Why machines stopped needing your interface
Miles · 0:00Right now, I mean, the whole internet is basically built for eyeballs.
Nora · 0:02Human eyeballs, specifically.
Miles · 0:04Exactly. Like, think about the last webpage you visited. They probably had, you know, colorful buttons, maybe some drop-down menus.
Nora · 0:10Carefully chosen typography.
Miles · 0:12Right. Typography, layout, all of it completely optimized so your human vision would just naturally track toward that shiny buy button.
Nora · 0:21Or the annoying newsletter pop-up.
Miles · 0:24Yeah, exactly. Exactly. We have literally spent the last 30 years optimizing the digital world for human psychology. But according to Andrej Karpathy, who is a founding member of OpenAI and the former director of AI at Tesla.
Nora · 0:38A pretty serious guy to be saying this.
Miles · 0:39Yeah, very serious. He's saying all of that human-centric design is about to become completely obsolete.
Nora · 0:45It's wild to think about.
Miles · 0:46He actually stated that 99.9% of attention on the Internet is about to be LLM attention, not human attention.
Nora · 0:53So we're the rounding error now.
Miles · 0:55We are the rounding error. And the human reader is basically just a legacy feature on the network at this point.
Nora · 1:00Which, I mean, it totally reframes how we even think about consuming information, right?
Miles · 1:05How so?
Nora · 1:06Well, if an artificial intelligence agent is the entity reading a software documentation page or, I don't know, a recipe blog, it fundamentally does not care about your cascading style sheets.
Miles · 1:17No, it definitely doesn't.
Nora · 1:18It doesn't care about your pretty layout or your bold text. In fact, all that visual formatting actively gets in the way.
Miles · 1:25It's just noise to the machine.
Nora · 1:26Exactly. It wastes the AI's processing power, what we call tokens in this space.
Miles · 1:31Which brings us to the actual stack of sources we are diving into today. Welcome to a deep dive into what is arguably the sharpest, most aggressive technical debate happening in artificial intelligence right now.
Nora · 1:44It is a massive rift in the industry.
Miles · 1:46Yeah. The era of humans reading documents is ending. That much is pretty much a given at this point. But the engineers and the researchers building the replacement are just fiercely divided on exactly how machines should read for us.
Nora · 1:59The architecture is completely up in the air.
Miles · 2:01Right. So today we are looking at keynotes from tech executives, specifically Sam Altman and Satya Nadella. We've got a really fascinating breakdown of Andrej Karpathy's latest personal AI workflows from Deere.ai.
Nora · 2:15That one is so interesting.
Miles · 2:16It really is. Plus, we have a very sharp, critical developer essay from Anand Lahoti. And to kind of anchor all of this theory in reality, we have the product fact sheet for a document intelligence tool called DocuStrata.
Nora · 2:30Which sits squarely in the crosshairs of this entire debate.
Miles · 2:33Perfectly in the crosshairs. So the mission for this deep dive is to explore two massive unresolved debates. First, is current AI retrieval just a temporary crutch? And second, should AI write our knowledge bases or just search them?
Nora · 2:46And the stakes here, I mean, they are structural. The architecture the tech industry settles on right now is going to dictate how you, the listener, interact with your own recorded knowledge for the rest of your life.
Miles · 2:55Seriously. To really make you feel the weight of this, let's look at a prediction from Microsoft CEO Satya Nadella.
Nora · 3:01Oh, the Excel comment.
Nadella's Excel comment: apps as a database in a costume
Miles · 3:02Yeah, the Excel comment. He is looking at the standard model of software, you know, software as a service, SOSABs, and predicting its absolute collapse. He used Microsoft Excel as the primary example. example, literally asking an audience, hey, why do I need Excel?
Nora · 3:19Which is an insane thing to say coming from the executive whose company basically prints money via Microsoft Office subscriptions. Right.
Miles · 3:26Predicting the death of the traditional application interface is a massive provocation. But his underlying logic points to exactly what Karpathy is talking about.
Nora · 3:35Nadella is just looking at traditional business applications and stripping them down to their mathematical reality.
Miles · 3:40What do you mean by that?
Nora · 3:41Well, most software, whether it's a CRM system for sales or an accounting app, is really just a database, specifically a CR-viewed database.
Miles · 3:50Let's break that acronym down for the listener just because it's so vital to understanding why the interface is dying. C-R-L-U-D. Create, Read, Update, Delete.
Nora · 3:59Right. Every single action you take in a software application boils down to one of those four operations.
Miles · 4:04Like adding a new client.
Nora · 4:06Exactly. You create a new client profile. You read their past purchase history. You update their phone number or you delete a duplicate entry.
Miles · 4:14And for decades, we've needed a graphical user interface, a UI, so a human could safely do that.
Nora · 4:21Yeah. So you could execute those four commands without accidentally destroying the underlying database. The UI is basically just a padded safety wrapper with pretty buttons.
Miles · 4:29But Nadella's point is that in this disruptive agentic AI era we're entering, the AI agent becomes the user.
Nora · 4:36Not you.
Miles · 4:37Not you. If I can just say to my AI, hey, find all the clients who bought our software in 2023 and email them a 10% discount code, the AI doesn't need to open Excel.
Nora · 4:47It doesn't need to open Salesforce or whatever CRM either.
Miles · 4:50Right. It doesn't need to click buttons. It just reaches directly into the raw backend database, executes the read command to find the clients, executes the update command to log that it did it, and sends the emails. The app UI is entirely dead.
Nora · 5:02The human interface just recedes into a simple chat box or even just a voice prompt.
