Mar 17 '26, Tuesday
Mar 18 '26, Wednesday
Mar 19 '26, Thursday
Mar 20 '26, Friday
26:51·2 SPEAKERS·3468 WORDS
AIt's. How's it going?
BHi. Hi. I'm good. How are you?
AI'm good. Is it Srajan?
BYeah, that's perfect. What about yours?
AMihaly.
BMihaly. Hi. Nice to meet you.
ANice to meet you too. So I don't know too much about you except that like we got some indication that we should have a chat. So why don't you tell me a bit about yourself and like what you're looking for.
BSure, yeah. I just give you a brief like history and maybe then you can like ask questions wherever you want to, like discuss more. I graduated doing computer science from IIT Bombay and that's where I was friends with Sumit who introduced us. Did a bit of like machine learning research during that time. And then I interned at Jane street which essentially like intoxicated me with like public markets. And I was like super interested in just learning more about them, which is why I got into like the HFT industry for a couple of years. And during that time we were like running some good trades, but I just wanted to learn more and build like more of a theoretical background for what I'm doing. Which is why I did like a master's in financial engineering here in the US at UC Berkeley, after which I joined another famous high frequency trading firm called Jump Trading. I was there for about a year. I realized that there were a couple of things not working out. I was a little tired of just doing abstract research and making a small number go up. And I also wanted to work on something more collaborative. Like working at a HFT is pretty cutthroat. It doesn't give like it's. It doesn't feel cooperative enough. Which is why I moved to like a startup called rippling. It's a B2B SaaS company. I was working there as a software engineer and that's really when like AI started getting popular, like AI assisted coding. And that's basically. I saw it happen at my job at Rippling, where users started like engineers started using cursor, myself included. And yeah, I felt like people weren't using it as much as they should. And I was like, I would consider myself to be an early adopter there. I was like, probably also very high on like the token count usage metric they were tracking at that time. And yeah, like I could see how this is going to change everything. And which is why I wanted to move closer to something AI specific, AI adjacent. Which is why I joined Microsoft in their Copilot tuning team. This is more of an internal product right now. We're dogfooding it. Are you aware of Mistral's recent Forge launch?
AMistral, Forge, very briefly, yeah.
BIt's basically like giving enterprises the ability to fine tune their own models without having a specialized team that does it for them and it ingests everything about the enterprise so that they can have models that work the way they do essentially. And I thought that Microsoft is very well placed for building something like that because they host a lot of enterprises already. A lot of enterprises data resides on Microsoft servers and they would trust Microsoft for this kind of AI to be built with them. I still think Microsoft is very well positioned to do that, but so I've been here for like three months now and I realized like it's very slow like even compared to rippling which wasn't like, which is a fairly big company. It's just way too slow. Like I get bogged down in processes more than actually being able to do work like getting reviews, getting accesses, lots of permissioning issues and things like that just slow you down much more than they should and just hard to like work at the speed that I want to. Yeah, that's pretty much it. And I read about amp Sumit told me a bit about what you're planning to do and sounds very exciting. I think. It's getting more and more clear that a few individuals with the right ideas and can leverage compute to essentially scale their idea. So we need a mechanism for making it easier for people to do that.
AYeah, no, I totally agree. That's. That's what we're trying to do essentially here like you said. So just to give you like a little bit of context about amp since I don't know exactly what you heard. So trying to like contextualize it, we as of yesterday put something up on amppbc.com anjane put it on X as of yesterday. So that's like a more like easy to reference like public facing thing that is now available. But like in general, yes, we are trying to align capital and compute for small teams so that they can scale both without having to like build a whole infra team and spend a whole lot of time thinking about the compute itself. Because generally to scale that you either become very inefficient or you have to spend a lot of time thinking about it. Right. And so neither one is great. Especially if you're spending like hundreds of millions of dollars on something you don't want to be inefficient with it. So that being said, like we are like near the start of Our journey in doing some of this, obviously there's a lot to go on with Onge being like very connected in like the Frontier Lab ecosystem. And like myself and Sebastian recently joined from Google, so we were working on some of the like AI infrastructure there. We're each kind of leading a different part of that and working together there. So we're like, oh, we should keep working together here. So that's kind of like the engineering team so far and we're getting started on like helping these labs with compute and then like continuing to build on top of that, like actual software that can, that can take it beyond, you know, what we would do initially in a spreadsheet or something. So it's very early, I guess, like the context that we weren't really given that I'm sort of deducing from what you're saying is like are, are you sort of like looking for a job to see if we have any openings? Are you looking for some kind of like. Could you clarify that?
BYeah, I am looking to work in a, in an early AI based company, essentially something that's in sf, which is why I'm reaching out here. And I think AMP is like, I like also personally feel like eventually compute should be a commodity right now it trades like real estate essentially. You have two year contracts and like, yeah, it's an illiquid and fragmented market. So yeah, I'm essentially looking for to join a team which is working on something, something that's required and something and a team that's small enough to move fast and I have like high leverage.
