What Surprised Us Most In 2025

I think perhaps the thing that most surprised me is the extent to which I feel like the AI economy stabilized. We have like the model layer companies and the application layer companies and the infrastructure layer companies. It seems like everyone is going to make...

Many episodes ago, we talked about how it was... felt easier than ever to pivot and find a startup idea because if you could just survive if you just wait a few months there was likely going to be some like big announcement that would completely make a new set of ideas possible and so like finding ideas is sort of returning to sort of normal levels of difficulty. Welcome back to another episode of The Light Cone.

Today we're talking about the most surprising things that we saw this year in 2025. Diana, you found a pretty crazy one. It's sort of a changing of the guard almost in who is the preferred LLM at YC during the YC batch. Yes, in fact, we just wrapped up the winter 26th election cycle for companies. And one of the questions we ask to all the founders that apply to YC is what is your tech stack and model of choice?

And one of the shocking things is that for the longest time, OpenAI was the clear winner for all of last year, last couple of batches, though that number has been coming down. and shockingly in this batch the number one api is actually anthropic came out a bit more than open ai which who would have thought i think when we started this Podcast series back then, OpenAI was like 90 plus percent. And now, Anthropic. Who would have thought? Yeah.

And, you know, they've been hovering around like 20, 25% for most of like 2024 and early 2025. And then only even in the last three to six months. did this sort of changing of the guard actually happen? They had this hockey stick with the growth of over 52%. Why do you think that is? I think there's a couple of things in terms of the tech stack selection. I think as we've seen this year...

There's been a lot of wins in terms of vibe coding tools that are getting built out out there. And coding agents are so many categories that this ended up being a bigger. problem space that actually is creating a lot of value. And it turns out the model that performs the best at it is actually models from Anthropic. And I think that's not by accident. I think from hearing the conversation we had with Tom Brown not too long ago, he came and spoke. That was one of their internal evils.

on purpose made them their North Star. And you can see it in the model taste as a result of what's the best choice of model for a lot of founders building products is Anthropic. the vast majority of the use cases people are using it for though is not coding so i wonder if there's like a bleed through effect where people are using claude for their personal coding

And then as a result, they're more likely to choose it for their application, even if their application is not doing coding at all. Because you'd be very familiar with the personality of Claude Opus or whatever they're choosing.

Sonnet, I suppose. How about Gemini? How's Gemini doing in those rankings? Gemini's also pretty much has been climbing up pretty high, I think. Last year was probably single digit percent or even like two, three percent. And now for winter 26 is about 23 percent. And we've personally been using also a lot of Gemini 3.0 and we've been impressed with the quality of it. I think it's really, really working. I mean, they have all different personalities, don't they? Me too. Yeah.

It's kind of the classic where open AI sort of has the black cat energy. And almost like Anthropic is kind of more the hoppy-go-lucky, a bit more very helpful golden retriever. At least that's what I feel when I talk to them.

And how about Gemini? It's kind of like in between. Harj, you prefer Gemini, actually. Yeah, I switched to Gemini this year as my go-to model. I think even before 2.5 Pro came out and just seemed better at... reasoning for me it was just like the increasingly i replaced my google searches with gemini and i just sort of trusted that google's i think like the groundings api and its ability to actually like use the google index to give you like

real-time information correctly i just found it was better than i personally i found it was better than all the other tools for that and it was better than perplexity on it too like plexi would be fast but not always accurate and gemini was not quite as fast as perplexity but was always pretty accurate if i asked it about something that happened today even if you use gemini as the reasoning engine in perplexity

I have not done that. Interesting. Yeah. So it's hard to know how much of it is the tooling and how much of it is the base LLM. That's fair. Yeah. I mean, what are your guys' tools of choice? I haven't switched off of ChatGPT. I mean, I find the memory very sticky.

It knows me. It knows my personality. It knows the things that I think about. And so I'll use perplexity for fast web searches or things that...

you know, I know is like a research task because I think ChatGPT is still like a little bit of a step behind for searching the web. I don't know. I think memory is turning into an actual moat for like that consumer experience and i don't expect gemini to ever have the personality that i would expect from chat gpt it just feels like a different like entity you know this thing i'm still surprised about is why they just aren't more um

consumer apps around like all the various things we do like if i think back one of the big changes for me this year is just the amount of prompting and context engineering i do for like my life like and we bought a house recently and like the whole thing like i just had like a really long-running chat gpt conversation stuffing it full of context of like every inspection report or like wanting it to be like level the playing field between me and like the realtor