Miles · 5:07You no longer open an application to do the work.
Nora · 5:09You ask the agent to do the work and the agent interacts with the raw data.
Miles · 5:13And think about what Karpathy advocates for to make this reality happen on the web. He is actively pushing for new web standards, specifically this protocol called LLM.txt.
Nora · 5:25That is such a perfect example of preparing the world for this 99.9% LLM attention economy.
Miles · 5:31Yeah, for anyone familiar with web development, you know that the default page of a website is usually called index.html.
Nora · 5:37Right, which contains all the visual instructions for a web browser.
Miles · 5:40All the stuff for a human eye. But Karpathy is suggesting that websites should start hosting an lms.txt file at their root directory instead.
Nora · 5:48Or alongside it, yeah.
Miles · 5:49So instead of pointing a visual browser to a graphical page, it points an AI agent to a clean, highly condensed, machine-readable, markdown version of the exact same info.
Nora · 5:59No menus, no graphics, just pure structured text.
Miles · 6:02It is a complete inversion of 30 years of digital design. And it extends way beyond web pages, you know. PDFs, images, video files, complex legal contracts. All of these formats were intended for uniquely human consumption.
Nora · 6:14And right now, they're basically locked up.
Miles · 6:16Right. So the premise of our deep dive is pretty much set. The machine reads everything. The machine does the clicking. The machine knows your data.
Nora · 6:23But the conflict, and it's a huge conflict, arises when we ask how it actually does that under the hood.
Miles · 6:29Yeah.
Nora · 6:29I mean, if you have 100,000 files in your Google Drive, how does the machine actually read them?
Miles · 6:34Which brings us to the format for today. We're going to take opposing sides on exactly how this read-nothing-know-everything future works.
Nora · 6:43We are not backing down on this either.
Miles · 6:45Definitely not. I am going to defend the visionary, massive-scale approach championed by executives like Sam Altman.
Nora · 6:52And I'll be taking the pragmatic architecture focused side. I'm arguing for the durability of current retrieval systems.
Miles · 6:59The tech industry is literally split down the middle on this, and we're going to debate the two core unresolved questions. First, is current AI retrieval a permanent bridge or a temporary dead end?
Nora · 7:10And second, should AI author our knowledge bases or only ever search them?
Miles · 7:16Let's dive straight into debate one, the architecture of reading at scale.
Nora · 7:20Let's do it.
Miles · 7:21To ground this for you, the listener, I want you to imagine your own digital footprint, like a lifetime of emails.
Nora · 7:28Decades of tax returns.
The archive you already have
Miles · 7:29Yeah, tax returns, contracts, project notes, maybe research papers from college. Let's just call it around 100,000 documents. Yeah. If you want an AI to actually comprehend all of that, you run into an immediate physical limitation today, which is the context window.
Nora · 7:45The dreaded context window.
Miles · 7:47Right. It's basically the amount of working memory the AI has for any single conversation.
Nora · 7:52I always think of an LLM as this brilliant savant who wakes up with severe amnesia every single time you talk to it.
Miles · 8:00That's a great way to put it.
Nora · 8:01So the context window is just the stack of papers you are allowed to hand it to read before you ask your question.
Miles · 8:06And right now you cannot hand it 100,000 documents. The context window is just way too small. Even the most advanced models today can handle, what, maybe one or two million tokens?
Nora · 8:15Which might equate to a few thousand pages, yeah.
Miles · 8:18So if you try to paste your whole life in there, the system either crashes or just flat out refuses the prompt.
Nora · 8:24And this specific limitation is what gave birth to the current industry standard architecture, a technique called RAG.
Miles · 8:32Retrieval Augmented Generation.
Nora · 8:33Exactly.
Miles · 8:34Now, I'm going to make my stance on RAG incredibly clear right up front. I think RAG is a dead end.
Nora · 8:41Wow, okay.
Miles · 8:41It is. It is elaborate, clunky scaffolding built around a purely temporary hardware limitation. The moment context windows get big enough and compute gets cheap enough, Reg A vanishes overnight.
Nora · 8:55I completely disagree, but okay.
Miles · 8:56Well, before I tear it down, why don't you explain how the scaffolding actually works for the listener?
Nora · 9:00Gladly. The architecture of Reg A is actually quite elegant in its efficiency because you can't feed the AI the entire 100,000 document archive all at once. You have to pre-process it.
Miles · 9:11Right, you chop it up.
Nora · 9:11You take all those documents and you literally chop them up into small manageable chunks, maybe a paragraph or two each. Then you run those chunks through what is called an embedding model.
Miles · 9:21Okay, this is where we need to explain the math because embedding model sounds like absolute magic, but it's really just geometry.
Nora · 9:26Precisely. It's just mapping. An embedding model takes human text and translates it into a string of numbers. That string of numbers represents a specific coordinate in a massive multidimensional space.
Miles · 9:38So think of a standard graph from high school algebra, right, with an X and Y axis.
Nora · 9:43Exactly. That's two dimensions. Now imagine a space with 1,536 dimensions.
Miles · 9:49Which is literally impossible for the human brain to picture.
Nora · 9:52Completely impossible to visualize, but mathematically trivial for a computer to navigate. The embedding model places every chunk of your text into this multidimensional space based on its semantic meaning.
Miles · 10:03Its underlying concept.
Nora · 10:05Right. So a paragraph about a corporate financial audit and a paragraph about tax liabilities will be placed very close to each other in this space, even if they don't use the exact same words.
Miles · 10:13Because conceptually, they are neighbors.
