The future of software development is being transformed by AI agents, which are enabling teams to work at unprecedented speeds through radical transparency and real-time collaboration. Developers must adapt by building AI literacy, understanding that software development is fundamentally a people problem rather than a technology problem. The key to success lies in maintaining human judgment, accountability, and the ability to identify useful problems to solve, as AI tools become increasingly capable but cannot replace human creativity and strategic thinking.
Install our extension to search inside any video instantly.
The Future of Software with Steve Yegge and Ajit Banerjee, moderated by Milkana Brace.Added:
Welcome everybody.
Thanks for coming into AI House. My name is Audrey and I'm the events and community manager here. I'm just going to go over some quick housekeeping stuff. Uh first, do not go onto the deck. The doors shut behind you. You're going to get stuck out there. Uh the bathrooms are down the hall to the left.
Uh don't run into the clear glass doors.
It's going to be really embarrassing. Um and you might have a bloody nose. Uh really quick also for the raffle. I don't know if you guys saw on the front uh the QR code. Uh the top 10 answers get a signed copy of Steve's book and then the rest uh we actually we have uh some AI house shirts over there. So everybody that participated is first come first serve. So if you want to run to the table afterwards and grab an AI house shirt uh those are there. And I'm going to turn it over to Taylor, our new AI house director.
>> Thank you Audrey. Hey everyone, I'm Taylor Soaper. I'm the new uh AI house director. Uh previously Thanks. Yeah.
Uh previously was at GeekWire uh for over a decade covering our ecosystem and super excited to be here at AI House and help build off the amazing work that Eion, our managing director at the A2 incubator, and Audrey, our community events manager, have built over the past year uh to build up AI House. Uh I just want to say thank you. Thank you for showing up. I know everyone's super busy. Feels like some of us with all these agents were busier than ever. uh we're building, we're selling, uh token maxing, you know, all all the above. So, thank you just for showing up and um uh just being here and being part of the community, um sharing your knowledge and hopefully getting inspired to build build more, launch a startup and all the above. So, just thank you for being here. Um I want to tell you a bit about u A2 incubator. So, AI2 incubator, we are upstairs. Uh we're one of the leading early stage uh funds for AI founders. uh we invest up to 600,000 uh in companies. Uh this could be have two co-founders, you could be a solo founder. Um but we're there at the very early stage and we help build and scale companies as they grow and accelerate.
Um we this is AI house as you know this is kind of where our community comes to life. Um we have events here uh every week for sure almost every day this week. Uh last yesterday we had pitch please which is our pitch competition event. We have five entrepreneurs pitch in front of a panel of venture capitalists. Really fun event. Um, tomorrow we're doing our open house.
This is our first time doing this concept where we're inviting founders uh to pitch their ideas to hang out to work out of here. So, we're going to be doing that every Friday. So, if you aren't going to be there tomorrow or on Friday, please just come back on on Fridays, especially in the summer. It's a great time to be here. Um, yesterday we were going to have an AMA with Tim Porter from Madrona, but that got postponed um in light of very sad news. Uh you may have heard um S Sigar from Mad Drone Adventure Group unfortunately passed away this week. Um it's terrible news.
He uh is such a pillar of this community. Um I just think of S. I remember reporting for GeekWire. Um he would always respond to my emails like whether it was a quick comment for a story or a longer interview. I just always remember someone being super generous with his time. And I think as you read some of the tributes to him, generosity really just is what he was really all about in his life and his work. So, um, we just wanted to express our condolences to his family and his colleagues and our friends at Madona.
So, um, yeah, S was a great guy. Um, I just want to say a little more about, uh, this space, uh, AI house, as I've said like 10 times in a row, but the AI house, uh, it's a public private partnership, which is really unique.
We're supported by the city of Seattle, as well as our partners at JPM, JP Morgan, and Google Cloud, and our community partner Ada Developers Academy. Uh none of this uh would have been possible a year ago to get this off the ground without our partners to build a place where builders can can be and and grow their ideas. Finally, uh if you're looking for a place to work, Waterfront Workspaces has you covered.
They're the ones who run the office space here. Um if you want to work alongside other startups like Sage, who's just down the hall. Um they'll they'll help you out. And if you mention AI House or AI2 incubator, they'll help you out even more. Um I'm going to stop talking. Thank you again for being here.
Really appreciate. And I know a lot of people first time here so thank you can come back. Um and I wanted to give a big thank you to Sage Ox for for being here and Steve. Um really appreciate you all and I'm really excited for this conversation about the future uh future of software. So thank you. I'm going to kick it off to Milana.
>> Thank you. Hi. Is this what I think?
Okay, great.
>> Thank you. Well, I'm Elconor Brace, co-founder and CPO at Sage. Welcome. So excited to have you here. I would say let's just jump in. And we have two amazing amazing thought leaders who will have a lot of interesting insights to share. So let's grab seats. Don't step on the chair rest like foot rest.
Apparently it's not safe.
Okay. So I would like to introduce Steve and Ajit. uh probably many people uh already know them but they have been around the Seattle uh tech area for a very long time for well who's counting for a very long time they worked together uh way back when maybe two decades ago in early Amazon days but since then um Steve has spent time at Google he is a very prolific writer who has been writing for a very long time well before AI about tech culture and technologies and languages. And more recently, he has really become at the forefront of the AI revolution. He wrote the um book on vibe coding, which some lucky people here will get a signed copy of. Uh he is also the creator of Beads, Gas Town, Gas City. I can't keep track of all the other things. Steve, you can talk about them, but we are so lucky to have you. There was one week last month when I opened New York uh New York Times and you were profiled and then there was an HBR article about you and then you were on Tim O'Reilly and this was like within a few days. So you're a very busy dude. We're very lucky to have you.
Welcome.
And then we also have a Jeet Banner, also a legend. Uh someone that I've worked I'm very lucky to work with uh here at Sage. He is the founder and CEO at Sage. Um this is not his first rodeo.
This is his fourth startup. Um he's had several exits before. Your most recent startup before Sage. What was uh ZedHub acquired by Hugging Face. Um and I think the technology you built is currently powering over 100 pabytes of data and used by millions of developers on a daily basis, which is incredible. Um, and yes, you've made um you you've made rounds at Meta and Apple and you were on the very original uh EBS team at AWS uh all these many years ago. So, you have a lot to offer to this conversation and I think you guys have a very different perspective in a way on what's happening. So, I'm looking forward to hearing where you agree and where you disagree.
So, with that, Steve, let's start with you. Um, you each have a mic, so hopefully it all works. You have to turn it on. So, Steve, what was your leap into AI? How did you get started on this path?
>> Ah, whoa.
She said, "Shove it into your face."
Okay. Boy, it's a It's a good mic. Good mic.
Can you make it make me sound more manly? Is there like a setting or so?
