The web was built for human eyes
Nora · 0:00You know, usually when we think about urban planning, I mean, we imagine building a city for people.
Miles · 0:05Right. Yeah. It's completely anthropocentric.
Nora · 0:07Exactly. You laid on sidewalks so people can walk or you put up street signs so human eyes can actually read them.
Miles · 0:14You design storefronts with big glass windows. Right. Because we build for our own biological senses.
Nora · 0:19Yeah. Every door handle is placed at a specific height just because that's where a human hand naturally reaches.
Miles · 0:25It reflects our limitations. Yeah. Like our walking speed, our field of vision. I mean, our cognitive load.
Nora · 0:32But imagine if, just overnight, a completely new population moved into the city.
Miles · 0:38Like who?
Nora · 0:39Let's say millions of autonomous delivery drones.
Miles · 0:42Oh, wow. Okay, yeah.
Nora · 0:43A population that doesn't walk on sidewalks, right? They don't read street signs with optical lenses.
Miles · 0:48And they definitely couldn't care less about your glass doorfronts.
Nora · 0:51Exactly. So suddenly the sidewalks are completely useless to them. The street signs are invisible or just irrelevant.
Miles · 0:58I need digital beacons or specialized loading docks, aerial corridors.
Nora · 1:03Right. The city's physical human-centric infrastructure is abruptly obsolete for the entities actually doing the most moving around.
Miles · 1:12That is a really fascinating way to frame it because that's not just a metaphor. That is literally what is happening to our digital infrastructure right now. We are the legacy residents in a digital city that is being rapidly repopulated by non-human actors.
Nora · 1:26And that is exactly what we are unpacking today. Welcome to the Deep Dive, everyone. We have a stack of sources today, interviews, white papers, critical essays, even a product fact sheet that basically prove the bedrock of the Internet is shifting.
Miles · 1:39The fundamental nature of how digital files are structured is changing right under our feet.
Nora · 1:44Because we are moving from a world where humans are the primary readers to a world where machines are the primary readers. And we're looking at this through the lens of five specific architects of this new infrastructure.
Miles · 1:55We've got distinct converging angles from Andrej Karpathy, Sam Altman, Jensen Huang, Satya Nadella, and Aravind Srinivas.
Nora · 2:03And they aren't collaborating on a single manifesto or anything, but they are all staring at the exact same structural collapse from different vantage points.
Miles · 2:11It's a massive convergence story. I mean, you have the people building the raw code, the people managing the scale, the hardware manufacturers, the enterprise software giants, and the search paradigm disruptors.
Nora · 2:23All pointing to the exact same phenomenon. Right.
Miles · 2:26The machine is becoming the reader, and that changes the entire supply chain of information.
Karpathy: documentation written for machines, not people
Nora · 2:32Okay. Let's unpack this, starting with the sheer friction of our current architecture. Andrej Karpathy recently put out this premise that practically set the tech world on fire.
Miles · 2:41Yeah. He was looking at software documentation.
Nora · 2:43Right. And he stated that 99.9% of attention is about to be LLM attention, not human attention.
Miles · 2:50Which forces a complete rethink of how we publish information. I mean, think about a standard documentation website today.
Nora · 2:56Okay.
Miles · 2:56If you are a developer looking up how to use a React library or a Python module, you go to a web page. That web page is heavily structured with HTML. It has a navigation sidebar on the left, a sticky header at the top, a search bar, interactive drop-down menus.
Nora · 3:13All the pretty stuff.
Miles · 3:14Yeah, intricate CSS styling to make the code blocks look distinct from the text.
Nora · 3:19It's the glass storefront. I mean, it's built for my eyes to stand and my mouse to click.
Miles · 3:23Exactly. But if 99.9% of the entities reading that documentation in the near future are AI agents like bots trying to ingest that library to write code, they do not care about your CSS.
Nora · 3:36They really don't.
Miles · 3:37They don't care about your sticky header. In fact, all of that HTML markup is actively detrimental. It's a computational friction.
Nora · 3:45Let's dig into the mechanics of why it's detrimental, because I think people assume an AI just kind of looks at a web page the way we do.
Miles · 3:52Right, like it have eyeballs.
Nora · 3:53Yeah, but it's actually reading the raw code underneath, right? And that code is incredibly bloated.
Miles · 3:58It is massively bloated. You have to remember how these models actually process text. They don't read words. They process tokens.
Nora · 4:05Tokens, right.
Miles · 4:05And every single token costs compute. It costs money. It takes up space in the model's context window. And it requires mathematical operations to process. So if you feed a standard web page into an LLM, a huge percentage of those tokens are going to be consumed by HTML tags. Like div class equals sidebar. Nav ID equals main menu. Span style equals color hex code.
Nora · 4:30Wow. Wow. So if I have 128K context window and I drop a web page in there, I might be burning, what, 30,000 tokens just on the invisible scaffolding of the page?
Miles · 4:39Easily.
Nora · 4:39Just on the stuff that tells a human browser where to put the text rather than the text itself?
Miles · 4:44Precisely. You are paying for the model to read the scaffolding. It's an incredible waste of the attention mechanism.
Nora · 4:49And it creates noise, right?
Miles · 4:50Exactly.
Nora · 4:51Yeah.
Miles · 4:51The model has to calculate the relationship between the word function and the HTML tag that happens to be sitting next to it in the raw text. The model has to work harder to extract the actual semantic meaning from the visual formatting instructions.
Nora · 5:04Which is why Karpathy says this whole paradigm is obsolete. I mean, his argument is that documentation shouldn't be a web of interconnected HTML pages at all.
Miles · 5:11No, it should just be a single massive text file, specifically a Markdown file.
Nora · 5:18Yes, Markdown.
Miles · 5:19Markdown is the perfect medium for this transition. It's plain text with incredibly lightweight semantic formatting, just asterisks for bolding, hash marks for headers.
