Ashwin Gopinath on Building Organizational Memory for Enterprise General Intelligence at Sentra
What we talked about
Ashwin Gopinath is the Co-Founder and CEO of Sentra, where he is building organizational memory for what he calls Enterprise General Intelligence (EGI). His thesis is that AGI will be commoditized, and the real differentiator will be how well a team retains and compounds its own knowledge, since every re-litigated decision is a memory failure. Before Sentra, Ashwin was an Assistant Professor at MIT, a research scientist at Caltech and Google X, and co-founded Palamedrix, a proteomics company acquired by Somalogic for $52.5M in under two years.
Show notes
Ashwin Gopinath left a career spanning MIT, Caltech, and Google X, and a proteomics startup acquired for $52.5M, to build Sentra, a platform for what he calls Enterprise General Intelligence. His bet is that AGI will be commoditized and the moat will be how well a team retains and compounds its own knowledge, turning meetings, messages, and documents into searchable organizational memory.
What we covered
- What Enterprise General Intelligence means, and why Ashwin believes every re-litigated decision inside a company is a memory failure.
- The problem he watched teams keep running into that made him decide Sentra had to exist, and how the product turns meetings, messages, and documents into searchable memory.
- His path from MIT professor and proteomics founder to building an AI platform, and what made him make the jump.
- How Sentra flags missed commitments and conflicting priorities before they escalate, and how the team figured out what counts as a real risk versus noise.
- Why his thesis that AGI will be commoditized changes how companies should be investing in their own knowledge today.
About Ashwin
Ashwin Gopinath is the Co-Founder and CEO of Sentra, building organizational memory for Enterprise General Intelligence. He was previously an Assistant Professor at MIT, a Senior Research Scientist at Caltech and a Research Scientist at X (Google’s moonshot factory), and co-founded Palamedrix, a proteomics company acquired by Somalogic for $52.5M. His work lives at the interfaces of biology, physics, and computation.
- LinkedIn: https://www.linkedin.com/in/ashwin-gopinath
- Website: https://www.sentra.io
Episode 161 of the PreVetted Podcast.
Full transcript
Federico Ramallo (00:00) Welcome back to the Pre-Vetted podcast where we spotlight extraordinary people and remarkable talent reshaping our world. Today I am joined by Ashwin Kopinath. He’s co-founder and CEO of Centra, where he’s building organizational memory for what he calls enterprise general intelligence. Before Centra, Ashwin was an assistant professor at MIT, a research scientist at Caltech.
and Google X and co founded Palametrics, a proteomic company acquired by SOMALOGIC for fifty two point five million in their two years. His career has lived at the intersection of biology, physics and computation. Ashwin, welcome to the show.
Ashwin Gopinath (00:49) I should be here.
Federico Ramallo (00:52) So for people meeting you for the first time, can you tell us a little bit more about what you’re doing today?
Ashwin Gopinath (00:59) Yeah, so so the thing that I’m working on right now is asking the question at a at a high level, you can think of it as how do you make sure that if you have a group of people, how do you make sure that they are all working towards the same goal? How do you make sure that there is least friction between them? And to, in some sense, get them to behave as a collective intelligence or like as an entity that is larger than themselves.
Like, how you get them to kind of have least friction between them so that if somebody learns something, everybody sort of learns from that experience? How do you get a group of people to think and work towards one single North Star with least friction between them? That’s what Centra is essentially intended to solve. That’s what we end up calling as enterprise general intelligence. You know, more popularly nowadays.
People are talking about it more as like company brain. How do you actually make sure that everybody sort of has a single sort of understanding and working towards the same goal?
Federico Ramallo (02:03) Right. Right. Very interesting. Yes, you you you talk about Centra being an enterprise general intelligence, right? and you’re kind of describing what w what it’s about. And what do you think is unique about this this software?
Ashwin Gopinath (02:22) Yeah, I d I wouldn’t necessarily think of it as a software. I would think about it more as a tool that plugs into pretty much every piece of information generation tool, that is meetings, messages, emails, whatever other tools that one might be using within an within a team or an organization, and extracts what is important for each individual.
Have access control so they are there so so that people who are not privy to certain information don’t get it, and like surface it as they are working. So I think people are working with agents a whole lot more. People have are working with other AI tools all the time nowadays, and I think that is going to just keep continuing. Even otherwise, if people are on a typical day when they are working, they are trying to understand information from different tools and different places.
Federico Ramallo (02:53) Right.
Ashwin Gopinath (03:13) We want to reduce the friction so that the information that is necessary for them at that time automatically sort of shows up for them. And there is and they can continue doing their work without ever having to fall out of, like go and look for context. So the information that is necessary for them is always available for them.
Federico Ramallo (03:36) Right. Right. I I understand. The I mean I the the the closest and probably forgive me if this this is not the right comparison, but the closest thing that this reminds me is confluence, having a wiki, organizational wiki for for for the company. Of course this is pre agent, right?
Ashwin Gopinath (03:56) Yeah, I I I I mean so so so wiki and all of these information, so people always had systems of record. It is not as if systems of record did not exist. Systems of records always existed. That is, there is a single source of truth where people need to go and look. But somebody needs to maintain it. Like all companies have got documents that gets generated that never gets read. Nobody maintains it. And when you read a document, you’re looking for only
pieces of that document you don’t read the entire document there’s something relevant for you that you need to go and look at so now there’s a lot of friction people have like I think people have done these analyses where like 95 to 97 percent of all documents that are generated nobody reads okay so so it so how do you actually make sure that the right thing gets gets to you and somebody maintains it and it is actually accurate all of these things need like that like
So so confluence and wikis always exist, but somebody needs to maintain it. Like what we want to do is have the system sort of ephemerally exist in all pieces of data that you’re working with, all tools that you’re working with, and constantly extract the information so the system knows who is working on who what, what ideas have been generated, what ideas didn’t get used, how things are transcribing or like transpiring in the organization.
You know, and and you know what didn’t work, what worked, and who worked on it, what decisions were made, you know, and and it’s just exists there. So as you’re going about doing your work, it just pops up and it just corrects you or like gives you the information that you need.
