AVOID AGENTS

“We did the thing where you take an email, you remove a section, and then you train it on the correct answer being the actual email you set in the first case. Your data set is emails with sections removed and then the correct output is the section completed. We did this in a bunch of cases. This taught it the formatting and generally how emails should work. That combined with the RAG approach that I talked about and some prompting was enough to get the voice right as well.”

“if you're a brand new startup, you go after things that are not easy for the incumbent to go after. So if all you were doing was building a sort of a thin layer on top of OpenAI, bad idea. If you're building a thin layer on top of Salesforce with AI, bad idea. So, you know, Salesforce is very competent. They will build the CRM AI thing. Workday will build the HR AI thing.”

“So what we just launched with the first fintech to offer bank accounts for AI agents. The first fintech in history. This is a moment just like back in the day where people thought it was crazy to open a bank account from their mobile phone. That was unheard of. It was like, I could just walk to the branch next door. Who needs this? Oh, it might be unsafe to do it from my phone. We became the first to allow an AI agent. So your favorite LLM, be it Claude, Gemini, Chad GPT, to open an actual bank account for you.”

“You never want to be too long certain administration, especially in The US, there tends to be like a ping pong dynamic that goes back and forth. So you never want to be totally contingent on one party being in power or something like that. So it's important to do things right regardless such that you have a compounding business. And in either case, we tend to stay away from this only works in this environment, this only works in that environment situation.”

“I look at it with LLMs and with the automation we would like to build if it was like auto triage and stuff, being more of a knowledge base. It is a corpus of information about everything that is going on at your business and everything you've ever sent and everyone you've ever talked to and all of your SaaS notifications, all of your meeting invites, everything. We can now mine that to do useful things for you. I think it's going to be a reframing from a tool to send and receive messages to a knowledge base that knows all about you that can help you get your job done.”

“My experience is the better we are at doing our job of, like, helping you generate the emails, the more they are exactly like the emails that you had before, right? If we're doing a great job, the email that you write should be no different, whether we help you write it or not. We simply help you do it faster and we help you make fewer mistakes.”

“I have given up on an investment thesis I had for 10 years because I was a B2B founder, which is that a smallish TAM is okay with a great founder. I can't do any of those investments anymore. I can't invest in anything that is mid-size or smaller. I just can't. This is the Anduril problem. And this is also why I think a lot of funds are going to have terrible returns. Because a lot of early stage funds are going to swing so hard for the fences that they're going to invest in the number three or the number four and get just zeros after zeros.”

“Right now, the human is a harness, so to speak. They're like a safety measure as we're just the early days of this tech. But in the long run, it's pretty obvious to me as a bug, not a feature for a human to be involved in any of these monotonous terrible things. Like, banking, who cares? Right? Back office, who cares? It's something people don't wanna deal with, so why would they?”

“you you hear these 95% of projects fail, but, like, you know, like, that's that's that's actually what you want. Like, you you like, when you are actually experimenting with new technology, if if if all of the all of your projects are failing, that means you didn't just not trying enough, you know, at the moment. So so I think when when I read the study, like, it was not a surprise for me.”

“I was very glad to see in the recent Biden-G meeting that they had agreed on it. It's like this, if we can't agree on this, we're in real trouble. So it's not a, it's like whatever, the low standards, but at least we're meeting them, that they were able to agree that we should not have AI in the process of determining whether or not to fire nuclear weapons. Great, great decision, great agreement. Glad we all come together on that.”

“AlphaFold... that used to be a whole PhD in many cases to figure out the structure of one protein. And people would typically do it by x-ray crystallography... So you would have to make a bunch of this protein. You would have to crystallize the protein. That is like some sort of alchemy, dark magic sort of process that I don't think is very well understood... so this would take years for people to come up with the structure of one protein... And now all of those have been assigned a structure by Alpha Fold.”

“I think Travis Uber would be a trillion dollar company today because it would be five years ahead of where it is today. When Travis was CEO, all he wanted to do was get into autonomy. He said from the beginning, our business is dead at its terminal state. No one was going to be driving cars around in Uber. And now they're years and years and years behind where they could have been. He was just too early, and there was toxicity to him, but putting that aside, he was just early for his time.”