How did I get into AI? The same way you all did on November 1stish 2022 chat GPT came out and I immediately made it tried to write Emacs code Matt and it kind of did. I was like whoa. Yeah. And uh I would say the sort of pivot points from there were uh I mean sort of realizing that it was um viable realizing that I was racing ahead of everybody else like everybody was in the enterprise was focused on completions and I was doing chat and then like I got into agents and everyone else was using cursor and chat and then I got into multi- aent and everyone else was like cursor still you know co-pilot and uh and then I had the death of the junior developer. You know, there was like a kind of a realization. I was the head of engineering at source graph uh when all this went down. They were a code they are a code search company which is still a pretty good business to be in because we're generating 10 times more code than we used to, right? So, you need a lot more search, right? But AI dropped right then and we had to pivot to be an AI company and like we had no idea what to do, right? So, we built a coding assistant and um I'll be honest, like I'm not a I'm not the kind of person that gets jealous, you know, of other people's ideas or anything, but clog code was such a good idea. I mean, gh I thought I was so close, right? I was talking to my boss about it at Source Graph at the time, right? Because I, you know, I could tell what was going on. We they needed to use tools. They needed to get the human out of the loop, you we needed to get a for some reason I thought we were going to do an Emacs, but the reason really that I missed it was that I'm old school and I grew up when bytes mattered and so I'm a little too frugal to have thought of an idea as crazily pigish as cloud code. It's like let's just take tokens and put gasoline on them and light them on fire and see what happens, right? And that's been like that the form factor changed from completions to chat was like a 10x token burn, right? at least and then chat to agents another 10x and then agents to orchestrators running bunches of agents every time we change the form factor it's because we find a way to spend tokens 10 times faster right so anyway that's how I got into it um we could go anywhere from there >> well thank you let's go to Ajet and hear how he got into it and then we'll we'll take it from there what was your path into agentic engineering or vibe coding I don't even know what language to use this crazy world >> so the first thing I want to talk about is that I would have never dared to be in front of 200 people next to this legend unless this world was on. And what I mean by that is that I have this term now called BC, before Claude.
Nothing about my life before December 2025 is relevant to my intellectual thinking anymore. I'm like this quiet infrastructure guy. I would never be here. And you know it's almost like I've seen the light and that is why I have no problem being here now because I think what we are talking about Steve myself is something that is very important for people to hear and this is so animated in our team is that we challenge each other like hey is that statement BC or AD you can't build that? Are you sure?
Is it because you're like scared or something? You know, this device was launched in March.
We came back from our fundraising trip on the weekend after we came back. Most people are exhausted after fundraising.
We're jumping up and down because we're going to build this and we're going to be building this and recording on this.
So these kind of ideas where you can do things which are unimaginable is something that happened in the recent past and I want to give credit to a lot of people like Steve who kind of opened our mind about what's happening. So that's the first thing BC kind of concept.
The second thing is you know Ryan and I have worked with Steve way back in our Amazon days. We are like really really old Amazon parts. I was in Amazon 2003.
I don't want to talk to people about when Steve joined and when Ryan joined.
From the beginning what we have been focusing on on Sage has been teams. We believe software and knowledge work always will be a team sport and I think there the collaboration with Steve has been very insightful from the beginning and this is the part which is going to make everybody uncomfortable.
What we have found for a team to cook the way we are cooking and I think which are many other teams are going to start cooking is a level of radical transparency which makes incredibly uncomfortable like experiences. When I told Milana that she was joining us and we'll be telecasting everything all our decisions her reaction was literally to cover her stomach. That is the level of discomfort one feels when you're at that level of and it has been very um it's been very interesting for us to see that Steve has also been telling us that this idea of like actually putting your dirty laundry while you're thinking and sharing it with your team is essential if you're going to be doing this level of speed up. The final thing that I want to just say that I've been having the time of my life So, we raised our seat round $15 million. You would think that that is something that I'm excited about. I left Hugging Face in May of last year. I walked away from three years of vesting and the team, the energy that we have in our team is something else to watch. You know, Steve comes in, we should be excited about this talk and I'm dragging him to show us some demo. We're jumping up and down. There are days when Milkana has tears when we are talking about this thing. We are in a place >> of all and joy. I want to be clear.
>> Yes, there are days where you would be looking at us and we are like kids in a Lego store and that is something which for all the talk of efficiency not many people are talking about which is that when you start cooking in this way you have fun and I want that to be kind of said loud and >> let's double click on what does it mean to cook this way? what does it mean to work in this AI space? Uh what is the SOX way of working? And then I'll turn it over to Steve. Like how can Steve stay on top of all of his projects and how does he work? But how do you work?
>> So we basically have two small rooms in this space. And you would think that we would keep our teams separate across those two rooms because they're small, but yeah, packed. And we actually sit right next to each other at this distance. and Ryan sits next to me and G sits next to me and the delight happens when we are working. We don't talk, right? We have our standup and the standup is in the morning and then you go into your deep deep kind of concentration space and then suddenly something shows up which is like Ajit Ryan has figured out that you're working on the V1 of this code base and actually there's a V2 of this code base and that kind of drops like a like a little drop of wisdom on the side because we are having all these agents murmur to each other.
The first time it happens, you're just like, "This is this is surreal. This is like something like Hunger Games when the parachute drops in with a little kind of an idea." And then you get addicted to this level of flow to the point where you can do this. You know, you can almost anticipate what the other person is doing. I think I think there have been great teams like this which have been always able to live at this level where they complete each other's thoughts, where they can pass the ball to each other in a no look pass. I think this technology because it speeds everything up allows many more teams to have this experience. And I think when you're cooking with a team of people that you like to cook, there is not a price that you can pay for that. It is something sublime. And I want to really talk about the joy aspect so much because all of the AI conversation is there's a gloom and malaise and I hate that stuff. I just want people to know that you can do newer funner things with this which were not doable before and I would love for many more to have that experience >> and I think um some people are well familiar with sage others are not. So maybe for a little bit of context we do focus on the context of the team shared context of the team. So when Ajet was describing these murmurss that happen uh when he is in his sessions with Claude, they happen from other agents that are working on the team that are emitting um updates on what it is that they're working on and that's how in real time you can see what's happening across the team both across humans and across uh the agents that are part of the team as well. So, um I I think AJ didn't take enough credit for the fact that yes, Claude is amazing, but we've also built a lot of these experiences that actually make those connections to happen in real time. So, Steve, how do you work? You have an army. You have a fleet of uh agents that you you use yourself, but tell us more. How do you manage? How do you keep it all? I know you built obviously tools to help you do that, but like what does your day-to-day look like?
>> Yeah. So I, you know, I'm just curious like in the audience here like who here uses like kind of like one agent at a time typically uh one or two agents throughout the workday?
Yeah. Okay, good. Okay. So, who here uses like I don't know three or four agents more.
Okay. So, uh we've got a pretty pretty AI literate audience, right? And uh >> we are in AI house.
>> Yeah. It's which is great. Um the way that you know the way that I work is uh you know it's it's I've shared it with the world, right? I made beads and that was the way I worked and I liked it. And then uh Gas Town kind of like doubles down on that. Uh Gas Town is all about swarming. Um you know, so I mean the basic workflow of Gas Town is you create work and then you do the work and then you create more work and then you do the work. And so you'd be like to create work you'd be like do a do a design right and then once you've worked out the design you say okay file a bunch of beads right file an implementation plan go review it blah blah blah you've just created a bunch of work and now you can say swarm it right and it'll go grind away and that's just that that cycle right of um of grinding through stuff and I love it right I think it's a great way to work um but I encourage people to like write their own orchestrators right like as soon as you got about four agents, it starts to get really hard to keep track of which one's working on which and then like you accidentally text the wrong agent and you like give it this big project and it like tries to make you happy by doing it in the wrong place and then you got to clean up and right like the cognitive overhead is just really high. So like when you get up into that space you you start to build your own you know your own stuff right and I think that's where a lot of people are right now. Um so yeah that's uh that's kind of where I'm at. Can you talk about the cognitive overload?
You've written about it and we've chatted about it quite a few times, but how do you stay sane to the degree you can stay sane?
>> Yeah. Well, look, I mean, for starters, right after I published Vibe Maintainer, like like when I was in high school, there was this one dude in in track, he was a freshman, who just like he just took off running ahead of the whole pack, right? just sprinting in this like we were running like miles and he just like flops over after a quarter mile, you know, completely. That's how I felt right after like beads and Gas Town and the wasteland and I got all these pieces out and and honestly writing the blog posts was as hard as writing the code because you got to have characters with costumes and like right jokes.
