Should we break down five wildly obvious fixes that will explode consumer LLM adoption? They don't want you to know this over at the big labs, but I have some ideas. Basically, everyone's been really focused on agentic coding and the SaaS apocalypse and what's happening in the business-to-business world and the enterprise world. I've just been sort of like thinking back on, you know, basic improvements to the chat apps that I use all the time, because there's some really obvious stuff that I think is in the works. I think it's coming, but I wanted to just sort of like get it all down in one place to think about what the next iteration and the next breakout moment when people are like, 'Oh, I'm using them even more. I'm having a better experience. What would that look like?' So, the first thing is that I realized that I've asked ChatGPT just when was OpenAI founded three different times. It's the exact same query. Like it doesn't need to light the GPUs on fire for that question. The answer literally never changes. You can cache the result and that's what Google does with those knowledge queries, knowledge panels, and there's a whole bunch of different ways to deliver results that are sort of pre-cached. And so if you just look down, when an LLM launches, basically every question has never been asked before. But now there's a lot of people that are just showing up with the exact same question. Give me the history of the Roman Empire. Give me the history of this company. And you might not be the first person to ever ask that question. But also, if you do a little bit of fuzzy search over it, there are probably hundreds of thousands of people that have asked the exact same thing. So cache those results, give them to the user instantly. And I think this instantaneous feeling of LLMs, like they felt slow for a really long time. They actually got slower. Like it was always sort of slow. You watch the token stream in, but then once the reasoning models and the thinking models and the deep research and the o3 Pro came out, it was like really slow. It was like close your phone and come back in 20 minutes. That doesn't have to be the end state. And I don't think it will be. And I have no better example than the number two on my list, which is cerebrus inference. So ChatGPT currently has a model called 5.2 Instant and it is not instant at all. I fired off a prompt of 5.2 Instant and I said, 'No reasoning, tell me the history of LLMs.' It took 38 seconds to deliver the full response. So for like all the tokens to stream in, and it does a good job. It shows you images and stuff and it is a cool illustration. But then I went over to Codex Desktop and I fired up GPT 5.3 Codex Spark Low, which is a crazy name which we'll get to. And it responded in under two seconds because it's, from what we know, Spark is incredibly cerebrus and it's very, very fast. And so everyone's obsessed with like the fast models in the agent coding world because you're waiting a half an hour for something to get back to you. You're waiting five minutes for something to get back to you and you're actually losing your train of thought. But I think that applies in consumer as well. And I think the interaction of sending a message and then just immediately getting a response before you actually think, 'Oh well, like it's waiting. I'll close the app. I'll check my messages. Oh, I got an Instagram notification. Let me go over there.' Like instant responses will keep people in the apps longer and user minutes will actually increase once that rolls out. So pretty simple implementation for I think most companies. My big question is like, I know Google has a huge advantage with TPU, but I don't know if they have like an answer to Cerebrus specifically. And Nvidia just brought Grok, which can do I think some of the same things. So I'm curious to know how every lab solves the like fast responses question because that feels like an important piece of the puzzle. It's not the only piece of the puzzle, but it's an important feature and I think we're going to see it rolling out to consumer LLMs very, very soon. And I do think it'll be an interesting moment for people to both ask a question and just boom, it's as fast as going to Wikipedia and just seeing like, okay, everything's rendered, it's thoughtful, it's what you want. And on the flip side, I think that it could make people a lot more chatty with them, like actually asking follow-up questions because you don't feel that cost of like, oh, if I ask you to follow up and tell me more or go a different direction, like I have to wait. I have to wait, so I might just close the app. Speaking of GPT 5.3 Codex Spark Low, no more model names. Like truly no more model names in consumer AI LLM chat apps like ever. Like just bury them so deep in the UI that you never see them. And people will complain. People will be like, I wanted it to be easier to pick. I like picking between Pro and Thinking and Fast and Instant. I know what I want for everything. People complain, but it will all inspire the model routing team to grind harder. And the model routing team has a hard job to do, but they will eventually figure it out. And eventually you should be able to just talk to the model. You can already do this in chat. You can say, 'Hey, think really hard about this question and give me a really thorough answer.' And it'll switch from instant to thinking. I don't know that you can trigger Pro from that. I haven't actually experienced that. I did try and trigger a deep research report. I said, 'Hey, please deep research the Roman Empire for me.' And it does not fire off a deep research report. Deep research is buried under like a plus button and you have to select it and say, 'Okay, I actually want you to do this thing.' And then it takes you down the deep research workflow, which I understand is like for inference reasons. They don't just want you firing off deep research reports all the time. But I think in the future, the model router should be very intelligent about, okay, this is a question that people have asked thousands of times. Let's just go get it from a database, which is crazy to think in the age of AI you wouldn't even be hitting a GPU, but I think that's going to be real. And then I think on the other side, it should detect like, okay, this person wants something that's far beyond anything that we've ever worked on before. We got to go search the internet. We got to write some code. We got to do a whole ton of stuff. So let's fire up deep research. Right. Fourth, ads. We've talked about this, but we got to get them in the LLM. We got to get them everywhere because I was thinking about the death of Google Reader. I don't think you were ever a Google Reader guy, were you? But it was amazing. You could take all these RSS feeds from all these different blogs during the blogosphere. You could put Marginal Revolution, Tyler Cowen's blog, all these different stuff, all these different things in there and just kind of scroll through them really quickly. And Google wound up killing it and everyone was like really upset. And the reason was, I think, because they never really got on the Google ad flywheel where there was real like revenue generation and...
Totally. Totally. But my takeaway from Google's surface area of products that are successful and loved, Google Search, Google YouTube, Google Maps, Chrome, Android, like these are all direct funnels for the ads flywheel. And so you can see that they're driving the bottom line. There's a whole bunch of folks on the team that are getting excited when they're hitting their numbers, when they're making more money for the business. And so they just get more and more resources, more and more engineering effort. Everything gets better. And I think that not only are ads the best way to deliver high quality products to the broadest possible audience, but they just make products better top to bottom. And yes, there's the stated versus revealed preference thing. And yes, you might want to pay to not have ads like you do on YouTube. Many people do, but I do think that ads flywheel is going to be really, really important as the inference gets really...
