← All conversations
People & ideas

DevReal: Parth Sareen Interview

Parth Sareen ↗With Alexy KhrabrovFeb 20, 20254:16

FunctionalTV interview or Q&A with Parth Sareen.

Listen

Download MP3 ↓

Follow the words

Read the transcript

hello everybody I'm Alexa kro the hello everybody I'm Alexa kro the founder and organizer of the longest founder and organizer of the longest running deest technical met up in the running deest technical met up in the world bya AI here we on location at world bya AI here we on location at Cloud flare we have an meet up about Cloud flare we have an meet up about ageny structured output and with us we ageny structured output and with us we have barin from AMA who is a software have barin from AMA who is a software engineer uh working on structureal engineer uh working on structureal yeah welcome thanks for having me so yeah welcome thanks for having me so super hot everybody's using it to host super hot everybody's using it to host their own things how did you end up in their own things how did you end up in Alama how and how why did it become so Alama how and how why did it become so successful yeah I guess that's a great successful yeah I guess that's a great question um I've only really been here question um I've only really been here for a few months I joined back in for a few months I joined back in November um I knew Michael and Jeff were November um I knew Michael and Jeff were the two co-founders of olama for a while the two co-founders of olama for a while and uh one conversation led to another and uh one conversation led to another obviously Lama's been doing pretty well obviously Lama's been doing pretty well so I've just it was just super exciting so I've just it was just super exciting for a new opportunity uh prior to this I for a new opportunity uh prior to this I was working on my own startup um around was working on my own startup um around agents actually and was using oama and I agents actually and was using oama and I loved it so it it just kind of made loved it so it it just kind of made sense interesting and before the startup sense interesting and before the startup like what did you do in software yeah um like what did you do in software yeah um I've done many many different things um I've done many many different things um ranged all the way from like iot ranged all the way from like iot Hardware to going to um software and Hardware to going to um software and distributed systems at like Tesla as distributed systems at like Tesla as well as worked on Siri at Apple for a well as worked on Siri at Apple for a couple couple different roles um through couple couple different roles um through internships um all super fun and yeah internships um all super fun and yeah wow okay super cool so like you've seen wow okay super cool so like you've seen the whole spectrum and so you ended up the whole spectrum and so you ended up like this foundational layer right why like this foundational layer right why like what's exciting about it what's so like what's exciting about it what's so interesting yeah that's that's a great interesting yeah that's that's a great question um for that we'll have to go question um for that we'll have to go back a few years so back in 2019 I was back a few years so back in 2019 I was working on natural language processing working on natural language processing um and this is like back in when you um and this is like back in when you were importing Spacey in Python oh yes were importing Spacey in Python oh yes very old stuff um now but remember the very old stuff um now but remember the Burke model like coming out and looking Burke model like coming out and looking into embeddings and encoding different into embeddings and encoding different things and what I really didn't like things and what I really didn't like about was just how poor the tooling was about was just how poor the tooling was and I like pretty much switched paths and I like pretty much switched paths from ml to distribute systems after that from ml to distribute systems after that um and eventually kind of Drew back into um and eventually kind of Drew back into more low-level programming and that's more low-level programming and that's pretty much what we do at AMA um and the pretty much what we do at AMA um and the first time I used AMA it was just first time I used AMA it was just tooling which for ML which just worked tooling which for ML which just worked and I was just super excited about that and I was just super excited about that mhm and uh how did you you guys like mhm and uh how did you you guys like decide that search your output is decide that search your output is something important for Al to do oh for something important for Al to do oh for sure um I would say that lm's best are sure um I would say that lm's best are used when you're working with used when you're working with unstructured data and structuring that unstructured data and structuring that data using llms because they have this data using llms because they have this understanding of the world but also what understanding of the world but also what you're passing in is so critical and so you're passing in is so critical and so structured outputs really just allows structured outputs really just allows you to go from something with no you to go from something with no structure to a very structure Json with structure to a very structure Json with everything you can need yeah no I mean everything you can need yeah no I mean so so now I understand you joined in so so now I understand you joined in November in December I met you at November in December I met you at unrapped yes and you get this talk and unrapped yes and you get this talk and like you look like one of these guys like you look like one of these guys whove been there forever he was few whove been there forever he was few weeks of the job so good job than than weeks of the job so good job than than so that's exciting right that's a so that's exciting right that's a Serendipity we in Bay Area do things and Serendipity we in Bay Area do things and then one thing is list to the other and then one thing is list to the other and now you're speaking to this med up right now you're speaking to this med up right so which is spreading the good word uh so which is spreading the good word uh very cool so you know uh what's kind of very cool so you know uh what's kind of uh on the road map for you is structural uh on the road map for you is structural boot an easy thing is there a lot of boot an easy thing is there a lot of work to be done there what's kind of an work to be done there what's kind of an in store yeah that's a great question so in store yeah that's a great question so more recently we've been kind of been more recently we've been kind of been working on ama's new engine uh which is working on ama's new engine uh which is purely written in go um and so with that purely written in go um and so with that I'm also working on a new structured I'm also working on a new structured outputs module uh to do sampling um and outputs module uh to do sampling um and so we should have some exciting things so we should have some exciting things soon hopefully around that so hopefully soon hopefully around that so hopefully faster and more accurate outputs Ino yes faster and more accurate outputs Ino yes wow this very cool so I see other things wow this very cool so I see other things than python take in charge right this is than python take in charge right this is great well thank you so much par we great well thank you so much par we looking to your talk and great to have looking to your talk and great to have you here at the middle thank you thank you here at the middle thank you thank you

Recovered English captions. Automatic transcription may contain errors.

Keep exploring

Follow the guest, their work, and the ideas behind this conversation in the Devreal knowledge graph.

Parth Sareen on Devreal ↗
Independent by design

Your player.
Your subscription.

One permanent feed. Listen in the podcast app you love, with the conversations always at home here.

https://struct.fm/feed.xml
113 audio episodes available in the feed.