 · The MAD Podcast (Matt Turck)】
单集:The Biggest AI Deployment Nobody Talks About | Samsara CEO Sanjit Biswas
片段:第 18/24 片(全集切片后按关键词筛出本片) · 本片命中:wafer、inference
说明:以下为该集逐字稿的**连续原文片段**(未删改;英文档为 YouTube 自动字幕/节目原文,可能有识别误差)。请只就本片内容判断是否构成**可交易**事件 —— 公司/产品/订单/产能/需求/合作/资本开支/高管表态等;泛泛的观点、预测、行业感想不算。若本片无可交易内容,判为不做。

--- 片段原文 ---
So maybe the compute combined with the algorithm combined with just like smarter model architectures will make what would have been like a one-day task a 1-hour task or you know even faster than that.

And if you suspend disbelief a little bit, what do you think you would be able to do uh in like a year or two?

You know, not 10 because obviously who knows.

Um but uh given the progress, so right now you're able to do um uh you give the example of warranty, you give the example of uh sort of pre-planning a day >> uh for for a driver.

What what else do you think you can do in the next year?

I I think a lot of our ability to predict what happens next year or two is by looking at what is like barely possible now and then what will the sort of um cost curves look like or or capability curves look like and um something that we talked about at our beyond conference last month was this idea of a of a driver ride along.

So, in operations, it's quite common to basically have a manager sit with you over the course of your day, like drive around with you for 8 hours.

And what they're doing is they're not looking for, you know, uh, how fast you're going.

They're looking for your habits of like, do you check your mirrors, like, you know, are you alert and are you aware?

That kind of thing.

That's basically a massive amount of, uh, video computation, right?

So, you can run a tokenizer.

It just turns into like a lot of compute.

You can do that today.

It's it's pretty expensive and costly, but it it works and it's doable.

We're pretty optimistic that the you know cost per million tokens is dropping so fast and the capabilities are rising that we can deliver that at better and better cost over time to our customer.

So I think in a year or two that will be possible and I have to say like you know just over last weekend I was playing um Sarah which is one of those big like wafer scale chip companies.

they have uh an inference model that you can run on their chip which is basically Gemma 4 but like hyper accelerated like that's a great example of like that is nonlinear in terms of jump right like what you get out of these big models today if you run Gemma 4 on your GP you might get like 100 tokens per second if you have a fast card if you run it in their cloud you get anywhere from 800 to 1500 tokens per second so call it 10x faster um that is the kind of thing where it unlocks these new use cases that we couldn't get to because it would have been too either expensive or too slow.

>> Not to promote the podcast on the podcast, but by the time we release this uh the prior episode will be precisely an episode with Andrew Feldman of Serial Brush if anybody missed it >> and of and of commercial.

I'm I'm glad you mentioned the the the ride along because I think there's a fascinating aspect to uh the whole dash cam almost from a societal standpoint.