Saining Xie: Silicon Valley is "Hypnotized by LLMs" — Building World Models Beyond Language
Zhang Xiaojun Podcast, March 16, 2026
In a snow-covered Brooklyn apartment in early 2026, Saining Xie is making one of the boldest bets in artificial intelligence. The NYU professor and freshly minted co-founder of AMI Labs alongside Turing Award winner Yann LeCun has just closed a record-breaking $1.03 billion seed round at a $3.5 billion pre-money valuation. Yet for all the superlatives, Xie insists he's "the normal one" — not the chosen genius, but someone who stumbled through career decisions with what he calls "mystical logic," guided more by people and curiosity than ambition.
This seven-hour conversation reveals a researcher who has spent over a decade quietly building the scaffolding for what he believes will be AI's next leap: world models that truly understand physical reality, not just the shadows cast by language on the walls of Plato's cave. It's a contrarian stance in a Valley he describes as "completely hypnotized" by large language models, and one that puts him on a collision course with the industry's most powerful narratives.
The Fundamental Flaw in Language Models
Xie doesn't mince words about LLMs. "I don't see LLMs as a success story for the Bitter Lesson — in fact, they are a rebellion against it," he says, referring to Rich Sutton's famous essay arguing AI should minimize human-designed structure. His reasoning cuts to the core of what language actually represents. "Language is not a thinking map, it's a communication tool," he explains. When humans write, they compress vast experience into symbols optimized for transmission between minds, not for capturing the full texture of reality.
This matters because LLM scaling laws, in Xie's view, contain significant "water" — what he calls inflation. "The Scaling Law of language models is based on a representation of knowledge," he argues. Models must memorize factual trivia to seem intelligent, bloating parameter counts. But a true world model wouldn't need to remember every Wikipedia entry. "You don't need to memorize all this knowledge — you need good understanding capability to filter and process information," he says. The decision itself matters more than the database.
The technical argument goes deeper. When you model P(y) — predicting labels or next tokens — your probability mass concentrates in a low-dimensional semantic space. But when you model P(x|y) as video generation models do, "the likelihood you assign says which phenomena actually exist in the world, which things are more likely," Xie explains. This forces the model to learn why four-legged cats are more common than three-legged ones, why motion should be smooth, why physics constrains possibility. "What you need to learn now carries far more intelligence than what you get from modeling P(y)."
Even more fundamentally, language operates in discrete tokens while the physical world operates in continuous, high-dimensional, noisy space. "If you process vision the way an LLM does, I would need to tokenize every frame, string it into a very long sequence," Xie says. "You're taking this world representation behind it — there's actually some kind of global state — and you serialize it into very redundant tokens." Transformers must then pay equal attention to every token, which "itself is unreasonable" for capturing spatial structure that humans process in parallel across cortical regions.
The Representation Thread
To understand where Xie is headed, you need to trace where he's been. His research journey reads like a systematic exploration of representation learning across modalities, objectives, and architectures. At FAIR from 2018 to 2023, he worked with Kaiming He on the MoCo series, pushing self-supervised contrastive learning to briefly match or exceed ImageNet pretraining. Then came MAE (Masked Autoencoders), which He led, taking a simpler denoising approach that also showed promising scaling behavior.
Both ultimately hit walls. "We found, hey MAE is incredible, MAE should at least win a best paper award," Xie recalls. "Turns out it didn't. And scaling up MAE would solve all problems — turned out it didn't scale up either." The vision community couldn't crack what language models had: a clear path from more data and compute to better results. "Right now vision doesn't have a Scaling Law," Xie admits, though he immediately pivots: "And more and more I think vision doesn't need a Scaling Law."
The breakthrough insight came from his 2024 work on Representation Autoencoders (RAE). Instead of denoising in pixel space or compressed VAE latents, RAE operates in the high-dimensional representation space learned by self-supervised models like DINOv2. "Your representation layer will now become very high-dimensional," Xie explains, "and you want to do denoising and image generation on this high-dimensional representation." A student questioned whether such high dimensionality might hurt training. Professor Ma Yi at HKU stood up and told everyone: "You must not be afraid of high dimensions. High dimensionality is an extremely important cornerstone in all machine learning."
This work convinced Xie that representation is the foundation on which everything else should build. "There's only one thing in this world that is important: how to learn this representation," he says. Language models will "gradually degrade to a simple communication interface" rather than driving intelligence. Video generation will similarly build on top of strong representations. "When we have such a foundation, we can easily decode it into language, easily decode it into pixels and generate videos, and also decode it into some kind of action."