Nora · 5:30So simple.
Miles · 5:31It strips away all the visual presentation layer and leaves only the structural hierarchy of the information.
Nora · 5:36And Karpathy's vision is that you design the documentation so it drops directly, cleanly, and instantaneously into the context window of the model with zero token waste.
Miles · 5:46Right, which directly ties into the new web standard that Jeremy Howard has been pushing, this lmls.txt concept.
Nora · 5:53Yeah, I've seen that. Anthropic and others are already starting to support it.
Miles · 5:55Exactly. I mean, we all know index.html.
Nora · 5:58Right. The default file a browser looks for.
Miles · 6:00Yeah. When a human navigates to a domain, it tells the browser how to paint the visual scorefront. And we know robots.txt, which tells legacy search crawlers what directories they're allowed to index.
Nora · 6:14So lmls.txt is an entirely new architectural layer.
Miles · 6:19It's the loading dock for the delivery drones.
Nora · 6:21That's exactly what it is. It's a file sitting at the root of a domain that says, Hello, AI agent. I recognize you are not a human. Do not bother requesting the HTML. Do not parse our CSS. Do not execute our JavaScript payload.
Miles · 6:34Here is a pristine, hyper-dense markdown file containing the entire semantic knowledge of the site, perfectly formatted for your context window.
Nora · 6:42You bypass the human UI entirely.
Miles · 6:44Entirely.
Nora · 6:45But it's not just code documentation, is it? The sources point out how much of our broader human knowledge is locked up in pre-LLM era format. Oh, it's everywhere. We have built decades of digital infrastructure using formats that were uniquely designed to stimulate human biology.
Miles · 6:59Think about the PDF.
Nora · 7:00Oh, the PDF.
Miles · 7:01The portable document format was literally designed to perfectly preserve the visual layout of a pretty piece of paper on a screen.
Nora · 7:07Right.
Miles · 7:07It prioritizes visual fidelity over semantic structure. structure. To an AI, a complex PDF with multi-column layouts, embedded images, and weird fine codings is notoriously awful to parse.
Nora · 7:20The text extraction often comes out totally mangled.
Miles · 7:23Right. And the same goes for video and audio. A video file is a massive opaque blob of pixel data, unless you run a separate compute-intensive transcription layer on it.
Nora · 7:32So we've trapped our own history in these anthropocentric formats. We built a world of PDFs and React apps and interactive dashboards. And now that the primary reader is a machine, all of that human-centric design is just friction.
Miles · 7:46And if we remove that friction, if we shift to formats like lml.txt and raw markdown that the machines can digest natively, the volume of reading, they're going to do scales exponentially.
Nora · 7:57Which brings us to the actual volume we're talking about here.
Miles · 8:00Yeah, the scale is wild.
Altman, token maxing, and the trillion-token context
Nora · 8:01Right, because if the friction is gone, the floodgates just open. This is where Sam Altman's perspective from the Sequoia Capital AI Ascent event becomes so critical to this deep dive.
Miles · 8:11He painted a picture of machine consumption that genuinely breaks the human comprehension barrier.
Nora · 8:16To understand the velocity of this shift, Altman compared OpenAI's earliest days to today. In 2016, they were a tiny research lab.
Miles · 8:25Right, they weren't building chatbots back then.
Nora · 8:27No, they were trying to get reinforcement learning agents to play video games. The scale was microscopic compared to now.
Miles · 8:32Fast forward to today, and he introduced this metric about their heaviest power users that is just staggering.
Nora · 8:39The phenomenon of token maxing.
Miles · 8:41Yes, token maxing. A few years ago, when LLMs were just starting to gain traction, their absolute heaviest, most obsessed user was burning through maybe 100,000 tokens a month.
Nora · 8:54Which, if a token is roughly three quarters of a word, is about the length of a solid novel. 100,000 words a month.
Miles · 9:00That's a heavy reader, but it's human scale.
Nora · 9:02Exactly. It's a biological scale of consumption.
Miles · 9:05But today, that 100,000 tokens a month is a rounding error. It's the per capita average for a casual user.
Nora · 9:12Yeah.
Miles · 9:12The new power user, the modern token-maxer, is burning through 100 billion tokens a month.
Nora · 9:17I need to stop on that number. 100 billion tokens in a single month.
Miles · 9:23It is the equivalent of reading tens of thousands of books every single day.
Nora · 9:26No human being could consume that in multiple lifetimes.
Miles · 9:30Not even close.
Nora · 9:30And this isn't a theoretical limit. This is happening right now every month by individual users leveraging AI pipelines.
Miles · 9:37They are feeding raw data into machines, having the machine summarize, synthesize, generate and analyze at a scale that is completely divorced from the physical limits of the human optic nerve.
Nora · 9:49It represents a fundamental decoupling of knowledge processing from human biology.
Miles · 9:54Yes.
Nora · 9:54And when you realize that this is just the current state of token maxing, you start to see why Altman's vision for the endgame of this technology is so radical.
Miles · 10:03Yeah, they asked him at the event what the platonic ideal of AI looks like, the ultimate manifestation of the technology.
Nora · 10:09And his answer wasn't some dystopian global supercomputer running the electrical grid.
Miles · 10:14No, it was incredibly personal, but architecturally mind bending. He described the Platonic ideal and a very small, highly efficient reasoning model equipped with a context window of a trillion tokens. A trillion tokens. And into that trillion token window, you place your entire life.
Nora · 10:33Just for one user, meaning every single email you have ever sent, every Slack message, every text thread, every book you've read, the transcript of every meeting you've ever attended.
Miles · 10:43Your entire digital exhaust.
Nora · 10:45And continuously appending to this singular context window in real time.
Miles · 10:50And here is the crucial architectural distinction he made. That model never needs to be retrained on your data.
Nora · 10:56Okay.