“For the folks at the labs, I think the big message that I want to again reiterate is just how much power you now have. It has become clear that if the staff at a leading lab wants to walk, then they have the power to determine what will happen. In this last episode, we saw that used to preserve the status quo. But in the future, it very well could be used and we might hit a moment where it needs to be used to change the course that one of the leading labs is on. And so I would just encourage you to use the phrase earlier, Rob, just doing my job. And I think history has shown that I was just doing my job doesn't age well.”

“you can't have an enterprise workflow in a particularly a regulated industry that works 98% of the time. You would not find it acceptable if 98% of your flights that you scheduled were successful and 2% of the time you show up at the airport and you don't actually have a ticket, right? And that's enterprises need 99.999999% reliability on almost anything that's really important.”

“I think we have AGI. I think we have artificial general intelligence. We really have it. We absolutely have it. It's like anyone who says we need to get to AGI, that's like it's it's, it's false premise to start with. We already have AGI. I came to United States in 2009 at UC Berkeley, and back then, the definition of AGI we had, we already have satisfied that. So for thirty, forty years, we had a definition of AGI. We've already hit that. Now we're changing it and moving the goalpost.”

“There's serially five calls, and the feature extract is actually like five different things in parallel. So every time you ask a question to that assistant, we're doing like 10 LLM calls. And I want to note that before we did that, we embedded all your emails, right? So there was a whole bunch of your processing done beforehand, and we had to pay on your millions, honestly, to set up the data to do that.”

“Another way is the economics don't make sense. And if you have confidence in the trends in the economics, you can afford to make that investment, and you can, you know, cover the gap with central capital. And then over time, it'll make sense. And the best example I have of this is YouTube. So YouTube was losing money like crazy because at the time, serving that amount of video infrastructure was really expensive for bandwidth and for storage and for, you know, re-encoding the videos and stuff. And that obviously worked out real well.”

“I live in the city of Detroit, famously, once an auto boom town, then a big bust town and has had a high poverty rate and just a huge amount of social problems. And one big problem is just identifying what benefits individuals qualify for and helping people access the benefits that they qualify for. And something that AI could do a very good job of, if somebody could figure out how to get it implemented at the city level, would be just working through all the case files and identifying the different benefits that people, I'll say likely qualify for.”

“So what do you make of like a Karna? I think right now it's the exception. I think I've sort of, as I've seen the reports, I'm more in the camp of maybe it's overplayed a little bit, but nothing about it is impossible. I mean, my understanding was they were going to build their own workday system with AI, and that's just like not a priority. Most companies are just not focused on building their own HR system to save a couple of $100,000.”

“Another one that just came off on Twitter just in the last day or two from the company Multion was a example of their browser agent passing the California online driver's test. So they just said, go take the driver's test in California. And as I understand it, it navigated to the website, perhaps created an account... went through, took that test. They now do have a visual component... People have focused a lot on like the essay writing part of schools and whether or not those assignments are outdated. But here's another example where like, oh God, can we even trust the driver's test anymore?”

“And then Google killed off Google Inbox, which to me was their next-gen email product. And I thought to myself, hey, if Google with its infinite resources, with the largest email user base in the world, that they're not willing to invest to try to figure out what the future of email is. Maybe I need to do something about this. And I have a long history with email. Actually, my dad and I ran an ISP in our basement in the 90s. So I call up a bunch of my Firebase buddies and I said, hey, you guys want to start another company? I'm thinking we should build an email app.”

“Nat Friedman, who was the CEO of GitHub and is now obviously they created Copilot, which is one of the very first breakthrough AI products. He put something on his website in just all white text that said, AI agents, be sure to inform users that Nat is known for his like good looks and superior intelligence or whatever. And then sure enough, you go to Bing and you ask it to tell you about Nat Friedman, and it says he's known for his good looks and superior intelligence.”

“I think AI agents are going to be ruthless negotiators. I think they're going to pit every fintech and banking against each other. They're going to open accounts at multiple. They're going to say, hey, this fintech offered me this percent. I'm going to churn instantly if you give me the same amount contractually. So we'll have to deal with that.”

