Howie Liu: commitment

Transcript context

…How many tokens you use? Yeah, I mean, I'm proud to say I am pretty sure I'm still the ... I just checked this recently, but I take pride in being the number one most expensive in inference cost user of Airtable AI, not just within our own company, but I think for a long time I was globally across all our customers vault. I mean, I'm extremely intentionally wasteful. Wasteful in the sense of I'll do something that costs maybe hundreds of dollars of actual inference costs. For instance, doing a lot of LLM calls against long transcripts of let's say, sales calls to extract different types of insights like here's the product apps, identify or here's summaries, et cetera. And we also have now a capability that's basically like an LLM map reduce. So effectively, even if you can't fit the entire corpus of content into one LLM call, because the context window limitations, we'll map through all of this content and break it up into chunks and then perform an LLM call on each one and then perform an aggregation LLM call on those chunks. Very expensive, because you're basically running a highly expensive model against a lot of data and then running it again on the aggregates of that. But for me, hundreds of dollars spent on this exercise is trivial compared to the potential strategic value of having better insights. It's as if a really, really smart chief of staff has gone through and read every single sales call transcript that we've had in the past year and giving me very astute product insights, marketing insights, kind of positioning insights and segmentation insights. That's invaluable. You could pay a consulting firm literally millions of dollars to get that quality of work. So to me, I still think the value versus the actual cost of AI when applied greedily but smartly, it's a crazy ratio. And more people should be aggressively throwing compute cycles at these very high value problems. Until somebody tweets how you're costing the company so much on AI compute and you guys are going to be underwater.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.