What Is a World Model?
Xie grew animated when pressed on definitions, acknowledging that "world model" has become an unhelpfully broad buzzword. Technically, it's a predictive function: given current state S_t and action a_t, predict next state S_{t+1}. This formulation dates to 1943 when psychologist Kenneth Craik first proposed humans maintain internal models to forecast action consequences. Control theory adopted it for lunar landers in the 1960s. Reinforcement learning embraced it with Sutton's Dyna architecture.
But the concept has fractured. Video generation companies call themselves world model companies, focused on rendering compelling visual sequences with some consistency — what Xie calls "world simulators." Fei-Fei Li's World Labs pursues explicit 3D spatial representations as a frontend for asset generation, attracting $200 million from Autodesk for industrial applications. These are valuable but incomplete.
What Xie and LeCun are building is something he calls a "predictive brain" — intelligence itself rather than just simulation or spatial scaffolding. "We're doing a path that cannot lead to true intelligence," he says of current approaches. "Someone needs to define new problems on this larger scale." The core requirements: understanding the physical world, large associative memory, reasoning and planning capabilities, counterfactual inference, and crucially, controllability and safety that emerge from prediction rather than post-hoc alignment.
This last point matters enormously to Xie. "LLMs are not controllable, not safe either, because they don't have a true world model," he argues. Current safety comes entirely through fine-tuning on examples of what not to say or do. But "a true world model actually you don't need to do this because you can predict what consequence your action will lead to." A robot holding a knife can foresee the danger of turning toward a person without needing thousands of training examples of knife accidents. The world model enables safety through understanding rather than memorization.
The Data Problem
If language models feasted on the internet's text, world models must somehow ingest reality itself. "The past era was about downloading the Internet," Xie says. "Now the era is about downloading humanity." A four-month-old baby has seen more visual data than all the tokens used to train the best LLMs — some 30 trillion tokens equals roughly 30 minutes of YouTube uploads. The scale is staggering.
YouTube becomes the obvious first target despite copyright and terms-of-service battles. "One side is tightly guarding against data collection, blocking you from scraping. The other side is trying every means to get more data," Xie describes the cat-and-mouse dynamic. ByteDance's massive advantage through TikTok and Douyin becomes apparent, though they too face legal challenges.
But video alone won't suffice. Xie envisions always-on wearables — AI glasses that see everything their wearers see, creating infinite token streams of first-person experience. The closest current analog is fitness trackers monitoring heart rate and sleep, making simple predictions about stress and recovery. "If we imagine there were some kind of wearable device that could truly be always on and could see in real time everything we can see," he muses, "I think this has enormous potential." Medical monitoring, athletic training, elderly care — all benefit from models that understand visual context and temporal patterns.
Robotics provides another data avenue and deployment target. The industry suffers from what DeepMind's Jie Tan calls "extremely uneven development" — physical capabilities now exceed humans, but cognition lags far behind. "Nobody is building the brain, nobody is building a robot brain," Xie says. Everyone from DeepMind to Physical Intelligence relies on VLAs (Vision-Language-Action models) that route through language models like Gemini. "This needs a second half of pre-training," he argues, invoking former OpenAI researcher Shunyu Yao's formulation. "Who will do this pre-training? I don't think it can be done in any country including in the Bay Area."
The DiT Backstory
Xie's best-known work, Diffusion Transformers (DiT), emerged from what he calls research's necessary "non-linearity." In 2022 at FAIR, he and intern Bill Peebles (now head of Sora at OpenAI) started exploring diffusion models to understand the connection between discriminative and generative learning. The original question came from Kaiming He: "Is this a discriminative model or a generative model?" It's a profound puzzle, since denoising performs discriminative prediction at each step but generates through their composition.
For two months they explored representation quality, finding diffusion models learned decent features but "much, much worse than self-supervised learning — completely not competitive." Then came the pivot. "In the final month we discovered, hey, by the way, our new architecture is indeed more efficient and scalable than U-Net." From a code perspective, DiT embodied what Xie calls Minimal Description Length — elegance through simplicity. "If your code is short and can achieve the same purpose, your method will typically be better than one requiring thousands of lines."