AOkay, so why don't we then talk a little bit about like you told me a little bit about your background. Right. Why don't we talk a little bit about like that in some depth. Right. Like kind of like what exactly you built in these places or like what you worked on or like what that looked like just to set some expectations. Like we're not really hiring right now since we just got off the ground. We may be hiring soon. Right. The way we're thinking about our team is the same that we think about like every, every team that we invest in. Right. So what you were just saying about like small talent density teams being able to scale capital and compute. We also are aiming to have a small talent dense team that is able to scale. You know, we want, we want to sort of dog food that we don't want to have like a thousand people just because.
BRight.
ASo the way that we're thinking about hiring is a little bit based on like, okay, what are we going to build when what are our needs? Like what's missing on the team, what can we automate versus not automate, like that kind of stuff and then hire to like fill in the needs more dynamically. Which is why I'm saying like even though we're not hiring right now, we may be hiring very soon as we, you know, get farther down a certain path or other. Okay, so does that sound good?
BYeah, that sounds good. Makes sense to me. Yeah.
AYeah.
BSo.
ASo then why don't you tell me a little bit more about like what you worked on. It sounds like at Microsoft you're pretty early so maybe like that's not useful but like you, you said you worked on at a couple high frequency trading places and then also at Ripling.
BYep.
ASo maybe just give me like a first overview. Like what was your role like?
BYeah, at a high frequency, like the first high frequency trading firm, it was called Tower Research and that was like pretty broad. I was, I used to trade commodities, oil and gas at local exchanges. Local being usually meaning Shanghai and Mumbai. So you would have signals from, from cme, the Chicago Mercantile Exchange and you would use those to trade trade those in these local exchanges because CME usually dictated the prices and these would follow. It was a relatively standard lead lag trade. And yeah, my job was quite like broad in that I had to develop the signals which would be based on like the book structures, how to combine those signals to get actual alpha, how to define order entries based on that alpha based on a simulation framework that we had and then also monitor each of those trades and see how they're doing in real time. Are there slippages occurring? Where is the difference between simulation and reality and ironing those out and just tracking how well those trades are doing. So it was pretty end to end. That's essentially what I did at Tower Research mostly the commodities trade at jump trading I was responsible for. So they were like trying out a new experimental team which was going to be news based trading. So you know how like the Department of Labor publishes reports at 7:30 in the morning and essentially it was high frequency trading but applying like front running that news, like could you get to that news as fast as possible and then push it into an alpha pipeline that parses it and understands how to react on the market. That was more of an infrastructure project where I was involved in understanding how the government is actually pushing these out and like monitoring what time they actually go out, what CDNs they're deployed at. Can we get servers co located at those particular locations. We already had a big spread of servers and just wanted to know we could just ping with those servers and find out where and try to like locate where, how the distribution is happening across the CDN and just. Yeah, it was more investigative in trying to understand how like the deployment of like how the publishing of this news happens. I wasn't involved on the quantitative side there which was to actually use that news to generate a signal, but it was more of a investigative infra based project there. I was also like on another team there which was doing some options based curve modeling and stuff. But that was fairly generic work of just modeling and like how do you fit, how do you fit a curve? How does the curve move with like different movements in the underlying stock price and stuff like that. At Rippling it was relatively generic software engineering work where I had to like build payment based tracking solutions for users. So one of the things that. Are you familiar with Rippling?
AWe actually use Rippling for amp, so.
BOkay. Okay, cool, nice. So one of the things we need to do as part of payroll is pay taxes to particular countries. And one thing that is really flexible in Rippling is you can hire anywhere and we handle the taxes in that jurisdiction for you. So a lot of the times those processes are not automatable because of the government regulations and a human has to submit these payments manually. So my job was to how can we make this as easy as possible for those people doing those payments and tie those manual records that were created into our actual database so that it's easier for our actual auditing to track all of that. So we iterated over multiple designs. The tax team obviously did not want to move away from the JIRA based tracking that they were already doing and we wanted it in a certain way. So I built like an adapter around JIRA so that they wouldn't have to change anything and we would still get our data. In the database as we needed it. And yeah, this is one of the things that I did. I could go into more details on other projects if you like, but at Rippling I essentially just learned how to use AI based tools and speed up my developments velocity a lot more.
AOkay, yeah, cool. You don't need to go into too much detail. I was just kind of trying to get the overview of your roles. That sounds good. So I think I have like a pretty good picture of like what you've worked on and stuff so far. And like how you describe yourself. Is there anything else that you want to chat about like at this point or should we just say, like, let's circle back when we're actually hiring?
