Welcome to the changelog, where we have deep technical conversations with the hackers, leaders, and innovators of the software world. I'm Jared Santo. On this episode, Adam and I are joined by Sid's brandage, founder of GitLab, who led the All-in-One coding platform all the way to IPO. In late 2022, Sid discovered that he had bone cancer.
That started a journey he's been on ever since, a journey that he shares with us in detail. Along the way, Sid continued founding companies including Kilo Code, an All-in-One, Agentic Engineering platform which he also tells us all about. But first, a big thank you to our partners at Fly.io, the platform for devs who just want to ship, build fast, run any code fearlessly at fly.io. Okay, Sid's brandage, back on the changelog.
Let's do it. Well friends, I'm here again with a good friend of mine, Kyle Goldbraith, co-founder and CEO of depot.dev. Slow builds suck, depot knows it. Kyle, tell me, how do you go about making builds faster?
What's the secret? When it comes to optimizing build times to drive build times to zero, you really have to take a step back and think about the core components that make up a build. You have your CPUs, you have your networks, you have your disks, all of that. It comes into play when you're talking about reducing build time.
And so some of the things that we do with depot, we're always running on the latest generation, or ARM CPUs and AMD CPUs from Amazon, those in general, are anywhere between 30 and 40% faster than GitHub's own hosted runners. And then we do a lot of cache tricks, both for way back in the early days, we first started depot, we focused on container image builds, but now we're doing the same types of cache tricks inside of GitHub actions where we essentially multiplex uploads and downloads of GitHub actions cache inside of our runners so that we're going directly to blob storage with as high of throughput as humanly possible. We do other things inside of a GitHub actions runner, like we go on off portions of memory to act as disk so that any kind of integration tests that you're doing inside of CI, that's doing a lot of operations to disk, think like you're testing database migration since CI. By using RAM disks instead inside of the runner, it's not going to a physical drive, it's going to memory, and that's orders of magnitude faster.
The other part of build performance is the stuff that's not the tech side of it, it's the observability side of it is you can't actually make it build faster if you don't know where it should be faster. And we look for patterns and commonalities across customers, and that's what drives a product roadmap. This is the next thing we'll start optimizing for. Okay, so when you build depot, you're getting this, you're getting the essential goodness of relentless pursuit of very, very fast builds near zero speed builds.
And that's cool. Commonness team are relentless on this pursuit. You should use them depot dev free to start. Check it out.
One line of change in your GitHub actions depot dev. Well, friends, we're back with a good friend of ours. It's been sit way too long. Way too long for sure.
Way too long. The last time we talked to you, the last time I talked to you, at least was when you were leading GitLab 2 IPO. This is pre IPO. This is in 2022.
It was speculative. That was obviously the North Star for you. You've accomplished that mission. You've had a cancer journey.
You've had an investor journey. You've had the American dream journey. And kind of back again, in a way. So welcome back to the show.
It's been way too long to see again. My pleasure. Thanks for having me. Looking forward to this.
And we look at your story from someone say, you know, this, the shadow of GitHub from way back in the day to now, how do you, how do you personally reflect on your journey? I would say, in tech, but also as a CEO, as a just a normal person, as an investor, how do you reflect on this long journey force? Well, it's been amazing kind of living living the American dream 2015. I came to the US.
We raised money for GitLab and we set our sights and grown the company, becoming a public company. And then six years later, we took the public, which was amazing and we were very fortunate to. Six years to public, huh? Yeah.
Very positive. That's the speed. And you were very much a leader too, with the way you opened the company, your open dogs, a lot of the ways you hired, the ways that you showcase culture. A lot of that was even not so much not in public, but it was, you were very much a star of that.
And then you were also a company that IPO. That's, that seems like maybe not the way it should be done, but it's obviously the way you did do it. Yeah. We tried to lead their company with a ton of transparency and we believe that helped us be an old remote company, but at the same time being super on the same page.
And it's paid its dividends both in coordinating us, hiring great people, but also getting customers excited about the company itself and how we would partner. Are you still involved in any capacity with GitLab? Yeah, for sure. I stepped down as exec chair last year or stepped down as CEO and now became an exec chair, which means that I still am an employee of the company.
So I still have two engineers reporting to me. We're taking on projects that I'm super excited about. And right now we're working on adding observability into GitLab. So really, really exciting to work on that.
So when you left GitLab as CEO, your CEO, correct? That's what you stepped down from. Yep, correct. That's the position.
It wasn't necessarily because you were done or because you were, you've done your thing and you're ready to, you know, get mojito and go to the island. You actually had a health problem, a major health problem. I mean, a health crisis in many ways around the end of 2022. And you've been on quite a separate journey, which is amazing to me.
Some of the details that you've gone through in order to be here today were very happy that you are here today. Can you unpack a little bit of what you've been through in the last three years, four years with regards to your health and the journey you've been on to fight cancer? Yeah, for sure. So, end of 2022, one year after the IPO, I discovered I had bone cancer, 6-centimeter tumor grown from my spine.
And we had to do surgery really quickly. I was in a lot of pain and we removed the vertebrae and most of the cancer and did a spinal fusion with a titanium frame with radiation, chemo. And I did my first single patient, I indeed, since 2017, I invested in one other work company, an over time that guy became my best friend, his wife became my wife's best friend. And every time he was fundraising, the investors wouldn't put any money in.
And every time I said, this doesn't make any sense to me that they're reluctant to put money in, I'll finance the company. So, over time, I became his biggest investor. And what he's doing is combining drugs with binders through click chemistry, allowing you to target different parts of the body. And when I became so very sick, he made a...