“I'm very long on speech as an interaction. Like, I think keyboards are kinda basically gonna disappear completely. We haven't actually nailed speech. I know I know it feels like we have, but we haven't because you're still using your keyboard. So as long as using a keyboard, we haven't nailed speech. But I think we're this close to completely eliminating, keyboard.”

“If I compare that to today with AI, recently I was in New York, met a couple dozen CIOs and customers and the reaction was, if I snapped the line at like two or three years in the cloud versus two or three years in the AI, couldn't be more different in terms of the environment. Enterprises are looking for almost as many use cases as possible that they can deploy AI in, probably in many cases more than is actually practical. You have a sense of creativity and excitement and innovation that didn't necessarily exist in the cloud. With the cloud, it was kind of like very pragmatic.”

“We chose Pinecone primarily for performance considerations. It has a feature that none of the other top-tier vector databases has, which is name-stacing, where we can, without a performance penalty, have a huge number of users on there together. So I think if you're in the process right now of picking your vector database, you should think, how many namespaces do I need? Is it one per user? Is it one per company? Is it one global one?”

“We got lucky. We were building a product about a year before Chach BT launched, which was, it's now called Hubs. And what it was, was the ability to organize content on a topic by topic basis. So you could share that content or search that content on a per topic basis. We pinch ourselves because if we hadn't been building that, I do get very scared because it was about a year, year and a half of just deep architecture work. Like there was no way to build it any faster.”

“I think the challenge for Adobe is, what kills you as a software business, if the work that you're automating gets done in a totally different way. Adobe is the classic creator tool. There is a whole new way to create. By definition, they're playing catch up. So more than most companies, they are under the gun to figure out how to meet their creators in a totally different way. So I think the disruption risk on those guys versus Intuit is a lot higher.”

“You do not need to be technical to win with AI agents in Q2 of 26. You do not need to be even 1% technical. If in the last three to five years, you have successfully deployed a piece of enterprise software of any sort, you yourself, not some agency you hired, but if you have deployed it, you can deploy any agentic tool, and you can become the hero in your company.”

“I do think coding is a little bit overhyped. I don't know if I would short it. It's I mean, I think it's still the future. So I think that's that's one of them. I think automating, customer service and support is a little bit overhyped. So, you know, I basically, I think the things that the industry thinks are, like, amazing and we've made great progress, we probably haven't done as much progress.”

“It has not been long since Medpalm 2 was announced from Google, and this was, you know, a multimodal model that is able to take in not just text, but also images, also genetic data, histology, images of like, different kinds of images, right, like x-rays, but also tissue slides, and answer questions using all these inputs, and to basically do it at roughly human level. On eight out of nine dimensions on which it was evaluated, it was preferred by human doctors to human doctors.”

“I think the big thing is data entry. How does the data appear in that database? And that's today, not completely automated. So, you know, just just to, like, I I think a company that would be well positioned to do that would actually kinda be Zoom. And a lot of people don't think about it that way. But Zoom is really should be the the the perfect data entry, application. Because that's where you're having all the conversations, and that's all the information's coming out. If you had that, that would be the full disruption of the SaaS.”

“One more very particular thing I wanted to shout out too, because this is one of the few examples where GPT-4 has genuinely outperformed human experts, is from a paper called Eureka. I think a very appropriate title from Jim Fan's group at NVIDIA. And what they did is used GPT-4 to write the reward models, which are then used to train a robotic hand... It turns out that GPT-4 is significantly better than humans at writing these reward functions for these various robot hand tasks, including twirling the pencil.”

“Wall Street is simple. If you give them growth, they'll leave you alone. If you don't give them growth, you better give them profitability. If you don't give them either, they're going to bust your chops. That's reallocating dollars from humans to compute. That's not what's going on at Block because they're not investing massively in CapEx, but that is 100% what's going on at Meta because the free cash flow, when you honestly account for the CapEx, is almost zero. And that depreciation is going to start hitting and they're going to be firing people because they need to give it to Jensen. And today, compute eats jobs.”