Miles · 10:57Its underlying weights the neural connections form during its initial pre-training never change. It just sits there, dynamically reasoning across the entire continuum of your life's information within its active memory.
Nora · 11:08Okay, I have to interrupt here because I can hear the computer science listeners screaming about the attention mechanism.
Miles · 11:13Oh, absolutely.
Nora · 11:14How is a trillion token context window even mathematically possible? The attention mechanism in a standard transformer model scales quadratically, right?
Miles · 11:24Yes, the O of n squared problem.
Nora · 11:27Right, so if you double the size of the context window, the compute required doesn't double, it quadruples.
Miles · 11:33Because every new token has to mathematically relate to every single other token in the sequence.
Nora · 11:38So if you scale that up to a trillion tokens, the compute required just to read the prompt would require the energy output of a small star.
Miles · 11:46Which is exactly why this is a platonic ideal and not a product launching next Tuesday.
Nora · 11:51Fair point.
Miles · 11:52To achieve a trillion tokens of active dynamic context, we need massive architectural breakthroughs beyond the standard dense transformer.
Nora · 12:00Are we seeing any of those breakthroughs yet?
Miles · 12:01We are seeing early signs of this with things like ring attention, which distributes the context across a mesh of GPUs or specialized KV cache offloading strategies, or even entirely new architectures like state space models like Mamba.
Nora · 12:14Oh, right, Mamba.
Miles · 12:15Which attempt to compress the historical context into a fixed state rather than keeping every token actively in attention.
Nora · 12:22So Altman is essentially banking on hardware and algorithmic breakthroughs that solve the quadratic scaling problem, allowing an AI to hold an entire human lifetime in its working memory simultaneously.
Miles · 12:36Yes. It knows everything about you, not because it was fine-tuned on your data, but because it has simply read everything you have ever generated, and it holds all of it in active memory, ready to synthesize and answer the moment you ask a question.
Nora · 12:49It's like we've spent 30 years building beautiful, ornate libraries for humans to walk through. We meticulously organized the shelves. We built card catalogs. We spent hours browsing the spines. And suddenly we realized that browsing is too slow. The biological interface is the bottleneck.
Miles · 13:05So we are ripping the doors off the library.
Nora · 13:07Tearing down the shelves. Replacing the entrance with massive loading docks sized for industrial forklifts because the machines are coming in to ingest the entire building in a single gulp.
Miles · 13:17The transition from human browsing to machine ingestion. Yeah. And, I mean, think about your own digital footprint in this context.
Nora · 13:25Okay.
Miles · 13:26Think about the thousands of saved PDFs you have in a folder, the forwarded email treads you skimmed and archived, the research papers you downloaded to Zotero but only looked at the abstract.
Nora · 13:36Oh, I'm guilty of that. We are digital hoarders.
Miles · 13:39Everyone is.
Nora · 13:39We collect information at a pace that vastly exceeds our physical ability to read it. We save it with the intention of reading it, but the friction is too high.
Miles · 13:48Altman is saying, you don't have to read it anymore. The machine will read the hoard.
Nora · 13:53But if the machine is doing all the reading and it's doing it at a volume of 100 billion tokens a month for power users, that introduces a profound economic problem.
Miles · 14:03Someone has to pay for the compute to process those tokens.
Nora · 14:06Right. If reading is no longer a biological activity but a computational one, how does that change the software industry?
Miles · 14:13Which brings in Jensen Huang, the CEO of NVIDIA.
Nora · 14:16Because if Altman is the one demanding the compute to process a trillion tokens, Huang is the one selling the silicon that makes it possible.
Miles · 14:24And Huang has a very specific industrial way of looking at this shift.
Huang: the data center as a token factory
Nora · 14:27He does. To Huang, data centers are no longer just places where you store static files for humans to download later. The cloud is no longer just a hard drive in the sky.
Miles · 14:37Right. He calls modern data centers AI factories.
Nora · 14:40Manufacturing intelligence.
Miles · 14:41Exactly. And the base unit of that economy, the currency of the factory, is the token. The token. For Huang, reading is no longer a slow biological bottleneck where a human sits down with a cup of coffee and stares at a monitor. Reading is a near instantaneous machine operation. It is a discrete compute cycle.
Nora · 15:02He gave this amazing example about agents interacting with new tools. He said that if an AI agent is confronted with a piece of software or a device it has never seen before, it doesn't fumble around clicking buttons to figure it out like a human would.
Miles · 15:15It just reads the manual.
Nora · 15:16It instantly ingests the documentation, presumably that llms.txt file, and in a millisecond it goes from zero knowledge to complete expert.
Miles · 15:25And because of that capability, Huang makes a prediction that completely inverts how we think about the software market.
Nora · 15:31What's the prediction?
Miles · 15:32AI agents are about to become the most voracious software customers in the world.
Nora · 15:36Okay, wait. We always assume software is sold to human users. We price it per seat. $20 per user per month.
Miles · 15:44But a human user sleeps. A human user goes to lunch. A human user can only click so many buttons a minute. An AI agent can read documentation, use API tools, process data, and generate tokens 247 without ever stopping. They will consume software services, cloud compute, and API calls at a rate no human workforce could ever match.
Nora · 16:05So if you are a software company, your biggest, most profitable client in the future isn't a Fortune 500 enterprise full of human employees. It's a swarm of autonomous agents orchestrating workflows across your infrastructure.
Miles · 16:17Which forces us to look at the architecture of the software itself. If the customer is an AI agent and the entity doing the reading and interacting is an AI agent, what happens to the human interface?
Nadella: what an enterprise app actually is underneath
Nora · 16:27Ah, this is where Satya Nadella drops a bomb. We have the CEO of Microsoft, a company whose entire empire, its massive valuation, is largely built on enterprise software.
Miles · 16:38Specifically, the software as a service model. The entire Microsoft 365 suite.