The paper was rejected from CVPR. "Not enough novelty," reviewers said. "You came up with a very simple structure and even though you got good results, the reviewers weren't convinced." Yann LeCun tweeted about it. They submitted elsewhere unchanged and won an Oral Paper. "This is a completely random process," Xie shrugs.
FAIR initially wouldn't let them list the company affiliation since Xie had already left. But after Sora explicitly cited DiT as its backbone, the architecture became the foundation for essentially all major video generation models. The irony is rich: "Several researchers came to me saying 'I worked on this for two weeks but my manager said you can't do this anymore, we have product cycles coming up.'" Industry knew DiT was the right direction but lacked freedom to pursue it.
The Kaiming Influence
No one looms larger in Xie's research formation than Kaiming He, ResNet's creator and FAIR's legendary researcher. Their collaboration began almost accidentally when He joined FAIR mid-internship in 2016. Xie's manager handed him off: "Kaiming, you come guide him." With one month left, they decided to compete in the ImageNet Challenge, producing ResNeXt — named partially for "Xie's ResNet," He's playful credit-sharing.
What Xie absorbed went far beyond papers. He describes He's workflow in almost spiritual terms. "Kaiming's magic is he can take all very ordinary things and turn them into gold-like valuable ideas," he says. The discipline was monastic: talks finished a month before deadlines while others pulled all-nighters, every line in papers checked to ensure no more than 40 percent whitespace, obsessive baseline construction before attempting innovation.
"The ceiling of your research depends on how good your baseline is," Xie learned. "If your baseline is weak you can easily fool yourself." This counterintuitive principle — make the comparison harder, not easier — enabled He's detection work (Fast R-CNN, Mask R-CNN) to dominate by starting from infrastructure that already exceeded mediocre CVPR papers. "Our baseline was already stronger than yours, so if I take one more step up, of course I'm going to go even further."
But the deepest lesson concerned research philosophy. He gave Xie a copy of the Diamond Sutra, the Buddhist text teaching that "all phenomena are like dreams, illusions, bubbles and shadows." The resonance with Western philosophy — Kant's thing-in-itself, Schopenhauer's representation versus will — points to a core insight about research taste. "When you're reading a paper, what matters is to break through the illusion the paper presents and question what lies behind it, what kind of substantive essence it contains," Xie explains.
This filters into every decision. The first lesson for FAIR interns was mastering Excel spreadsheets to track experiments. "You really have to make some decisions: what metrics should I focus on, what columns should there be, how should I define control variables, how to make each experiment as informative as possible." Run too few experiments and signal stays unclear. Run everything blindly and you drown in noise. The craft lies in designing the spreadsheet itself.
Turning Down Ilya Twice
Xie's path not taken haunts the conversation. In 2018, fresh out of PhD, he interviewed at OpenAI with John Schulman in a "small dark room for five or six hours working on one problem." The handwritten questions on pencil-marked paper moved him. But FAIR had Kaiming, Piotr Dollar, Ross Girshick — "the three pillars of computer vision." When the OpenAI offer arrived, "Ilya called me and I said almost nothing and rejected it."
Ilya Sutskever was angry. "Why didn't you even discuss it before rejecting? Is the money not enough?" Xie's tone was apologetic but firm. At the time in 2018, for top PhD students, FAIR was the more certain bet — more open, more academic. "Unless they really wanted to do what OpenAI was already doing, I think most people would still lean toward FAIR."
The second rejection came in July 2024, just after Sutskever founded SSI. He emailed asking Xie to join. This time they discussed how to give AI the ability to love. "The conclusion is that this is very important," Xie recalls. "Because without it we face a very uncertain and very dangerous future." But when Xie asked about multimodality, Ilya's response was decisive: vision was "already solved well enough."
That fundamental disagreement sealed it. "SSI has its own language-based approach, and that approach is not the path I want to pursue," Xie explains. He doesn't see this as bitter rivalry. "Brothers climbing a mountain, each making their own effort," he says, quoting a Chinese proverb. "And without the recent developments in LLMs, there might not have been the current state of computer vision." But the paths diverge sharply on what intelligence requires.
The Yann LeCun Partnership
The AMI Labs founding story verges on the mystical that characterizes Xie's decision-making. A manager suggested in late 2025 that Yann LeCun seemed unhappy at Meta. Xie's first instinct was disbelief — the godfather of AI, leaving? Two weeks later in a one-on-one meeting, LeCun opened: "Don't tell anyone yet, but I've decided this. What I want to do should be done outside. I want to build a company."