Federico Ramallo (05:36) And I think that one of the issues we had with Confluence is that it’s always outdated. Even for a day, an hour, you know, by design it’s always outdated because we we humans make decisions that we are not so good at going back to the document and putting it back again, right? so we have to kind of remind ourselves or I have to remind the team, my team, to go back and
Ashwin Gopinath (05:56) Correct.
Federico Ramallo (06:03) Remember to update, right? And that that was a lot of friction, right? And then we when we moved to tickets, w we used confluence in the tickets, right? for tasks to to track tasks. we had a lot of documentation there that sometimes was not going back to the to the confluence as well, right? So that that has been a l another friction. what we ended up doing at one point is we start doing for onboarding video recordings of of
you know, if you are the new person I’m going to show you how we do things. And that became the new knowledge base, if you like, right? But I mean it it’s imperfect in many ways, but for somebody new then we can send them to a list of videos to watch, right? It was not very efficient, but it it kept the record of knowledge updated, right? so what what I I I see you’re doing with Century, you’re fixing that problem.
Ashwin Gopinath (06:38) Mm.
Federico Ramallo (06:57) because you you now have agents that can collect that information and summarize it, so you’re always keeping that relevant information safely stored.
Ashwin Gopinath (07:09) Correct. I would I would the only place where I want to correct is it’s not summarization, right? It so memory is not summarization. I think this is the problem that I think most people have in a they they confuse memory for knowledge. Okay. Memory is not knowledge, and no like memory is not being able to retrieve anything perfectly. Memory is different. Like a good way to understand what memory is.
And how it’s different from knowledge is. Let’s say you have three people looking at the Golden Gate Bridge. Okay, let’s say you have a designer, an engineer, and a biologist, just for example. They’re all standing at the same place, same time, looking at the Golden Gate Bridge. Knowledge would be if somebody takes a picture of that, right? Memory is what each of those people took away from that experience. The designer might remember the color of the bridge. The engineer who might remember the sort of
The cables, the 90,000 miles of cables that were there. And and and the biologist might completely forget the bridge and might remember like some dolphin or the whale. Now, why is that? That is because that’s that’s because it mattered to them. Memory is part of the knowledge that matters to a system or a person. So, like from an organizational memory point of view.
You don’t need to remember everything that was said in a meeting. You need to remember what was important in a meeting. Who attended the meeting? Was the meeting about a particular project? Okay. What changed about the project? What decisions, why did you change it? Who said what? Those are the things that are important. Everything else is like you can forget about it. It’s like
You know, so so so that’s the core difference, right? If you’re summarizing, you’re summarizing also noise, like all these other non-important things. So memory is not just knowledge. Like memory is not just storing everything in a in a in a in a file system and just forgetting about it. The way to think about it is memory is what happens when you it’s like a scholar reading a library and understanding things that are important. It’s not just the library.
Federico Ramallo (09:26) Right, right. I I see it’s a it’s a nuanced but powerful distinction, right?
Ashwin Gopinath (09:33) Correct, correct. See, because because each person in an organization cares about different things. The salesperson might look at the same piece of data and remember something different than the CEO or a product manager. And it matters, like otherwise, what happens is you you need to go searching for things. A true memory will bring the information for to you and will learn.
In some sense, the true memory is somewhere it’s not a database, it’s not a model, it’s somewhere in between.
Federico Ramallo (10:09) Right. Right.
So this also clear puts a clear distinction between having and again I apologize for the root comparison, but having a bunch of markdown files in a folder, right? Because what you have there is just a storage of text and information, right? which could be more or less relevant, but the difference
with centra is that you’re saving the information that is relevant for that particular context and then you can query it and make make it much more faster to for agents to get back to the to the inf to r relevant information they need to make the decisions, right?
Ashwin Gopinath (10:53) It’s kind of correct, but I think the way to think about it is it takes the data and then it structures in a way wherein later on, if an agent or an individual queries it, depending upon what the agent wants, it will re- it will extract and create the context that is relevant for them. So a good way to think about it is this, right? Like an again, example, right?
So if you look at a sales complaint, a customer complaint, right? An email, you can just store it there. But then the question is the system basically, if a sales associate asks about that customer, it will just tell them what is relevant for the sales associate. It basically says, This is the customer who has an ACV, how much what they are spending with you, and why, what what was wrong with the deal that that caused them to that that that complaint to occur. But if the same
Pro a piece of customer complaint was being pro like how it affects a program manager is the or the product manager is the product manager will understand the feature that did not satisfy the customer. They are two different things, they are in the same piece of data, but it’s relevant, different parts are relevant for different people because that’s what they care about.
Federico Ramallo (12:13) I see what you’re saying, yeah. So you see the information from different lenses depending on what’s relevant for each I would call it agents, but also humans, right?
Ashwin Gopinath (12:24) Yeah.
Yes, yes. So so the way in which we think about it is the memory basically, depending upon the ontological lens that you project onto it, the memory changes for you. Like this is again another way to think about it is like I love giving this example. It’s like if you are going on a hike, right? A boulder, okay, is just a boulder. Okay. You don’t you might not even remember it. But if you were tired, that boulder becomes a chair. You sit on it. Okay. So you will remember it as like a place where you sat.
not a place, you just not a random piece of boulder. So it it changes. F the same piece of information can change depending upon your perspective.
Federico Ramallo (13:09) Right. Right. I see I see what you’re saying. That’s very interesting. I I go ahead, yes.
Ashwin Gopinath (13:13) And right now and right now
there are no tools that does that.
Federico Ramallo (13:20) Right, right, because all of the tools focus on storage and retrieval without the analysis.
Ashwin Gopinath (13:29) It’s storage and retrieval, but typically they are for a process, for some purpose. Okay. So when you when you’re doing it for a purpose, you’re basically selecting what are the aspects of the knowledge that is important for that purpose. But it is not for all possible or like a large number of reasons why a company might need it. Right? So so so that that’s the nuanced difference.
Federico Ramallo (13:35) Right.