Nora · 16:44Right. And Nadella asks a question that should send shivers down the spine of anyone currently building traditional enterprise apps. He asks, hey, why do I need Excel?
Miles · 16:54It's an incredible thing for him to say.
Nora · 16:55It's like the CEO Ford asking, why do we need cars? Excel is the lifeblood of global finance.
Miles · 17:02But his logic is impeccable. When you break down what an enterprise application actually is, he foresees a disruptive era of agentic AI that will aggressively collapse the traditional app model.
Nora · 17:14To understand why, we have to strip away the visual layer and look at the underlying architecture of business software. Let's do that because he uses this phrase. He says these apps are essentially CR-rood databases with a bunch of business logic.
Miles · 17:25Right. CRDU stands for Create, Read, Update, and Delete.
Nora · 17:29Okay.
Miles · 17:30At their absolute core, almost all business applications, whether it's Salesforce for managing leads, Workday for HR, Jira for project management, or Excel for financial modeling, are just relational databases.
Nora · 17:41They are digital filing cabinets that allow a user to do those four basic operations on a piece of data.
Miles · 17:48Okay. So you have the database at the bottom. On top of that database, developers build the business logic layer.
Nora · 17:53These are the rules.
Miles · 17:54Yes. If a user updates a sales lead to closed, automatically update the revenue projection in this other table. It's the workflow routing and the permission structures.
Nora · 18:03And then on top of that, you have the front end, the user interface, the buttons, the dashboards, the pie charts, the colorful menus.
Miles · 18:10Exactly. But why does that UI exist?
Nora · 18:13For us.
Miles · 18:13It exists strictly as a translation layer because humans cannot easily read or manipulate raw database tables or JSON payloads. We need visual metaphors.
Nora · 18:23Many folders and windows and interactive buttons to help our biological brains understand the state of the database.
Miles · 18:30But wait, so if the user is an agent, the agent doesn't need visual metaphors. No. It doesn't need a pie chart to understand data variance. It can just read the raw JSON array.
Nora · 18:40Precisely. And this is how Nadella's vision of enterprise software collapsing perfectly mirrors Karpathy's vision of web documentation changing.
Miles · 18:48They are both shedding the human-centric interface.
Nora · 18:51face. Nadella argues that in the agent era, the AI itself becomes the unified logic layer. The agents will be multi-repo CRUD. Meaning they don't care about the boundaries between different apps. Right. An agent doesn't need to log into Salesforce, export a CSV, log into Excel, run a pivot table, and then paste the result into a PowerPoint. That's a human workflow. Exactly. The agent just sits across all the underlying databases databases simultaneously via their APIs. It reads the raw data from the CRM, runs the mathematical logic in its own AI tier, and directly writes the update to the presentation tier. So the UI,
Miles · 19:28the actual app part of the app that humans interact with, becomes completely vestigial.
Nora · 19:34It's an evolutionary leftover from the era of manual data manipulation. If I want to know my
Miles · 19:39sales projections for Q3, I don't open an app, navigate to a dashboard, and adjust a bunch of sliders. You just ask the agent. And the agent reads the database, processes the logic, and gives me the final number. The entire concept of an app collapses into the AI tier. Hey, why do I need
Nora · 19:55Excel? If you have an orchestrating agent acting as what Nadella calls the UI for AI, you just tell
Miles · 20:01it what you want. It writes the Python script, queries the data lake, performs the matrix multiplication, and returns the insight. The traditional SAWS interface is dead weight. This This is a massive architectural paradigm shift. It's the death of front-end engineering for enterprise tools.
Srinivas: answers with citations instead of ten blue links
Nora · 20:16But it introduces a really crucial problem about the human interaction model. If the UI collapses, if I'm no longer opening files, clicking through dashboards or browsing websites, how do I actually interact with the information?
Miles · 20:28Enter Aravind Srinivas, the CEO of Perplexity.
Nora · 20:32Right.
Miles · 20:32His whole thesis is that the traditional search engine model, the list of 10 blue links, is fundamentally broken for this new era.
Nora · 20:39Because a list of links still assumes the human is the reader. Google says, here are 10 URLs that have high vector similarity to your query. Go click them, parse the HTML, find the relevant paragraph, and abstract the answer yourself.
Miles · 20:54It outsources the cognitive labor back to the biological user.
Nora · 20:57Serenovas argues that the future is the answer engine. The interaction model flips entirely. You move from searching and clicking to asking and receiving.
Miles · 21:07You ask a natural language question. The AI goes out, reads the relevant documents, whether that is the open web or your company's internal multi-repo databases, and synthesizes a direct answer for you.
Nora · 21:17You delegate the reading entirely.
Miles · 21:19But there is a massive caveat here that Shrena Vest brings up. And I think this is where we have to pause. Because if you are delegating the reading to a probabilistic machine, you are walking into an epistemic minefield.
Nora · 21:30Oh, an absolute necessity for this to work is provenance, verification.
Miles · 21:34Because LLMs, fundamentally, are prediction engines. They're not databases of facts.
Nora · 21:39Right. If they don't have the exact information in their context, they will statistically interpolate. They will hallucinate a highly plausible, grammatically perfect fiction.
Miles · 21:49Therefore, Srinivas insists that every single answer generated by an answer engine must carry rigorous inline citations back to the original human-created source material.
Nora · 22:01It has to show its math. Here is the answer to your query, and here is a footnote linking to the exact sentence in the raw database where I extracted this specific metric.
Miles · 22:11So if you doubt the AI synthesis, you can click through to the ground truth.
Nora · 22:14Without that strict verification layer, delegating the reading is just a recipe for organizational chaos.
Miles · 22:20You would be operating your business on blind faith in a statistical model known to fabricate data when compressed.
Nora · 22:27Okay, I need to jump in here and push back on this entire narrative. I want to play devil's advocate for a second.
Miles · 22:31Go for it.
