The Model That Cannot Show You What It Saw

At two in the morning Beijing time last Tuesday, an AI token I had never heard of doubled. I found the reason on page four of its documentation: a citation to a Google DeepMind paper about a "vision-first" road to artificial general intelligence. There was no arXiv link. There was no author list. There was one sentence, and that sentence was worth somewhere near two hundred million dollars in market capitalization before the sun came up over Zhongguancun. I have written before that you should follow the fear, not the chart. That night the fear had nothing to do with missing a pump. It was older and simpler. A market was pricing a claim that nobody in the market could check. Eighteen years of watching this industry has taught me that we fall in love with nouns — decentralized, trustless, intelligent — long before anyone asks what a noun actually does at the level of the code. So I went looking for the paper. What I found, and what I did not find, is what I want to write about. Here is what we know, with the adjectives removed. Google DeepMind and Harvard have jointly proposed a research direction in which visual learning and multimodal integration become the main trunk of the road toward general intelligence, displacing text as the primary teacher. Not a model. Not a benchmark result. A proposal. That distinction carries more weight than the headlines suggest. In research there is a category of document called a position paper. It does not report an experiment. It argues that attention, funding, and graduate students should flow in a particular direction. Position papers are how a field argues with itself about what to build next. They are serious documents, written by serious people. They are also, by construction, promises rather than evidence. The intellectual case for vision-first is not weak, and I want to be fair to it. The argument runs like this: language is a symbolic system built on top of physical experience. A child learns the word heavy by lifting things, not by reading a definition. Pure text training gives a model the statistical residue of human descriptions of causality without the thing itself. Video, by contrast, carries time, occlusion, contact, momentum, and consequence inside every frame. If you want a model that understands the world rather than one that describes it, vision is the more honest teacher. DeepMind has spent a decade walking toward that intuition — Flamingo for vision-language alignment, Genie and Dreamer for world models, the RT series for embodied control. Harvard contributes visual neuroscience and cognitive science. The pairing is credible. What it is not, yet, is demonstrated. And here is the detail the chart left out. The proposal arrived without a paper title, without named authors, and without a single reproducible result. For a research agenda, that is normal. For a two-hundred-million-dollar move, it is not. Hold both facts at once. The tension between them is the story. I start where I always start, with the plumbing, because plumbing is where claims go to die. In 2017 I spent my nights reading Gnosis Safe's Solidity line by line and submitted twelve logic flaws in its multi-signature implementation. Every one of them was invisible at the level of the pitch and obvious at the level of execution. Vision-first AGI deserves the same treatment. If you can pull the claim apart, you should. Begin with token counts. A ten-minute 1080p video at 24 frames per second is roughly fourteen thousand frames. Patch-tokenize each frame the way a vision transformer does and you are no longer comparing a video to a novel. You are comparing a video to a small library. Text corpora have been measured in the trillions of tokens for years, painstakingly filtered. Video is effectively unbounded, and almost none of it has been curated at all. If visual pretraining becomes the main trunk, the compute curve bends in a way that makes today's training runs look like a warm-up. The order-of-magnitude figures engineers use privately run one to two exponents higher. No published budget comes close. But compute is the easy complaint, and I want to make a harder one. Video training is not primarily FLOP-bound. It is bandwidth-bound. Moving petabytes between storage, preprocessing, and accelerators, keeping them synchronized, and doing it without starving the chips is an interconnect and data-pipeline problem long before it is a GPU problem. I watched this pattern in a different costume during DeFi Summer in 2020, when Compound's governance token collapsed and took my savings and my study group's savings with it. The failure was never in the elegant part of the diagram. It was in the assumption that the system would keep behaving the way the diagram said it would. Now the question that should interest this industry more than the AGI headline. Can you prove what a model was trained on? For text the answer is uncomfortable but tractable. A corpus can be hashed. A dataset can be committed to. Membership can be proven without revealing the document. For video, every one of those properties degrades. Footage is fragmented across platforms, licenses, cameras, and phones. It contains faces, plates, doorways, and children. Attribution is a legal fiction more often than a technical fact. This is exactly the problem my small team at Verifiable Truth has been building against — zero-knowledge proofs over