Right. Right. And the
The other question that comes to my mind is how do you manage access control? I mean if because in coming from a a more traditional approach, I would have teams having access to different types of information, right? You have a different scopes, right? so marketing people would not know exactly the the database, production database, they don’t need to.
But but eventually they they do need to have shared information, right? Because if you isolate it then you can make the wrong decisions, right?
Ashwin Gopinath (14:33) Correct, correct, correct. So the way in which we do the access control is very similar to how so the the core data, like the the source of the memory has access control. So well, let’s say it’s from an email or from a meeting, you know the people who were in the meeting, and whoever owned that meeting can decide, you know, what is the scope of that memory. The only difference is like tip the I think it’s like it’s
It’s more a cultural and psychological change. Wherein like right now, when you are having a conversation with somebody or when you’re in a meeting, you don’t think about it as everybody has access to information that is there here. Okay. So I think the company changes when you start saying, you know, like any conversation on any piece of data that is being generated, a lot of people’s, you know, it’ll impact them.
Okay, and then you can basically change the way in which should it be private, should it be public? And that that that that nuance can be also set up in a way wherein the data itself, in the meeting itself, if you were to say, okay, this is private, okay, then centra sort of like flags it and changes it. So so think about it this way, right? The best way to think about it is centra is think of it as centra is basically sitting in on all pieces of information.
And think of it as a person that like that an individual with infinite bandwidth who can be in all pieces of information, all pieces of data, all meetings, all messages, everything, and understand it.
Federico Ramallo (16:09) Right, right. That that’s why I was thinking about security because if it has to everything, it could possibly leak some information, right?
Ashwin Gopinath (16:19) Yeah, so to the to a to the to the extent possible, we are taking all the access controls that are there in file systems and transferring that onto this system. So part of the reason why we have been spending a lot of time developing it is to kind of go through all of these issues. And and in and in in the several design partnerships and like you know rollouts that we have been doing, and we have about half no about a dozen of them right.
dozen team that are basically using this. The problem never is access control. Like access control is least of the problem. I think, I think the the it’s not even issue, right? It’s just things that used to take weeks or days gets collapsed into minutes. So like if a KPI changes, okay, in like not only do you know that the KPI changed, but you can also now say
What decisions that somebody specifically made is most likely the reason for this KPI to change. Typically, things that would have taken like days of to kind of track down gets collapsed into like tens of minutes. As soon as the KPI changes, if you have flagged it, the system basically goes through all the things that happened and says this person, like Bob or Jack, said something or like there was like some supply chain delay.
that caused this KPI to change. So the learnings become so much faster.
Federico Ramallo (17:48) Because you’re not
I I think you you’re you’re exposing a very powerful concept that is all of the systems I mean I I I I go back to confluence because it’s what I’m I’m more used to, all of the system track the decision already made, right? What you’re tracking is how decisions are being made. Right? And then from there you can you can
Ashwin Gopinath (18:14) Yes.
Federico Ramallo (18:18) you you can tr have trussability towards all the decisions that are being made. and that’s much more powerful
Ashwin Gopinath (18:23) Correct, correct. You you so so you can go back
Federico Ramallo (18:28) go ahead.
Ashwin Gopinath (18:29) I mean, so one place is like KPI changes. I mean, that has been v highly impactful for like one or two clients that we have. Another client that we have wants to use this for or like are are are are using this for performance revenue, which in most organizations takes a lot of time every quarter. Okay. here you can just basically like the system has seen everybody doing everything. They it is track it knows who committed what into GitHub.
You who told who talked what discussed what in what meeting, what was discussed in Slack, all of that. So you can just immediately say this changed occurred, which was positive for the company. Who exactly did it? Like why did it exactly happen?
Federico Ramallo (19:15) Wow That it’s so powerful it
Ashwin Gopinath (19:18) And pr exactly, I mean the reason
why we are we are doing this is because like fundamentally all companies are like all companies are just a collection of people working with each other. It’s we are in a people business. Okay, so we wanted to actively work on problems that accelerated fundamentally human problems and solve fundamentally human issues.
Not really, I mean, yes, it can be used for you know speeding up things, you can be used for you know reducing token counts. There are a whole bunch of things that one can do, but fundamentally, we wanna basically augment and improve teams that are composed of humans. So we wanna solve human problems, fundamentally, human problems with AI.
Federico Ramallo (20:06) Right, right. that’s very interesting. And I can see how it’s it’s basically a git blame for business decisions.
Ashwin Gopinath (20:15) Yeah, that’s I mean, we call it like GitLog of the company in some sense.
Federico Ramallo (20:19) Right, right, yes, yes. That’s very, very powerful.
Ashwin Gopinath (20:22) But how you use it
how you use it is completely different.
Federico Ramallo (20:27) Right. Right. And are you tracking the decisions being made by humans and agents?
I mean it differently.
Ashwin Gopinath (20:36) Yeah, so we have
been trying to yeah, so we treat so we think about it this way, right? Like we think about everything that happens in a company to have like classes of objects, right? You have actors, okay? So that would keep that would be people, teams, companies, including agents. You have interactions, those are class of things, entities. So interactions would be meetings, messages, like agent human conversations.
Okay, then some decisions get made in all of those interactions. Something gets decided. Okay. You will decide to do something, and there is a rationale behind that decision. So the rationale and decisions and other things. Then you have commitments, wherein like discussions can be there, and then there is commitments to do something. Okay. And then there are value-creating objects. So these are services, products, something. Okay. So these are the classes that are important for the company. So agents are a class of.
Actors, humans and agents are agents and agents will interact with each other. Some decisions get made, and that decisions could be just you know completely you know non impactful to a value creating object. Okay, it might never affect anything. Okay, you might be just two people decide that you know they might you know share the workload or something, but if it doesn’t affect a product or a service, something that gives them money.
Like value creating objects, then the in the from the company’s perspective, it’s not that important. So every one of the so so these are the things that we are tracking. Like in all the messages, all the emails, all the agent-human traces, agent-agent interactions, we are tracking these things and creating a continuing record of it. So you can always trace back and say, Bob or Bob’s you know, Claude session, what were the decisions that were made? What were the discussions?