So I got all that done and it was just I mean I was getting hammered by 50 to 100 PRs a day, right? people are just like so I did I did the the sane thing right which is I got help right I turned it over to a bunch of people who are you know maintaining it now and uh and and then I was able to like step back and the first thing I did interestingly was I when I took a break was it was like right when Opus like started having issues do you guys remember that about two months ago everybody was it was all the meta onx us was like unusable for a while and I just wasn't coding at all. I was like kind of out there just trying to figure out what's next because that's like like what I'm always trying to do. I mean I I can tell you what's next is just absolute, you know, pandemonium, right? I mean like the next five years is just going to be an incredible amount of change compared to the previous like 20 years, right?
Just absolute madness. And and what I realized was that I was racing like way too far out ahead of everyone, right?
And I was blocked on basically enablement, right? Like like adoption like uh AI literacy. Yeah. So I want to share I want to share with you this story. I want to share this story from Ezra Savard at Netflix. Okay. He's the guy that trains everybody at Netflix on AI.
He's been doing it for three years. and he shared with me just the most absolutely like gamechanging stuff, okay, from their training program. They actually teach our book, me and Jean's book, the vibe coding book. All right, and they they have like line chefs and they teach you how to do it vibe coding in in cohorts and whatever. over the period of you know what with a lot of data driven uh analysis they found out that there are basically three cohorts of AI literacy that you need to know about and care about in the industry in all places. All right, cohort number one and there's probably none of them in this room are the ones who spend zero million tokens.
They're the kind of non-users, right?
They chat once in a while. Cohort two is single agent synchronous throughout the workday.
Okay, they have achieved the first level of AI literacy which I would argue is the baseline that your whole company needs to get to before you can even think about pivoting, right? And then and it turns out that you can get people from the zero cohort to the one cohort in four and a half hours. If you bring their manager and their team and they're on the clock and it's blessed and they bring their actual work with them and they have a good trainer and they sit there and there are no more than six to 10 people. All of these have to hold true. But then five hours later they will be move they will they will be like I get it. and they'll be using an agent single syn sing single agent synchronously to do their work and their mind will be opened at that point right because they'll realize oh I can I can do stuff right and then the second cohort and the only one the only other one that matters is 12 to 15 million tokens the first cohort is four million tokens a day roughly second cohort is when you finally trust your agent enough to let it go work by itself and you spin up multiple agents and you just start like reviewing their outputs that's uh 12 to 15 million million tokens a day and you are officially like basically trainer level good with AI.
Okay. And any more token spend metrics above that is is just vanity metrics and and and and BS, right? You'll get nothing from it. In fact, what's what it appears is that in know at first you have to get everybody up into the second cohort. You just you just have to and by the way you can after a few weeks I don't know exactly how long Ezra knows like a couple of months in cohort one you take one more class same format same whatever and they'll get you into cohort two forever right so it's it's it's totally solvable right the training problem but we face this training problem headon right I mean like enterprises need to and it's a culture problem too like companies are filled with people who like their job is to say no and that's their whole identity and they're like I get to say no a lot. I knew a guy who just like he just like yeah the buck stops with me and I'm like okay buddy you know stop a lot of bucks then. Um those kinds of people will kill your company as you try to pivot to AI right because the shape of your company is going to be really different. Yeah. You know uh like because the way that you do things is going to be different because the way that you build things is going to be different because you're probably going to be working like these guys. Did you did you all read the um did you get a chance to read my post called the uh the AI hive? No, the anthropic hive mind that one there's also AI vampire the anthropic hive mind that was based on these guys and anthropic right because I talked to a bunch of people at anthropic like 40 different people across the company and executives founders sales Boris everyone they're a hive mind like you can argue about whether they're a cult or not you know semantics But they're definitely high mind, right? Like in in the sense that like there's there's none of that none of that stuff that your other company your company, you know, where there's hierarchy and there's decisions and there's planning and I don't know, grown-ups like organization. None of that stuff exists in Anthropic. They are just like buzzing like all together, right? And making stuff. It's pretty nuts. So are they, oddly enough. It's really strange.
I mean, like, when I come by, they're all just like tripping out going, "We're so happy." And I'm like, "Is this safe for you guys?" Because it's they're not acting a fully normal, right? The grin is a little too wide. I just watched smile and it was like But anyway, uh, so I think got there.
>> I feel seen.
So Steve, you just gave us the recipe or the playbook for how enterprises should adjust to this world of the training and the cohorts and whatnot, but what about individuals? What about like the people in this room or their friends who are trying to figure out ah is my job at, you know, at risk and how do I change?
You know, a line that I've heard and I've kept saying is that you are either at the beginning of your your career right now or you're at the end of it. So you are either getting on this train and it doesn't matter how many years of experience you have. Uh this is the beginning for all of us. Or if you miss the train, you miss the train. It's the end of your career. So what advice do you have for individuals who are not going to wait for their company to train them? What do they do? Where do they start?
>> Yeah. Choo choo choo. The train's leaving the station. You better get on and be left behind. And that train is anthropic. No, I'm just kidding.
Although that's what you would think if you spend any time on X at all. Um or if you ever go to San Francisco. Um it's uh yeah, it's a it's a crazy it's a crazy time to be sure. One thing to be aware of is that you're not any further ahead or behind than anyone else.
You're not right. We're all roughly plus or minus six months from each other.
None of this stuff is really that hard to figure out. So like if somebody else makes a bunch of breakthroughs, you can catch up to them pretty quick, right?
And uh I mean look it does take realistically it takes a year to build like your AI muscle right about a year I think is that Jean Kim has been sharing metrics there's been some studies where to where you build up enough like trust with models that you can actually like predict what they're going to do pretty reliably and then like you know what I mean you're not going to do stupid things and you're not because it's tough right they'll screw you any chance they get those models will right any chance uh and this is this is a real problem people don't realize this right now right I mean like okay so like okay we're at the beginning of like lots of big services catching on fire and burning right like GitHub was the first of many to come right because everyone's like doing exactly the wrong thing and and leaning too hard into it and vibe coding stuff into production and trusting their LLMs and and it catches fire and burns. And so, I mean, look, if you're looking for an opportunity, I predicted a year and a half ago that there would be a new profession called the fixers. All right? The Winston Wolves that come in and just fix your screwed up vibeccoated mess, right? You that you right, you all could seriously just like we should just print shirts.
You guys are already you're already qualified for that, right? Because like you know, you're already out there experimenting and you understand how dangerous this stuff is. I don't know.
there's some sort of transference problem where CIOS and CTO's play with at home and then they go we're going to use this for prod, right? And then all of a sudden planes are crashing, right?
So, uh it's I mean like so this is why I mean honestly this is why I I've I've had lots of stuff in the works. Lots of I've got a lot of cool projects. The gas planet ecosystem or whatever you call it is it's got a life of its own, right?
Um, I mean really surprising like arms of it are getting funded and and and it's cool, but uh I keep getting styied because everywhere I go people are like not like you. If they were all like you, life would be so much easier, right?
like they're they're just like I spent an hour on a phone with a company who shall remain nameless but they're kind of French and um and uh they're they spent 45 minutes on the phone with me because they're going to bring me because I get p my job now is I get paid to come tell everyone that everything's going to be all right.
I mean whether it is or isn't right.
They pay me to come in and say it will.
Right.
>> Do they pay you more if you say it's going to be really all right? Like extra great. All right.