BYeah, no, that, that, that might be a good. Yeah, this might be a good point to stop. Maybe just one thing that I might want to know is, like, when you do start to hire, what does your current project roadmap look like? What kind of engineers would you be willing, likely to hire? Is there a specific kind of workflow that you're targeting that we begin with tests and that's the end result we want to have and derive all the code from there, or is there a specific framework that you're thinking about?
AYeah, so I think ultimately the things that we want to build, there's a pretty good mixture of things. And a lot of it also is going to be depending on direct feedback from the labs that we're working with. So I'm not going to share the full. Here's the exact product roadmap and the details right now, but I can give you a general overview of what that looks like. For example, as you might imagine, a lot of these labs aren't using their compute efficiently because they just don't have something to run at a certain point in time and then a bunch of other labs do. And so one of the main things that you could, you could look at across the system, like amp or like if we develop a system, is like pooling together a bunch of demand and filling in the gaps, like, efficiently. Right. So if someone's not using their compute, how can someone else use it? And vice versa, in that there's a lot involved, right? There's. You kind of have to build a scheduler. You have to work with, like, customers to figure out, like, just aspects of it should look like where they need to like, have a touch point or like a surface that they interact with. And then you need to understand how the scheduler works and so on. You want to make sure it's super secure and that doesn't affect their performance or anything like that. So there's like a bunch of aspects like obviously testing, benchmarking, watching for failure rates, watching for like, utilization anomalies. Like, you, you can really branch out into like a thousand different things. So I, I wouldn't say, like, there's like a specific thing there that I would highlight that it's like, oh, we're only going to do this. Part of the initial scope is like, how do you basically get like into the scheduling orchestration kind of layer? Right. Because that is a, that's a way that you can like, make it work across labs as opposed to if you're trying to get it to like, you know, optimizations of like a kernel or something like that. That's a lot more like workload and lab specific. Right. There's more to build on from there, but I would say that's probably a good starting place for that specifically. We already have, like myself and Sebastian have worked on similar systems at Google and so we have some of the, like how to get started there, but like what specific parts of that end up resonating very well with the labs taking off. We need an owner for like that kind of stuff. It's, it's kind of like determined a bit on the go.
BI see.
AWe're also generally, I think, going to be looking into how to procure more compute for these labs. Not just like how to procure, but like making that really efficient and more automated and stuff like that. So that's kind of like the overview. But again, right now I, I don't think like we are ready to hire for any role until we kind of like have a more burning need for that, which may happen in the near term, but like, I just don't have that visibility yet.
BOkay. Yeah, sounds good. Yeah, thanks a lot for taking the time and telling me about amp and yeah, hoping to hear from you in the future, like when you do like plan to expand.
AYeah, no problem. Great to talk to you, Srojan, and good luck at. You're at Microsoft right now. I know you're looking around and it's slow. I've been there as well at Google. I hope that it changes for you. I think it ever. Every company there's like, you know, teams that are slower, teams that are faster as well. Which, you know, not speaking as like ampere, but more as just like a guy that also worked at a big tech company until very recently. I've had, I had a lot of experience on my first team at Google that I also felt was like kind of infused, deteriorating of like it feels slow. It feels like nothing really is worthwhile is getting done. And like, I don't really feel like I have agency over any of the changes or anything like that. Right. But I actually left that team and went and did a like rotation as a product manager. And then I was like, just kidding, I don't, I don't want to be a product manager. And then so I was kind of lost. I was like, oh, where do I go? What do I do? And I ended up deciding just to like find like a small team within the company where I felt like I could have more of an input on, like, what gets done. And also because it's small, there's a lot less, like, red tape. Right. And so I went to, like, a really small team. It was just, like, one other engineer year, and then she left within, like, two months of me joining. So it was just me. And that ended up being, like, a really fun team. I. I was there for about four years after that, and, like, we grew the team to, like, 12 people or something like that, but. But the product itself just, like, exploded, so. And there I felt like, even from the start, I had a lot of agency and a lot of, like, ability to move quickly, get things done, balance priorities and whatnot. Right, Nice. It feels very different to work on that team as opposed to, like, any of the first things that I worked on, even though they're within Google. So I. I would say, like, consider that as an option as well, as you're, like, poking around and looking at these things. But no, I totally get it. Like, it's really frustrating.
BYeah, yeah, no, I'll keep that advice in mind. Thanks for. Thanks for that.
AYeah, yeah, no problem. All right. Have a good weekend.
BYeah, you do. Nice to meet you as well. Bye. Bye.
There is a human-machine superorganism growing, and we can't even see it because we're inside of it. The way that lichen grow in rings and circles even though the individual flecks don't know where they are. Claude is a big circle and it's clustering energy and resources on the planet and the people who work for it.
Mar 21 '26, Saturday
Mar 22 '26, Sunday
Mar 23 '26, Monday
Open this page on desktop.