And you can trace back and there’s like you can constant you at any given point, you can connect to any one of these classes, okay, and you can basically make like go back, like what are the effects? So you say this decision, what happened from that decision onwards?
Federico Ramallo (22:43) Right. That’s very interesting. and the th the reason I was thinking about make making the distinctions i was because of of value for the business, right? I mean how how much value should we put to an agent making a decision versus a human making a decision, right?
Ashwin Gopinath (23:04) Right. I mean ultimately you could if you if you have this system, one of the things that we wanna do over time is, you know, if this is like six months or a year from now, as these kind of dozen companies or dozen teams that we are part of that’ll grow, okay, we can go back and ask the questions, you know, did the agents actually affect the bottom line, or was it like humans that made the decision and agents just acted on it?
Okay, you like so so, in some sense, you can think of this as like beginning or like tools like this and capabilities like this, and we are not the only ones who are doing it. Others are also developing it. So I don’t want to say that we are the only ones who are working on it. Other ones that like play like we have a certain approach to it that nobody else is taking, to our knowledge, but but others are also in and around this theme. But either way, this is like beginning of things like outcome AI. Like one way to think about it is like you can go back and you can play the play.
Play the tape backwards and say, you know, all these good things that resulted. Okay, why? Who was responsible for this? Was it an agent? Was it Claude? Was it you know, Chat GPT, or was it like somebody in the company that basically made these decisions? You can go back and say, you know, ROI of everyone and every decision within an organization. If it is sitting in the
If you have that entire data and structured in this way, we can actually start asking those questions.
Federico Ramallo (24:29) Right, right. I in in in my book where it The Invisible Distance, which I’m finishing almost finishing on the editorial part, I I told tell a story about a software engineer that built a feature but but but was laid off before the feature was launched, right? And the and the accolades of that feature that had an impact on the business.
the accolades was to the team to the current team. So kind of my my the moral of the story is that the system is designed to reward who remains on the team, right? and you know even though this person was you know the main contributor to that she was not able to get the recognition and by by that time
if I send the recognition, you know, personally to her, it was just opening a wound rather than just, you know, acknowledging her her contributions to the to the team, right? Or to the company, right? so when you you know when you were talking about this this idea of recording the decisions, remind remind me of of this story, right? Where we we’re we’re you know the si the current systems are not really good at at rec
Ashwin Gopinath (25:23) Mm.
Federico Ramallo (25:43) the recognition, you know, there’s too much pol politics around that. It’s a very inaccurate system, if you like, right? so the
Ashwin Gopinath (25:51) Completely
c I completely agree. I mean, it is one of the this is one of the reasons why one of the folks that we are working with very closely wanted to use this for performance tracking. Because they were like or or because typically what happens is the loudest voice gets the you know credit. You know, there are lot of people who contribute, but that contribution is completely lost because they are not like actively
Federico Ramallo (26:03) Right.
Yes.
Ashwin Gopinath (26:16) pushing it forward. I mean they might be like introverted. They are not like like they are not, you know, writing all of those things in the performance review. So that happens all the time. Like and and and and what we want is to do that truly at like a organizational setting. Like the good choices that you make, whatever it might be, we want to kind of lean into it. So you can think of it as as the data sort of like collects itself.
We can start really thinking about building a model of the organization. Because that entire trace can be used to build a model of the organization. And then you can start asking even more interesting questions like do a reinforcement learning on the right decisions that you made. Like if you hired the right people, okay, they might not be the like, like they’re not necessarily the loud person, not the most political person, but what are the teams that actually led to the right decision? It might be some.
Quiet person somewhere in the in the meeting might not even speak, might be just like like sending messages on Slack. Maybe they are the ones that are actually important. You see that all the time in organizations, you in teams. You have like this one amazing person who’s very not not necessarily very loud. and and and and finally, one of the other things that you can start asking is: is AI actually useful? Like we ask, we we go around and have this debate on you know what is the impact of AI in organizations.
Okay, things like what we are doing sort of starts answering questions of that nature. You know, after six months, after a year, you can ask, you know, was this, you know, was all this money spent on Claude or ChatGPT or whatever other tool? Okay, was it actually useful? Did it actually produce something that changed the top line or improved the bottom line?
Federico Ramallo (28:04) Right, right. I I am exploring on building this zero employees company to to see how you know to learn and see how it works. And what I’m learning, and we’re on the early stages so you know, but what we’re what we’re learning from the community is that there’s all this incredible, you know, s zero employees companies or IG companies that
are generating zero revenue, right? Because they have a great system but haven’t actually, you know, got into revenue yet, right? Or or they haven’t figured out how to do that yet, right? so I I I see what you know, I I I see the value of what you’re talking about. the the the other thesis that I I I I describe in my book is this idea that you know we can have AI arbitrage
We can have labor arbitrage, but the largest leverage that we can have is large context. If we have humans that have been in the company for many years, they know the all the decisions of the business in a much more you know, finer detail than somebody that just joins the company, right? So having a long change your team has a much bigger impact in the outcome than
any agent could have right now having said that if you if you provide to these people agents that can help on on on the load then these people can manage the agents in a much more effective way right which supports your thesis that having those having the recorded decisions provides and and knowing what works and what doesn’t work provides much more value for the business right
Ashwin Gopinath (29:42) Correct. I mean and and and if I were to be like so bold to basically say, right, it’s just sort of we are so while I respect and I admire the desire to have like zero employee organizations and all of that, like like Centra and me personally, I I I think we are still far away from that. I think the world is too messy and people’s desires and
Because you are ultimately the value creation is gated by some human somewhere. Like somebody needs to buy or somebody needs to do certain things. The value creation has to be some individual or group of individuals need to basically make a decision as to if something is important or not. It’s not an AI that is made, there’s no there’s no objective value creation. Objective is some somebody says, I will pay this much dollars or so this money for something.