Nora · 10:15Exactly. Meanwhile, a paragraph about your dog's veterinary appointment will be placed thousands of miles away on the total opposite side of the space.
Miles · 10:24Okay, so you have mapped your entire life's archive into this giant mathematical constellation. What happens when I actually ask the AI a question?
Nora · 10:32Let's say you ask, what were my tax liabilities in 2022? The system takes your question, runs it through that exact same embedding model, and turns your question into a coordinate.
Miles · 10:44It drops the question into the space.
Nora · 10:46Right. Then it simply looks for the chunks of text that are physically closest to your question in that mathematical space. It retrieves those specific paragraphs and only those paragraphs.
Miles · 10:57It literally extracts the closest mathematical neighbors.
Nora · 11:00Correct. That's the retrieval part. Then it takes your original question, bundles it with just those three or four retrieve paragraphs, and drops that tiny little package into the LLM's context window.
Miles · 11:11And then the AI reads those specific snippets and generates your answer.
Nora · 11:15Retrieval augmented generation.
Miles · 11:16Okay, I admit, it is a brilliant workaround for a tiny context window. But it is still just a workaround. Let's look at what Sam Altman described live on stage at the Sequoia Capital AI Ascent event.
Nora · 11:28The platonic ideal speech.
Miles · 11:29Yes. He was asked directly about the future of context windows, and he laid out his ideal state for how AI should function. He envisions a highly capable reasoning model paired with a context window of a trillion tokens.
Nora · 11:44A trillion tokens is several orders of magnitude larger than anything currently in production.
Miles · 11:50It's a massive leap. But in Altman's ideal state, you do not chop your life into little chunks. You do not run them through an embedding model. You don't map them into a 1,536-dimensional space.
Nora · 12:03You just pour it all in.
Miles · 12:04You just pour your entire life into the model's active memory. Every email, every book you read, every single conversation you've ever had, it all just continuously appends to the context window.
Nora · 12:14So it just knows everything.
Miles · 12:15The model never has to retrieve anything because everything is always in its active working memory.
Nora · 12:19I mean, it is a compelling science fiction vision. I'll give you that.
Miles · 12:22It is the inevitable trajectory of the hardware. Hi. Think of your archive like a massive physical library. RAG is essentially a slow, clunky card catalog.
Nora · 12:32A very fast mathematical card catalog.
Past the card catalog: what vector search changes
Miles · 12:34But still a card catalog. You walk into the library to find an answer, but your brain, the context window, is too small to hold all the books. So you have to go to the catalog, you look up the mathematical embedding, you write down the coordinates, you walk deep into the stacks, you pull off three specific books, and you read them to answer your question.
Nora · 12:50It is systematic, it's targeted, and it's highly effective.
Miles · 12:54It is tedious. And honestly, it relies on the embedding model being absolutely perfect. What if the answer to your question was split across a book in the biology section and a book in the history section, and the math just didn't pull them both?
Nora · 13:08Then you get a partial answer, sure.
Miles · 13:09REG breaks in those edge cases. Altman's trillion token ideal is like walking into the library and instantly downloading the text of every single book directly into your cerebral cortex.
Nora · 13:20Neo in the Matrix style.
Miles · 13:22Exactly. You don't need a card catalog anymore. You don't need to search the stacks. You burn the card catalog. Once the context window is functionally infinite, RAG is completely unnecessary. I truly believe we will look back on vector embeddings the way we look back on dial-up modems.
Nora · 13:38Dial-up modems, really?
Miles · 13:39Yeah. It's just a weird clunky noise we had to tolerate to get on the internet before broadband existed.
The economics nobody mentions: what reading 100,000 documents costs
Nora · 13:44Okay, the analogy is vivid, but it willfully ignores the brutal economic and physical reality of how artificial intelligence actually functions at scale. How so? I am aggressively taking the opposing side here. Our RAG is not a dead end. It is the clever, durable architecture that makes read everything actually work, not just today, but permanently.
Miles · 14:07But if the compute catches up, why keep the card catalog?
Nora · 14:11Because of the marginal cost of cognition. Look at what Jensen Huang, the CEO of Nvidia, has been saying.
Miles · 14:16The guy building the actual chips.
Nora · 14:18Exactly. Nvidia builds the physical chips that make all of this possible. Huang frames modern data centers as AI factories, and he explicitly states that tokens are the new currency.
Miles · 14:29Right.
Nora · 14:29Generating tokens and processing tokens in a massive context window requires immense computational power. It requires electricity. It requires cooling water. It requires physical silicon.
Miles · 14:40But compute always gets cheaper over time. I mean, Moore's Law is held for decades, right? The chips get smaller. The processing gets faster. The cost approaches zero.
Nora · 14:50Approaching zero is not the same as being free. Altman himself admitted at that very same Sequoia event that compute cost is the number two complaint he hears from developers.
Miles · 15:01OK, fair.
Nora · 15:01Let's look at the physics of your library analogy. Let's say you could instantly download the entire library into your brain every single time someone asks you a question. Doing so burns a massive amount of metabolic energy.
Miles · 15:13So we're talking about energy efficiency.
Nora · 15:15We are talking about basic economics. Let's say you have that 100,000 document personal archive. Pushing all 100,000 documents through the neural network of an LLM every single time you ask a simple query like, what did I say to Sarah about the marketing budget last Tuesday, is an astronomical waste of compute.
Miles · 15:33You're passing the whole library for one email.
Nora · 15:35You are asking the model to process the entire works of Shakespeare, your 10 years of tax returns, and every spam email you ever received, just to find out what you said to Sarah last week.
