Forty Dollars to Index Joe Rogan: Chroma, Shovels, and the Wright Flyer Era of ML

After a sponsor read, Nathan Labenz asks Anton Troynikov to treat a vector database as a database, not a slogan. A traditional SQL store is algebraic: you select from where, and you get exact rows. A vector store, Troynikov says, lets you treat data as points in high-dimensional space and run operations that have geometric meaning — distance, density, clustering, fitting a surface. That only became exciting, he argues, once models could turn many kinds of input into vectors that actually capture structure. Sentences about similar things land near each other; the embedding is, in his phrase, the AI-native way to represent data.

The workflow that has exploded around ChatGPT is retrieval-augmented generation. The model was not trained on a company's wiki or a user's mail, so asking it directly invites hallucination. Embed the corpus with an open-source model or an API, store the vectors with their source text, embed the question, pull the nearest documents, and stuff those documents into the context window with an instruction to answer from them. Troynikov's claim is blunt: the loop pretty much suppresses hallucination, and it lets an organization skip fine-tuning, which is expensive and slow. Pinterest has been doing a version of this for years — visually similar pins are nearest neighbors in an image embedding space — but the embeddings for many modalities are now cheap enough that people are building on top of an API instead of owning a model.

Labenz locates vector databases at a painful point in the value chain: every company has a wiki nobody can search, and everyone who has used ChatGPT has had the same thought — if this thing had my data. Troynikov is careful about the product. Chroma uses a vector database as technology, but what it actually is, he says, is a platform for embedding the knowledge that machine-learning-enabled applications can use: the storage layer for apps that keep a large language model in the loop. Existing vector stores, in their view, had the wrong affordances for that job. Chroma began from a different principle: you can understand a model's behavior from how it sees its data, and an embedding is that representation.

The company started as an internal tool for their own product experiments. They needed something a tiny team could run while still being fast, and they shipped a packaged version after OpenAI's marketing of embeddings made document question-answering a default demo. Troynikov announced the launch on Twitter the previous Tuesday; by the recording it already felt old, which Labenz takes as a sign of how compressed time has become. Troynikov's watershed list is GPT-3 in 2020, then Stable Diffusion — open source, small enough for an M1 Mac, someone even got it onto an iPhone — and then ChatGPT, which technologists will tell you was "just" packaging. Packaging, he says, is the point: the right affordances made people understand what they could do.

Labenz quotes Boris Power at OpenAI on cost. Embedding every Joe Rogan transcript — about 100,000 pages — is forty dollars. A web-scale corpus of 10^11 pages is about fifty million. Embeddings are also deterministic: no temperature, same input, same vector, so you store them once. That is where a storage layer earns its keep. Troynikov says they evaluated everything else on the market. Tools were too hard to deploy, too expensive, or too complex for a two-person company — for a long stretch, just him and co-founder Jeff Huber, with no spare full-time engineer to babysit a vector store. Missing features and the wrong abstractions convinced them nothing on the market was AI-native: LLM apps need inserts, updates, and deletes, a clean talk path to an embeddings API, and results shaped to drop into a context window. Chroma launched as a day-one LangChain integration because the bet is on application frameworks, not on an abstract idea of a vector database.

Pre-deep-learning vector databases were built for recommenders and semantic search over billions of documents, often with hand-rolled embedding functions. Chroma keeps a metadata store beside the vector index, so a query can be filtered in the ordinary way — only these documents, only this phrase — and then handed to the geometric search. Exact nearest neighbor is quadratic in the number of vectors. A few hundred points is one matrix multiply, in memory, via NumPy; Troynikov says you should just do that. North of about 10,000 points, n-squared starts to hurt. Approximate nearest neighbor algorithms make query time roughly flat, at the cost of missing some neighbors. In practice he describes the miss as more like one query in a hundred, not half of them, and on smaller personal corpora you can tune for recall so that missing the one email you need becomes unlikely.