Okay, so value creation is fundamentally with humans. I mean, in some sense, and in because of that, I do believe that humans are the ones, at least for the foreseeable future, going to make decisions. Agents might basically go and execute it, but fundamentally it’s gonna be teams of individuals that are going to basically make these decisions and continue to work. So I wanna basically make individuals’ ability to work with each other, maximize like maximize that rather than replacing.
I don’t think that replacement is going to happen anytime soon. Just being honest. Very like I’m being brutally honest with that, right? Because we have tried this and it does it. It’s like gets you about like 90-95% of the way there. The last 5% is where all the problems begin. So, so so yeah, and fundamentally somebody needs to make those decisions. And and and and and and it’s for that group of people that Centra is basically being.
Federico Ramallo (31:21) Yeah.
I completely agree and I’m I’m reaching to the same conclusions with these zero employees companies, where it gets to a point where it’s running loops on on lack of decisions or or get we get to a point where it says we need you know, we because I’m acting we’re acting as as board members, right? So we’re not supposed to be involved in the operation. But then you see how it gets locked and it asks for
for a decision to the board members, right? so it’s it’s kind of what’s the point of of building all this automation if it comes back to you, you know, for all these, you know, little little questi little questions, right? Or little decisions, right? What what’s the point, right? But it’s it’s a it’s an that’s why I call it an experiment. and I’m arriving to the same conclusion that at the end of the day the human decision becomes much more important, right? And
Ashwin Gopinath (32:03) Yeah.
Federico Ramallo (32:20) And actually I can a and that’s why my my my my thesis with with the longer context is that those decisions it it’s what makes the difference between having a a human managing agents that has all that context versus a human that doesn’t have that context because it’s it’s it’s a new hire, doesn’t know how to manage those agents. And yes, we can have higher speed, but
If you have higher speed and you go in the wrong direction, then you’re not moving forward and faster, right?
Ashwin Gopinath (32:51) Yeah, completely agree.
Federico Ramallo (32:54) Right. And and then the the other thing that I’ve been thinking about, you know, this idea of agents and humans is about accountability, right? which who who has the accountability of the actions of the agents, right? I mean, in my view, it should be the human that supervises agents.
Right, because the agents don’t have accountability by themselves.
Ashwin Gopinath (33:21) Correct, correct, correct, correct, correct. So I completely agree, right? Like I think I think people are so so like ultimately when something goes wrong, I don’t think the AI companies or the agent companies are going to take responsibilities. They are not going to basically and they’re not going to. I mean in their terms of service they’re going to be like ultimately whoever bought it is responsible for it.
Okay, so but within an organization, okay, within an organization, then sort of you you start asking the question who’s liable? Like who are you going to fire when you something goes wrong? Okay. Like, because that that I mean from a management management perspective, that is what happens. Like ultimately, when something goes wrong, like we saw this in that in Amazon and we saw it in a few other places where you have AI systems that have been rolled out and it causes mistakes and like AWS goes down or like the system goes down.
Federico Ramallo (33:57) Right.
Ashwin Gopinath (34:14) Or the system gives a deal to somebody that they can’t honor, okay. Who is responsible? You can’t say that an AI tool is responsible. Yes, then are you going to fire the engineer who built it? Is that the VP who authorized it? Is it the company? And like who pays, like if there is monetary laws, who is responsible for it? Or do you just have like proliferation of AI insurances, sort of like Corgi and others, who are basically going to you know insure these things?
Like, and and something to keep thinking about here is this, right? Like, and this is the problem that I keep thinking about, right? Like right now, we have a certain number of decisions that are being made with AI, and there is a certain probability of errors. Let’s say, just for a number’s sake, you’re making 100 decisions, and there is 5% probability of error, which means that statistically, five mistakes are being made. Now, a conventionally people will say that as AI improves,
the error goes down. But what you forget is that as the error goes down, the usage goes up. Okay. So let’s say now like the error goes down to 1%, but now you are useling you are making 10,000 decisions with that. So now what happens is you are making like percentage-wise, you have improved, but absolute number-wise, you are making a lot more mistakes.
Federico Ramallo (35:36) Right. Right.
Ashwin Gopinath (35:37) So so now the
question is where are and and these are complex systems wherein errors sort of propagate. Like like and and we see that in in in like in in dynamic systems all the time, wherein like some small variable changes and it has it sort of like has an outsized impact. So I think that when people worry about like job loss and all of that, I don’t think
They need to worry so much because ultimately the AI is not going to be responsible. You can’t blame an AI and basically say somebody needs to like AI is fired or like Anthropic is now going to pay for this. It’s not happening. Okay. So so ultimately some human needs to sit there and say, I checked the AI’s decision and I am going to be the person responsible when this particular decision goes wrong.
Federico Ramallo (36:12) Right.
Right, right. And the more I I learn about you know running you know, implementing automations within organizations, the more I learn that software engineers are going to be a much more relevant role because now you need somebody that can manage all that, right? on the technical side, on the business side, right? and as a software engineer building software, it gives you much more freedom on
what’s what you can build, right? N not freedom but productivity, right? I mean I used to you know work on on the syntaxes and make sure that it it was correct and all that stuff and now I can think on higher levels. You know, I want to migrate from this technology to this other technology. So let’s go and let’s plan that. Let’s put stages on it. So I can make these higher decisions that it’s more of a engineer manager managing a team of
agents that are actually building the software.
Ashwin Gopinath (37:24) Correct, correct. I mean the the the completely agree, but but I really like to sort of I wish more people thought about the accountability issue. Like who takes accountability for when things go wrong. I mean you typically think about it only in regulated spaces like me like like medicine or you know law and stuff like that. But the reality is this is true in pretty much every area. It’s just a question of the liability of when things go wrong.
You just don’t think about it that way. You don’t think about it as, you know, if I do something wrong, like I’m responsible for certain things in a company, and the CEO is responsible for it. Like, and and and junior developers are responsible for certain small parts of it. And when they go wrong, you can fire them. But you can’t necessarily do that. Like, how do you do that for an AI? Because it’s not as if they don’t make mistakes. They do make mistakes. They’re not, they’re not they’re not perfect. And they’ll never be perfect. No system is going to be perfect.