to understand kind of all the dynamics and things that are going on and it just feels like there should be an app for that but simultaneously i'm sure you took the uh pdfs and just like dropped them into gemini and said like well summarize and tell me what's important for me i guess i worry about i worried about i still don't trust the models enough to be accurate without lots of prompting and it's a high value transaction so you don't want to like get incorrect

data out of it so i still feel like you need to put in the work and it feels like there should still be apps that just do all the work for you did you see carpathy release like sort of uh LLM arena of a sort, which I mean I do by like hand right now using tabs. It's like you have Claude open, you have Gemini open, you have Chechipiti open, and you give it the same task and then you take the output from each and then I usually go

to Claude at that point. And I'm like, all right, Claude, this is what the other one said. What do you think? And check each other's work. Actually, I think that particular behavior at the consumer level that we're doing... startups are doing as well they are actually arbitraging a lot of the models i had some conversations with

a number of founders where before they might have been loyalists to, let's say, OpenAI models or Anthropic. And I just had some conversations recently with them. And these are founders that are running. Larger companies like Series B level type of companies with AI, they're actually abstracting all that away and building this orchestration layer where perhaps as each new model release comes out, they can swap them in.

and out, or they can use specific models that are better at certain things for just that. For example, I heard from this startup, they use Gemini 3. to do the context engineering, which they actually then fed into OpenAI to execute it. And they keep swapping it as new models come up and the winner for each category or type of agent work is different. And ultimately...

They can do this because it is all grounded based on the evals. And the evals are all proprietary to them because they're a vertical AI agent and they just work in a very regulated industry and they have this data set that just works the best for them. I think this is the new norm.

right now where people are expecting yeah the it's cool that the model companies they're spending all this money and making intelligence faster and better and we can all benefit let's just do the best it's almost like the era of

intel and amd with new architecture would come up people could just swap them right yeah it feels at the highest level that angst around where's the value going to accrue is it going to go to the model companies or like the application layer ie the startups feels like that

ebbs and flows in either direction a little bit throughout the year to me. Like I feel there are moments where like your clawed coat amazing launch and it was like oh okay like the model companies are actually going to play out the application layer but then to me at least as all vibes based like gemini surge especially over the last few months just feels like it returns us to a world of where exactly that like the models are all essentially commoditizing each other

It's just like the application layer and the startups are going to set up to have another fantastic year if that continues. I'm curious what you think, Jared, with a lot of perhaps... the negative comments on twitter around is this a bit of a bubble uh ai bubble yeah when i talk to undergrads this is like a common question that i get is like oh like i heard it's a big ai bubble because like

There's all this crazy round-tripping going between NVIDIA and OpenAI. No, this is great for you. Is it all fake? Yeah. No, this is fantastic, right? People look at the telecom bubble and it's like, there's just billions of dollars, tens of billions, hundreds of billions.

like sort of sitting in a bunch of telecom back in like the you know 90s actually that's why youtube was able to exist right like if you just have a whole bunch of extra bandwidth that isn't being used and is relatively cheap the cost is low enough for like something like youtube to exist like if there wasn't a glut of telecom then

like maybe YouTube would have happened and just would have happened later. And then that isn't that like sort of what we're talking about here. Like, how do we, we have to accelerate, right? We have. the age of intelligence the rocks can talk they can think and they can do work and you just have to zap them more and you get like smarter and smarter stuff at this point

I think the argument to college students is actually like, because there will be a glut, there is an opportunity for you. And if there was not a glut, then there wouldn't be as much competition. The prices would be higher. The margins lower. the stack would be higher right and then you know what's one of the big stories this year like nvidia suddenly is on the outs like i think their stock is today is like around 170s or something you know i think i'm still a long-term buy and hold honestly but

For the moment, people are like, oh, well, Gemini is so good. And nobody seems to be NVIDIA only now. And everyone's buying AMD. And TPUs are working. So, you know. at the moment it looks like there's you know what does that mean like there's competition and uh it means that there will be more compute not less and then that means that

probably a little bit better things for all of the big LLM companies, like sort of the, you know, the AI labs. They get a little bit of power, but, you know, they too are in competition with one another. So then what does that mean? Well, then it's, you know, go up another level in this stack, right? Like as long as there are a great many AI labs that are in deep competition with one another, then that's even better for that college student.

who's about to start a company at the application level. Yeah, I think that's exactly right. It's like people are asking this question like, is it a bubble?

that's maybe a question that's really relevant if you're like the equivalent of like comcast like if you're nvidia that's a very relevant question like oh are people over building gpu capacity but like the college students they're not comcast they're actually like youtube if you're doing a startup in your dorm room it's like the ai equivalent of like youtube and like kind of doesn't really matter that much maybe nvidia stock will go down

next year i don't know but like even if it does that doesn't actually mean that it's like a bad time to be working on an ai startup yeah it's what zuck said on a podcast earlier this year i think right it's like meta may end up over investing like a significant amount in like the capex and infrastructure but like they essentially have to the big companies have to do it because they can't just like sit on the sidelines and in the case like demand falls off a cliff for some reason