We asked for an exception to treat me with his drug and we got that exception. And we did all the treatments. And then two years later, in 24, the cancer started growing again. And the doctor said, you're done with standard of care.
There's no more treatment we have for you. Maybe there's a trial somewhere and I started looking around. The weird HLA type, which disqualified me from a few trials. And I have a bone cancer osteosarcomic, a rare disease.
So, there wasn't anything for me. And it was my... If I don't move, I'm going to be that soon. I had to step down as a CEO because I needed to focus my full attention on this.
I started talking to anyone, traveling anywhere. I went to the medical conferences. I went as far as China to get scans. And I started doing maximal diagnostics, making my own treatments, doing treatments in parallel, and trying to scale that for other people.
And so, that's been the journey I've been on over this last year. And it's been a lot of ups and downs, but also extremely interesting. And I've learned a lot, but I've also think that there's a lot of things that could be better. And I'm trying to change those to be better and making the path I've been on easier for other people to follow.
I understand the reasoning for parallel directions or testing because you only have so much time before it's too late. And so, I totally understand that. It seems like as a software guy, you want to try to isolate and do one thing at a time and then see if it works and then do the next thing versus having multiple factors. I suppose if it's working, it doesn't really matter which of the five or six is working.
If it's working, but certainly for reproducibility and for helping others, you do want to be able to have some sort of result that you can come out of it. So tell me about the parallel side and some of the trials that that introduces and how you're maybe solving those problems as you go. Yeah. If you want to be able to reproduce, you need to isolate the change.
So when I started doing parallel treatments, one doctor told me to my face, but what if it works? We don't know what would have cured you. And then I let the doctor know that I wasn't interested in finding what cured me. I was interested in getting cured.
That's right. I'm not research. I'm a real person. That's right.
The combination of treatments, there's very little incentive for farmer companies to kind of spend a billion dollars. It's now a billion dollars for a successful trial to spend a billion dollars for an approved medicine to research a combination of drugs. So there's no incentives. So all you see on the market are these drugs taking in isolation and doctors are reluctant to combine them because it's never been done before.
But if you look at the first principles, you can say, Hey, these treatments are probably okay to combine. What you don't want to do is have like combine two things that both hit the kidneys because your kidneys are going to be over capacity and that's not good. So every time we do we combine treatments, I have, for example, a pathologist in my tumor board to look at like, Hey, can we combine these things? And we also look at my own markup, including genetics, but also histochemistry, where you kind of, you color slides in and we do multiplexing.
So we can call them multiple things. For example, I told you I went to China, China three months ago, and I did that to get the first B7H 3 scan available to me. Because of what we learned there, we were able to kind of change the development of one of the drugs I'm developing to remedy what we saw. So in China that I have higher B7H 3 expression in the liver than any one of the 20 patients they saw before, and we decided to combine that medicine with something that's rarely expressed in the liver.
So it's not going to destroy my liver going forward. So you got to be kind of, what I'm trying to do is kind of be first principles about it instead of kind of relying on evidence for the ability to combine because that evidence, it doesn't exist because no one has the incentive to do that. Right. And most people die because they, not because they run out of treatments, they run out of time.
So why China is a technical thing is a regulatory thing here. There's more advanced in certain areas like why why to go all the way there for this particular test. Yeah. So do we clear most of my treatments and most of my tests are happening in the US, but Europe and China also part of it.
And what's remarkable about China, I found remarkable that I was the first, it was my first time traveling to, to that hospital. Within two hours, I was checked in, they formulated the diagnostic scanning gradients, they gave me the scan, they printed the results, they talked me through it, and I was outside again. I've not seen that speed in the US. One other thing that's happening is that these hospitals have a lot of patients.
So instead of running a trial at 10 different hospitals where it's a lot of work to combine all these investigators and to train all of them, you can run it in a single hospital, the same hospital where the investigator is at, the researcher is at. So they have a much higher speed of running trials. And you're starting to see a giant shift in the literature of trying to do more trials and making more medicines. So where do you stand?
How's your health? How are you feeling? How are you doing? What's the future look like?
Yeah, I'm doing well. Right now we cannot detect my cancer. So that's great. That's a really good spot.
It's a great spot to be compared to a year ago. I was growing and we had zero treatments today. I have a therapeutic ladder as in if things get worse, we'll escalate the medicine and there's 30 medicines on it. So we have options now.
And I'm continuing the development of 10 different drugs and diagnostics that make a lot of sense for me to have. I've also started in the meantime, six companies to scale treatments for other people. Treatments that I've done, but also treatments I've not done by just technologies we came across that we thought are under invested in. Yeah.
Do you know what has made your cancer undetectable then? I know that you weren't being that research lab before that research right before, but like how do you have you pinpointed what may have worked with these 10 different medicines you're developing? Yeah, for sure. I might have been one of the first people to do single cell sequencing as an individual patient outside of the context of a trial.
It's kind of a strange name because you're not sequencing a single cell, you're sequencing thousands of cells, but you're kind of you're able to see the genetic markup of each cell. And you're able to see what the what the tumor is looking like, like which cells are in there. How do you behave and what's their genetic markup? That's an incredible powerful technology.
We compared that to an atlas of kind of what bone marrow should look like. And we zoomed in, we were able to kind of zoom in on the cancer cells. And one thing we noticed is that they had a lot of fibroblast fibroblast is kind of scar tissue. And we were aware of a treatment in German in experimental treatment that has an FAP binder that binds to fibroblast.