Who profits from the machine-reading story
Nora · 22:32Because we've just spent 20 minutes listening to five of the most powerful men in tech tell us that machine reading is the inevitable, glorious future of all digital interaction.
Miles · 22:42Let's look at the incentives.
Nora · 22:44Exactly. Look at who these five witnesses are. Jensen Huang is the CEO of NVIDIA. What does NVIDIA sell?
Miles · 22:51They sell the H100 and B200 GPUs that are strictly required to run these massive language models.
Nora · 22:57Sam Altman is the CEO of OpenAI. They sell access to their models metered strictly by the token. They are literally selling the reading process itself.
Miles · 23:06Satya Nadella is the CEO of Microsoft who owns Azure, the cloud infrastructure where all of this astronomical compute actually happens.
Nora · 23:14They are the infrastructure cartel.
Miles · 23:15They are.
Nora · 23:16They aren't just predicting a gold rush. They are the ones selling the picks, the shovels, and the maps.
Miles · 23:20They desperately want us to believe that AI agents reading trillions of tokens is the future. Because for that future to arrive, every enterprise on earth has to spend hundreds of millions of dollars buying their compute, their tokens, their cloud storage.
Nora · 23:34Right. To a normal company, say, a global logistics firm or a retail chain, tokens are not a magical new currency. Tokens are a massive, exorbitant operational expense.
Miles · 23:45Every time an agent reads a document, it burns a fraction of a cent. Multiply that by 100 billion tokens a month and you bankrupt a midsize company.
Nora · 23:53So isn't it entirely possible that this whole 99.9% of attention narrative is just a highly coordinated, self-interested hype cycle designed to inflate their massive capital expenditures and stock prices?
Miles · 24:08That is the necessary skeptical lens. We absolutely have to view their claims through the reality of their financial incentives.
Nora · 24:14Because token maximalism is, without a doubt, a sales pitch.
Miles · 24:18O' When Altman talks about the platonic ideal of a trillion tokens of continuous context, he is simultaneously describing a scenario where OpenAI is getting paid to process a trillion tokens per user indefinitely. The margins on that are astronomical.
Nora · 24:32It's incredibly convenient that the utopian future requires infinite consumption of their specific product.
Miles · 24:38O' However, and this is the crucial synthesis we have to make, we must separate the hype about the scale of the infrastructure from the reality of the underlying behavioral shift.
Nora · 24:46What do you mean?
Miles · 24:47Even if the infrastructure providers are hyping the agentic future to maximize their margins and justify their data center buildouts, the change in how users expect to interact with information is organically cementing itself.
Nora · 25:01You mean the toothpaste is already out of the tube?
Miles · 25:04Think about the generational divide in AI adoption that Altman noted. Older knowledge workers still tend to treat AI like a flightly conversational Google search.
Nora · 25:13Sure.
Miles · 25:13But look at 20-year-olds in university right now. They are treating LLMs as localized operating systems. They are dumping 50 PDFs of complex course material into a context window, having the model synthesize the disparate arguments, cross-reference the citations, and draft comparative outlines.
Nora · 25:31They aren't clicking blue links.
Miles · 25:33They're asking complex, multivariable questions. That behavioral baseline is real, regardless of whether NVIDIA's forward PE ratio is overvalued.
Nora · 25:41The expectation that the machine will do the heavy lifting of reading the document is the new standard for the next generation of knowledge workers.
Miles · 25:48That makes perfect sense. The economic reality of deploying this at enterprise scale might result in a brutal awakening when the CFO sees the Azure bill for agenda compute, but the user experience of ask, don't read is far too compelling to abandon.
Nora · 26:03Exactly. And that brings us to the most practical, fascinating application of this entire convergence.
What this means for the documents you already have
Miles · 26:09What happens when you take these massive macro concepts, carpet these LLM optimized formats, Altman's hyperscale context windows, Huang's token economics, Nadella's collapsing UI, and Srinivas's answer engine?
Nora · 26:23And you point all of them squarely at one single person's lifetime of unstructured
Miles · 26:27data. This is where we look at the fact sheet for DocuStrata.
Nora · 26:30Yeah. This isn't some enterprise behemoth built by Microsoft. It's a specialized tool designed to solve a very modern, very specific problem.
Miles · 26:37And its tagline perfectly encapsulates the shift we've been talking about. Read nothing, know everything.
Nora · 26:42That tagline is essentially the Karpathy thesis compressed into four words. DocuStrata is an AI document intelligence layer.
Miles · 26:50Its origin story starts with a pain point that is intensely relatable, just scaled up to an extreme degree.
Nora · 26:56Yeah, the fact sheet details that the founder built this because they had a personal archive of over 100,000 documents stored in a legacy note-taking app.
Miles · 27:05100,000 PDFs, web clippings, scanned receipts, email threads, voice memos.
Nora · 27:12And then the legacy app raised their subscription pricing by an order of magnitude.
Miles · 27:16Now, the legacy response to that problem is to find a cheaper digital filing cabinet. You export everything to a local hard drive, or you migrate to a cheaper cloud storage bucket. You preserve the filing system.
Nora · 27:29But the founder realized something fundamental about human behavior. Having 100,000 documents is not a storage problem. It is a reading problem.
Miles · 27:36No human being has ever read or will ever reread 100,000 documents in their personal archive.
Nora · 27:42We save things because we have a fleeting thought that says, I might need to know the specific parameter in this white paper someday.
Miles · 27:48But the friction of finding it, opening the PDF, scrolling to page 47, and rereading the text is so high that we almost never actually retrieve it.
Nora · 27:56So instead of finding a cheaper place to hoard documents they will never look at, they built an architecture that reads the archive for them.
Miles · 28:02The archive ceases to be a passive filing cabinet you have to manually rummage through. It becomes an active intelligence you interrogate.
Nora · 28:10Let's walk through exactly how this works under the hood, because it touches on all the technical hurdles we've discussed. The first phase is ingestion.