>> Uh, you know, I look, I usually have a way to explain to them what their their best way forward is, and they usually buy it. So, so far it's been it's been working out. Uh, but anyway, the Frenchies, it took them 45 minutes to say this.
They talked around and around and around it. But what they were saying was our CEO hates AI.
That was that was it. And they wanted me to come give a talk that would fix it, right? Uh so that's that's my my job, right? Is to go fix it. But it's a it's a serious problem, right? Because like the CEO is just like a lot of people.
Code is a craft, right? AI makes swap and blah blah blah blah blah. Right. say that with a French accent, right? And it it's a it's a it's true at every every company's got skeptics, right? So like this the the world I mean like the job market that you're in and the world that you're in.
First of all, it's just so incredibly unstable right now. I mean like um I mean I'll be honest with you, we're all grown-ups here.
Uh and I'm not employed anywhere that could fire me. It's really great. It's really great being independent. I can tell you just not having a boss. Like you you tweet something bad and you wake up in the morning, you're like, "I'm fired." Oh, I'm not. Um, uh, and now I've forgotten what I was going to tell you guys.
>> And okay, how do we move on from here? How about this? Uh, you just talked about how you make money in this space, right?
You've created a lot of value that you've shared openly with people and one has to wonder, well, what do you get out of it? So you just talked about uh people or companies pay you a lot of money to um to follow your advice. Um what about you Ajit like you've taken a very very different approach which is hey there is a big company to build here. You just raised $15 million. How do you see the opportunities in this space? There probably a lot of founders here. So very curious to hear your perspective.
So I think we have to go back to one year ago and you know I just sold my company to Hugging Face. I've got a four-year west in front of me. What the hell am I doing quitting and leaving that place and ending up in this place, right? No health insurance, no salary.
What happened was that I started leading this guy and the words that were coming out of his mouth did not make sense to me.
He started saying things like I have not seen the code that I have written for beads. I got another person who was a co-founder of a company called Inkto which was my first company which went down in a crashing flame. He he created this thing called Ador. It was the precursor of Clark. He started writing things like eder is writing itself. Now May of last year, those words didn't compute.
And one of the things that I've learned in 25 years of tech is when smart people start telling you stuff that the words don't compute, try to get to the bottom of it. So, this was one of those best decisions of my life, right? Um, launched Zetub. If you go on any hugging face page, you'll see a Zet logo on it.
And then Ryan and I basically got together and said that what would it be if we think about the world of knowledge work about agents as if we knew nothing about the last 25 years and Ryan being Ryan basically just said that hey I have never been an intern let's you and I the youngest and the second youngest principal engineers of Amazon of our time in 2004 just go in and clean the toilets and do the bottoms up work in startups. So Taylor over there was one of our earliest clients.
He worked for a company which is based here. We had another person and when people ask me about how are you so sure that this is going to change everything the answer is that we kind of were in this mindset where we were looking in a complete open way and December 4 2025 when Opus 4.5 drops like it was one of those 911 moments for me. Ryan and I are sitting together and Ryan's like, "Bro, this is not anything other than a full-on engineer now." And both of us just sat down. And I think this is where a lot of people here have had that experience of just getting kicked in the stomach. But I think what is very unique to the sage kind of team is that from the beginning we knew this was happening in a team for us.
So the question which comes immediately is that okay we go to any place we starting to see these clouds pop up all over the world right and we first thought we were losing our mind and then thankfully we had Steve telling us that we actually were losing our mind. So we have a photograph on LinkedIn in December 17. We're calling ourselves the first claolics. The word cla had not been invented.
But the insight that we had was that everybody's focusing on this individual productivity, right? And I've got 50,000 work items and they're just finishing up in so long.
Where does that developer get what to work on next?
And I think the most important insight that I want to kind of take from Milana's question is that always we believe in teams and what we also believe in is that what is interesting about AI is not just an execution engine for the past but suppose you are a company looking for product market fit.
How can you move that much faster?
If you can move 40 times faster, like I think you can, and by the way, it might be 400 times. We don't know how much further it can go. Then what you can do is get through the yak shaving part, which is also a word that Steve has taught us. And get to the point where you can actually see if your product is causing delight.
And if we can make a formula out of this right where we can go from idea to kind of production and just roll through this the first kind of idea is that okay this whole world of linear and GitHub and postth hog it all kind of collapses into some new kind of a category but what I'm again most interested in is how many of these groups of four or five people can start doing problems that they wouldn't have dared in the past And you know there going to be tons of new databases, compilers, devices built because people are fearless and I think I want to be a part of that that new economy and that is basically how I plan to make a lot of money by making a lot of productivity in the world happen. So, you know, I think there are a lot of founders in in the room here and one question we got uh pretty consistently as we were fundraising was why aren't you going to be roadkill for the anthropics and open AIS of the world and and I think in our in the case of Sage, we have a pretty um defensible answer um in that we work across LLMs and across um different stacks and we we enable the context for the team to be available whether you're working in claude co-work or in um you know in any of the the coding agents and assistants whether they exist today or down the road. So I think in our case we had a pretty defensible answer but a lot of teams are facing the reality of like code is free.
So if code is free how can you build business around it? You can build a consulting business in a human focused business of which we're actually seeing quite a few in a very surprising way or you can build a product that has some modes um that are that are built in. So curious to hear your thoughts. I have many thoughts about what are the modes that one can build around but curious to hear what modes you see uh founders here might build around and really protect the value they're creating.
Yeah, the first the first thing you got to do is make sure that uh you know exactly what happens to your product as models become smarter.
The bitter lesson, right? Uh so take a look at your product and if it's a bunch of personas or a bunch of skills or a bunch of prompts or something like that, then um the model's going to have all that built in next time around and your company's going to be out of business.
A lot of founders don't seem to know the bitter lesson.
Um, you've also got to build something.
I think these days, look, everyone can build stuff. Like, Gas Town wasn't actually that hard to build. It took about, I don't know, 17 days once I had beads. That took, you know, another two weeks. So, whatever, right? It took a month. You know, anyone can build stuff like that now. So, like, if you're in SAS, it's really, really hard to build something that other people are going to buy because they can build stuff, too.
And you got people like Mitchell Hashimoto up there on X saying take all your dependencies inhouse immediately uh for security reasons and for other reasons, right? Um and uh pair them down to just what you need and uh own them, right? Well, that's pretty anti-SAS, right? CFOs are going to be tightening their purses and it's really hard. And so like what what can you do as a founder to sell into that hostile market?
And the answer is you you build something that they look at it and go, "Well, I mean, we we could rebuild the Death Star, but that's kind of a big project. Maybe we could just rent a Death Star, right? And then you got yourself some viable SAS, right?" And uh they are effectively I mean like they're effectively building something that like jams a wire into his head and then into my head and we're like connected at a you know what I mean while we're working together because it's an extra channel.