Federico Ramallo (38:08) Right.
Ashwin Gopinath (38:20) Yeah. So when they go wrong, how how do you deal with that? And and the scale of the problem is only going to keep increasing.
Federico Ramallo (38:29) Right, right. The the automation that allow us to do more also allow us to make more mistakes. And we need to have ha better tools to to have traceability of those mistakes, right? And I think that’s where your tool actually fits, right?
Ashwin Gopinath (38:42) Correct.
Yeah, I mean I our tool our tool fits there not necessarily for pointing out mistakes. That’s not necessarily the intention, but it in general it’s the it’s it’s to basically say that humans are going to be part of the system. And humans are I mean and and there are gonna be more and more humans who are going to be needed to do more and more intellectual job of who’s deciding, why are they deciding, or like to have the context to basically back up their decisions.
and so so fundamentally we wanna solve quintessentially human problems, which is teams of people working together. And this is something that we’ve been doing for like since we were apes in some sense, okay? like groups of people working together, and you can make that better. So, like, so that
Federico Ramallo (39:28) Yeah.
I I I love your analogy, yeah. yeah, because at the end of the day we’re still, you know, we’re still you know, imperfect animals in a way, right? yeah. Even though we try to you know become more
Ashwin Gopinath (39:37) Ha ha.
Yeah.
Federico Ramallo (39:55) Herkey higher but in at the end of the day we’re still animals with our biases, our instincts and our, you know, flaws, right? Right.
Ashwin Gopinath (40:03) Sure, sure. Yeah, exactly. Exactly. I
mean, so so so so it’s sort of like so it’s sort of like let I mean and and and and and I’ve at this point I have seen like enough teams, like whether it is in academia, whether it is in outside academia, in like research setting, in company settings, all of that. Ultimately one of the things that I have walked away from is that you can have
I think the team, a group of people, a great team can always make up for weak technologies, weak market, all the other problems. If you have a great team, okay, it’s a joy to work with them and they can actually like do wonders. But then you can have like individuals who are fantastic on papers, but like really shitty teams, okay.
And they’ll basically have all the money in the world, all the things in the world, and they still mess up. So the question is: how do you actually get people to work together and like you know, like coordinate and be the best version, not of themselves, but as a team? Like, and that is a problem I think that we can solve. And we see things of this nature, right? Like, in some sense, I think about it as like social media is a great example of AI.
Sort of getting people to work together. Like you have this concept of groupthink and things like that. In some sense, Twitter and Facebook and all of that are sort of bringing these people together and getting them to think in the same way to a certain end. I mean, they’re it’s not dip, it’s not, it’s not, they’re not doing it, they’re not doing it purposefully. Okay. However, that the algorithms sort of like get group them together and like it’s like self-organization that has happened. So now the question for us is.
How do we do that internally? How do we make sure that like if you’re part of a team, how do you make sure that all of them are like working towards the single goal of creating the most beautiful product that the company is building or the best service and like have least friction and like moving towards them, like moving as a single entity? You see, so so this is not fundamentally impossible. You see this all the time.
Federico Ramallo (42:16) Right. Right. And and com coming back to what we’re talking about is it’s all those little decisions and b little or you know, small or big decisions that lead towards making a great product that is successful. Right. right, right. and and I understand your point that Central is not designed to find, you know, you know the culprit.
Ashwin Gopinath (42:32) Precisely.
Federico Ramallo (42:40) you know, who made that mistake particularly, but it it’s it’s the knowledge of all the decisions that can make difference between achieving the outcome expected outcome or or or or wouldn’t, right? So being able to find that so you can fine tune it, right?
Ashwin Gopinath (42:53) So so so the best
Correct, correct. The best way to think about it is even the blame assignment is kind of different, right? Like for instance, if you have a driver and you have like a car with like lane discipline or like lane you know, when you veer off the lane, it just buzzes or whatever. Lane assist, right? So the point is do you see lane assist as blaming you? Like you’re not waiting for somebody to fall asleep and crash the car. Then you can say it was the driver’s fault. But if you had lane assist,
Federico Ramallo (43:09) assist, yeah.
Ashwin Gopinath (43:24) As soon as you start at drifting, it it corrects you. So you don’t need to make the mistake before you know you correct yourself. So Sentra in its perfect form in an organization is if somebody makes a less than ideal choice, they get informed and the system corrects you to move. So at that point, you’re not even made a mistake. It’s just like a like an autocorrect. It just basically pops in and basically, like do this thing. I mean
Yo your your understanding about the problem is not right. Or like you and somebody else are going orthogonal to each other, probably talk to each other and correct. So all like it changes the way in which a w like an organization sort works.
Federico Ramallo (44:04) Right, right. And I I think I told you already this this example, right? It’s it’s like this movie, Brilliant Mind from you know, with John Travolta, right? Where he he becomes smarter out of s out of suddenly and they and he’s able to build design the the crops for for his friends you know, farm, you know, because he’s able to see all the connections and everything, right? And
What you’re describing is the same thing. You know, it’s you have a an assistant that set tells you, well, if you go that route, you’re going to crash. If you go that right, it’s less efficient. So, you know, let’s go back to where you want to go. You human are still driving, but you have the assistant that that, you know, give you the the foresight, right?
Ashwin Gopinath (44:48) Yeah.
Correct.
Federico Ramallo (44:54) That’s amazing. And then the the other thing that we haven’t mentioned, and I think it’s important to mention, is token efficiency, right? Because now this is for the humans, but for the agents, you’re also providing that reinforcing feedback where you can guide the agent to find the most efficient path, right?
Ashwin Gopinath (45:15) Correct, correct, correct. I mean, so so the best way to think about it is this, right? Like sometimes you give a lot of information, a lot of context in there. And if you know where the information is located, if you know like like what is important for your particular task, then the amount of context that you need to give for the agent comes down dramatically, which means that the yeah and it doesn’t need to go and search all these different places and construct the context to before answering the question. So you’re
Your effectively every single task that you’re giving, the amount of tokens needed goes down. Like in our cases, like depending upon the complexity of task complexity of tasks, like sometimes we see like 12x reduction in token needed for the same task. So if you’re going to give the entire context or ask Claude or like ch like GPT or you know FordX or whatever agent tick system that you’re using.