Xie asked about business model. "Then I realized wow, this is completely aligned with what I'd imagined." What was it? "I think you can call it world models, or the logic behind this is: what I want to do, in the current environment in any country in the world, I don't think it can be done, including in the Bay Area."
The conviction was simple: the work requires research depth but can't happen in traditional academia (insufficient resources) or closed industry labs (insufficient freedom to define problems). "We need to return to what we discussed at the beginning," Xie says. "In this kind of finite game, in this highly competitive environment, every company seems to have lost its ability to define problems." OpenAI once defined problems. Now even OpenAI gets swept into races it defined.
Xie agonized for about a week. LeCun's pitch was straightforward: this was a better choice for Xie personally and for the company, offering autonomy to build teams and roadmaps. "I also incredibly, incredibly grateful that Yann could give me that trust," Xie says. Six co-founders signed on, including a CEO, COO, VP of World Models (the original JEPA team director), and CRIO (Chief Research and Innovation Officer, Pascale Fung from Hong Kong). Several walked away from tens of millions in unvested OpenAI, Google, or Meta stock.
Why? "I think maybe we're all just a little crazy," Xie laughs. But beneath the self-deprecation lies conviction. "This is the only place where we can do this." The mission isn't chasing leaderboard points or product cycles. It's pre-training the second half — building the brain that robotics lacks, that wearables will need, that embodied AI requires.
The Silicon Valley Critique
Xie reserves his sharpest criticism for the Bay Area's cultural moment. "Silicon Valley is very LLM-pilled," he says flatly. "The Valley is deep in the trenches of Large Language Models — it's completely hypnotized." This isn't mere trash talk. He spent years part-time at Google's Gemini GenAI team and saw the constraints firsthand.
"The entire AI industry right now has this enormous value chain," he explains. "At the very top there's Bitter Lesson, there's a narrative of AGI and LLM. This has defined a series of benchmarks. So you compete on leaderboards. The leaderboard defines resource allocation." If your goal is topping LMSys Arena, you can only pour resources into that race. "Under this value chain, resource allocation means they can't do this part of the research," even when researchers know it's important.
Video understanding is his example. "Neither academia nor industry is doing much of it, or people are doing it but not with a fundamental World Model angle." Why? "Without exception, regardless of which company, all of them have been assigned to a video generation model team." That's the only position within the value chain where video expertise fits. "Even though they all know we haven't solved this problem, we need a better World Model-based video understanding model, people won't have space to do such exploration."
The resistance to academia compounds the problem. "People look down on the work academia is doing. They don't think academia's work can truly generate any kind of impact." Even within companies, research groups can't share what they're doing with product groups. Several Google researchers told Xie after his RAE paper: "I worked on this for two weeks but my manager said you can't do this anymore, we have product cycle one, two, three coming up."
So AMI Labs plants itself in New York, not Palo Alto. "The Bay Area is really a pure tech bubble," Xie says. "Everyone can be very focused on doing one thing. And New York is a more real world." That realness provides grounding. Walking through Washington Square Park to his NYU office, seeing street performers, mothers with strollers, old men playing chess, students studying — "I find the world is much bigger than we imagine. Not everyone cares about what AI is. They have their own lives."
This grounds the question: what should AI actually be doing? "Do we have some kind of social responsibility?" Xie wonders. "More contact with people, more contact with people living in this world, helps me understand what AI is and how to build the next generation of AI in new ways." It's the conversation Ilya wanted to have but Xie hadn't reached yet.
The Scaling Law Heresy
Perhaps Xie's most contrarian technical claim is that vision doesn't need and shouldn't want LLM-style scaling laws. "The model won't be that large," he predicts. "You don't need to remember everything, all the subtle details in the world that you can see." A world model doesn't solve definite equations in high-dimensional space to determine whether apples fall. "It doesn't need the highest level of human intelligence. It needs good understanding capability to filter and process information."
This directly contradicts the Chinchilla scaling paradigm where compute C roughly equals 6 times tokens N times parameters D. "I think the Scaling Law of language models is based on a representation of knowledge," Xie argues. "It may have a relatively balanced ratio — your number of tokens and your parameters need to be roughly 1:1." But world models optimizing for decision-making rather than fact retrieval could have radically different compute distributions.
The biological analogy is precise. Humans take in roughly one billion bits per second through all senses combined — vision alone processes enormous bandwidth. But behavioral output is merely 10 to 100 bits per second. "What kind of model is our brain that at twenty watts of power takes in one billion bits per second and converts it into 10 bits per second of behavioral output?" The filtering system is paramount, not the database.
