The Ledger
Every statement that passed quotation and attribution checks. Mix any filter with any other: certainty 1/5, debate potential 5/5, or both at once.
why aren't all 44 resolved? a statement only gets an assessment when the public record can support or contradict it. opinions and what-ifs never can, and 1 checkable ones are still open, waiting for their date. predictions held up or didn't; assertions are supported or contradicted. on every card: ▮▮▮▮▮ certainty · ▮▮▮▮▮ debate potential. speakers are clickable
Srinivas: The median human software engineer is probably worse than modern AI
“Today, it's pretty obvious that most human software engineers, at least, like, the median human software engineer is probably worse than an AI today.”
Srinivas: No genuine differentiation exists between major AI chatbots right now
“I'll just say it as blunt as it can be is there's not really a genuine differentiation between ChatGPT or Anthropic or Gemini or Grok or Meta AI. Right now.”
Aravind Srinivas: Google will never make AI search the centerpiece of core Google
“Google has the least incentive to bring out AI native search or agents right there on core Google homepage or Google apps. It can be hidden in a mode or like sometimes firing some for some queries, but it's never going to be the central piece of it.”
Srinivas: Instagram is only fine because TikTok was banned in India
“And the only reason Instagram is still fine is because TikTok is banned in, in, in many countries and particularly in India.”
Srinivas: Simple AI ideas scaled with compute outperform complicated architectures
“What matters in reality is making things work. And it's often the simplest ideas that work in practice, especially when thrown a lot of compute at them. The simplest ideas typically outshine the complicated ones.”
Aravind Srinivas: Ilya Sutskever Truly Made Neural Networks Work Through Scale
“And I would say the forefathers like Lacan or Hinton, Benjia, they did a lot of work to establish the foundations, but one guy single-handedly, you know, with, of course, with a group of amazing engineers who worked with him, truly made it work. I'd say it's I…”
Srinivas: AI will not automate physical service jobs anytime soon
“Not anytime soon. Which is funny because that's not paid as much as someone gets paid to write code today, right? So it's a, it's like, it happens in the reverse way. Like everybody wants to think what they do is the one that's the one that's going to be taken…”
Srinivas: AI differentiation in 2025–2026 will come from agentic behavior, not Q&A
“But I feel the, this year in 20, 25 and six, the differentiation is going to come from more agentic behavior. Where like the question answering, like answering questions will be seen as a commodity.”
Srinivas: Many reasoning models will exist, but few products will integrate personal context well
“There's gonna be a bunch of great reasoning models, but there's not gonna be a hundred products that really package, personal context all the API integration, services integrations native integration to your phone to be an assistant really well.”
Srinivas: Perplexity offers Deep Research 10x cheaper than OpenAI via DeepSeek
“It's actually pretty expensive to serve deep research for us. We still price it at 20 dollars a month. I think OpenAI's one is, like, slightly more detailed on some queries. But it's priced at 200 dollars a month. And you can see, right, that that's simply bec…”
Srinivas: Hyper-optimizing consumer AI margins right now is a mistake
“Hyper-optimizing for margins now would be the wrong tactical move.”
Aravind Srinivas: Meta's ad business will flourish even more as AI improves
“I feel like they're very well positioned to keep their existing ad business strong or even, or make it even stronger in a world where AIs actually work. It's a kind of a interesting position to be in for them where their ads business is going to flourish even …”
Srinivas: Meta has a stronger moat than Google due to network effects
“Meta has a bigger moat than Google because Google's moat on distribution comes from their deals with carriers, OEMs, and like all, you know, all these people but Meta's moat is this raw network effects, like, Nobody pre-installs Instagram or WhatsApp on phones…”
Srinivas: 10M users paying $1,000/year creates a multi-hundred billion dollar AI company
“Like, just like an assistant that is so personalized to you, and does a lot of work for you gives you daily briefs, updates, does market research for you, without you even asking for it, is massive. Like people would pay like hundreds of dollars a month for su…”
Srinivas: Google will remain dominant for years via transaction capture
“Another thing I'll tell you why they'll, they'll still continue to be dominant for a few years is let's say you do your research on like what microphones to buy, whatever, or like, you know, Let's say the best headphones for podcast recording. You're buying eq…”
Aravind Srinivas: Google is the only company full-stack independent of Nvidia
“The only one who's managed to do this, I would say, is Google. They built their own chips, they built their own software around it called JAX, and then they built their own accelerated linear algebra library called XLA. And then, you know, they have their own …”
Srinivas: India must train foundation models and build a DeepSeek competitor
“India should definitely train its own models. And not... India should have its own, like, deep seek like company that, that trains models and like competes, not just on Indian languages, but on global benchmarks.”
Srinivas: Indian IT giants will hire fewer people going forward
“They're not going to hire as many people going forward.”
Srinivas: Clients will demand faster delivery and lower IT fees, not cancel contracts
“I feel like humans will still trust other human businesses to do stuff for them, but they'll just push them to, like, hey, like, now that AI can use this, why do you guys need, like, three months to get it done faster? Like, why do you guys need to charge us t…”
Srinivas: Platforms for deploying and sharing personal AI apps will be massive
“I think that layer is still not taken off, but it's certainly something that's waiting to happen because as you can clearly see, software creation is getting a lot easier. So someone's going to be able to be that platform for deploying all these things in a se…”
Srinivas: AI agents will eliminate the need for engineers to build apps
“There's this thing called Replit or Bolt, where you can just go and describe an app you want to build and the agent will build and deploy it for you. And I think that's where things are heading to. Bolt, B-O-L-T or Replit, R-E-P-L-I-T. And sure, it's not going…”
Srinivas: AI compute access will not be democratized globally due to cost
“I think what won't be democratized is access to compute, mainly because it takes a lot of money. And that really depends on which countries choose to invest early on and later on in the process.”