If you’re you if you’re going to ask it to go construct the context and answer a question, and if you were to kind of use us or like Centra as the memory, then you know the the the token usage difference can be as large as like 12x difference or or or depending upon like for coding tasks, we are seeing between 20 and 60 percent reduction in tokens, while the accuracy is the same and sometimes higher on the same benchmarks that like
you know, GPT 5.5 or you know opus four point eight uses for you know showing their coding skills like like we have a we have we have like these these are things like terminal bench and deep sue like SWE benchmarks where we are able to show that token like apples to apples without if with and without centra everything else being the same we can see easily between 20 to 60 percent reduction in tokens
While the accuracy remains the same or slightly higher. Because it think about it, right? You’re just removing the noise. So it knows exactly what needs to be used. So you can basically go like thick things get things get like improved.
Federico Ramallo (47:01) Wow.
Right, right. Because without memory does the Asians don’t know don’t don’t know anything about the business. It needs to look everything, yeah.
Ashwin Gopinath (47:22) It needs to look at everything. Yeah, or
yeah, and it’s not just business, right? You can point it. So so we originally started out with the idea of using this for reducing complexity with within the organization, within the within like different systems. So we build a single memory layer that basically understands the like it’s agnostic. It doesn’t ha it doesn’t care really about what the data is. So we started pointing it towards coding agents simply because in like co-repos, like code repos.
the reason we did that was our token cost was going through the roof. So about two months back, we were like, let’s point this to our repo. Let’s see if it actually improves. And we started seeing it improves our systems. So we decided to kind of optimize it. and now it’s got to a point, and we that we are we are nowhere near, you know, the best it can be. It can be all far better than what it is right now. But we are already at like
Between twenty and sixty percent, depending upon the complexity. Some of these code bases are very small, so you don’t need you don’t have that effect. But some of the other ones that are very large, the larger it is, the more complex it is, the more important memory becomes.
Federico Ramallo (48:30) Right, right. And what you’re describing basically is you allow the agents to find a gradient, what is the most efficient direction they should follow for highest return of of of for maximum return basically.
Ashwin Gopinath (48:44) I mean from an analogy point of view that is totally fine. It’s not like technically it is not exactly right. I mean I don’t wanna I don’t wanna say that it is doing some kind of a gradient descent and like learning or something like that. It’s not doing that, but from an analogy point of view, yes, that’s what you’re saying is kind of true.
Federico Ramallo (49:00) Right. Yeah, what what you’re doing is providing the the context and the decision so the agents can make the the final decision.
Ashwin Gopinath (49:08) Correct. Like one way to kind of think about it is this, right? Like when you have these coding agents, when you’re touching one part of the code base, typically they don’t it doesn’t automatically know what is a blast radius. What that means is if you make a change at a single point in a particular file, how will it affect all the other things in the in the in the repo? That information is typically not there. So what happens with the coding agent is it’ll go make a change, then it’ll try to run it or compile it or do something with it, and then it’ll get a bunch of errors.
Okay, then it’ll look at the errors and go and correct it. And then that’ll have like an other set of effect. So you need to kind of go back and forth. Now, if it already knew what each piece is, what it’s supposed to do, what each part of the code is supposed to do, how it affects everything else, then it knows that when I’m touching one place, I need to also change things in all of these other places.
Federico Ramallo (50:01) Right, right. I I mean what what I’m doing to work around that is I make changes and then I run audits processes where basically I go through the whole thing. Yeah.
Ashwin Gopinath (50:11) Yeah, but that’s token usage.
That is token usage. Like if we if you had infinite tokens, if you had infinite tokens, then that’s a that’s a different matter altogether. But we are not in that world. So or you can think about it as an efficiency point of view, right? Like so we so structure a improves the efficiency.
Federico Ramallo (50:15) Yes.
Right, right. I mean the the the reason we’re doing that is because when we’re building something we want to make sure that everything works correctly. So we have to go back and go back because it doesn’t have all the context to make all the necessary changes to to fix everything. Yeah. Yeah.
Ashwin Gopinath (50:44) Exactly. Exactly. Exactly.
But if you have that, everything becomes a lot easier. I mean, and I’m not saying that this will solve all problems. But and like and and and and there’s a lot of knobs that we can turn to kind of improve this. We are nowhere near as efficient as we could be.
Federico Ramallo (51:01) Right. But as I understand you have the the research done so you can you know which direction you need to go to improve the system.
Ashwin Gopinath (51:11) Yeah, like yeah we we we have we have we have some we have we have we have like we have a we have a clear understanding as to what directions we we let’s put it this way. I know what directions are wrong. Okay. That is provably wrong. Okay. I don’t know whether the directions that we are taking is the most optimum one. So we are just not doing a bad we are not just not we just let’s let’s for the time being let’s not make the mistakes that are like obvious. Okay. And you know, hopefully we’ll basically continue to move down that path of like, you know, just don’t do the
Federico Ramallo (51:23) Right.
Ashwin Gopinath (51:39) Don’t make bad decisions. That’s all.
Federico Ramallo (51:41) Right. Right. Amazing. And what is the future ahead for Central? What are what are you planning to to do next? Which milestones are you are you trying to achieve?
Ashwin Gopinath (51:50) So
so so so right now we are we are trying to do this right rather than going fast. we have a bunch of like great investors and like design partners right now that we are working with and trying to kind of answer core enterprise questions as we are sort of scaling up. The reason is that, you know, there are so many enterprise like AI pilots that are going on, and you know, they all fail because of the fact that, you know.
They’re not able to show return on investments. They’re not able to show the value creation or like do it right with all the you know bureaucratic details of an enterprise. You can forget about it, like there is an idea that you know startups will not think about all of these things. That’s all fine, but ultimately push comes to shove, you need to sell into the enterprise. I mean that’s where the real money is. And these are the large organizations and they have bureaucracies, they exist for a reason, and we have to play the game and we have to kind of change it.