And then the other on the other side you put a radioactive element, Lutecium or Etinium. And it did two treatments there with incredibly good results. First of all, no side effects. I couldn't detect any.
And we saw 60% necrosis and 20% shrinkage, which considering it was only two treatments, that was a great result. My cancer is in my back near my spinal cord. Because of the shrinkage, it detached from the spinal cord, allowing a really talented surgeon to go in and go after it and remove most of it surgically, which was impossible before. This is because it would attack the scar tissue.
What was the name for the scar tissue is because you were able to pinpoint that and allow it to detach from the scar tissue. Yeah. So what we were able to do, because of the scar tissue, we had a binder for that. It was called FAP or experimental treatment in Germany had that.
And you combined a binder with a radioactive stuff. So a lot of drugs, you just give them to the entire body. Yeah. A standard chemo drug is devastating.
This drug is also a really nasty thing. Lutecium is radiating you from the inside. I was like 10 times as radioactive as an airplane at altitude from the outside, like what was happening inside of me. Oh, wow.
They have to keep you in isolation because you can be around people. You can be around people. But because it was on this FAP binder, it targeted my tumor especially. And there's also a little bit of risk for like, for example, your loss of taste and smell, but I had non-side effects there.
So it was very specifically bound to the tumor tissue, which is just amazing. Like most cancer treatments are way worse. And yeah, enabling surgery was a big deal. If it's right up against the Dura, your spinal cord, that's not great.
But in this case, it was detached from there. Because it was amazing. Yeah, it was really great. How long were you radioactive?
You were really active for a week, but only two days in isolation. And then you get a letter and then you walk at the airport. Maybe you've seen those things in the US now. You walk into the airport and there's a black thing.
I guess what? Beep, beep, beep. We once had three people with scanners around us. And they had to call kind of Washington because their sensors were showing plutonium, which wouldn't be good.
Wow. But it was literally TCM after all. But they had to have a scientist in Washington. You were a person of extreme interest there for a little bit.
Yeah, as soon as they knew it was me and not my luggage, they were super chill about it. They were super chill. That's hilarious. Wow.
And then I had a type of cancer that's sarcoma. There's like 70 different types. And I asked this question because I have a dear friend of me and my wife who passed away probably 14 years ago. And she has sarcoma in her lungs.
She was actually studying in Houston, which is a very popular place for cancer research, at least here in the States, is one of the capitals of cancer research. And she was studying to become a doctor in this field and actually not through her lab time or anything like that, but she came to actually get the cancer. I don't know how to describe getting cancer if it just develops or if it grows or what actually gives the cancer to you. But she developed the same sarcoma cancer that she was studying and it was just really, really bizarre.
Yeah, quite bizarre. Like there was no real connection there. The same doctor that she was studying under was her doctor during her process. She obviously ended up passing away.
It was about a year, maybe a year-ish and devastating to our life and our friends. But can you speak to the sarcoma type of cancer and what that type of cancer does or does not do? You know what to do? But there's 70 different types and I'm not even that familiar except for I know she has sarcoma cancer and you mentioned you have a version of it as well.
Yeah, like I'm not a doctor. This is no medical advice. I know very little about that type of sarcoma. I do know sarcoma are genetically pretty diverse.
So there's a lot more studying to be done. For example, the standard genetic reports showed like TFE3 amplification. But when we really started digging, it wasn't the driver at all. So because of all that copying of the genes, it's harder to analyze.
But there's also more special things about it that allow you to do more targeting. You mentioned Houston. Big fan, Houston Methodist is especially great here. I know very little about long sarcoma.
What I do know is that for bone cancer, the biggest risk is that it spreads to the lungs. If you have these cells going on the loose, the first place to likely end up are your lungs also because your lungs are very tiny vessels. So they get trapped. The cells get trapped there the fastest, but also because it's kind of a similar cancer that can grow there.
So we were watching for that. And at a certain point, I did a kind of a regular pet scan where you look for activity that is out of the ordinary. And I remember vividly I was talking with my direct reports and I get a message, a text from my GP and he says, it's positive, not subtle. Like, it's positive.
Okay, great. And I'm like, oh, positive in a medical setting. That means something else. It means positive for cancer and not subtle.
That sounds bad. And I opened up the report and I read my death sentence. It spread to 50 sites in my lungs. My lungs were lighting up like a Christmas tree.
And the radiologist gave an undifferentiated diagnosis. And he answered, spread to the lungs, way metastatic. 50 places means like surgical, your host. There's no way to do it with surgery.
And so I worked out of the room, collected my thoughts for 10 minutes, called my wife, called in my CFO, CLLO. I think this is a situation. What do we do? We're going to tell or not, started to tell.
And we all took a long walk together, started to, my doctor started coming in. So sorry for you, so sorry for you, so sorry for you. And then the last doctor basically to chime in was like, yeah, 60% of the chances isn't cancer. I'm like, I get a call like, Chalo, what do you say?
Why now? He's like, yeah, the lymph node pattern, how it spreads, this doesn't seem like lung cancer to me. And this is a guy who've run the most osteosarcomatrons or the most sarcomatrons in the nation. So he might not, he's just seen so many patients and seen so many people who just happened to.
So okay. I went to now, I was a dog with a bone and within a few hours we cleared it up that this was the remnants of COVID. It wasn't because we were so afraid of this happening, including the radiologists. Everyone had a bit of tunnel vision, but it quite the experience to tell your team you're gone.
Yeah. And then the exact opposite, when you're like, you know what, I think I'm in a mistake. It was just a little bit of nuts. Cause that's, that's exactly like it went to her lungs and brain quickly, very, very quickly.