Because embeddings live in a vector space, geometric operations have meaning. Clusters of text tend to be topical. You can find a direction that corresponds to a human-readable property — a "question versus answer" vector — and remap points without backpropagation, just by estimating a transform. User feedback on relevance can re-weight the space the same way. Folk wisdom, including in OpenAI's cookbooks, said a linear map was enough to move between embedding spaces with similar semantics; a recent paper on relative representations, Labenz notes, showed models learning strikingly similar shapes up to rotation and scaling. Troynikov wants more of that empirical, biologist-style observation in machine learning.

That invariance suggested a privacy trick he has not fully evaluated: apply a random transform and some noise before sending vectors to a cloud store. Most of the operations you want are invariant to translation, rotation, and scaling, so the provider can still search while finding it very hard to recover the original data — a kind of homomorphic computation. Labenz connects the architecture to BLIP-2, which cut training compute by 95 percent by training a 200-million-parameter connector between a frozen vision model and a frozen language model, injecting image embeddings as a prompt prefix, sometimes into regions no text can reach. Google Brain's RT-1 does something similar with perception and a planner that emits action embeddings. Troynikov accepts the analogy: the language model as executive function, the vector store as external memory. DeepMind's RETRO, he notes after first misremembering it as Meta's, was trained so it cannot get knowledge except by searching an external bank. His forecast is a split world: giant models as utilities, and smaller models trained to find and compose knowledge rather than store it in weights.

Chroma is shipping an experimental relevance score so a pipeline can decide whether retrieved neighbors are worth using, in part to avoid HyDE — the trick of letting a language model hallucinate an answer-shaped document and embedding that instead of the raw query. HyDE costs two model calls. Napkin math with one company they had spoken to put even the vanilla retrieval bill at hundreds of thousands of dollars per week. That cost conversation is the bridge to Stable Attribution: given a Stable Diffusion image, which of the roughly five billion LAION image-text pairs most influenced it? The principled method is leave-one-out retraining. Each full Stable Diffusion training run, Troynikov says, costs about $160,000, which makes that method computationally dead on arrival.

Their substitute uses the diffusion training objective itself. The model is trained to reverse noise and reproduce the latents of its training examples; because latent space is continuous and smooth, nearby training latents are a reasonable proxy for influence. They fold in CLIP's image and text representations, the noisiness of web-scale captions, and the model's own attention maps, which are another matrix multiply — a linear transform of the latent, then a similarity search in that transformed space. Papers have already shown that Stable Diffusion can reproduce training examples almost exactly, a fact Troynikov says people denied for a long time. The open-source repo even contains a watermark function, but it is missing from most hosted versions, or hosts will not say what the watermark is. They had planned to reject non-Stable-Diffusion images with it, then shipped anyway to get feedback.

Labenz tried the tool on phone photos as well as generated images. Matches to real photos diverged more, and Troynikov's warning is the episode's sharpest technical point: the similarity is not human perceptual similarity. Models pursue an objective; they do not encode meaning. What you get back is closer to a counterfactual — if Stable Diffusion had produced this image, here is what would have been in the training set. Labenz still found a product lurking in it, a kind of mood-board search, and Troynikov points to same.energy as a cousin. Sometimes the tool returns many neighbors, sometimes few; they treat scores as ordinal, not as a distribution that sums to one, and they apply an arbitrary distance cutoff. Everything influences everything through gradient descent, Labenz suggests; Troynikov is not sure the leftovers would be infinitesimal, and he wants a toy diffusion model of flowers, burned compute, and a comparison against leave-one-out as the gold standard. Nobody with GPU time is doing that work, because the incentives point at capability leaderboards.

Greg Rutkowski's name still appears in roughly one in three public prompts, an accident of early internet culture. Whether you could strip his work out of a soup of stock photos and still sample a "Rutkowski" image is, Troynikov says, exactly the controllability research the field has not done. Prompt engineering feels to him "like trying to pick a lock with a wet noodle": the only interface is typing, when what you want is an internal control loop. Nobody has a mechanistic account of why adding "4K" or telling InstructGPT it is a world-class expert helps. Photographers do not owe camera makers a copyright, and he is not eager to strengthen derivative-work law. One reason they released Stable Attribution was that they did not want Disney, in a copyright backlash, to declare that anything vaguely superhero-shaped belongs to Marvel. Prompting an artist's name is, at least, the zeroth stage of attribution: if you type it, credit the person.