I mean, I don't know about you, but I I I talk out loud when I'm coding with Claude. No, it isn't. You know, oh, no it isn't. I type it, right? Well, now you got some you got a puck sitting here listening to you, right? Hardware assisted, you know, hive mind here. So, you you look at that and you say, "Oh, wow. Well, this is just going to be more effective as the models get smarter, right? And they'll just be better at noticing that we need each other's stuff and Right. So like if you build something that rides that wave up, hell, you can even build stuff that's too hard for the models today. That's actually a really smart used to be a smart move until the exponential curve stopped because of mythos, right? Because they're holding it back. But you know, it didn't really stop. It's just held up for the general public. So you can still aim ahead. You you aim by what I mean by aim ahead is you build something that doesn't work.
like you're just yelling at the model going why don't you understand right and it's just like it's trying but it's too hard but you know that the next iteration it'll get it right that's what I did with Gastown and I was building with like Opus 4 right when I first started and maybe even Sonet 3 something and then by the time four five and four six came around right the thing actually started working that's what I mean by aiming ahead of the models so yeah you gota I mean like where you're at right now where we're all at I mean you gota I mean it's safe to Hey, you got to hustle, right? I mean, this is like this is a really hard pill for a lot of sw people to swallow. A buddy of mine like mentioned that about 20 years ago, CS became computer science became the money degree, right? The money major. And so, everybody switched from going to doctor and lawyers to going to computer science because that was what the money was. And and uh and so people weren't going into computers because they were geeks and hackers that had like, you know, done it since they were kids. They're doing it because the money. And so a a great portion of our industry is filled with people who are literally only doing it for the money and they don't really love it at all. And so when you come along and say, "I'm having the time of my life," they're like, "Well, I wouldn't be, you know, and uh it's a tough, right, tough crowd, tough tough cell."
They don't want to learn anything new.
They don't, you know, they don't want to give up their craft and they they don't want to um you know, they don't want to give up their identity of, you know, who who who they are and how they can produce it, right? So, um, you know, my advice to everyone is, you know, the world's changing real real real fast now. It's really fast. And so, you got to be super neuroplastic.
Nothing like fear and hunger to make you super neuroplastic if you need any uh motivations. And, uh, but there are other ways. Um, thinking that you're going to make a lot of money makes you incredibly neuroplastic. I found this out in college. It was a cheat code for studying for my physics tests. God, I hated physics. But I would just tell myself, I'm going to be a world famous physicist and NASA is going to pay me a bunch of money. And I would just like soak it all up and do really great on the test the next day, right? So find what works for you, but you're going to have to hustle because you got to know the old world and you got to know the new world. You're going to have to navigate both of them for a little while.
>> I love to add a bunch of stuff. I think um tough is what Steve talked about, right?
I think if you just think about the world of the last 20 years, one of the benefits that we have by being older farts is that we have seen some of the previous really nasty crisis.
And for me, one of my nasty crisis was that I was in the com boom in the Bay Area. I was in a company called Intim.
Inktomy was up there with names like Amazon, eBay, AOL, etc., etc. Now, one day I went to a party from my undergrad university and there was this president gold medalist, a guy very famous guy called Dr. Rajiv Motwani who's written a book on randomized algorithms and he found out that I was from Inktomy and he was the adviser for these two people you might have heard of Larry Pageige and Sergey Brin and he just creeped me because that's what you know he was doing. He's like, "You're from INC to me. You got this search engine.
It's called Hotbot. Do you know anything about math? You don't know and this and that." I was in another division traffic server. So I came back and told my bosses that look there's this new company called Google. What do you think? And you know Berkeley, Stanford, some rivalry. They basically said ah we're going to kill them. Two years later I was out of a job and I'm on an immigrant visa.
The lesson that I learned then is that when the world changes this much, nobody in your vicinity actually has a clue about what is happening. So when you are cooked, your friends are cooked with you, right? And so this is where like you know really good decisions happen because of trauma. So if you look at my career 2007 I become a principal engineer. I'm in the EBS team. I quit the moment I become a citizen to do my first startup. Go through this be like the worst crisis that happens right again. I'm in meta. I do another startup. The point is this thing that Steve talks about about decision making in ambiguity.
It is a muscle.
And if you've just worked in a big company, that's not a muscle that you get to develop.
If you have never had decisions where every decision is something that could tank your company, you've not built that muscle. And so, one of the challenges that I have when I tell people that everything is different is that they're like, I'm going to wait till it stabilizes.
And this is across the board. You might be early on in your career, you might be later on in your career.
There is no stabilization in our foreseeable future. So the empirical aspect of just you have to go and try things out is the best advice I can give you. And all your friends who are telling you that it's just business as usual may also be cooked.
>> So both of >> Yeah. Go ahead. tell one really quick story because Ajet reminded me when he said we were really old and we've seen a lot of stuff like everyone complains that AI causes huge devastating, you know, wreckage, you know, outages and stuff, but I was reminded uh I was at an Amazon old-timers party a couple days ago and we were talking about how once they were messing with the uh uh the customer database and they needed to update some customer's address in Japan.
Okay, so they said, you know, update, you know, customers where, you know, they said update customers set address equals blah blah blah blah blah. And then they all looked it over and they like checked it over and then they said, "Yeah, it looks good." And they sent it and they had forgot the wear clause.
Okay.
I learned what that meant that day. And so every single person in Japan suddenly had the same address and all shipments going to that dude's house right from that moment on. And we were like, you know, it was just like, you know, big mess, right? So, I mean, you know, AI can't outdo us. We we did way worse.
So, both of you guys talked about how the next five years are going to be really challenging. How there is no stabilization, you know, it is chaos, it is different, you know, what practical advice do you have for people here or for their friends?
What can they do? How do they navigate the the turbulence that we are in and that's ahead of us?
>> So I think for the first thing which I'll talk about is everybody thinks there are only four people who have kind of got it made. People who work in Nvidia, people who work in anthropic, people who work in OpenAI and maybe one other company, right? Everybody else has got no agency. I think that's the first thing that I just like despise.
And you know, you got agency, you can go and there are so many SAS companies that haven't done for 20 years. you can figure out an angle based on your domain knowledge and you can attack them.
I think the hardest part which I'm seeing with a lot of people is that sitting for two or three months with a group of people and working really hard on an angle is something that requires a level of kind of calm and uh agency which in this level of whiplash. Oh, what is going on in hormones? What is going on with the latest model? It's just hard to focus and do that. My best advice is like do what Ryan and I did.
Lock ourselves up in his mother's apartment for a month and don't emerge till you have a working idea and then pitch people like you and G to join us.
I don't have anything better than that.
Like you literally have to go into a hole and think about what do you have that is worth bringing to the world and um work your ass off.
>> Steve, anything to add?
>> Uh I would add that you uh you know you want to team up with other people. Uh you can you can accomplish more together I think than by yourself. It's really tempting to lock yourself in your basement with 18 agents and try to take over the world. I know a lot of people doing that actually and I don't think any of them are going to take over the world. Um so yeah, team up with others.
Um um looks maxing.
I kind of I kind of joke. I hear AI researchers are all doing that because they have no other way to differentiate.
Um of course they're also the same researchers that are starting smoking because lung cancer is going to be cured in a few years anyway. So I don't know if I'd follow what everything they do.
Um let's see. Maybe I'm not the best person for advice, but >> well, there's so much going on as you've alluded to many times, right? There's so much hype. There's so much also real stuff under the hype. So, how do you two stay up to date on all of the breakthroughs like I was I was at a conference last week and the one of the leaders of Anthropic said even she didn't know what Anthropic was doing.
You know, how can you keep up with all the different players and all the different models and breakthroughs? What do you pay attention to?
>> There's this hack that I just learned from Steve on December 17th and he mentioned it in this part. He talks about almost an option portfolio of problems that he has in his back pocket and he keeps throwing the problems at the AI agents and you know he would see them getting better and better and better.
What I took away, I don't know what Steve wanted me to take away from this, was that until you try things yourself, there is just so much hype on X and LinkedIn or wherever you're getting your news from, it's almost impossible to understand what's going on. So, I'll give you a very p practical advice. The CTO of um HuggingFace recently said, "Hey, these open-source models are becoming really good." So, we got VC money. The first thing that we got in our company is M5s and I gave it a shot and they are like in my experience something that I'm running in the batch mode. But the point here is that people have a lot of opinions about where things are going and this surface of capability is very strange with this thing. And what I mean by strange is that at some point it acts like a fain man level genius and then the next question it acts like the biggest that you have ever seen. And you have to figure out that contours on a dayby-day basis and you just have to build an intuition around it. I don't have a better um answer on that.