So literally the two places that tends to spread. It may begin lower in your liver and your back like you were, but it spreads quickly to other places and it sort of takes you quickly. If it's not, if it's not COVID, if it's something different. Right.
Probably the happiest COVID diagnosis of all time. Thank God it's COVID for sure. I wasn't super happy about the diagnosis though. That's really just an I, maybe not going to be friends.
Yeah. I mean, obviously they put you through turmoil and terrible things. I had a nowhere near as serious and yet relatable diagnosis with my knee when I was told that I had torn my ACL, I would never play basketball again by a physician's assistant. And then the actual knee doctor came in.
He's like, no, you're good. I'm shortening the story. But I had like 15 minutes of like, oh, I love my whole life changes in small ways. Yours is obviously in a huge way.
But then him being like, nah, she didn't understand this thing and she messed up and it's not as bad as she thought it was. And I was like, let's fire her or something, you know? And do you think it's not trivial to diagnose? Yeah.
It's not easy. So it's undetectable. And in this moment, you just share this dramatic story arc that you've gone through. You're recalling it here in this moment, this podcast is the expectation at this point with you and your medical staff that it's not going to come back or it's just undetectable and you're still actively trying to just work on it or treat it.
Like, what is the current status of how you're dealing with it? Is it just gone for now, wait, or are you handling things? Yeah, we assume that it's not gone. And the only reason it'd be gone, we know that the surgery didn't cure it, that they left positive, so-called positive margin.
So they left cancer. The only thing that was remarkable when we, the cancer we removed, we were able to flesh freeze some of it, which is amazing for people with cancer. It's way better to flesh freeze stuff because it allows you to do many, many more diagnoses on it later than if you do the standard, which is FFPE, which is like, preserves it, but it precludes you from doing a ton of tests. So if you're going to have an active role in your treatment, you've got to convince the hospital to flesh freeze some of it so that you preserve optionality to run these diagnostics later.
And what we saw was that the T cell infiltration went up a lot. It went up from 20% to 90%. So the place was buzzing with T cells, which is a good thing. And we think that is a combination of the two checkpoint inhibitors that I'm taking, that kind of unleash your immune system.
And one experimental treatment that is an oncolytic virus, like modified comical virus, that drops the TGF data and the TGF data, something that cancer uses to hide from the immune system. And it kind of tears away that invisibility cloak. So that's good news. We don't superthink it.
That's curative. But I'm gearing up in two weeks to take an mRNA vaccine. And that's like a 50-50 of curative. But there's a reason we have like 10 drugs in the pipeline.
We're trying to cover all our bases and rather have a direct too many than one too few. Yeah. It sounds like you're in a by-time conditioner. Am I reading that right?
You're buying time to research it more, to attack it more, to have an mRNA opportunity for 50-50 chances that we're at? Yeah. So the mRNA is potentially curative. So in that case, you're kind of, you're done with this.
But it's certainly like buying time. You got to have enough treatments to make it to the new development, both science advancing and some of these works that I'm developing to just take years, many years to develop. So you want to make sure you make it to the next spot and hopefully as good of a state as you can with like no, hopefully no metastatic cancer, hopefully no fright. This is the year we almost break the database.
Let me explain. Where do agents actually store their stuff? They've got vectors, relational data, conversational history, embeddings. And they're hammering the database at speeds that humans just never have done before.
And most teams are duct-taping together a Postgres instance, a vector database, maybe elastic search research. And that's our friends at TigerData. Look at this and said, what if the database just understood agents? That's agent at Postgres.
It's Postgres built specifically for AI agents and it combines three things that usually require three separate systems, native, model, context, protocol servers, MCP, hybrid search and zero copy forks. The MCP integration is the clever bit your agents can actually talk directly to the database. They can create data, introspect schemas, execute SQL without you writing fragile blue code. The database essentially becomes a tool your agent can wield safely.
Then there's hybrid search. TigerData merges vector similarity search with good old keyword search into a SQL query. No separate vector database, no elastic search cluster, semantic and keyword search in one transaction. One engine.
Okay, my favorite feature, the forks. Agents can spawn sub-second zero copy database clones for isolated testing. This is not a database that can destroy. It's a fork.
It's a copy off of your main production database if you so choose. We're talking a one terabyte database, fort, and under one second, your agent can run destructive experiments in a sandbox without touching production. And you only pay for the data that actually changes. That's how copy on right works.
All your agent data, vectors, relational tables, time series metrics, conversational history lives in one queryable engine. It's the elegant simplification that makes you wonder what we've been doing at the hard way for so long. So if you're building with AI agents and you're tired of managing a zoo of data systems, check out our friends at TigerData at TigerData.com. They've got a free trial and a CLI with an MCP server.
You can download the story experimenting right now. Again, TigerData.com. What does day to day like for you? I guess as maybe the last six months, how is your life in terms of how you live?
What are the things you focus on? Are you traveling a ton to obviously do research or get treated? What's the day to day in your life right now? Yeah, it's a lot of working.
I like working. I used to last few weeks. I've worked in 10 hour days. A lot of meetings, mostly the whole day is meetings, 25 minute meetings, like 15 a day.
I've also done some traveling for pleasure. We thought that the beginning that this might be my last year. So we celebrate it the 25 year anniversary of me and my wife, thanks to my friends and family. We have this great, also a lot of traveling, going to medical conferences, going to treatments on multiple continents.
But also doing fun stuff, adding observability to GitLab. I started an AI company called Kilo. I'm still creating these companies around open source projects with open core ventures. And then since September, starting to biotechs every month.