Miles · 15:45The argument, though, is that eventually, processing that much data will cost fractions of a penny. So the inefficiency just won't matter.
Nora · 15:53That assumes the demand for compute stays static while the supply increases, but demand is scaling exponentially alongside it. No matter how cheap reading becomes, physics dictates that reading three relevant pages will always cost exponentially less than reading 100,000 pages.
Miles · 16:10So why burn the energy of a small city to process the whole library when a mathematical index could just instantly point you to page 42 of one specific book?
Nora · 16:20Exactly. AG is an efficiency mechanism, and efficiency never becomes obsolete.
Miles · 16:25Okay, I see the economic argument. If you are running an enterprise software company with, say, millions of users, doing full context processing for every single query would bankrupt you in server costs overnight.
Nora · 16:36It would vaporize your profit margins instantly. But, you know, my defense of ROG isn't just about money. It is about a fundamental requirement for the future of information.
Miles · 16:45Which is?
Nora · 16:46Grounding.
Miles · 16:46Grounding, as in ensuring the AI isn't hallucinating or fabricating facts.
Nora · 16:51Exactly. Consider Aravind Srinivas, the CEO of Perplexity. They are currently leading the charge in AI-driven search.
Miles · 16:59Right. They're huge right now.
Nora · 17:01Look at the core design commitments of enterprise tools like DocuStrata, too. They are absolutely militant about one feature. Citations.
Miles · 17:11Showing their work.
Nora · 17:12R-Agg, by its very architectural nature, forces the system to pull exact source documents.
Miles · 17:18Because the model only ever sees the three paragraphs the embedding search hands it, it physically can't pull from anywhere else.
Nora · 17:24Yes. So when the AI generates an answer, it can point precisely to the snippet it was provided and say, I derived this fact from this specific email on this specific date. It creates a verifiable chain of custody.
Miles · 17:36And the trillion token model breaks that chain.
Nora · 17:38It completely obliterates it. If you rely on a massive context context window without retrieval, if you just pour your entire life into a trillion token blender, the AI turns into an opaque black box.
Miles · 17:49It's just stirring all the concepts together.
Nora · 17:51It synthesizes across so much data simultaneously that you cannot easily verify where a specific thought came from. Like, did it get that 5% growth projection from the finalized contract signed yesterday or from some rough graft email you sent two years ago?
Miles · 18:06The model just knows it. Sort of the same way a human remembers a fact without necessarily remembering the exact book they read it in.
Nora · 18:14Right. But we don't trust human memory for legal contracts. In an enterprise setting or a medical setting or a legal setting, delegating the reading to an AI only works if you can verify the work. ROG preserves the citations. It forces the machine to show its work.
Miles · 18:30So your position is that ROG is a permanent architectural feature of trustworthy AI. It is not scaffolding. It is a load-bearing pillar.
Nora · 18:38It is the only way to keep the compute costs sane and the absolute only way to keep the outputs verifiable.
Miles · 18:43Okay. I will concede that the compute argument is mathematically sound today.
Nora · 18:46Yeah.
Miles · 18:47And the need for citations is obviously critical for enterprise trust.
Nora · 18:51Okay.
Miles · 18:51What if forcing the AI to read raw documents at all, whether through RAG or a trillion token window, is just the wrong approach entirely?
Nora · 19:01entirely. What is the alternative to reading the documents? Predigestion. Oh boy. Hear me out.
Miles · 19:07What if the AI just rewrites your entire archive perfectly beforehand? What if it synthesizes all the information into a persistent knowledge base so that when you ask a question, the AI doesn't have to scan raw, messy emails at all. It just reads its own perfectly structured summaries.
Nora · 19:25We are moving right into the debate over write time versus query time synthesis.
Miles · 19:29This is debate two, and honestly, it gets to the very heart of how we construct a source of truth.
Nora · 19:34It really does.
Right-time versus query-time synthesis
Miles · 19:35The core argument here is about when the AI does the heavy cognitive labor of understanding your data. Should an AI compile all your raw documents into persistent AI-written summaries? Or should it synthesize answers completely fresh, reading directly from the untouched original documents every single time you query it?
Nora · 19:52It is a debate about the permanence of AI thought.
Miles · 19:54And I am passionately taking the side of right-time synthesis here. To back this up, we really have to look at the dayer.ai breakdown of Andrej Karpathy's personal knowledge-based system. Let's get into it. This is an absolutely brilliant workflow. Karpathy built a 100-article, 400,000-word personal wiki covering deeply complex technical topics. And he didn't write any of it himself.
Nora · 20:19Not a single word.
Miles · 20:20The AI authored the entire wiki.
Nora · 20:22Which, I mean, 400,000 words is a massive amount of text to let a machine author autonomously. That is the length of like four standard novels presented as factual truth.
Miles · 20:32But the architecture behind it is incredibly elegant. It operates in four continuous phases, and it actually solves the exact compute problem you were just complaining about. Let me walk you through how it actually works.
Nora · 20:41Proceed. I am highly skeptical of autonomous AI authoring at this scale, but let's break down the mechanics.
Miles · 20:46All right, so phase one is ingest. Karpathy takes raw data from all over the internet. He uses an Obsidian web clipper to grab web articles. He downloads massive, dense PDFs of academic research papers from RGIF.
Nora · 20:58Code repositories from GitHub.
Miles · 20:59Yeah, code from GitHub. He just dumps all of this into a raw staging directory. It is a messy, unstructured pile of source documents.