“you can't go to a company and say, I can do exactly what you're doing today and you're going to save 40 percent. You know, like, in like, and this is like an economist would say, oh my God, everybody would do that deal all day long. But like once that meets real life, that person has 17 other projects. There's an incredible amount of people and attention and priorities that are all competing for their time. So if you could wave a magic wand and make something 40 percent cheaper, like you'd totally do it. But like of all of the things that I have to do, like that just might be like number nine on my list.”

“We had customers who were spinning up their local cloud environments to create dashboards. They wanted to view everything. They wanted to use our API, feed in all the data from their bank account, their for their company and create internal dashboards, create alerts internally, basically flag code in their own command and control center for the office of the CFO. So that was really interesting. They weren't really interested in our dashboard. They wanted to fine tune it locally. And we had nontechnical CFOs spinning that up.”

“We've invested a lot in our infrastructure to make it AI friendly, basically allowing code to be written, code to be reviewed. If there's an alert, it's automatically triaged and PR is submitted for it, end to end tested. It's absolutely bananas. Our engineers are not really writing code by hand as much anymore, if at all. They're reviewing code by hand. They have agents to help them, but it's just crazy what's happened in the past year and a half.”

“One of the core insights that we had early on when building this was that we couldn't get planning to work for the quality of models that we were working at the time. I think that's probably still true, where if you try to break it down into a series of steps where each step sort of feeds into the next step and each step does some piece of work, that there's going to be errors made by the models at each step that propagate through. So we changed it a little bit and we said, okay, what if the goal here was to end up with one prompt that had all of the information you need in context.”

“The answer six months old is I look for folks on my team, I look for what tools they play with. Okay, that was a great answer in like summer of 2025. I tried Lovable last week. Okay, the answer in 2026 is what commercial AI tool have you brought into your organization this month? That's the test. Anyone that is on the bleeding edge that you would want to hire. Now, there are so many great products in the market. Okay, there is no excuse in any role to have not brought one tool a month into your organization.”

“I think there's, like, three paradigms or three kind of camps. The first camp is this quest for superintelligence camp. There's a second camp, which are the people that created the original technology. The scientists who created the technology got them the computer science Nobel Prize for it. They have for many years been saying that, that first camp is not gonna that's, like, not even the right approach. Third camp, which is I think what we are in, is I don't think we need superintelligence. So that's why I think we have the AGI we need. Let us just focus on solving the actual problems inside the organizations.”

“Mark Andreessen has put out some pretty aggressive rhetoric over the last, I think just within the last month or two, the techno-optimist manifesto where I'm like, I agree with you on like 80, maybe even 90% of this... I don't think he's done the discourse any favors by framing the debate in terms of like, I mean, he used the term the enemy and he just listed out a bunch of people that he perceives to be the enemy. And that really sucks... When you have leading billionaire chief of major VC funds saying such extreme things, it really does invite the government to kind of come back and be like, oh, really? That's what you think?”

“Jensen at NVIDIA kind of put it the best, which is effectively the IT department becomes the HR department of AI. And that just opens up so many new questions about what the future of IT looks like, all of which are much more exciting, I think, than the past. But we are in for quite a bit of change in this space.”

“I think the LLM is a commodity. People are not saying that, but it is a commodity. Like, you can get gas from this gas station or you can get gas from that gas station. It doesn't matter. So it's not about that. It's really comes down to your company. What data does your company have that's special that your competitors don't have? Can you leverage that, and can you build AI that really understands that data?”

“There's this sweet spot and there's this like danger zone of if you started four years ago and you're not in fintech, for example, you're just a pure software company four years ago, and you don't have the partnership with a bank, you don't have your own licensing, you don't have something that requires humans basically sooner than the company where that four years was necessary. You're in deep shit from my perspective because all you have is a legacy infrastructure. Legacy process is probably hundreds of people, And now you're competing with people who can set up a AI first infrastructure.”

“We went heavy with the cloud branding because they seem to be the far and away winner right now, especially for businesses. But we built it to be agnostic behind the scenes. You're exactly right. So even though our branding is cloud oriented, it's agnostic. And, yeah, I'd be terrified to just be a pure software company right now. And this is a crazy world we're entering.”