It's their CapEx, not the startup's CapEx. And there's still going to be tons of infrastructure and ideas to still continue building. There was this book written by this economist called Carlota Perez.

who studied a lot of tech trends, and it studies a lot of technology revolutions, and it summarizes that there's really two phases. There's the phase of installation, which is where a lot of the very heavy... capex investment come in and then there's the deployment phase where really ripples it where it rips and then everything explodes in terms of abundance

And during the initial phase of installation is where it feels like a bubble. There's a bit of a frenzy because it starts first with a, there's this new technology that's amazing, which happened with the ChatGPT moment in 2023. Everyone got super excited about the tech and then everyone got super hyped and got into investing into a lot of the infrastructure with buying a lot of GPUs and...

all the giant gigawatt data center built out. And then people say, but what is the demand? What are going to be all the applications to be built out? I think right now we're in that transition, which is actually really good news for startup founders because they are not involved.

into the building, the data centers, but they're going to build the next generation of applications in the deployment phase when it really proliferates. And what happened, just coming back to the analogy with the era of the internet. Before the 2000s, there was a lot of heavy CapEx investment into the telcos, right? Those were giant projects that college students wouldn't be involved, but they were very heavily invested. And in some cases were...

I mean, there's a whole thing with dark fiber and some pipes that are not used, and that's fine. The internet ended up being still a giant economic driver. And what that means is startups like... The future Facebook or the future Google are yet to be started because those come in in the deployment phase. Because right now, I think things are still getting built up. I do think the foundation lab companies and GPUs very much are falling into the bucket of infrastructure.

Yeah. I mean, it's interesting to watch how this stuff is evolving a little bit. So do you remember summer 24, there was a company called StarCloud that came out and was one of the first to come out and say, we're going to make data centers in space. And what was the reaction when...

you know people laughed at them yeah on the internet yeah right they said that's the stupidest idea ever you know i guess 18 months later uh suddenly google's doing it elon's doing it in every interview now apparently is that right it seems to be like his top talking point yeah and so i mean why is that like i feel like one of the aspects is that like part of the um infrastructure build out right now that's so intense is like we literally don't have

power generation boom supersonic instead of making supersonic jets right now is on this good quest to create enough power for a bunch of these AI data centers that are being built right now. They use jet engines. And even those are so bad. The supply chain for jet engines to generate power for data centers is so backed up that you would have had to have ordered these things.

you know two or three years ago just to even have it in two or three years from now you know these constraints uh end up like influencing like fairly directly what the giant tech companies need to do to win the game three or five years out. Like suddenly there's not enough land.

You know, in America, we can't build. The regulations are too high. In California, we have CEQA, which is totally abused by the environmental lobby to stop all innovation and building housing, by the way. We just don't have enough. like to just do the things that society sort of needs right now so you know the escape valve is like actually let's just do it in space

Yeah, come to think about it, we kind of have the trifecta of YC companies that are solving the data center build-out problem. Well, you need fusion energy. Yeah, well, we have the company that's solving the... no land problem by building through data centers in space. We have Boom and Helion, which are solving that we don't have an energy problem. I just funded a space fusion company that just graduated called Zephyr Fusion. Oh, yes, that's a cool one. And they actually had a great seed round.

out of Demo Day. They're in their 40s. They're national lab engineers who their entire careers, they were building, you know, tokamaks and fusion energy. And they came into the lab one day. They looked at the physics. you know, looked at the math and the models and they said, you know what, if we did this in space, it would actually pencil.

And so they're on this sort of grand next five, ten year quest to actually manifest it, to actually create it in space, because the equations say that it is possible. And if they do it, it's actually the only path to gigawatts of energy up there in space. So it might be an even more perfect trifecta shortly.