So how many companies are you currently involved in here in December? I started 30 companies over the last three or four years. What made you want to get into open source, agent code? Even you have all these medical companies, you've got GitLab stuff going on, you want to work on this?
Why do you want to work on this particular problem? I saw the rise of open source, agent coding companies. And obviously, agent coding is very cool. It's the future.
And it was cool to see that open source had a chance here. I kind of, before I thought it's all going to be closed source. And I saw the rise of open source and it felt okay. Time to join that game.
Are you deploying a similar playbook that you deployed with GitLab back in the day? Yeah, the kilo is, it has great agent coding, but the best reason to use it is that it's only one. So not only can you do the coding, it can also review your code. Not only can it do that, it can also do the security review.
Not only can it do that, it can also deploy it. So it's available everywhere. The OscogetBrains, CLI, IndiCloud, and the mobile app. And it can do the full life cycle, including deployment.
We're working on more there. So just like GitLab, the idea is to do it all in one solution where you don't have to switch applications the whole time to do what you need to do. And to be open source where people can contribute improvements. Right.
So what's the open core aspect of kilo? So you mean what's the page functionality? Yeah, we're the lines drawn. What do you think?
We're thinking if a feature is more appealing for managers or executives, that's where they need to pay. So for a team's functionality, for example, to see AI adoption throughout an organization, that's page functionality, 15 bucks per user per month. And then we also have an enterprise version with features that appeal to executives, like centralized, being your own keys that they can send it to bedrock in their VPC. That makes a whole sense.
So are you working on models as well? Are you just a model like NOSDAQ, which you're looking at aspect of the world? We're not working on models ourselves. We're working very, very hard on supporting older models.
So hundreds of models work over 500. We're the most prominent launch partner for new models. So a lot of new models, a lot of stealth models are first launched on kilo. And our users, a lot of our users use free models like Grock, Min, Max, GLM with kilo.
Interesting. I think I haven't even heard of a couple of those. I've heard of Grock, but I mean, Max, I haven't even heard of. So this is either underground or upcoming or when you say new, bleeding edge, perhaps.
Oh, yeah. It's all launched in the last two months. There's a lot of Chinese frontier lamps. And they want to quickly get their models more popular.
So they offer incredible pricing. Like Grock, it's got a good model and it's been free for months now. It's an interesting space because you've got a lot of things happening around Google in particular. There was an announcement today for a flash model.
I think it's called Gemini flash, if I can recall correctly. But when you look at that, they can essentially offer their models in the API layer for virtually free or a loss leader because they're trying to build their platform while opening eye and anthropic are burning cash, making money for sure. But when you're fighting a competitor that essentially can outlast you because they already have so much income at cash or value elsewhere in the chain, it's kind of hard to battle somebody who's willing to lose or break even when you have to profit. What's your take on the current landscape of just the competitive nature from nation states, like China, et cetera, and that's just companies there to the battle that's playing out here in the US?
Yeah, I think the leading coding model, that's an incredibly competitive space where the always feels a bit like OpenAI is a better model and then Google has a better model and Tropic comes back with the best model by far. For sure. But also the best free model is incredibly competitive with DeepSigmin, Max, DLM, all these providers are one-upping each other. What we offer our users is you're going to keep using the same application.
You don't have to sign up for a ton of stuff. It's just our thing. There's no charge. It's completely free.
And then you have the ability to use any model you want. Soon you can use different models. You can use free models, state models, whatever you see fit. So in Kilo, you can also kind of define different profiles because a lot of people prefer to open the eye models for planning and architecture that they prefer something else like I do so this four dot five for hard book troubleshooting.
I do is grok for kind of the smaller tasks that are easier. So it's kind of a mix and match approach. I think we're going to live in a multi-model world instead of one model that's the best at everything. For sure.
One of the challenges I remember for GitLab, which I think you all overcame was because you wanted to be all in one and you were providing the entire software development lifecycle tooling. Almost every aspect of that hat you had like startups and businesses that like their entire focus was that one thing, whether it was like bug tracking or observability or deploying or whatever it is. I feel like you're going to have the same situation here with Kilo where it's like there are people attacking like separate slices of what you're trying to do and I wonder how you're going to do all the things well. Yeah, the same way we did it with GitLab, go incredibly fast.
If you look at like where we've come from, like we only started in March and now we have like code reviews, deployment, all these things in one package. So we're moving extremely fast also thanks to agent decoding plus contributions. And I think we can even improve a little bit on working with the wider community. And yeah, just like it was like when GitLab grew up, it was like, oh, but Travis CI has already dominated the CI space.
And I'm like, hey, code rabbit is already dominating code reviews. Just, yeah, let us cook. We'll go faster than anybody else. You got the same people working on it or how do you know you can cook so fast?
I mean, they worked last time, but was it the same team because talented people make good products? There's some people who after GitLab went to work for other companies and some of them have joined, but the majority of the team has never worked at GitLab. And they joined Kilo because they want to go incredibly fast. They know the bar is high.
They know they got to work really, really hard. For example, at some point, we once a quarter we come together, we do focus week and we all work our butts off for a week and we do that together. And a week before focus week, it was clear that we weren't recruiting fast enough. So I put a LinkedIn post out.
I said, look, fly by Monday. We'll make a decision by Wednesday. And then the next Monday, you're expected an Amsterdam. And four people did that and they all shipped the initial prototypes of their functionality by the end of the week.
And all of them launched their functionality December 10th, less than two months after focus week. And it's all held up so far. And now they're all going to make it better. But it's incredible speed.