Miles · 28:17Getting all that messy human-formatted data in.
Nora · 28:20Right. And it handles everything. Messy email files, sprawling Excel sheets, legacy PDFs, raw voice notes. It pulls from cloud drives and even has mobile hardware scanners for physical paper.
Miles · 28:32But ingestion is useless without the second step, reading and embedding. This is where we solve Karpathy's problem with human-centric formats.
Nora · 28:41If you upload a photograph of a crumpled signed vendor contract, DocuStrata doesn't just save the image file.
Miles · 28:47Well, it runs optical extraction, pulling the raw text out of the pixels. It runs transcription on the audio. It converts the messy DOM of a web clipping into clean markdown.
Nora · 28:57It normalizes everything into a machine-readable format.
Miles · 28:59map. And then it takes all of that extracted text and embeds it into a vector index.
Nora · 29:04Okay, we need to spend some time explaining vector indexing because this is the absolute secret sauce of how these AI systems actually remember 100,000 documents without having to keep all of them in active memory at once. Which would blow up that attention mechanism we talked about. Exactly. How does a vector index actually map meaning?
Miles · 29:22It requires a shift in how we think about search. A traditional search engine like what we've used for the last 20 years relies on lexical search.
Nora · 29:32Exact keyword matching. Right.
Miles · 29:34If you search the word canine, it looks for documents that contain the exact string of characters, C-A-N-I-N-E.
Nora · 29:41If the document says dog or puppy, traditional search misses it completely unless the author specifically tagged it.
Miles · 29:47Exactly. But a vector database doesn't look at character strings. It maps semantic meaning. When DocuStrata ingests a document, it chops the text into chunks. chunks. It passes each chunk through an embedding model. That model translates the semantic meaning of that text into a dense array of numbers, a vector representing its location in a massive, multidimensional, latent space.
Nora · 30:10We're talking 1500 dimensions of meaning.
Miles · 30:12So it's not just a 3D grid. It's a 1500 dimensional space where concepts are plotted based on their relationship to everything else the model knows.
Nora · 30:22So the concept of dog gets assigned a specific mathematical coordinate in this high dimensional space.
Miles · 30:28Right. The concept of puppy gets plotted extremely close to it because their vectors are mathematically similar. The phrase canine companion is also in that immediate neighborhood.
Nora · 30:38Even a document that never uses the word dog but talks about taking my golden retriever to the vet for his rabies shot gets plotted in that same semantic cluster.
Miles · 30:47Exactly. So when I go to DocuStrata and I ask a question, I'm not searching for keywords. I'm searching for concepts.
Nora · 30:53Yes. When you ask, what was the name of the contractor who fixed the roof in 2019, and what was the warranty period? The system embeds your question into a vector.
Miles · 31:03It places your question in that 1500-dimensional space.
Nora · 31:07Then, using a mathematical calculation called cosine similarity, it instantly measures the distance between your questions vector and the vectors of all 100,000 chunks in your archive.
Miles · 31:17It retrieves the chunks that are mathematically closest in meaning, even if they don't share the exact keywords of your prompt.
Nora · 31:23It finds the exact PDF invoice from 2019 and the email thread discussing the warranty instantly.
Miles · 31:30And then we hit the third step, interrogation. This is the answer engine at work.
Nora · 31:34The system takes those retrieved chunks, drops them into the LLM's context window, synthesizes a plain language answer, and provides exact citations back to the SORF documents.
Miles · 31:44It completely inverts the software product category. I mean, for decades, tools like Evernote, Notion, Apple Notes, they were all built on the unstated assumption that you are the reader.
Nora · 31:55Their entire value proposition was helping you organize your data, giving you folders and tags and nested hierarchies so you could find things and read them yourself later.
Miles · 32:04But organizing is dead labor. Tagging files is dead labor. If the reader is a machine that can instantly traverse a 1500 dimensional vector space, you don't need folders.
Nora · 32:15The value of the software migrates entirely from the act of organizing to the act of answering.
Miles · 32:19And this dictates DocuStrata's business model, which is the purest, most direct expression of Jensen Huang's AI factory economics that we've seen.
Nora · 32:29The fact sheet lays out their core operating principle. Subsidize ingestion, monetize interrogation. Break down how that works in practice because it's the exact opposite of how cloud storage works today.
Miles · 32:42Think about Dropbox or Google Drive. They monetize ingestion. They give you a little bit of storage for free. But if you want to store 100 gigabytes of data, you pay a monthly fee. They are charging you for the digital real estate.
Nora · 32:54DocuStrata flips that. They don't care how many gigabytes you store. Putting your documents into the system, the storage, the vector indexing, the optical extraction is incredibly cheap. They practically give it away.
Miles · 33:03Because storing static data on a hard drive is a solved, commoditized problem that costs fractions of a cent.
Nora · 33:09Exactly. The metered, valuable unit, the thing you actually pay for on a variable basis, is the question.
Miles · 33:15You pay per interrogation. Because asking a complex question requires the AI to perform synthesis. It requires tokens.
Nora · 33:22It requires the GPUs in the AI factory to spin up, consume electricity, perform matrix multiplication, and manufacture intelligence on demand.
Miles · 33:31If storing data is cheap and reading is what machines do natively, then the act of asking a question and receiving a verified synthesis is where the live extractable value lives.
Nora · 33:42They aren't selling you a digital warehouse. They're selling you a metered research assistant.
Miles · 33:47And when you look at the personas this applies to, it's vast. The academic researcher who has 600 PDFs in the reference manager, but has only deeply read 40 of them.
Nora · 33:57The corporate lawyer who needs to find a specific indemnification clause buried in 10 years of varied vendor contracts.
Miles · 34:04Or even just an estate executor who is handed a literal cardboard box of a deceased relative's paperwork and has to figure out where the life insurance policy is.