The other thing which I'm also very much realizing is that you have to get more confidence in your taste in your gut feelings more than you were in the past and I have just become as I said I would never have been coming up and talking to you people but once you start having when in the past whenever you start something um people don't understand how messy the beginning of an idea is. So as I said Zetub got acquired by hugging face in 2020 when I was talking about this idea everybody was saying that storage is cheap what is your problem right and the answer is that for AI systems and things like that these kind of mechanics make a difference and now a lot of these blob stores are coming up with mechanisms similar to the ideas that we came up with but the problem that I'm going to say is that in the beginning you have to trust your gut instincts and that is something a lot of people have a problem with once something is proven they can follow rules um you know everybody talks about serviceoriented architecture but if you actually go back in time to 2003 when you know Steve Ryan myself and all of us were in the world when this was going on it felt very messy and I think the the current world where you have to understand where the hype is and where it's not you really have to build your own smell And that comes with playing around with things.
>> Question.
>> Oh, well, the question was about how do you I'm sorry.
>> Well, how do you how do you follow what's happening in in the industry and how do you distinguish between what's hype and what's real and how do you stay on top of developments?
>> Yeah. So, uh, my first instinct is to just completely ignore everything until I've heard of it about 50 times, right?
Uh, which has worked out pretty well so far.
Uh, I think I'm supposed to look at Obsidian next. I think I've heard of that 48 times.
Um, uh, look, I think we've reached a point where you're there's going to be so much software out there, right? so overwhelmingly much software that you're not going to be able to pay attention to it all and you're going to need search engines and aggregators and you're going to have to have a personal agent at some point probably that knows what you like and could go out and find stuff for you, right? I think we're all going to need a chief of staff at some point. I think some people may have some success with this with OpenClaw right up until their bank accounts get drained, but I don't know, steps in the right direction, I guess.
Um, look, uh, you know, the most important thing is that you're you're you're trying to build stuff yourself and, um, you you've got a feel for how the models are responding today, you know, because that's that that changes over time and and you, uh, uh, you got to develop your own style and your own, you know, your own taste, your own vision. But I mean, like, yeah, we're we're we're entering a world where, like, uh, like you say, you can build anything. Software is free, right? uh you're going to get differentiated on taste. And so we're we're entering a world where everything is a show.
Everything is a spectacle. It's an attention economy. Your software doesn't just have to be good. And it doesn't just have to be sexy to agents. It also has to be packaged really well. So, you know, like it or not, you if you want other people to use your stuff, you're all marketers now, right?
But I wouldn't, you know, I wouldn't sweat it too much, right? Most of the stuff most of the stuff right like I mean like spectriven development be mad I don't know there were a bunch of technologies two months ago that everyone was talking about that nobody talks about today right so it's changing really fast so that's that's why you wait until you've heard about it 50 times right wait for a few months I think beads is finally ready does everybody here use beads no not using beads okay I would I would actually go so far now that it's been I don't know eight months beads is pretty good Right. That's a that's that's a long time in AI land. It's got I think it's close to 25,000 stars on GitHub, too, which is nice. Uh yeah, Beads is just like a way for you to keep track of stuff. It replaces the built-in task tracker and memory system, and it's just way better. And then once you get into beads and you realize you can fire hose work items, like as many as you want, anything you ever wanted to do, just make a bead for it, right? And then that's that's when it starts to get addictive. And like then I think you break into that new space where you're building your own orchestrators and yeah >> actually want to add something more to that in a very strange way the desire for human connection has become much more in this world and for example you talk about kind of you know learning about software every meeting with Steve has been a step function in our thinking at Sage jocks. And what I'm getting at is that I think the desire that humans have for real connection is going to be exponentially increased because of the craziness that is happening. And I'll give you a example. You talk about all this explosion in software. You know, one of the things which happened in the last month was that there is this new protocol called media or quick. And I found about it.
You were warned.
>> That's what happens.
Don't use media war if you want to sit next to Steve. But the point is I found out about this by hanging out with my friend from Meta. I don't think it's an accident that we are looking for each other here because of the whiplash that all of us are feeling. So I think in a very I'm not a very social person. I do not want to be in this spotlight. I want to be in front of like a dark space in front of my screen as you know very well.
>> I do.
>> But in a very weird sense in the ad world I think these kind of groups and actual face-to-face interaction figuring out what is working for you, no no tweet mess, I think is going to become even more important for us. And I love this space for that.
I'll ask you guys one last question and then we'll open it to the group. But um how do you think roles are changing uh for people who are building products together? You know obviously coding agents really were you originally and primarily used by developers but we are increasingly seeing uh designers, product managers, other knowledge workers really relying on agents. So, how do you think if you were to fast forward like a couple of years, how do you think work and working together looks like with agents?
>> So, first of all, we are in the middle of figuring this out. Um, but one of the things which I kind of will be very blunt is that worldclass thinking has always been scarce.
Right? There's been a lot of people who kind of thought that they were worldclass thinkers and I think that the people who are able to start thinking from first principles work backwards they didn't ascribe a label to themselves like I think that if you're starting to say that you're a product manager or a designer you are not going to be so successful in this world because you should be able to figure out business design product and tech at the same time. And I think one of the hardest things that people are going to be able to have to do is that that self-label that they have. I am an infrastructure engineer. I am a dev rail engineer. I'm a front-end engineer. I'm a de I'm a GPU systems optimizer. I think those one of the other things which I learned from Steve was that we have to go through a cycle of grief.
and just obliterate that conception of ourselves and just say that okay what is it that I can contribute based on the tools that we have and I think the people who are going to be able to make that transformation are and also are able to work with different people I think neurodiversity is very important I I may have mentioned it I really believe in team sports being the thing you have to collaborate with people who are different from you who have their LLM trained with a different worldviews And if you're able to play well with others and have this egolessness and fearlessness, you may have a chance.
>> Thank you. Well, we have about 20 or 30 minutes for a question. So, Taylor, any particular Oh, go ahead. Go ahead, Steve.
>> Well, I could answer the roles question too if you want. Um, I just had one really cool thing to add, which was that there is a venture capitalist who uh famously tweeted a month ago or so. You probably saw that there are only four roles in the new knowledge work world.
Uh you know what these roles are, right?
Yeah. First one is anyone know what they are? The first one is builder, right?
Your it's your PM, engineer, builder, agent, wielder, get stuff done, right?
Number two is your s sur maintainer type, right? Number three, hot people. There will always be a role for hot people. And number four is grown-ups.
Those are the four roles, right? To which I reply, "Thank God I'm hot."
Right.
>> Thank you for adding that.
>> Yeah. Welcome.
>> All right. Questions?
>> Oh, okay. Excellent. Thank you so much.
>> You just want to maybe just say it out loud.
>> Okay.
for learning new things.
All this comes in a context, a business context or selling. What do you want to tell your chief counsel or not?
