Amazon’s CTO built a meeting-summarizing app for some reason

In a post on Werner Vogels’ personal blog, he details Distill, an open-source app he built to transcribe and summarize conference calls.
© 2024 TechCrunch. All rights reserved. For personal use only.

How does Amazon CTO Werner Vogels — a man worth untold millions, who during the COVID-19 pandemic bought outright the small Central Amsterdam Airbnb he’d been living in — spend his days? From the looks of it: building AI-powered meeting-summarizing apps. Go figure.

In a post this week on Vogels’ personal blog, he details Distill, an open-source app he built with his “OCTO” (Office of the CTO) team to transcribe and summarize their conference calls. Distill takes an audio recording of a meeting (in formats like MP3, FLAC and WAV), analyzes it, and generates a summary along with a list of to-do items. It can optionally spit out that summary and list to platforms such as Slack via custom integrations. 

An example summary from Vogel’s Distill meeting summarizer, powered by Amazon tech.Image Credits: Distill

As one might expect of an app from Amazon’s CTO, Distill relies conspicuously on paid Amazon products and services to do the computational heavy lifting. AWS Transcribe carries out Distill’s transcription; Amazon S3 provides storage for the meeting audio files; and Bedrock, Amazon’s generative AI development suite, handles summarization.

But why create a meeting summarizer when there are countless tools out there that would satisfy the purpose? Well, I have to imagine that Vogels thought, why not? He has tons of resources at his disposal and seemingly enough spare time for hobbyist programming projects. Per the blog, he’s already trying his hand at porting Distill’s codebase from Python to Rust. (Being the CTO is nice work if you can get it.)

One unique thing about Distill is that it lets you select which AI model performs the meeting summarizing. By default, it’s Sonnet, a mid-range model in Anthropic’s Claude 3 family. (Amazon’s large stake in Anthropic might’ve had something to do with that design decision.) But any model hosted in Bedrock will work, like Meta’s Llama 3 and models from AI startups Mistral, AI21 Labs and Cohere. 

Vogels doesn’t promise that Distill won’t make mistakes. 

“Remember, AI is not perfect,” he writes. “Some of the summaries we get back … have errors that need manual adjustment. But that’s OK, because it still speeds up our processes. It’s simply a reminder that we must still be discerning and involved in the process. Critical thinking is as important now as it has ever been.”

I’d argue that having to be “involved” in summarizing kind of defeats the point of an automatic summarizer. You might as well hire a stenographer. But you’ll never catch Vogels badmouthing the tech his employer’s selling. And that, I’d wager to say, is why he’s still CTO. 

 


Comments

Leave a Reply

Your email address will not be published. Required fields are marked *