Nora · 34:12In all of those scenarios, the core problem is identical. The raw volume of unstructured information vastly exceeds the human capacity to process it within the necessary time frame.
Miles · 34:22DocuStrata is the sharp point of the spear. It is the convergence of all our witnesses pointing at that specific human limitation.
Nora · 34:29Okay, this all sounds incredibly compelling. A utopian vision of frictionless knowledge retrieval. I upload my messy, chaotic life into the machine. The machine reads it perfectly. I ask natural language questions, and I get pristine, sighted answers. But if we are delegating the entirety of our reading, processing, and synthesis to probabilistic language models, what is the failure mode? What happens when this architecture quietly breaks?
Knowledge base poisoning: when AI writes into your archive
Miles · 34:56And that takes us to the most critical part of this entire deep dive, because the failure mode of delegating reading to an AI is not a loud, catastrophic crash. You don't get a blue screen of death.
Nora · 35:06The failure mode is a slow, silent poisoning of your own reality.
Miles · 35:11We have this incredible critical essay from Anand Lahoti titled The Hidden Flaw in Karpathy's LLM Wiki. And he points out a structural danger in this machine reading pattern that almost no one in the hype cycle is talking about.
Nora · 35:23Let's go back to Karpathy's vision of a personal knowledge base for a moment. Karpathy advocates for what system architects call write-time synthesis.
Miles · 35:31We need to define the technical difference between write-time and query-time synthesis because the entire risk profile hinges on this difference.
Nora · 35:39In a system using write-time synthesis, the LLM doesn't just read your documents, it actively writes permanent summaries of them.
Miles · 35:45When you upload a raw document, the LLM ingests it, instantly writes a clean formatted summary, extracts key concepts, creates links between ideas, and saves all of this newly generated text permanently into your wiki. So your wiki isn't just a
Nora · 36:01database of your original files anymore. It becomes a hybrid mix of your original raw documents and the LLM summarized interpretations of those documents all
Miles · 36:10sitting side by side. It's like having a librarian who doesn't just catalog the new books coming into the library. The librarian reads every new book, writes writes their own heavily condensed summary book about it, and puts that newly authored summary book right on the shelf next to the original text.
Nora · 36:25And because the LLM summaries are so clean, so perfectly formatted and marked down, and so dense with internal links, when you ask a question later, the vector retrieval system is highly likely to grab the LLM summary instead of the messy original document.
Miles · 36:41It's mathematically more attracted to the search algorithm.
Nora · 36:44Which sounds efficient on the surface. What's the problem?
Miles · 36:46The problem is what Lahoti calls knowledge-based poisoning. Let's use his brilliant concrete example to illustrate exactly how insidious this is. Let's say you are a business owner and you upload a raw vendor contract into your system.
Nora · 37:00The original legally binding contract specifically states, payment terms are net 30 with a 2% discount if paid within 10 days.
Miles · 37:08A very specific, highly factual clause with distinct financial metrics.
Nora · 37:12The LLM ingests that contract at write time. It decides to update a master concept article in your wiki called Vendor Payment Terms.
Miles · 37:20It reads the clause, compresses the information to save space, and writes into the wiki. Standard agreements use net 30 terms with early payment discounts.
Nora · 37:29It dropped the specifics. It summarized a little too aggressively.
Miles · 37:32Exactly. It's a reasonable, natural-sounding summary. It isn't overtly hallucinating fake numbers, but it is lossy compression. The specific 2% figure and the 10-day window are permanently stripped from the summary document.
Nora · 37:47Now at this point, the original PDF contract is still in the system. The ground truth hasn't been deleted.
Miles · 37:52But the summary is what the Vector Index is going to retrieve.
Nora · 37:55Yes. Six months later, you are doing financial planning and you ask your system, what is our typical early payment discount for vendors?
Miles · 38:03The Vector Index searches your knowledge base. It finds the vendor payment terms article first because it's the central hub of information. It's heavily linked, and it perfectly matches the semantic intent of your query.
Nora · 38:14The LLM reads that article to generate the answer to your question.
Miles · 38:18But the article doesn't contain the 2% number.
Nora · 38:20Precisely. So the LLM is stuck. It reads early payment discounts, but it has no specific metric. So what does a probabilistic prediction engine do when it lacks specific facts, but has been fine-tuned to be relentlessly helpful?
Miles · 38:33It hedges. Or worse, it interpolates.
Nora · 38:35Exactly. It might say standard discounts vary, but often range around 5%, pulling from its general pre-training data across the Internet rather than your specific documents. Or it might just confidently hallucinate a number that sounds plausible in the context of the sentence.
Miles · 38:52And here is where the architecture becomes terrifying. Let's say you run an automated AI health check on your wiki to make sure everything is internally consistent and accurate.
Nora · 39:02The health check will pass with flying colors because the LLM might have written another summary document last month about financial obligations, and it based that summary on the flawed vendor payment terms article.
Miles · 39:13So now you have multiple LLM authored articles in your database that all perfectly agree with each other. They are internally mathematically consistent.
Nora · 39:22They just happen to be consistently wrong. They are consistently completely unmoored from the actual underlying ground truth of the original signed contract.
Miles · 39:30And wait, I can hear people listening to this thinking, if the AI wrote the summary, shouldn't it be highly accurate? Why would it drop the 2% discount number?
Nora · 39:38And it comes back to how these models work. They don't know what is legally important. They optimize for linguistic fluidity and brevity. The 2% might not have seemed statistically important to the summarization prompt at that exact moment.
Miles · 39:52And how do you even know your database is poisoned? There is no moment where something something visibly breaks. There's no 404 error message. The wiki just becomes increasingly
Nora · 40:02self-referential over time. Future LLM responses reason on top of prior LLM summaries. The chain of custody back to the original source quietly frays and then breaks entirely. You are living in a hallucinated consensus. It's an epistemic closed loop, a game of telephone
Miles · 40:19played by an AI with itself over months and years, slowly drifting further and further further from reality. And because the answers always sound confident, well-structured, and syntactically flawless, you never suspect a thing.