Right. So >> yeah. So the question is broadly yeah how do you how do you do how do you roll this out responsibly? How can you actually like tell your executives that you are uh are rolling it out in a way that is not going to bring things down and that can can continue to uh meet your obligations to your to your shareholders and your customers and your users and and all that. Yeah. Um this is you know this is the fundamental question and um you you got to look at like situations like the GitHub being on fire problem and think may maybe they didn't pay enough attention to those questions right um I I you know I truly think that there's this weird uh problem where there's some people who are too into it some companies are just like you know what I mean over adopting overconfident under careful. And then there are other companies who are just way way way overcautious and way over careful and just dragging their feet too hard, right? And striking that balance is just, you know, incredibly difficult, right? Because uh you you know, it's it's just it's just like teaching an individual. You get them really excited about burning tokens and eventually they're good at it, right? And then they turn into pigs, you know, and all of a sudden the CTO has to turn it off for everyone. And it's like whoa not like that right you know and so companies like have to sort of find their way forward here and the the advice that I give them is I mean like I I go in like let me give you an example concrete example right a concrete problem if you're if you're producing code 10 times faster or say say you're producing 10 times as much code and your defect rate is more or less constant which it probably actually isn't if you're Vive coding but let's just assume that it is you're still producing 10 times as many defects out in production, right? Do you have a plan in place for that? You know, how are you going to deal with that?
Right? These are the kinds of questions that, you know, you have to uh you have to introduce into the fabric that you're rolling out when you're when you're doing your AI experiments at your company. And I see companies are like really leaning in. I mean, like, you know, I saw one company um that rhymes with Feda Fog. I can't say who they were, but uh they're really big and they do data stuff and uh yeah, they um uh gosh, you know, I've completely forgotten what I was going to say because I had to obuscate their name.
Anyway, uh you get the idea. They uh they they were rolling things out very slowly in their SDLC at the edges and oh yeah, they spun up a bunch of 100 person teams. That's what it was, right? And so like they've got the innovation side of it which is which is great to see and the culture side of it, right? But like so so when I see companies like that, the first thing I do is tell them, whoa, time out, right? Put on the brakes. Are you guys really making sure that you're not like, you know what I mean? Like, you know, cramming stuff out into into production that you shouldn't. And another another problem that hits every company, like some some folks at a bank in Australia called me up and told me about this. They said that we made our engineers go five times faster. Okay.
Uh which is great, but the business couldn't keep up because the engineers were rolling stuff out too fast for the business. And so they went to the business and they said, "We're going to give you all AI so that you can be faster, too, right? So that you can keep up with the engineers. How does that sound?" And they said, "Whoa, whoa, whoa, wait a minute. Wait a minute. You want us to do five days of work in one day? you. Right. They were like not having it and you can't fire your business. There's going to be like 800 of them and they're like a labor union at that point, right? You can't, right?
And so like what do you do? I mean like rolling it out to the company is that's actually a good thing in some sense because it slows the brakes a little bit on the AI adoption to your point, right?
But I mean like striking that balance is it's the problem of our generation, right? Of our time of this year.
If yeah I would love to add some more color to that. Galax in the back is uh going to be doing a talk on AI tinkerers about what we are doing with BDD and I think our testing and the amount of tokens we bear uh burn on making sense of the surface of what we are trying to do has to be we're thinking about things in APIs and files and code while we should be thinking about the behavior that we want to maintain and I think there's a lot of work to be done. Uh the other person that I want to highlight here is um one of my colleagues. I can't see him but Rupac is around here and Rupac is a researcher from Maxplank Institute and he works on provable systems and one of the things that he's challenging us is that with in fact he's right behind you actually and u you know with these new systems you can start writing provable code from the beginning which used to be more painful. I just want to be very clear about one thing though. Um, you know, these are the early days, right? So, in 2006 or 2007, am >> I going to die?
>> In 2006 or 2007, if you went back to EC2, the only people who are using EC2 and S3 were startups like SmugMug and things like that. 2009, Netflix comes in using this for benchmarking and stuff like that. Capital One comes in in 2011.
Nobody remembers this part. Nobody's going to use this new stuff right away, right? And so yes, use your judgment. If you're Capital One, please don't go crazy, right? But if you're a startup like us, the burst in kind of getting something out is worth the risk. And you sh you have to make that calculations based on your own judgment. And that again is the human taste that is never going to go away in my opinion.
>> I think there was your hand was up.
Yeah.
So I'm finding a multish but I haven't respons.
>> Uh yeah, you can put some Claude has some hooks now. Pre pre-tool use hook, that's your friend. Pre-tool use hook is like seriously it'll save your ass. And then you can just like put a whole big long block list of things you don't want it to do.
So when they uh you asked about a question when we did the registration and my question at first was more about u trust and trust being the achilles heel of AI right the less trust there is is what you just mentioned a minute ago as we went through the conversation I think I move from trust to I think it's more accountability I'm not an attorney by any means but it comes down to do that.
Accountability is at the end of the day.
That's again you're using a third party to build something that maybe you built, maybe somebody else built, maybe the engine built it itself, but we put ourselves in a contract, whatever the case might have been, and now things didn't go as we thought they would go.
Who's accountable? And my question to you is at the end of the day to address accountability because there really not a lot of models, right? As a human being, what are we used to holding accountable but people? So you put people places maybe not because you need them in that particular spot because somebody has to be accountable. So may you may be able to do without all of your design and maybe all of your cyber security because you put it on a fantastic skill that you built in the night that deals with cyber security but at the end of the day somebody has to be accountable and you may end up grabbing a person to put him in that role to be accountable for the skill that you built on cyber sec that is running the application. What would you guys do? So, >> you can always hire your agent, you know.
>> No, a human head has to roll. It's a good point, right? I mean, like, you actually do need a human at the top of every accountability chain. Uh, the CESO role, chief information security officer, is kind of designed this way.
They're sort of designed to get fired every couple years, right? Um, you you know this, right? This is, you know, by design. Um, so, uh, I don't know. What advice would you give a Jeep?
You know, I went through the pain of getting sock too for my previous company and it is crap. And so, one of the problems that I do want to highlight is that when people talk about the past, it's not as if the past was great. It is a bunch of form filling together for I think what you need to have is something where you can convince the world that you care about the at the level that you're advert advertising. So if you're S3, you're advertising at a certain level of security. If you're a bank, you're advertising at a certain level of security. And a sock 2 certification doesn't change that. It's the brand that you're going back after the accountability. And I think we have to come up with more nuanced uh approaches to you know telling people at the level of risk that they want to be willing to take with a alliance rather than something that lawyers can do because lawyers have only very blunt tools right now. I think we have to evolve and I think society has to evolve to get to this part where I can go in and create a product with a higher risk profile which is only attractive to a certain person and then when I'm willing to do the work to go through a level of security test then I can bring in other different kinds of customers. So agreed on your accountability but I think the communication has to be done at various more granular levels than the tools that are there right now. Going back to the other comment that I made, I do want to talk about you talked about code. I think we need also a version of GitHub where we kind of store more than the code the model that was generated the intent of what you were doing because these models are going to change all the time and you never know whether your code which is generated by model one was riddled with holes and not and you might want to go back in time. So I think the world of forensics and security and even like alerting it has to be uh completely rethought from scratch. I don't think it's an easy answer at all.
>> More questions here.
>> Yeah. Um, so I love using these tools to get a lot to do is still like a lot of agents like was faster.