The way I would look at it is it used to be that you needed like a team of seven people to do something. You have one person and kind of a whole team of agents working on something. So yeah, we have like 20 engineers, but in reality, they're all working with seven agent coders. And that's the expectation that you as an as one human, you ship at the speed of what two years ago would be a team of seven or eight.
That's interesting. So you've successfully parallelized your cancer treatments and your engineers are parallelizing their agents. I can still only keep one agent busy that time. Maybe I'm not ready for this new era.
I couldn't work for for kilo. How do they do it? You know, you got all these different, you know, down in the weeds of how they're managing multiple agents working on a singular feature, I suppose, or a set of features that support a larger feature and how are they staying productive and fast in that environment. That's common.
Like you're not alone in only being able to keep one agent busy, but I think 2026 is going to be the year that that changes our parallel functionality where you have kind of you farm it out to multiple agents and multiple work streams. That's just getting started a month ago, the engineer who made child vac one of the early ways to paralyze things joined us and using amazing work. There's still a lot to do. For example, kilo has a functionality that's orchestrate, but orchestra to say this is what I want to achieve and split it up into separate jobs.
So each of those jobs have small context windows. Because small context windows means it's cheaper to run, it's faster to run, but most importantly, it's not going to lose half of its context because it's kind of getting to the maximum context window that's effective. But in the coming weeks, we'll ship something where those different agents are now able to work in parallel. And you can do a run-off, especially if you're using these three models.
You might as well have two models work and then see which one has a better result. So I think you're going from you have one working directory today to in the future, you have multiple tasks running in parallel. And each of those tasks is running multiple types of agents because it's hard to predict upfront, which agent will be the best in any given thing. Right.
That sounds cool. It also sounds expensive. I suppose commodity pricing just keeps going down. Is that the approach here, just innovation and investment brings down the price of agent coding?
If you go to kilo.ai, you look at our home page. We see the 10 most popular models for kilo. I try to come to this morning and I think eight out of 10 are for free usage today. So that's the way to afford it.
Free as in premium or free as in free. Free as in free. Until you. Until what?
There is no catch. I told you a lot of times in my life and then eventually there's a catch somewhere somehow. It might be that this whole thing ends at some point where these models. The catch is temporary.
Where these models. But what we've seen over time, this has been going on for half a year now. And there's models that kind of launch for free and at some point they become paid. But the rate at which we models are launching is faster than the rate at which they're getting deprecated.
So there's now more free models on the market than they were after a year ago. That's interesting. Eventually you would think you'd have to make the money somewhere. Yeah, maybe this is like when VC money was plentiful and the rivers were really affordable.
VC dollars are plentiful and the leading labs are free models. Right. I use it while you can. You're a ski alert to burn.
Burn the VC money out. Burn the VC money. I think that's a legit strategy. The challenge with it is over time I think you do become hooked.
It's like drug dealers, the first ones always free. In this case maybe get a lot for free. But eventually it's like I don't know how to write software any other way. And then three years from now and that's when they turn on the old money making machine and now you're just subjected to renting your life away.
Yeah, I think what you're hinting at is that the cost will go up. I think the cost will go up. Not so much because the same quality model will get more expensive. The same quality model gets 10 times cheaper year over year.
What we've seen is that the price of a frontier model per token has kind of held steady. Every time you think it's going down there's like GPT 5.2. What is it? Thinking or pro?
A really expensive one. And so that's holding steady. But the amount of tokens you can burn as a human, that is shooting through the roof. Yes.
With this split up the roof. Even with my only one agent. I can burn some tokens. I imagine if I had six agents going, that's just six X my token ability.
Exactly. And the future we're going to live in is not one where you have like a $100 or like a $20 subscription a month. I think humans are going to burn $10,000 in tokens, maybe $100,000 in tokens per human. You think so?
And there'll be 100 times more effective than the humans of yesterday. Right. It'll be worth it because one human plus all that is more effective than exactly how everyone humans would cost you that 10 grand. Doing what?
That's the question. Is it just coding? Is it just playing? It's just next business, man.
It's coming out as ears doing everything. Every knowledge worker is going to have a collection of agents working for it. It's a wild world. It's interesting to think about how that plays out though.
It's almost unimaginable to think about. I couldn't imagine this year from last year, for example. I could, I suppose, but the vantage point, the perspective is it shifts so quickly and so much. Like, for example, I just mentioned Google and their flash model.
That's like news as of I believe today. Like in the last 15 hours, you know, every day or week we have a new model, new flip, a new change. And it seems to be like UX is the moat. UX is looking figured out.
I think he was an example of how you're trying to evolve or think about different UX around agent usage. It seems hard to predict where we'll go. It's incredibly hard. And like, we thought this like AGI where the AI would be as smart as you want, this is seminal moment.
And it happened in April. It just flew by. And we hardly even noticed it. You said we had AGI in April?
Is that what you just said? Yeah. I think in April, we, the models became as smart as the median human. Oh, is that way?
AGI is? Artificial general intelligence. My definition, I stole April from Tyler Cowan, which I think is one of the smartest people on Earth. I must have missed it.
I don't know how generally, I guess maybe you're saying because most humans are generally unintelligent. Well, also like, think about CHAT-GP-D. It might not know more than the expert in a field, but it knows something a lot about everything. The ultimate polymath.
Yeah. Yeah. I mean, like I was having some really interesting conversations with CHAT-GP-D that was very therapeutic. You know, I have been on the code side of things and that's where I'm comfortable with how it knows what to do.