Nora · 40:30Lahoti has this chilling sentence in the essay. He says, this is the difference between a system that can make mistakes and a system that launders mistakes into truth.
Miles · 40:38It's profound. So what is the architectural fix? If right-time synthesis is a slow-acting poison for a large-scale knowledge base, what is the alternative design pattern?
Nora · 40:48The alternative is strict query-time synthesis.
Miles · 40:51Exactly. In a query time architecture, the original documents are treated as sacred, immutable ground truth. The LLM is never, ever allowed to write a new summary document that gets saved alongside the originals. The librarian is strictly forbidden from writing new books.
Nora · 41:08But then how does the vector index map everything if it's not summarizing the concepts?
Miles · 41:12At the moment of ingestion, the LLM is only allowed to extract structured metadata. data. It can identify named entities, it can map relationships, can pull out keywords, and attach all of that as invisible structural tags to the original document. It creates a robust index card,
Nora · 41:27not a summary book. The extracted structure is merely a navigation aid for the vector search. So when I asked the question about the discount six months later... When you ask the question, the system uses that structural metadata to find the original raw contract PDF, and then it it performs the synthesis fresh, right then and there at query time, reading the actual
Miles · 41:46untouched clause that says 2% discount. It answers your question based on the raw ground truth every single time and discards the synthesis as soon as you close the chat.
Answer memory and the unresolved tension
Nora · 41:57Okay, this brings us to a massive unresolved tension regarding DocuStrata because when we look closely at their product fact sheet, they are walking on a razor's edge between these two
Miles · 42:06paradigms. They absolutely are. DocuStrata's core retrieval loop is designed around query time synthesis. You ask a question, it retrieves the original chunks and answers fresh with inline citations. That is safe. That preserves the epistemic chain of custody. But they have this
Nora · 42:22heavily advertised feature called answer memory. The fact sheet proudly states that when the system produces a complex, well-granted answer to a query, that answer is automatically filed back into the archive as a first-class searchable document. Meaning it is embedded and indexed
Miles · 42:37right alongside the raw originals.
Nora · 42:39Which is right-time synthesis sneaking in through the back door.
Miles · 42:43Exactly. If I ask a highly complex question and it spends compute synthesizing a brilliant answer from 20 different documents, it saves that answer as a new document to save compute next time.
Nora · 42:55And the next time I ask a related question, the vector index is going to retrieve that saved answer instead of the 20 originals.
Miles · 43:02So we have to ask the critical architectural question. question. Does this answer memory create a compounding asset where the system gets genuinely smarter, faster, and cheaper over time because it remembers the hard cognitive work it already
Nora · 43:15did? Or does it slowly poison the knowledge base with AI self-reference, exactly like
Miles · 43:20Lahoti warns? If the AI hallucinates even slightly in that first complex answer or drops a nuance and that flawed answer gets saved as a first-class document, you have just injected
Nora · 43:29poison directly into your own ground truth. It's an unresolved tension in the tech world right now. How do you balance the speed and compounding knowledge of right-time synthesis with the absolute factual rigor of query time synthesis? And honestly, the industry doesn't
Miles · 43:46have the definitive answer yet. It is the defining architectural debate for the next generation of
Nora · 43:52enterprise knowledge management. If you are building a personal wiki with 100 articles, maybe the right time risk is low enough because you can manually spot check it. But if you are deploying this across a global law firm with 10 million contracts, that error compounding is an existential liability.
Miles · 44:08OK, we have covered a massive amount of architectural, economic and technical ground today. Let's recap the scale of the inversion we've been tracking.
What we lose when we stop reading
Nora · 44:17For 30 years, the entire software industry from Microsoft Office to Google Search to traditional CRMs was built to help you read and organize. The human was the primary processor. The biological interface was the bottleneck.
Miles · 44:29And now, driven by the sheer scaling laws of LLMs, the falling compute cost of token generation, and the inevitable collapse of traditional user interfaces, we are entering an era where the digital infrastructure is built to read for you.
Nora · 44:42Your only operational job is to formulate the query.
Miles · 44:45But let's bring this back to you, the listener, for a moment. Delegating retrieval and reading to a machine is an undeniable massive time saver. No one actually wants to spend three hours hunting for a specific indemnification clause in a folder of PDFs.
Nora · 45:00But I want us to think deeply about what is lost in this transition.
Miles · 45:03There is a physical cognitive friction to the act of reading. When you have to sit down and read ten research papers just to find one specific answer, you don't just find the answer.
Nora · 45:14Along the way, you absorb ambient context. You encounter tangential ideas you weren't looking for. You begin to understand the nuance of the author's underlying argument.
Miles · 45:23You unconsciously build a mental map of the entire domain.
Nora · 45:26By delegating the reading to DocuStrata or to an orchestrating AI agent, you get the answer instantly. But you bypass the journey. You get the destination without having to traverse the terrain.
Miles · 45:37Are we losing deep comprehension in the blind corporate pursuit of operational speed? Knowledge isn't just a database of facts you can query on demand. It's a web of interconnected understanding held within your own mind.
Nora · 45:49If the machine reads everything and you read nothing, do you actually know anything? Or are you just an API endpoint for an external intelligence?
Miles · 45:58Which leaves us with a final provocative thought to mull over as we watch this infrastructure shift beneath our feet.
Nora · 46:04If the machine reads everything and your only remaining job in the knowledge economy is to ask a good question, question, what happens if over time, by not reading, we lose the context, the nuance, and the vocabulary required to even know what questions to ask in the first place?
Miles · 46:21If you don't know the landscape, you don't know where to look. And eventually, you don't even know what you're looking for.
Nora · 46:25A sobering thought for the agentic era. Keep asking questions.