The second is that now the most important So if I understand your question correctly, it's like hey a a I'm becoming coding slightly faster but like what I have to do is not becoming clarified anymore by any of these tools much and secondly I'm doing some things which are kind of throwaway and which might be distractions. So I think for the first one I think of a company which is searching for product market fit almost like a higher level agent. If you think about our room we have got to observe orient decide act across humans and agents. And if you think about it at the higher level of abstraction then what you have to do is to go outside go talk to people see what they might be willing to react to see if you can implement it and then see whether the customers reacted the way you wanted to do. But I think that's this that is the core of what we are talking about in Sage which is that the input and the output become the bottleneck when the processing becomes so good right and you have to spend a lot more time collecting uh maybe user stories or insights and also doing much more prototypes. The second part I have an interesting thing to kind of throw at you. You know in uh in computer chips there is this thing about predictive branching where you sometimes execute a bunch of your pipeline even though you don't know whether the branch is going to go yes or no right and I think that with this new world where experimentation has become so low why shouldn't you try out a bunch of things you know I think we are going to get to a point where in the past you some optimization which was maybe 5% and not worth your effort we are going to start attacking ing those problems. And this is one of those things where what I'll offer you is that consistently maybe ask yourself a question in this ad world is that are there experiences where if I'm willing to spend 10 times more tokens than what I'm doing right now, I can elevate that experience and you might be surprised by some of the things that might be throw away actually clicking with a segment of your customer audience and things like that. So I do want to kind of leave that space for serendipity for the ideas that you might not think has got legs in the beginning.
>> The only thing I would add is that you sound like you have yourself a project that I would consider I don't know I don't have a good word for it but a it's a like a wedge project. It's like one that like weeds out models right. Some models uh don't do well with it. Maybe all the models don't do well with it.
Right? I have a set of these like I have, for example, a React client that I'm trying to build for my GL game and it's still too hard for Opus 47, right?
Uh, you know, the the the the effectiveness of models is not by any means uniformly distributed. And so you're going to find certain problem spaces, they just suck, right? They straight up suck. And so what you do is you you just keep keep an inventory of them, keep them in your back pocket, right? And every time there's a new model drop, bring it out and be like, "Hey, let me show you something, right?
You know, can you do this?" you know, and usually it's no, but one day there'll be one that can do it, right?
Maybe it'll be Mythos, you know, or whatever. Uh, it's really important to have these these projects like these projects that are too hard for today's models like ready at hand where you are faster by hand than they would be, right? And they exist that the ceiling for for them is going up with the model drops, but there's it's still a ceiling, right? And the reason it's important to have this is that it's really hard for most of us to see how fast this is moving. You're only really looking back about six months and only looking forward about three months. But if you have projects that the models can't do and a year and a half goes by and all of a sudden they can do your project, you can feel it in your gut that this thing just got smarter. You see what I'm saying? So hang on to that project. You know, value it because it's going to be a good anchor. That's the word I was looking for. They're anchors to help you understand the exponential curve of the motion right over a period of years.
>> Go ahead, >> John.
>> Thanks.
>> Has your mind changed about who should be aiming for level eight? Do you think the highest leverage thing a software developer can do today is to try to spend all their time trying to get to level eight?
>> No, I think there's a there's a level nine. We know what level 9 and level 10 look like now. And you can actually kind of skip short circuit there. Yeah. Um level 9 is so level up through level eight is basically like how you use them for development like building stuff, right? But the next frontier immediately is is deploying 247 autonomous agents on your behalf to handle for you, right? That's level nine. soon as you've deployed your first agent pack somewhere that's like never sleeps and it's pulling a queue looking for some data thing and whatever doing some some ETL for you you've hit level nine right gas a toolkit for that right so you can uh you can build a bunch of those and I'm actually just in the process today of starting to wire up my own you know what I mean business processes with this thing I'm really excited uh but uh yeah you don't have to build your own orchestrator there should probably be someone on your team who does, you know what I mean? Um, but yeah, um, I don't I don't think everyone has to build one. I mean, if you get, as long as you learn how to use them.
Now I haven't always.
>> All right. So, for starters, let's revisit why Claude's program is giving you a 95% discount, right? It's a it's 20x usage. It's a 95% discount over the API token cost. And the reason is that is incredibly valuable training data for their next models. And it's why Anthropic is ahead of all the others.
And it's why SpaceX bought cursor just for their data so that they can train their model, right? So, uh, I forgot his question.
>> Well, what do we do when the music stops and there's no more subsidies at this level?
>> So, look, this trick is going to continue because coding is reasoning and reasoning is problem solving. And so, they're making general problem-solving models using by teaching them to be better and better at coding, right? So that data is going to continue to be valuable from humans producing it until they really can do everything at which point we're kind of all screwed, right?
So uh you know uh for the foreseeable future I don't see it shutting off.
Yeah. Now that said, you know, maybe maybe they change their minds and say, "Okay, we got enough or we only want to get data from a few people now, right?
And the general public, we don't care anymore." The nice thing is I'm hearing really good things about 55. I don't know if anyone's actually used it for coding, but I'm hearing pretty good things, right? Which which makes me happy because we don't want to be dependent on one company for the rest of our lives, right? Also, the open source models lag by about seven months. And so by summer, we're going to have stuff that's as powerful as Gemini 3 running on, you know, your local local GPUs, right? So, uh, there's hope on the horizon because the models only have to be as good as Opus is today for the whole game to have completely changed forever, right? It's an engineering problem at this point, right? See what I'm saying? You don't look convinced.
Yeah, but I mean you can get a lot done with peon models, right?
I mean like like it's like the head of the Mythos team said, Mythos is going to seem stupid next year, right? He he tweeted that last week. So I mean this is just a relentless the open source ones are going to be just fine. I mean I suspect you know there's another phenomenon happening right which is the distillation is allowing these giant algorithms to run on much smaller models. So like I I'm just I I'm tempted to just be like call me when like I can have like sonnet 3.5 running on my Mac mini. You know what I mean? Or that class >> 27.
>> Okay.
soon apparently.
>> All right. Well, I think there are more questions. Um Steve and Ajit will be around afterwards as well. So, grab them and and ask your question directly. And >> thank you, Milana.
>> Wait, we have one more. We have a We're going to do our raffle giveaway for Steve's book. So, uh Audrey's going to read out our fun answers.
>> All right. I'm just going to say your first name so I don't give away if you said something silly. Um the question was what's one part of building software you think will not change with the adoption of AI? Uh Paul said the fact that building software is a people problem and not a technology problem.
Winner uh by the way just pick it up over there after uh Anuj understanding and writing down clearly the problem to be solved. JP figuring out useful problems to solve will still be hard.
Jean, the need for humans to understand other humans intent. Thomas, the need for human in the loop, advocating for the best experience for the end user given the nuanced context and empathy of a given problem space. Mark, figuring out what the leadership team really wants.
Himant, uh, only the requirements. I've tried automating everything else and it works mostly fine.
Nice. Uh, Lindsay, the lack of understanding between dysfunctional PMs and dev teams.
Nicholas, the change is the only constant code answer. Rhett, screaming expletives at the computers.
Those are our winners. Thank you guys very much. I just wanted to thank again uh all of our panelists. We thank you guys so much. Um, stay around, network uh and enjoy the rest of the evening.
Thank you guys so much.
Related Videos
Agentforce NOW AMA: Build with React and Salesforce Multi-Framework
SalesforceDevs
490 views•2026-05-28
How agent o11y differs from traditional o11y — Phil Hetzel, Braintrust
aiDotEngineer
450 views•2026-05-28
Re: 🗣️📍theprophedu📍2026 GST 103 CLASS (E-EXAM REVISION)
theprophedu
636 views•2026-06-04
WEB TECHNOLOGIES UNIT-2 | Degree 4th sem BCOM Computers web technologies unit-2 full explanation💯✅
LearnwithSahera
1K views•2026-05-29
More tests are always better? How to use AI to identify tests that bring little value
Alliance4Qualification
335 views•2026-05-29
Search Algorithms Explained in 60 Seconds! 🤖💨
samarthtuliofficial
218 views•2026-06-01
People of Game of Thrones using JavaScript DOM
AltCampus
296 views•2026-05-30
Instagram accounts got PWNed
EricParker
13K views•2026-06-03
