How does Amazon CTO Werner Vogels — a person value untold thousands and thousands, who throughout the COVID-19 pandemic outright purchased the small Central Amsterdam Airbnb he’d been residing in — spend his days? From the seems of it: constructing AI-powered meeting-summarizing apps. Go determine.
In a put up this week on Vogels’ private weblog, he particulars Distill, an open supply app he constructed together with his “OCTO” (Workplace of the CTO) crew to transcribe and summarize their convention calls. Distill takes an audio recording of a gathering (in codecs like MP3, FLAC and WAV), analyzes it, and generates a abstract together with a listing of to-do objects. It might probably optionally spit out that abstract and checklist to platforms reminiscent of Slack through customized integrations.
As one would possibly anticipate of an app from Amazon’s CTO, Distill depends conspicuously on paid Amazon services and products to do the computational heavy lifting. AWS Transcribe carries out Distill’s transcription; Amazon S3 gives storage for the assembly audio information; and Bedrock, Amazon’s generative AI growth suite, handles summarization.
However why create a gathering summarizer when there are numerous instruments on the market that may fulfill the aim? Nicely, I’ve to think about that Vogels thought, why not? He has tons of sources at his disposal and seemingly sufficient spare time for hobbyist programming tasks. Per the weblog, he’s already attempting his hand at porting Distill’s codebase from Python to Rust. (Being the CTO is sweet work if you will get it.)
One distinctive factor about Distill is that it lets you choose which AI mannequin performs the assembly summarizing. By default, it’s Sonnet, a midrange mannequin in Anthropic’s Claude 3 household. (Amazon’s giant stake in Anthropic would possibly’ve had one thing to do with that design resolution.) However any mannequin hosted in Bedrock will work, like Meta’s Llama 3 and fashions from AI startups Mistral, AI21 Labs and Cohere.
Vogels doesn’t promise that Distill gained’t make errors.
“Keep in mind, AI shouldn’t be good,” he writes. “A few of the summaries we get again … have errors that want guide adjustment. However that’s OK, as a result of it nonetheless quickens our processes. It’s merely a reminder that we should nonetheless be discerning and concerned within the course of. Vital pondering is as essential now because it has ever been.”
I’d argue that having to be “concerned” in summarizing sort of defeats the purpose of an computerized summarizer. You would possibly as properly rent a stenographer. However you’ll by no means catch Vogels badmouthing the tech his employer’s promoting. And that, I’d wager to say, is why he’s nonetheless CTO.