Something else I feel like happened over the course of this year is the interest in starting model companies. I guess maybe at both ends, there's like... the people who can raise the capital to go and actually try and build a head-on competitor to OpenAI, which there are very, very few.

ilio with ssi but then more so within yc people trying to build like smaller models um i've certainly had more of those in the last few batches than before like whether it's sort of like a model strong on edge devices or maybe like a voice model specific to a particular language and i'm curious to see if that trend continues going back to this early era of yc actually we sort of saw the explosion of just startups being created and maybe especially sas startups partly what

What fed that was just knowledge about startups became more dispersed. There wasn't the canon of library information on the internet about how to start a startup, how to build software. And then over a decade, that just became more commonplace and that just exploded.

society's knowledge of startups and how to build things and it's maybe feels like maybe we're going through that moment in sort of the ai research meets like actually building things with with training models i think we are absolutely going through that right now yes like where it's going from being a very rare skill set to a more common one because like open ai a decade ago it was like a rare like you needed you need like a

unique combination of skills, right? You need like your researcher brain, your sort of like engineering brain, maybe like your sort of finance business brain. Wait, wait, wait. So did you just describe Ilya, Greg? You got it. There was like a rare team, right? There just wasn't that configuration of team around very much. And now, a decade later, there's like...

plethora of people who have like the research background, the engineering background, the startup capital raising background, or at least can be taught how to do all of that kind of stuff. And I'm curious if that would just mean we'll just see more. applied AI company starting and maybe there'll be like even more models to choose from for all the various specific tasks I think so I think the other thing that's even contributing and making this a very bigger snowball is because of RL.

I think there's all these new open source models that people are doing the fine tune on top of it with a particular RL environment and task. So it is very possible that you can create the best. domain-specific, let's say, healthcare model trained on a generic open source model by just doing fine-tuning on it and doing REL. It beats the regular big model. Actually, I've heard and seen a number of startups where...

Their domain-specific model beats OpenAI, let's say, on healthcare. There's this particular YCE startup that told me that they collected the best data set for healthcare. And they ended up performing better than OpenAI and a lot of the benchmarks for healthcare with only 8 billion parameters. I guess what's funny is that you do need to have a post-training infrastructure. We've also had YC companies where they had something to beat.

OpenAI, GPT 3.5, and they were doing fine-tuning with RL. But then, yeah, GPT 4.5 and then 5.1 came out and basically blew their fine-tuning out of the water. You have to keep going, yeah. Yeah, you got to keep going. Yeah, I mean, you actually have to continue to get to the edge. Anything else that really sort of stood out from this past year that jumps out to you?

It's funny, we started the year with one of our episodes that got a lot of views around vibe coding. I think we were talking about it more as observing a behavior that was happening from our founders. And I was surprised to see that this became like a... giant category there's lots of companies that are winning i mean we have replit there's emergence there's a bunch of them

Varun Mohan had gone over to Google. He released anti-gravity. And did you guys see the video? Actually, I'm sort of curious whether they actually used Nano Banana or any of these video gen things, because it's like a little too perfect, but Google has.

the budget to do the high production value video, but it's, you know, Varun at the keyboard and then, you know, Sergey is like right behind him. So I was like, it was very cinematic. Anyway, I think Sundar was, you know, also not only talking about... space data centers uh he was also talking about vibe coding and i knew that i was a little bit trolling back but knowing what we know i mean yes vibe coding is not

you know, completely usable and trustable for, you know, 100% of your coding period. Like this, you know, it is not true that you can... ship 100% solid production code today as of the end of 2025. Yeah, I was thinking about things that surprised me in 2025. And I think perhaps the thing that most surprised me is the extent to which...

I feel like the AI economy stabilized. I feel like when we did this episode at the end of 2024, it felt like we were still in the middle of a period of incredibly rapid change where the ground was shifting under our feet and nobody knew.

when the other shoe might drop and like what exactly was going to happen with startups and AI in the economy. Now I feel like we've kind of settled into like a fairly stable AI economy where we have like the model layer companies and the application layer companies and seem and the infrastructure they're coming. It seems like everyone.

is going to make a lot of money. And there's kind of like a relative playbook for how to build an AI native company on top of the models. I feel like things really kind of matured in that way. Which feels is all downstream of like the models themselves.

incrementally improved this year but there haven't been like major steps forward that have shaken everything up which is has a knock-on effect many episodes ago we talked about how it was felt easier than ever to pivot and find a startup idea because if you could just survive if you just wait a few months there was likely going to be some like big announcement that would completely make a new set of ideas possible and

create more opportunities to build things it certainly feels like that has slowed down and so like finding ideas is sort of returning to sort of

normal levels of difficulty in my experience in Office Hours. I agree. I'll tell you what's not a surprise. Do you remember that report, AI 2027, where it was just sort of like this doomer piece that said like, oh, well, society is going to start falling apart in 2027. But, you know, at some point they quietly revised it to say that it wasn't 2027, but they kept the title. Maybe it's not a surprise. Like I was always a little bit of a skeptic of like this fast takeoff.