Srinivas: Governments should regulate AI applications rather than foundation models
“I think like regulating models is not necessarily a great idea. And it's not gonna work in practice either. People are still gonna be able to download a model and use it. I think the best way is to regulate applications.”
Srinivas: Perplexity does not train on web data, unlike ChatGPT
“I mean, perplexity, that's why we sort, we attribute it to a source, like, we don't, like, say it's our content, and that way we give credit to the source, and we're not actually training on the data, but ChatGPT is different. They actually train on all the da…”
Srinivas: Perplexity routes every user query through four to five AI models
“Every query goes to a bunch of models but then they're doing different tasks, like one model rewrites your query into a more, like, easily understandable format for the AIs. Another model, like, does the chunking of the pages into, like, parts that gets consum…”
Srinivas: Perplexity built custom Nvidia runtimes and uses Cerebras chips
“A lot of the open source models that we serve ourselves with some fine tuning, ah, we, we've tried to serve it with extreme efficiency, like we wrote our own runtimes for NVIDIA chips, and we used other chips like Cerebras, and that helps us to, like, make the…”
Srinivas: Open-source model releases force AI labs to lower API prices
“Every three months or something like there's a new open source model out there and then that forces them auto labs to lower the prices because then nobody's going to use their APIs.”
Aravind Srinivas: Nvidia GPUs succeeded because AI scaled through neural networks
“If AI was not neural nets, then GPUs wouldn't have mattered. But AI happened to be just basically neural nets at scale. And so all the primitives they built, all the software stack they built ended up being, like, The core foundational building blocks for neur…”
Srinivas: Western AI labs deprioritize Indian voice and speech synthesis
“I think voice most of the AIs are pretty bad at Indian voices. The speech recognition and speech synthesis are not necessarily good. That's a place where you can make a clear difference because it's not a high priority for the Western labs to make it work.”
Srinivas: Core infrastructure and backend fundamentals remain essential despite AI
“I think it still helps to be very good at infrastructure, backend data centers, like floating point arithmetic, storage, all the core fundamentals are not going away. In fact, like, I would say they're very essential in a world where AIs are taking care of the…”
Srinivas: AI will cause significant short-term labor displacement
“The dystopian part of it is unfortunately, in the short term, there's going to be a lot of labor displacement. Not as many people are needed to get a work done anymore.”
Srinivas: Moving slowly on AI will cost hundreds of billions or trillions
“Moving slows is going to cost us a lot long term, and lot means like hundreds of billions or trillions of dollars. So it's best to keep accelerating right now.”
Srinivas: Physical proximity matters much less for mastering AI than hands-on usage
“Physical access matters way less anymore. I think it's more the amount of time you get to spend yourself with an AI model using these apps, understanding where they fail and talking to the best people.”
Srinivas: General intelligence requires one system learning thousands of tasks without reprogramming
“What was really on the frontier of science at that time when I was doing PhD was like, how can we figure out general intelligence? In a manner similar to a human, which is one system doing hundreds of thousands of tasks without explicitly being programmed for …”
Aravind Srinivas: Current AI Lacks Recursive Self-Improvement
“What you're suggesting is like an AI that not only learns and trains on stuff that the humans throw at it, but also like decides what to do next in terms of how to make itself better. Recursive self-improvement. That is not correct yet.”
Srinivas: Today's AI Combines 10,000 Knowledge Professions Without Hard Coding
“Whatever we are working with today is like, okay, it's some 10,000 knowledge worker professions in one system without any like hard coding.”
Srinivas: Training neural networks solely on daily stock opening prices is useless
“If you're training it on the raw stock price, let's say you just have a bunch of numbers of the stock price of Nvidia opening price every single day. Sure, it's not going to be useful on its own, because there are so many other factors that influence the price…”
Aravind Srinivas: Neural networks are the only ML method that truly scales
“There are so many other ways to do machine learning that are like, you know, support vector machines, linear regression, logistic regression, there's like a whole bunch of techniques, but it happens to be that neural networks is the one way to do things when y…”
Srinivas: Robotics AI generalization is constrained by scarce physical data
“Generalization across different physics settings is still like pretty bad. It's not like training on the internet. There's not enough data. So you actually have to build something that's truly intelligent so that it can learn with very little data.”
Srinivas: Scaling compute alone is useless without high-quality training data
“Compute alone is useless. Like, people have tried to reproduce these things with doing the same thing, and it doesn't work. You gotta throw high-quality data tokens at the problem, too.”
Srinivas: AI model reasoning requires training on step-by-step lecture and textbook solutions
“If you want reasoning to emerge in a model, it's good for you to, like, make sure you have YouTube transcripts of video like lectures MIT lectures, Stanford lectures and textbooks, like where you actually have problems, where it's not just a problem, but the s…”
Srinivas: GPU cloud provider CoreWeave will IPO soon
“Okay, so there's this company called Core Vive in, in the US. I think it's gonna IPO pretty soon.”
Srinivas: Data centers won't have high margins without software integration
“I don't expect it to be a pretty high margin business of its own unless you combine it with good software.”
Srinivas: Everyone will have an affordable AI personal assistant within five years
“I think we'll all have like a personal assistant. It's going to feel really amazing. It's not going to be a luxury thing anymore. It's not just a thing billionaires had access to. It's going to feel like an iPhone where the same phone that, that the president …”