And I guess on the human side where I have a conversation with it and it's not just a question and answer, it seems to know such detail about particular things. And I don't know how it does that, but it was a therapy conversation essentially. And it was just uncanny, the response type and the level of care. And it wasn't just how it outputs.
It wasn't just formatting. It wasn't just user experience. It was this depth that I never expected a machine basically and potentially even AGI who that knows. To give me back this feedback loop, I just never, like, I'm steeping this stuff.
I think we've been steeping this stuff for years and it is just quite literally unbelievable. Unbelievable. That can have that kind of conversation. I'm being vague to a degree, but it was very therapeutic.
It's incredible. This year it helped us a lot. Think about how much stuff I had to learn this year. And even yesterday I was like, I don't really know how a TCRT is different than a CAR-T.
And wow, CHAT-GP-T walked me through and all the pluses and minuses. But we're going to give a talk at OpenAI soon because we did some things that were super beneficial to us. We didn't RNA test of my cancer. We know what RNA was going on.
And we just sent a spreadsheet to CHAT-GP-T. And what came out of it was really insightful. Can you be specific in any way? Like what would your mind, I suppose?
What you're looking for with RNAs, like, not so much, you have the genetics, but what's actually happening at the cell level, what's being expressed, and what are potential pathways that are over expressed. And you can target. Because I know from the genetic, there's something that sounds like a party drive but isn't, MDN2 is over expressed. But I want to see how, like, the thing it acts on, the P53 division, how that's going in my cell.
And there's better examples here. But I'd have to ask CHAT-GP-T for them. I don't want to be distracted while talking with you. But like for sure, just incredible things.
And it's not the end-all be-all. Like, don't take the treatment without first checking it with a doctor. But if you have to get smart about a subject in a year, what an amazing blessing it is. It can connect dots.
That's where I like, and the dot connecting is across the board of all sorts of different disciplines and domains. That's why I said polymath is like, that's probably not even a great example. It's like, you know, Uber, Ultimate, Ultra-Think, polymath. I don't know.
No, polymath is great. And that's also the big benefit, right? I talked to hundreds of specialists this year. But when you start talking to people who are extremely narrow, because they're like the expert in the world in one specific thing, it's really hard to do the integration.
You end up in these meetings with like eight people and they're incredibly expensive and hard to organize. And something like CHAT-GP-T, which is like very knowledgeable about everything is incredible for integrating, and don't know like, giving you directions to go in. Yeah, that's where I really, maybe this, I don't want to tangent us by any means, but this is where I get really, really conceptual, I suppose, about where the future might go. Because we need, I think at least, even in this moment, we need humans to go super deep and super thoughtful into a very specific domain.
And like you mentioned, you have to have eight humans in a room, hard to organize. But we need those kind of folks to think about the long-term future humanity. And not so much like in this way, I'm saying it, but like, feed this algorithm, this LML, this model, it's knowledge, not to replace them, but to augment how fast we can iterate through a problem set that requires eight humans and a polymath that's above those folks to go so deep on a subject that's just basically impossible for a collective mind to do in one briefing room. Like it's just humans don't work that way.
We don't share that way even interpersonally, but a machine or an AI that's designed to do that can for us. And that's really where I, it's kind of like a headspace kind of thing, but that's where I'm at. So it's like, you got to have these super deep thinkers with this deep knowledge. And they're obviously going to be human because that's what we've been as a race to enable this AI to think so deeply and so vastly as a polymath would and connect us that we just wouldn't normally connect as humans.
And that's just saying if I've been able to look further, it's because I stand in the shoulder of giants and it is in a very narrow field, you're able to, some people are able to read all the relevant publications and to keep up with that. But it's incredibly hard work. And now with these AI things, you suddenly have a thought partner that not just read everything in the field you're in, but in all other fields. So you're not forming these people standing on top of each other in one discipline.
It's like it's as far as the eye can see every discipline you're up to date on the latest knowledge that it's just incredible. It allows you to run around where before if you moved outside of your field, the whole thing fell down and now you're just able to combine disciplines like never before. Yeah, it truly is amazing where it stands today and also where it's headed, which none of us can really know, but we can see that trajectory moving us forward. So you just got this engineering meeting.
All the ideas are great. And now you've got to go through your notes. This time consuming part, read them all, digest them all, add action items, put it in the right places, tag team members. This is all necessary work, but it's tedious.
Okay, so flip side that take that same position and flip it over into notion and using notion agent. Notion agent does the busy work for me. It's like having a project manager that keeps everything on track in the background while I focus on the bigger picture. Get my work done.
Do my best work being the artist. Notion brings as you know, all of your notes, all your dogs, all your projects into one connected space that just works seamless, flexible, powerful, and actually fun to use with AI built right in, you spend less time switching between tools and more time creating great work. And now with Notion agent, your AI doesn't just help you with work, it finishes it for you. Notion agent can do anything you can do inside Notion.
It taps into your workspace, the web, and connected tools like Slack and Google Drive to complete assigned actions end to end so you can focus on the hard decisions. It's like delegating to another version of you that knows your style, knows your workflow, and knows your preferences because it learns from how you work. With a single prompt, Notion agent forms a plan, executes it, and will even reassess and try again if it's a snag. You can do multi-step tasks like creating new pages or databases from scratch, summarizing entire projects, and does this all in minutes.
You will sign the tasks and your agent does the work. And since this is all inside Notion, you're always in control. You tell your agent how to behave, and it will remember and update automatically. Everything your agent does is editable and transparent.