Nora · 21:08Basically a standard data lake concept. You gather all the unstructured data into one place before processing. Right.
How an LLM wiki compiles itself
Miles · 21:13But phase two is where the magic, the right time synthesis happens. This is the compile phase. Okay. This is not RG. This system does not wait for a user to ask a question to do the work. Instead, the LLM systematically reads through all those raw files in the staging directory and incrementally builds a structured wiki.
Nora · 21:32But how exactly does an LLM build a wiki incrementally? It's a text predictor, you know, not a software engineer.
Miles · 21:38It is orchestrated through complex scripting. The LLM is prompted to read a raw research paper and write a concise index file summarizing its core claims.
Nora · 21:46So it summarizes first.
Miles · 21:47Then it is prompted to look at all its summaries and write overarching concept articles. Imagine damply researched Wikipedia-style pages on specific topics like transformer architecture or loss functions. But here's the most important mechanical detail. The LLM auto-generates a complex graph of backlinks.
Nora · 22:06Explain the mechanism of the backlinks for the listener. How does it know what to connect?
Miles · 22:10So as the LLM is writing a new concept article, the orchestration script feeds it a list of all the other articles that already exist in the wiki.
Nora · 22:19Ah, I see.
Miles · 22:20When the LLM writes a sentence about, say, attention mechanisms, it recognizes that concept from its list, wraps it in brackets, and creates a hyperlinked backlink to the existing article on attention mechanisms.
Nora · 22:33So it is basically weaving a conceptual web. It is acting as the author, the editor, and the librarian all simultaneously.
Miles · 22:40Exactly. It creates a highly structured, interwoven graph database of knowledge, which leads right into phase three, query and enhance.
Nora · 22:48This is when he actually uses it.
Miles · 22:49This is when Karpathy actually asks it questions. He has an agent he can ask complex research questions to. And because the wiki is so well structured, so deeply cross-linked, the answers generated by the agent are incredibly rich and accurate.
Nora · 23:02Because the underlying text is so clean.
Miles · 23:03Exactly. The agent isn't trying to piece together fragmented paragraphs from a RAG search over messy unstructured PDFs. It is reading the pristine, distilled concept articles it already wrote.
Nora · 23:16And if the agent discovers a new connection during the query?
Miles · 23:19The outputs from the queries get filed back into the wiki to make it smarter. And finally, phase four is lint and maintain.
Nora · 23:26Now, linting is a software engineering term for analyzing code to find errors. How do you lint narrative prose?
Miles · 23:32Well, the LLM runs health checks in the background while Karpathy is sleeping.
Nora · 23:37That's a little creepy, but okay.
Miles · 23:38It's efficient. It stands the Wiki articles for inconsistencies. It does autonomous web searches to fill in missing information. It looks for orphan pages that have no backlinks and figures out how to connect them to the broader graph. It is a living, breathing ecosystem.
Nora · 23:53So your argument is that this amortized, right-time synthesis is the superior model for managing information?
Miles · 24:00Absolutely. Absolutely. The efficiency is just staggering. Think about your compute argument from debate one.
Nora · 24:05I'm thinking about it.
Miles · 24:06The AI does the truly heavy lifting. The deep reading of 50-page academic PDFs, the connecting of disparate dots, the extraction of core concepts. It does all of that once.
Nora · 24:17During the compile phase.
Miles · 24:18Yes. Future queries are incredibly fast and incredibly cheap because the AI is not rereading 100 raw documents every time you ask a question. It's just referencing its own clean summaries. It is literally like having a team of PhD researchers pre-digest everything you care about and hand you the executive summary.
Nora · 24:37I mean, that sounds like a utopia of efficiency, I will give you that.
Miles · 24:40It is.
The assumption you make when you trust the compiler
Nora · 24:41But it rests on a terrifying assumption. You are trusting the PhD researcher to never, ever make a mistake. And you're trusting that their summaries never warp the original meaning of the text. I'm going to push back on this right time pattern aggressively. I am firmly taking the side of query time synthesis, and I'm going to use Ananda Lahoti's developer critique to completely dismantle the safety of Karpathy's wiki.
Miles · 25:03What is dangerous about a beautifully organized AI wiki? I mean, if the AI makes a minor mistake, the Lin phase catches it, or the user just manually edits the text, right?
Nora · 25:13You only fix mistakes you can see. Lahoti's critique centers on a phenomenon he calls closed epistemic loops.
Miles · 25:20Epistemic meaning related to knowledge.
Nora · 25:22Right. He argues that right time synthesis is a slow acting poison. It leads to a catastrophic failure mode known as knowledge-based poisoning.
Miles · 25:31Poisoning. That implies it spreads and corrupts the whole system. How does a single hallucination poison a 400,000 word wiki?
Nora · 25:40Because the AI is inherently probabilistic. It is predicting the next most likely token, which means any summary it writes is inherently lossy. It naturally compresses information. information. Let's make this highly concrete for the listener because Lahoti provides a brilliant, specific example of how this code level drift actually occurs in reality. Let's hear it. Imagine you are using Karpathy's right time wiki pattern for your company's operational documents. Okay, standard corporate setup. You ingest a massive complex vendor contract. The original raw legally binding contract states that payment is, and I quote, net 30 with a 2% discount if if paid within 10 days.
Miles · 26:17Very standard business term. You have to pay in 30 days, but if you pay early, you save 2%.
Nora · 26:22Correct. Now, your LLM runs its compile phase overnight. It reads that dense legal contract and decides to author a concept article for your corporate wiki, titled Vendor Payment Terms.