argument because even with the scaling law it is log linear so it is slower it requires like 10x more compute and it's still sort of you know topping out right And that's one form of good news. Another form of it's weird to call this good news, but human beings don't like change in our previous episode where we sort of blew up that. MIT report that said that 98% or 90% of enterprise AI projects fail. Well, it turns out that 90% of enterprises don't know how to do IT, let alone AI.

it's weird to say that that's a good thing but in the context of fast takeoff like that is a real break on the ability of this new really insane technology from actually permeating society i love to accelerate but like it's weird to say like oh well actually in this case maybe that's a good thing right like it is a shockingly powerful technology but

you know, between being log-linear scaling and human beings really don't like change, like organizationally speaking, society will absorb this technology. Everyone will have enough time to sort of... process it like culture will catch up governments will be able to respond to it not in like a frantic sb 1047 sort of like

you know, let's stop all the compute past 10 to the 26, right? Like just these knee-jerk responses to technology. We're excited about the ARC AGI prize is, you know, going to come in and do the winter 26 batch.

as a nonprofit. The funny thing about that is like, yeah, maybe there's a team right now that is climbing the leaderboard of arc agi and they're going to accelerate this thing again something that surprised me to relate to that with the startups is i remember around this time last year we were talking about how companies are getting to a million dollars ar and raising series a's without hiring like some cases

hiring anyone just the founders maybe hiring one person which just felt very unusual i feel like a year on

that hasn't translated into, okay, and then they went and hit like 10 million ARR or they scaled without adding any more people to. No, they turned around and started, and started hiring like actual teams yeah like post series a it actually largely feels like the playbook is the same and the companies might be smaller for the same amount of revenue but it feels it's entirely because they hit the revenue so

fast and there's still just bottleneck on how long it takes to hire people versus they have demand for less people. I still think there is like, you know, some effect, but it is not like open and shut. It is not like you don't have to hire executives anymore.

I think they're like, there might be a case of two foie gras startups, like one being Harvey and the other one being open evidence, right? Harvey, the founders are incredible. They were very early. And then there's this sort of idea of like...

for vcs you could just go down sand hill road and like the fixes in like you just sort of block out all of them and then all the people you know there may be 30 people who could write checks of like 10 to 100 million dollars and if you just sort of get all of their money

Like there's sort of no one who can actually come in and do the next Series A. And then basically you're safe. Like you have capital as a bludgeon is capital as a moat in that case. Right. So, yeah, Harvey is interesting because, you know. Lagora is coming fast for them. And obviously we have some skin in the game on Lagora, but we think that they have as good a shot at any. I guess that's one trend that we saw in 2025 is that there is like a first wave of like AI head of companies like Harvey.

who might have wasted a lot of money on fine tuning, actually. Totally. That like broke out really big in 2023 and kind of did a victory lap that, you know, oh, we've won the space. And now we're seeing a second wave of companies like Flagora and Giga. And it turns out that like, oh, actually like.

It isn't so simple. Yeah, the wood beneficiary of burning some non-trivial double-digit percentage of your capital stack on fine-tuning that... buys you no advantage is like basically the investors are the only winners there because they just own more of your company you know yes at least it relates to like the the hiring and team size i feel like of the two camps one being the ai is going to make everything more efficient you will need less people and the other

AI is going to reduce the cost of producing the time to produce things and so then the expectations from your users and customers will just go up and you'll need to keep hiring more people to satisfy the growing expectations. I feel like this year has been more in that second camp.

And I think that is what's driving the fact that the companies are still just hiring as many people as they were pre-AI. It's just like the bar for what their customers expect. And they're all in the, you know, like Lagora's racing with Harvey, Giga's racing with Sierra. Like they're all...

still competing for the same set of customers and they still ultimately are bottlenecked on like people and like i don't think anyone's bottlenecked on ideas but they're bottlenecked on like people who can execute really well i don't know i think that's like still it's exciting feels like an exciting phase I agree with you that like the era of the one person running a trillion dollar company is not here. Not yet. Yeah.

But I think it's going to trend that way eventually. That'll be a wild time. Maybe that's a prediction for next year. You think it's coming? I mean, I don't think it'll happen in 2026 either, honestly. I mean, I think you will have many stories of...

companies run by you know under 100 people that are making hundreds of millions of dollars so i mean gamma was interesting to see like uh one of the biggest things that they said in their launch that i think is a very good trend is they said they got to 100 million dollars in ARR with only 50 employees.

so which is very different it's you know such an inversion right like normally you have the big banner and the like little x thing you know image and it's like oh yeah like we raised all this money and look at all the people who work for us it's a good trend

the reverse flex which is like look at all this revenue and look how few people work for us well that's all we have time for this time we just wanted to wish you a really happy holidays and happy new year from all of us to you and yours see you next time