You can always undo changes so you can trust it with your important work. And Notion, as you know, is used by so many people, over 50% of Fortune 500 companies and some of the fastest growing companies like OpenAI, Ram, and Versailles, they all use Notion agents every day to help their team send less emails, cancel more meetings, and stay ahead. So try Notion now with Notion agents at Notion.com slash changelog, that's all lowercase Notion.com slash changelog to try your new AI team eight Notion agent today. And we use our link, you're supporting our show again Notion.com slash changelog.
You mentioned $10,000 a month, something like this, you know, with a human and some agents doing their thing in the future. I just saw a post today about cursor, some of paying $14 a month for cursor and someone saying that's ridiculous and the other person saying it's worth it, et cetera, et cetera. We can have these kind of battles, but let me think back to Kilo because I did see on your homepage, you're kind of attract trying to attract cursor customers. I think it's even a billboard somewhere in the valley calling all cursors over to Kilo.
Can you explain the angle, the relationship, I'm sure you're playing a cursor users, listening to the pod who are very familiar with that particular tool and maybe use that as a way to explain what Kilo looks like and how it works and all that. Chris, there's always a little bit vague about like what is less price and how much do you how much do you pay and make significant changes over time with Kilo. We pride ourselves on you can use over 500 different models, including a ton of like stealth ones and free ones. Also, if you're going to pay for a model, you pay less price.
We charge exactly what the provider lists as a price. So you know exactly what you're getting for your dollar. And I think as people start kind of consuming these AI tokens more and more, that is the way to go instead of like a subscription where it's a fixed amount. But if you go over, there's a chart which might probably hide on this price, but it's hard to exactly know.
We think a great mix of free models for the problems where you can use those plus paid models is exactly less price without any commission or overcharge is a compelling proposition. And so Kilo uses this orchestrator itself in order to pick models in addition to orchestrating your own tasks based on complexity and the particular needs. Right now people are still selecting their own models in Kilo. It's not trivial to select the best model and we also have their own preferences.
And what people especially don't want that to ski low, they don't want the model to switch up on them. They don't want auto model where you don't really know what it did. And they also want to be sure that they're using the exact same thing the whole time. So if I say I want that, I want exact that with the exact context window, because it's very frustrating if you're randomly using another model or somehow the context window is half of what you're used to.
Right. I think humans are really good at kind of selecting the right model for the right task, which you don't want to, you don't want the software to change it up unless they give you permission and then still you want to know why exactly what you used. That could be configuration in a way, right? You can automate that selection through configuration.
Obviously, we went conventional for configuration in a lot of cases, but that's a way where you can say, well, when you do these kinds of tasks, I'm cool with you being on the flash model, for example, or on the haiku model, where it's faster, more iterative, planning, whatever. And but when we're planning, we're going to go deep and we're looking at a geotechnical or a linear ticket or we're looking at a pull request. And we're looking at this. I want, you know, the better model, the more recent model to examine that for me.
And you have profiles and you kind of set your own profile. That's a combination of one of these models, but also one of the include these sprouts. You can even share these profiles across your organization. So you collectively work together.
Like this is a really good agent for upgrading our Java code or something like that. Because those are going to change too. And, you know, your mileage may vary. So you may use a particular model and you're a Rust developer or you may use a different model and you're a Java developer and you may get better results per your team with a model.
It's something you should probably be able to select from, but you're still getting top of your models, but in a config where you can say, well, I prefer Opus 4.5 versus the latest Gemini, for example. Yeah, and you're basically creating team members and just like in a real company, like it's helpful to have the same team member. Now, these profiles don't have memory yet. That's something for the future.
But they already have a collection of prompts that are super helpful for that company, for example, in simple way, your style guidelines are like, I tried to upgrade our Java before and it needed these and these prompts to be effective. So being able to share at least these prompts is super helpful. How does Kilo manifest itself to people? Like what does it look like?
How does it work? You mentioned it's in like all the major editors. So I imagine it's just in the case of VS code extension and there's a win surf deal. Can you explain kind of the service area of the product?
Yeah. So all the VS code based editors, all the JetBrains editors, it's available. Here's a CLI that you can run anywhere and then we also allow you to kind of run it into cloud agents. That's cool.
And then we're soon launching it. We're so mobile already, but we're soon launching a mobile app as well. Oh, explain how that would work. I think one of the cool things is if you have a thought, you can just kick off.
Yes. With that. So instead of kind of give it to me. Yeah.
Look, you're somewhere. You're not going to go to your desktop. You have an idea. Now you write it in your notes or whatever you do tomorrow.
You just kick off an agent and then you can pick up that agent whenever you get behind your notebook again. It's going to be available in JetBrains, in VS code, etc. So all your sessions, you have a complete overview of your sessions, whatever platform you're on. And I'm looking forward to also getting kind of a mobile notification when my agents are done because that's my primary thing.
Like the agent is done. I need a thing so I can see what it produces. This is something that Cloud Web is what I call it. It's just Cloud.ai.
But it's not Cloud Code. It's Cloud Web. They just enabled you to connect a repo on GitHub. And I think this is like early days of this feature.
And I didn't really have a repo to connect it to to think about. So I was like, I just want to do it in a isolated area where there's no damage. You know, I can't go right, code accidentally or some pull request or do something weird. And I connected to a brand new empty repo.
And I just kind of one-shot it in a way this idea that was really just a basic GoCLI. And in this case, it was kind of a coding session, but I was just trying to find experiments. Okay, if you have this mobile interface or this web interface that is not a coding environment at all, it's not Cloud Code in the terminal. It doesn't have terminal functions, it's a neighbor system or things like that.