training-data provenance, so a model can demonstrate the origin of what it learned without exposing anything proprietary. And I will be candid, because integrity is worth more than positioning. Proving membership is feasible. Proving that a clip sits inside a licensed corpus is feasible with the right commitments. Proving non-contamination — that nothing outside that corpus was ever touched — is not feasible at video scale today, and I do not expect it to be feasible soon. There is a gap between provenance and innocence, and most marketing in this space quietly assumes the gap away. This is the same error I audited in 2017, wearing different clothes. A multi-signature wallet looks decentralized until you read the upgrade path and discover that three humans can move the whole thing. A training pipeline looks verified until you ask who holds the admin key over the dataset manifest. Vision-first removes no chokepoint. It multiplies them, because the data is vaster, messier, and more legally entangled than any text corpus has ever been. There is a second chokepoint nobody is pricing. Verification infrastructure centralizes the way rollup infrastructure centralized. After Dencun, blob space was cheap and everyone built on the assumption that it would stay cheap. It will not. Within a couple of years the same squeeze arrives: a handful of provers, a handful of sequencers, a handful of relayers, and a fee market that quietly reintroduces the rent decentralization was supposed to abolish. If vision-first models spawn an industry of visual verification, expect the same arc. Cheap, then crowded, then expensive, then gatekept. And a third. Vision-first couples naturally to embodiment, because vision is not only perception — it is the substrate of action planning. That is where the safety surface widens. In text systems we worry about prompt injection. In a vision-first world the analogous attack is perception injection: a crafted image or a manipulated camera feed steering a model embedded in a physical system. We built a decade of defenses for language adversaries and almost nothing that scales for visual ones. Now consider what it would take to falsify the claim, which is the test we almost never run. A vision-first thesis should predict something concrete. It should predict that a model trained without language supervision outperforms a text-first model on intuitive physics, spatial reasoning, contact prediction, and occlusion. It should predict that video pretraining transfers to manipulation tasks better than any caption-conditioned objective. Our benchmark suites are overwhelmingly text and code. A credible causal-physical benchmark that a vision-first model could win and a language-first model could lose does not meaningfully exist yet. Until it does, the debate is philosophy with a compute budget. Here is the contrarian part, and it cuts against my own side of the aisle. The reflexive crypto response is to tokenize this within a quarter. Vision-first AGI becomes video-data DAOs, decentralized GPU marketplaces, agent tokens with world-model narratives. Most of that will be theater, and I say that as someone who builds in this space. But the deeper inversion is this: a vision-first path makes verification harder, not easier. The industry's founding promise — that crypto will make AI accountable — runs backwards here. Text is the more auditable modality. Text is hashable, searchable, and small enough to reason about. Abandon text-centrism and we walk into an era in which no one can prove what a model learned, and the people with the most data will be the least able to demonstrate it honestly. There is also a strategic reading nobody in the token groups mentioned. Alphabet owns the lab making the proposal and the platform holding the largest curated video corpus on Earth. A research agenda that elevates video to the trunk of intelligence is, among other things, a resource argument made inside a very large company. That does not make the science wrong. It makes the framing worth watching. If you can find the actual paper — the authors, the experiments, the ablations — send it to me and I will revise everything above. Beyond provenance, the privacy question will not wait for the research to mature. Video pretraining at scale means ingesting faces, streets, and private moments, and the compliance regimes that have barely been written for text will arrive for video with sharper teeth. Teams that design consent, licensing, and deletion into the data layer now will be the ones still standing when the rules land. That is slow work, and slow work is the only kind that survives a regulatory winter. What I keep returning to is not the architecture. It is the question underneath it. Every technical choice about what a system learns from is a choice about what the world is allowed to see, and what we are allowed to know about the seeing. Language left a paper trail. Frames leave none. We are being asked to trust a road to general intelligence whose first act may be to burn its own receipts — and to fund it with tokens that cannot audit a single frame. So I will keep asking the unfashionable question while the charts go vertical. Not whether the vision-first thesis is true, but whether we will build the instruments to tell. Follow the fear, not the chart. If you can make this provable, you will have built something more durable than any model — and more scarce than any token.