Miles · 26:34Makes sense.
Nora · 26:34The LLM compresses the dense legal jargon and writes the following summary. Standard vendor agreements use net 30 terms with early payment discounts.
Miles · 26:44That is a totally reasonable summary. It captures the essence of the contract without bogging the wiki down in legalese.
Nora · 26:51It is a reasonable summary to a human scanning for the gist, but structurally it is lossy. The exact 2% figure is entirely lost. The specific 10-day window is lost.
Miles · 27:02Ah, I see.
Nora · 27:03Now, flash forward six months. A new accounting employee joins your company. They query the AI agent. What is our typical early payment discount for vendors?
Miles · 27:12And the agent goes to search the wiki.
Nora · 27:13Exactly. Exactly. The agent doesn't RU's search over the compiled wiki. It instantly retrieves the vendor payment terms concept article because the AI authored it specifically to be the central hub for that exact topic. It's a perfect match. It is the perfect mathematical hit for the search retriever. It is heavily backlinked. But remember, the retrieved text does not contain the 2% figure. It only contains the phrase early payment discounts. So what does the AI agent do when
Miles · 27:41When confronted with a question it doesn't have the exact data for in its retrieved context.
Nora · 27:45It has a few terrible options, all of which stem from its probabilistic nature. It might hedge its bets and say, we offer early payment discounts, but the exact amount varies by contract.
Miles · 27:56Which is unhelpful, but, you know, not actively destructive.
Nora · 28:00True. But often the LLM will hallucinate. It is designed to be helpful, to provide an answer. So it looks at the prompt, looks at the retrieved phrase early payment discounts, interpolates from its broader training data on standard corporate practices, and confidently generates the answer, we typically receive a 5% early payment discount.
Miles · 28:20Ouch. So the new employee pays the invoice early, expecting a 5% discount, and throws the company's accounting into absolute chaos.
Nora · 28:27It gets worse. Remember phase 3 of your beloved wiki pattern, query and enhance.
Miles · 28:33Oh man, the outputs from the queries get filed back into the wiki.
Nora · 28:36Exactly. The system takes that perfectly coherent, confidently written, entirely hallucinated answer about the 5% discount, and authors a new document in the wiki, or updates the existing one.
Miles · 28:49Validating the hallucination.
Nora · 28:50Yes. And because the system is designed to trust the wiki as the ultimate source of truth, the background linting phase won't catch the error. Internally, the wiki remains perfectly coherent.
Miles · 29:01The knowledge base is literally laundering its own hallucinations into facts.
Nora · 29:05You now have a knowledge base whose internal consistency is a complete lie that points away from ground truth. The chain of custody back to the original immutable vendor contract quietly frays and then completely breaks.
Miles · 29:18And nobody notices.
Nora · 29:19Nobody notices because every individual output looks grammatically fine. It is not a sudden loud system crash. It is a slow, silent drift from reality. That is an epistemic closed loop.
Miles · 29:30That is genuinely terrifying in a corporate setting. Yeah. If you scale that up to thousands of contracts and tens of thousands of queries, the entire institutional memory of the company just slowly detaches from reality.
Nora · 29:43Which is exactly why the only safe architectural solution is query time synthesis. In a query time architecture, the LLM should never be allowed to author persistent narrative prose during the ingestion phase.
Miles · 29:55So what does it do during ingestion, if not summarize? Just nothing.
Nora · 29:58It should only ever extract structure. It can read the vendor contract and extract entities like vendor name or contract date.
Miles · 30:06Metadata stuff.
Nora · 30:06Right. It can extract tags. It can build a metadata index. But it must never author paragraphs that replace the original text. The original raw documents must remain immutable and authoritative.
Miles · 30:18So every time someone asks a question, the system has to retrieve the raw legal contract and read the dense legalese fresh.
Nora · 30:25Yes. All synthesis, the actual writing of the answer, must happen fresh over the original documents every single time you ask a question. It never references its past summaries. It only references the raw, original ground truth.
Miles · 30:37But that destroys the compute efficiency. I mean, it is so much slower and so much more computationally expensive to force the LLM to parse complex legal boilerplate for every single query instead of just reading a clean, pre-written wiki page.
Nora · 30:51It is slower and it is exponentially more expensive, yes, but it never drifts.
Miles · 30:56I guess that's the trade-off.
Nora · 30:58For a personal hobby wiki, maybe Karpathy's right-time method is acceptable because the stakes are low. But for a team knowledge base that needs to be legally, medically, or operationally trustworthy three years from now, query time synthesis is the only way to ensure the system doesn't launder its own minor hallucinations into established bedrock facts.
Miles · 31:17Okay, this is fascinating. We have these two massive theoretical debates happening simultaneously right now. Bridge versus dead end on how we retrieve. Write time versus query time on how we synthesize.
Nora · 31:28Huge structural questions.
Miles · 31:30These aren't just, you know, academic whiteboard arguments happening at OpenAI and Microsoft. These philosophies are colliding right now in a real commercially available product that is trying to solve this exact problem for you, the listener.
Nora · 31:43The battlefield where all of this theory meets reality is DocuStrata.
Miles · 31:47Let's introduce DocuStrata to the listener based on their product fact sheet. This tool is built specifically for anyone sitting on a mountain of unstructured documents.
Nora · 31:57Which is pretty much everyone.
Miles · 31:58Yeah, think about your own digital life right now. The scattered PDFs, the half-finished spreadsheets, the tax receipts, the endless email chains. You don't even know what you know anymore.
Nora · 32:08It's overwhelming.