This suggests smaller models processing vastly more data, with different efficiency curves. The model doesn't scale linearly with data volume because it's learning compression and prediction, not memorization. "We can proceed in several stages," Xie says. "Start with internet data, start with YouTube." But distribution of compute will differ — "a larger portion might be used on video" with combined generative and discriminative methods plus language, not just prediction loss.
The Human Arrogance Problem
Late in the conversation, Xie grew passionate about what he calls "human arrogance" — the assumption that human intelligence defines intelligence itself. He'd been reading "Are We Smart Enough to Know How Smart Animals Are?" by primatologist Frans de Waal, discovering that many supposedly unique human traits exist across species. Tool use, mirror self-recognition, theory of mind, even political maneuvering among chimpanzees described in de Waal's "Chimpanzee Politics."
Scrub jays bury food, but if they notice a peer watching, they wait for it to leave then dig up and rebury the cache elsewhere. Chimpanzees shown two boxes, one with a banana and one with an apple, will immediately open the apple box if they see an experimenter eating a banana — reasoning about hidden state. "Language is indeed unique, only humans have it, but that doesn't mean other animals don't communicate," Xie notes. Whales have their own language.
This reframes AI's challenge. Rich Sutton's podcast comment resonated deeply: "You think these things are great and impressive, that they're hard? I don't think so. I think building the intelligence of a squirrel is the hard problem." Once you can build a squirrel's intelligence — with its own goals, intrinsic rewards, emotions, social behaviors — "writing code, going to Mars, going to the moon, those things would be the easy ones."
From a human perspective this seems absurd. From the universe's perspective, it's obvious. "If you set aside human arrogance, building a squirrel's intelligence is actually a harder problem," Xie insists. And robotics becomes the test. Can we build something reliable enough to function like a 12-year-old child handling household chores? Current robots have superhuman limbs but infant-level brains. "Nobody is building the brain."
The Research Philosophy
Threading through everything is Xie's view of research as inherently non-linear. "The worst work is when your starting idea is the same as your ending idea," he says. "You didn't encounter any obstacles, any difficulties. This shows your idea is boring." Good work requires two months of flailing exploration where "you have this underlying melancholy because you can't see any light," followed by sudden convergence in the final month before deadlines.
The exploration can't be aimless. Kaiming He taught him to track experiments in Excel spreadsheets designed to maximize information gain — not too sparse (unclear signal) or too dense (meaningless noise). "Every experiment needs to give you what I just called this gradient," Xie explains. Failed experiments often provide larger gradients than successful ones. "As long as you don't know which direction to go — that's the most feared thing."
Ideas emerge from this process, not from armchair speculation. "If you want to come up with an idea by just thinking, it's definitely not a good idea," Xie says, channeling Kaiming. Either you're smarter than everyone (unlikely), or 10,000 people worldwide are thinking the same thing (you'll lose the race), or it's a bad idea others already tried (you'll waste time). Real ideas come from doing — reproducing baselines, making extensions, finding surprising observations.
The goal isn't impact, a word Xie "has a bit of an aversion to" as overly aggressive and masculine. He prefers Hannah Arendt's framing: research is for understanding itself, and spreading understanding creates resonance, a "feeling of family" with those who grasp what you've grasped. "Being able to understand certain things, and what I understand can somehow have some influence on the world's or AI's development — I think this is something I care very much about now."
This makes research an infinite game, borrowing from the book of the same name. You don't optimize average paper quality but maximum — your signature work. "You in your lifetime truly only need to succeed once," Xie says. A few times gets you to the top. This is why he can tell students he did five internships during PhD and half produced nothing, or admit he hasn't had a best paper since 2015's Holistically-Nested Edge Detection (HED) despite thousands of citations.
The Underdog Identity
Despite the record seed round, Xie insists AMI Labs is "completely an underdog company." Yann LeCun is "someone who holds firm to himself and always tries to do the next thing but that thing hasn't been proven yet." The company faces industry pressure, requires frugal resource deployment, and bets everything on a contrarian vision. "We may have raised a lot but compared to the resources LLMs are mobilizing now, this is just — I don't know what percentage, it's so far off."