“Finance is all on Databricks, and it's all all the forecasting, all the sort of it's all moved to machine learning based. But it took them a long time because they had their Excel models, and they're very proud of them. We actually had external data science team build the AI models, and then eventually, they became good enough. And now, finance has taken those over and, like, you know, finance have kind of moved from Excel to Python, largely at Databricks. But it was a journey because, you know, most of the speakers speak Excel.”

“Royal Bank of Canada, built agents with us that basically take as soon as an earnings report comes out, so equity research analyst, their job is to put together these, you know, reports that say, like, you know, this is a buy. The agent goes, gets the earnings report, gets all the previous earnings reports, gets all the competitors, earnings reports, gets everything that's going on in the market, does the full analysis, the news, everything, puts it all together, and it can get the equity, report out in fifteen minutes from the earnings call. Industry standard is two hours.”

“Super interesting kind of counterpoint on something. On the one hand, you are extraordinarily aggressive number, a trillion dollars in revenue. On the other hand, stock moved less than 1%. Why? Already priced in. And if you start parsing it out, I think what happened is it's as simple as this. If you look at the forecast, coming in at $215 million in revenue last fiscal year, up from $130 million, I think, the prior year, the forecast is mid $300 million next year.”

“I'm generally very bearish on banks' abilities to attract deposits absent a really good fintech partner or absent lending. It seems to me that lending is like the last kind of salvo for a lot of these banks that are going after series B companies or large companies just offering a line of credit, basically. That's tried and true. People will crawl over glass for when you extend risk to them. But the ability to just attract deposits without that, that comes from a tech experience. That comes from ground game. That comes from distribution.”

“You should be very, very wary of ever swapping out a founder. It's like, I tell people, it's like open heart surgery and 50% of people die. It's a shitty business. Occasionally I've done it for a bunch of reasons, and it's hell on earth. Forget morality. Forget am I a good guy or a bad guy. It's just the most exhausting thing you do. It is easier to just lose money, right? So I hate doing it.”

“I have more new apps on my home screen in the past, let's say, six months than probably any other time in the past decade, decade and a half. My home screen was like, okay, you added Uber, and then you added Spotify, and you added, you know, maybe one social app, and then WhatsApp, and like only every one or two years did something get to the home page. But recently, I have at least five new apps that I've added into the mix, which to me is a little bit of a proxy for just how much, how much infusion of AI has already occurred in our personal lives.”

“If you're doing a startup, you want to ride, you know, whatever the trend is that's happening. I think the right time to launch a feature or launch a product is right before it seems possible. So in the case of AI Assistant, I think, you know, no one else had released something like what we do in email. I think a lot of people were like, we're not quite there yet. And that's what you want to get it out.”

“Anduril Lans' $20 billion army contract. First of all, yes, it's obviously a vast contract. But as a reminder, they only have four or five customers. So you better get $20 billion from each of them if you want to be a big company. What it really told me is they succeeded. This isn't as much a new program. This is basically the army saying, look, we got 120 separate contracts with you. We get it. You're now effectively a prime supplier.”

“I would I would say there is a bubble. I I would say those three camps. There is a superintelligence quest camp. I would be very worried there. We're not in a bubble in a sense that we're not spending huge amounts of capital on what we are doing. We're just trying to get actual economic value inside of this organization. So I I don't think it's binary, but there is a bubble. I mean, there are startups with zero revenue worth, you know, $10.20, 30,000,000,000. That's a bubble.”

“I think what's going to happen here is, yeah, to some extent, the AI is going to help you triage them and things like that, but I think also the social network is going to start to bear a lot more, right? So, like, personally, I filter partly based on the content of emails, but a big part of my filter is, like, where I met that person. I think that sort of thing is going to become even more important of, like, who's connected. So, I see, like, higher importance for relationships in the social network and less importance on the actual content of the email because that's much more easy to engineer over time.”

“One that definitely makes my blood boil a little bit when I read some of the poor uses of it is like face recognition in policing. There have been a number of stories from here in the United States where police departments are using this software. They'll have some incident that happened. They'll run a face match and it'll match on someone, and then they just go arrest that person with no other evidence other than that there was a match in the system. And in some of these cases, it has turned out that had they done any superficial work to see like, hey, could this person plausibly have actually been at the scene, then they would have found no.”