Miles · 32:09DocuStrata is targeting researchers managing hundreds of academic papers, estate state executors who've just been handed decades of physical paperwork, or startup founders trying to navigate five years of chaotic institutional memory.
Nora · 32:23The common thread isn't a specific demographic or profession. It is a universal, modern situation. You possess a body of documents that is simply too large for a human to read and reread by hand.
Miles · 32:35And their tagline is the ultimate distillation of this entire deep dive. Read nothing. Know everything.
Nora · 32:41It is an incredibly bold promise, and it relies entirely on the AI architectures we've just debated.
Miles · 32:46They import everything. Emails, spreadsheets, PDFs, voice notes. They run optical character recognition, OCR, on scanned physical images so that a dusty old printed contract becomes machine-readable text.
Nora · 33:01Total ingestion.
Miles · 33:02But the real paradigm shift is how they structure their business model. I want to highlight DocuStrata's philosophy here because it flips the old Silicon Valley model on its head. Their operating principle is subsidize ingestion, monetize interrogation.
Nora · 33:17Which makes perfect sense when you look at the economics of compute we discussed earlier.
Miles · 33:21How so?
Nora · 33:21Well, in the old era of software, think of Evernote or Dropbox storage was the valuable commodity if you paid a monthly fee for gigabytes of space. But today, reading text and storing text is computationally cheap.
Miles · 33:33Storage is basically free.
Nora · 33:34Right. Right. Asking the complex question, forcing the LLM to process tokens and synthesize a unique answer, that is the expensive, valuable part.
Miles · 33:43Right. So DocuStrata lets you dump 100,000 documents into their system for practically nothing. They want all your data, but they charge you per query. They monetize the interrogation.
Nora · 33:52Makes sense.
Miles · 33:52But here is where the ultimate collision of our debates happens. DocuStrata has a flagship feature called Answer Memory.
Nora · 33:59This is the feature that forces the entire debate over write time versus query time completely into the open.
Answer memory, mechanically
Miles · 34:05Here is how AnswerMemory works mechanically. You ask DocuStrata a deeply complex question. Let's say you ask, trace the evolution of our marketing strategy between Q1 2023 and Q4 2024, citing all relevant budget approvals. A very heavy query. The system goes into your archive, retrieves the right documents, and spends a massive amount of compute generating a perfectly synthesized, heavily cited, beautifully formatted answer.
Nora · 34:30A standard, high-quality ARAG operation.
Miles · 34:33But then DocuStrata takes that pristine AI-generated answer and files it back into your archive as a brand new, first-class, highly prioritized, searchable document.
Nora · 34:42It treats the AI's synthesized output as equal to the human's original raw input.
Miles · 34:46Yes, and I am claiming absolute victory here for the right-time synthesis side. This proves that Karpathy's compounding wiki pattern is the future. Answer memory turns your static archive into a compounding cognitive asset.
Nora · 34:59Explain how it compounds, exactly.
Miles · 35:01Think about the economics. Why on earth would you pay the massive compute cost to re-read 50 emails and re-answer that deeply complex marketing question in September when you already paid the AI to answer it perfectly in March?
Nora · 35:15Because things might have changed.
Miles · 35:17But when you search for that topic in September, DocuStrata doesn't have to start from scratch. It just searches the answer memory you already generated. The AI did the heavy lifting once, and now that knowledge is permanently accessible for pennies.
Nora · 35:31I see where you're going.
Miles · 35:32It is peak efficiency. It is the right time dream realized in a commercial product.
Nora · 35:37I will absolutely not concede victory here. In fact, I counter that DocuStrata's architecture actually validates everything I have been arguing for.
Miles · 35:45What so?
Nora · 35:46First of all, let's look at the foundation. The system is fundamentally R-RAG-G based. It has to be. DocuStrata knows you cannot put 100,000 documents into a context window today, so it relies entirely on vector indexes and retrieving original chunks.
Miles · 36:00Well, yeah, for the first query.
Nora · 36:01Or-RAG is the durable bridge they are walking on. It is not a dead end. It is their core infrastructure. structure.
Miles · 36:07Fine. They use our RAG to fetch the documents initially. But the answer memory feature itself is pure write-time synthesis. They are compiling a persistent knowledge base.
Nora · 36:18And answer memory is exactly the slow-acting poison Anand Lahoti warned us about. Think about the mechanics of what happens over a timeline.
Miles · 36:26Walk me through it.
Nora · 36:27You asked that complex marketing question in March. The system generates an answer. Let's say it is 99.9% correct. Correct. But because it is a probabilistic LLM, there is a slight, almost imperceptible
Miles · 36:40hallucination. Like a tiny detail.
Nora · 36:42Right. A minor detail is dropped. A specific date is slightly shifted.
Miles · 36:46But DocuStrata relies on citations. It grounds its answers in the retrieved text.
Nora · 36:50The first time it answers the question, yes, it cites the original emails. But then it takes that slightly flawed synthesized answer and files it back into the corpus as a first class document.
Miles · 37:00Okay. So the answer is now in the database.
Nora · 37:01Now come back in September. Another user searches for a related marketing topic. The ROG system, doing its mathematical job perfectly, searches the vector space. It retrieves the March answer memory because that document is a perfectly dense, perfectly structured summary of the topic. It's mathematically closer. Mathematically, it is a much closer match to the query than the messy,
Miles · 37:25see fragmented original emails. The system prioritizes the AI's previous output over the
Nora · 37:32raw originals. Exactly. It guarantees that the minor hallucination from March will mathematically contaminate the new search in September. The new answer will be generated based on the flawed old answer drifting further from the true. It breaks the chain. The chain of custody back to the original emails is broken. You are no longer reading your documents. You are reading what the AI thought about your documents six months ago. Answer memory is a ticking time bomb of epistemic drift packaged as a feature. Wow. Okay, I see the horror story there.