But being underdog appeals to Xie's research identity. "The more you don't believe in me, the happier I am," he grins. "Can greatness not be planned?" he asks, referencing the book "Why Greatness Cannot Be Planned." The answer is yes — you need exploration, gradual idea formation, willingness to pivot. "I think this applies to startups too."
The journey already shows signs of the non-linearity Xie describes. They have "things to work on" with "quite promising results coming soon." An upcoming paper called Solaris (named for the Tarkovsky film about an ocean that materializes human memories) will demonstrate video generation, but more importantly, advance the representation learning foundation. By March they'll likely have more to show.
The team of 25 includes several who walked away from eight-figure equity packages at OpenAI, Google, Meta. "Why? I think maybe we're all just a little crazy," Xie says. "Everyone realizes this is actually necessary. On one hand it's somewhat of a counter-consensus view, but on the other hand over the past year it has gradually become a consensus." The consensus is simply: what exists now isn't enough. Someone must build the predictive brain.
The Skiing Metaphor
Xie recently took up skiing, finding in it a metaphor for entrepreneurship. "First, skiing is a sport about balance. Second, you have to be fearless and point your shoulders down the slope," he explains. Instinct screams to lean back away from danger, but that guarantees a fall. "Only when you completely abandon fear, with enough courage, and point your shoulders toward the slope, you actually become more stable."
The hymn of humanity is the hymn of courage, he quotes from JoJo's Bizarre Adventure. And courage grows from companionship. "You never walk alone," he adds, borrowing from his beloved Liverpool FC. Klopp's self-description as "not the special one" but "the normal one" resonates — someone who acts as a battery for the team, charging others rather than hogging spotlight.
In startups as in research, Xie believes, you need that battery role even when discouraged. "I feel discouraged every day," he admits. "I think it's become a kind of researcher's fate. Everyone has this underlying melancholy because the process is like groping around in a dark lightless place." Joy comes only in brief flashes when something works — maybe 5 or 10 percent of the time.
But the melancholy lifts when walking through Washington Square Park, watching people live lives utterly unconcerned with AI. "I find my most stress-relieving moments every day are this roughly 5 to 10 minute walk," he says. "The world is big." And perhaps that's the deepest argument for world models over language models — the former at least aspires to engage with that bigness, that bewildering physical reality, rather than just the compressed linguistic shadow it casts.
The JEPA Journey
Xie's evolution on LeCun's JEPA (Joint-Embedding Predictive Architecture) mirrors his journey from academia to startup. "I went through several phases: doubting JEPA, understanding JEPA, becoming JEPA," he laughs. At first it seemed like "yet another self-supervised learning algorithm." Then gradually he saw deeper logic. Recent work like LeJEPA with rigorous proofs showed that good representations must be isotropic Gaussian — a specific mathematical property.
But the real insight was realizing "JEPA is not a model, not a specific algorithm. JEPA is a complete cognitive architecture." In LeCun's 2022 paper "A Path Towards Autonomous Machine Intelligence," JEPA encompasses world understanding, prediction, planning — the full stack. "This gave me new insights into JEPA and I found that JEPA actually isn't a specific method people outside tend to say. It's not like that."
Instead JEPA is "a very vast ocean, and in this ocean there can be many ships sailing." LLMs are part of it. Video generation is part of it. The vision isn't to shrink down into LLM constraints but to "deliberately think big, have enough space for us to explore, to scale up, until we achieve some kind of new breakthrough." When exactly? No one knows. But it's a path Xie is "willing to invest my life in to walk."
The 42 Question
The conversation ends where it must: with the universe and fate. Is this world a giant world model? "Of course the world is a giant world model," Xie answers. Can you predict fate then? "I don't think so." Why not? "We don't have enough resources. You'd need a computer as large as the Earth, or the size of the entire universe, to tell you the answer about life, about the universe, about anything."
And that answer, like Douglas Adams promised, might ultimately be 42 — simultaneously precise and useless without knowing the question. Perhaps that's the point. What Xie is building isn't a system that predicts everything, but one that filters intelligently, processes hierarchically, and acts appropriately in an irreducibly complex world. Not omniscience, but competence. Not AGI as apotheosis, but intelligence as participation.
For now, he walks through Washington Square Park each day, watching the world be the world, storing that in whatever world model runs in his own brain, and translating it somehow into the next architecture, the next paper, the next step toward downloading humanity rather than just the internet. Whether the 25-person team in New York can actually pull this off against the industry's full weight remains to be seen. But the conviction that someone must try, and that current paths are insufficient, burns through every word.