If it has to change, it has to change slowly by being in a conversation with them. We can’t basically say we are going to do it this way and you need to kind of agree to us. So we are just work we are working with them closely to kind of answer these questions as we are slowly adding on like more design partners and like folks who want to kind of work with us. But you know, so and and and things once and and we are open to sort of working with more folks as they come along.
Federico Ramallo (53:12) Right, right. Interesting. so are are you close to production or are you still working on
Ashwin Gopinath (53:17) yeah, yeah. Both both of
no no no no no both of them are live. like like core centra is live, the core code memory is also live in the sense that we are just basically like trying to work with individuals to kind of get them on board one at a time rather than sort of like open it up for everyone. because you know memory is a very tricky like people don’t know exactly what memory is, or like there’s no consensus on.
People think that knowledge and retrieval is memory. That’s not memory. That’s only part of it.
Federico Ramallo (53:50) Right. Right. Yes, and and probably the it’s going to be a point where the analogy of memory is going to break down, right? Because we’re going to go past what our concept of memory, our you know, how we we we process memory is different on how you know the organizations and the agents and the humans should, right?
Ashwin Gopinath (54:12) Correct. And and and and ultimately here’s the thing, right? Like memory is some kind of learning. Memory is kind of a learning. Like you remember it’s not ma it’s not learning the way in which the weights are thought of as. the question. You remember something.
Because somewhere there is a utility function, you might not be able to define it, but some part of you or some part of the system determines that this piece of information is likely going to be useful in the future.
That is memory. And that’s very different from just remembering everything.
Federico Ramallo (54:45) Right.
Very interesting. we’re running out of time, Ashwin. I truly appreciate you being here today. Any final remarks before we wrap it up?
Ashwin Gopinath (54:56) Well yeah, I think I think but
So we have seen the LLMs sort of converge. Like LLMs, the core LLMs, their capabilities converging. We have we are seeing sort of like the agents sort of converging. Personally, and I think this might be a bit of a like a like I think I think this might be like us being like ridiculously optimistic, but I think that in the next 18 to 24 months, there will come a time.
Wherein people are going to be more and more excited by memory, the way in which I’m describing it. It’s like closer to a continuous learning. People might call it continuous learning, people might not like might call it something slightly different. But the essence is memory is not just remembering something. It’s basically like filtering it and learning, or like being able to choose what part of it is important for you in the future. And that system.
Will become more and more important and that dialogue is going to move towards that as the agentic capabilities sort of saturate. Because that’s what is going to make one agent better than another agent. And that’s only going to get louder and louder over the next several years.
Federico Ramallo (56:02) Right.
Right. And the n now that I’m hearing you about talking about memory, the other thing that comes to mind is this idea of experience. What is experience? Is all the mistakes we make through the years, right? So when we have a when we have a new problem, right, or an or a situation, having all that experience allow us to avoid those mistakes and we know where to go, right? And that’s what you’re building.
Ashwin Gopinath (56:19) Yes, yes.
Precisely.
Precisely. That’s that’s that’s a that’s a that’s a great way to put it, right? Like that’s a great way to put it. It’s basically try like memory and experiences and all of these things are they’re they’re not what you would think of as learning the same way but but as everything else, but but they are they have aspects of it.
Federico Ramallo (56:50) Right, right. And we we don’t think about it, you know, when we have all that experience, we don’t think about every single decision, but at the end of the day, it weights into how we make the next decision. Right?
Ashwin Gopinath (57:00) Precisely. So you don’t even you you you don’t even so and
this is the hard part, right? Like you can’t tell me why you would remember what you remember. Okay. Like it’s very hard. It’s very hard. But like some somehow you end up like there is a utility function that you have that causes you to remember it. Like and now now that’s what is now that that’s what correct.
Federico Ramallo (57:10) Right, right.
Right, we
We we call it intuition, but it’s actually this
undescri and undescribable way of of how we think.
Ashwin Gopinath (57:29) Correct, correct. And and and and and this is the reason why I don’t think memory can be solved for an individual. Like, because I cannot like like this is why when you are doing Claude or when you’re doing Char GPT, like Char G PT and Claude doesn’t know why you asked the question that you asked. It knows what are the conversations, but you don’t know what part of that conversation is important for you, and it doesn’t know why you would remember those.
Federico Ramallo (57:56) Right.
Ashwin Gopinath (57:57) Unless
it has a model of you in itself, it cannot. So so from a personal point, which is why we are not going down the path of personal memory. But for organizations, I know what is important for the organization. They care about people and their products. That’s all. And they want to maximize their profits. So it’s an easier problem.
Federico Ramallo (58:17) Right.
Right. I I think I dare to th to say that the reason you cannot build a memory system for an individual is because all the decisions happen within our minds. But when but when you have a team, the same decisions that happen within our minds, you know, a solo entrepreneur would talk with sales, would talk with marketing, within its mind, right, their mind, right? But now we when you have a team
All those decisions are exposed in the communication between people, right?
Ashwin Gopinath (58:48) Correct, correct,
correct. So like the way in which we think about it is it’s sort of we call it internally as chain of thoughts. Like conversations between people are chain of thoughts of the organization. And we know how important chain of thoughts are for like for these models to kind of reinforce and to kind of know when they make a mistake, you look at the chain of thoughts and you say this is why it made a mistake. And when you’re doing fine tuning, you go and try to change that. now the exact same thing can apply to humans.
Now it’s just the conversations. That’s the chain of thought of the organization. And once you capture all of that, you can learn a lot from.
Federico Ramallo (59:23) That’s amazing. Ashwin, I truly appreciate you being here today. I mean we can continue talking, right? and I think what what you’re building in Centra is amazing and I’m I’m looking forward where you take it next.
Ashwin Gopinath (59:29) Yeah, yeah, yeah, yeah.
Was was a pleasure. Was a pleasure. Yeah. Thanks, Federico. Yeah.
Federico Ramallo (59:40) Thank you, Ashwin.
Listen on your favorite app