Miles · 37:59If the AI is constantly writing its own source material, it's just breathing its own exhaust. Exactly. The data degrabes like a photocopy of a photocopy.
Nora · 38:08Precisely. The compute efficiency and convenience of answer memory is basically a high-interest loan taken out against the future structural correctness of the entire archive.
A standoff worth understanding before you choose a tool
Miles · 38:19We are at a complete standoff. And honestly, that is exactly where we are going to leave it. We are explicitly refusing to resolve this argument for you, the listener, because the brightest minds in the tech industry are currently split right down the middle on this.
Nora · 38:33Both sides have incredibly compelling, mathematically sound arguments, and both sides simply cannot be right in the long run. The architectures are mutually exclusive.
Miles · 38:42So let's lay out the victory conditions. What actually has to happen in the physical world of silicon and code for each side to win this architectural war?
Nora · 38:50If your side is going to win the massive context, right-time synthesis vision championed by Sam Altman and Andrej Karpathy, two fundamental laws of current computing have to break.
Miles · 39:01What is the first?
Nora · 39:02First, the physical constraints of chip manufacturing and energy production have to be overcome. Compute costs must plummet to near zero, and context windows must become functionally infinite without bottlenecking data centers.
Miles · 39:14So we need cheap, infinite compute.
Nora · 39:16We need to be able to load 100,000 dense legal documents into a model's active memory for fractions of a penny.
Miles · 39:23Right. It requires a complete decoupling of intelligence from energy cost. And the second condition.
Nora · 39:28For right-time synthesis to not destroy itself through epistemic loops, LLMs must achieve a near-zero hallucination rate.
Miles · 39:36That's a tall order.
Nora · 39:37They cannot remain purely probabilistic text predictors. They must become perfect reasoners. If they are going to autonomously compile our wikis and write persistent answer memories, they cannot make even minor mistakes.
Miles · 39:50Or else the poison spreads.
Nora · 39:51Right. If they do, the AI-compiled knowledge bases will just quietly poison themselves over time. But if models achieve perfect reasoning and massive compute becomes practically free, your side wins. Argi is dead and we just instantly download the entire library.
Miles · 40:06Conversely, for your side to win the pragmatic RA-driven query time synthesis approach, approach, the current physical constraints of the universe just have to hold steady.
Nora · 40:15Pretty much.
Miles · 40:16Token generation must remain a significant marginal cost. Reading 100,000 documents will always cost more electricity and silicon than reading three mathematically targeted documents.
Nora · 40:28And the hallucination problem must remain intractable. AI must remain at least somewhat probabilistic and inherently prone to drift.
Miles · 40:36It has to stay a little bit flawed.
Nora · 40:37If those two things hold true, if compute always costs money and models sometimes guess, then ARAJ and query time synthesis are not temporary scaffolding. They are permanent, necessary architectural features of any trustworthy AI system.
Miles · 40:53You will always need retrieval to keep the compute costs down. And you will always need immutable, raw, original documents to keep the AI from breathing its own exhaust.
Nora · 41:02It is the only way to build a system you can actually trust a business or a legal defense on.
Miles · 41:07So what does this all mean for you, the listener, sitting there looking at your own mountain of unread PDFs, forwarded emails, and saved articles?
Nora · 41:15It means we are standing on the precipice of a fundamental shift in how humanity relates to its recorded knowledge. For all of human history, storing information was easy, but retrieving it and understanding it was the bottleneck. Now the tools are racing to eliminate that bottleneck. The machine is preparing to read it all for us.
Miles · 41:33Which brings us to our final provocative thought, something we want you to mull over long after this deep dive ends.
If the machine reads everything, what is left for you
Nora · 41:40What is it?
Miles · 41:40We started by talking about the friction of human reading, the slow, tedious process of opening the file, scanning the text, and trying to comprehend it. Let's say the massive context vision is fully realized. The machine reads absolutely everything for you. It never forgets, it never sleeps, and it never misinterprets a clause.
Nora · 42:00Your only remaining job in that future is simply to ask the right question.
Miles · 42:03It sounds like an intellectual utopia. You have perfect recall and perfect synthesis at your fingertips.
Nora · 42:09But is it a utopia? Or is it a cognitive trap?
Miles · 42:13Think about the slow, painful process of reading a dense vendor contract or struggling through a difficult 40-page research paper.
Nora · 42:20It's not fun.
Miles · 42:21It's not. But the friction of struggling with a text, the confusion, the rereading of a difficult paragraph, graph. That friction is often where the actual learning happens. It is where human intuition
Nora · 42:33is forged. If the machine does all the reading and you only ever consume the perfectly synthesized, frictionless, bite-sized answers it hands you, what happens to the deep, intuitive understanding that humans used to get from the struggle of reading? Are we upgrading our technological
Miles · 42:49ability to retrieve information at the direct, irreversible cost of our own biological ability
Nora · 42:55to comprehend it. It is the ultimate paradox of agentic AI. We are building machines that understand everything so that we don't have to. Exactly. When the filing cabinet reads itself,
Miles · 43:05do we eventually stop knowing what's actually inside?
Nora · 43:08A critical question we will all have to answer very soon.
Miles · 43:11Thanks for taking this deep dive with us.