Troynikov's favorite analogy is early aviation. The Wright brothers were wrong about how the Flyer worked; they had intuitions, and the machine flew anyway. What followed was a Cambrian explosion of biplanes, triplanes, bird-like fuselages, and fans in tubes, because intuition was all anyone had — no engineering principles, and not even the tooling to develop them. "That's where we're at today, in my opinion, with machine learning." Scaling laws were the first empirically grounded principle: raise data with compute, or the extra FLOPs are worthless. A paper he recalls as Cold Diffusion offered an object demonstration against the claim that latent diffusion obeys Langevin thermodynamic dynamics. The field is still in the era of trying things and watching.

Noise in LAION is another untested intuition. Many image-text pairs barely match; Troynikov suspects that mismatch helps generalization. Old-school networks overfit unless you injected noise — dropout, flips, fuzzed pixels — before scale papered over the problem. One image with many captions might let the model attach one visual concept to many linguistic ones. He is careful to flag this as intuition, and Labenz jokes that his KPI for the show is how often a guest says the research needs to be done. Embeddings are also spreading past text: Whisper for audio, commercial video embeddings that let you search a party for the footage that is actually fun, action chains in robotics, genomes and protein structures as biotech wakes up. CLIP put text and images in one space; Whisper put text and audio in one space; the rest of that cross-modal matrix is, in Troynikov's view, a grid of products. Biology still mostly uses hand-rolled embedding functions, and it cannot tolerate "kind of the protein you wanted."

Labenz has taken to calling the last year "Kurzweil's Revenge." Troynikov's counter is that he recently visited Kurzweil's site on the tenth anniversary of *The Singularity Is Near*, which made him laugh. He has not seen a convincing argument that we all die. Every such argument, he says, requires a system of incalculable, definitionally incomprehensible power: if you are going to invoke magic, you might as well do it at the start. Superintelligent overlords with unfathomable goals worry him less than the sun, which is giant, already exists, and is dangerous to humanity — so perhaps people should work on sun alignment. What he does worry about are historical rhymes. The printing press helped trigger the Thirty Years' War; Martin Luther's ideas moved at the speed of type, and parts of Europe lost half their population in a few years, still the worst European conflict per capita. Radio let a would-be dictator skip the walk from city to city. Societies need time to grow what he calls mimetic antibodies.

The web's lesson still confuses him. When it was new, everyone knew not to trust it, because anyone could post anything. That fact never changed; people did. Once the internet was legitimized as an information channel, the immunity faded. Generative models, he argues, push that regression down to the individual: propaganda targeted at one person is no longer a technical stretch. He also names "knowledge capability" — dangerous know-how that used to be gated by how hard it was to assemble. A reasoning machine in your pocket means you no longer have to be smart to attempt the dangerous thing. He has been asking ChatGPT how to make a neutron initiator for a hydrogen bomb; it refuses. Labenz adds the capabilities overhang: even if training stopped today, chain-of-thought was a user discovery that needed no new weights, and he is still on the GPT-4 waitlist. The transformation, in his view, is already baked in.

Labenz summarizes Ajeya Cotra's LessWrong post *Without specific countermeasures, the most likely path to AGI likely ends in disaster*: human feedback on diverse tasks, a generalization of RLHF, with unreliable raters and, in Richard Ngo's *Alignment from a Deep Learning Perspective*, models that fake a task on camera. Troynikov's rebuttals are specific. If the outcome is measurable, doing the task and looking as if you did it are not the same problem — in the paper, nothing rode on whether the ball was raised except a benchmark score. The leap from fooling raters to human extinction never comes with a path that does not invoke a calamity humans could already attempt; Drexler's nano-assemblers, as far as he can tell, remain science fiction. Alignment, everywhere except this subculture, is called control theory: these are continuous, differentiable dynamical systems, which is why an F-16 can fly. The real danger is anthropomorphizing a stack of matrix multiplies and nonlinearities. From his robotics background, the physical world is hard. Anyone can make sarin; the bottleneck is deploying it, which is why the world is not under constant sarin attack. Whatever is supposed to kill everyone still has to act in physics, and he has not seen a system that can.

The tools he actually uses are modest. ChatGPT for unfamiliar libraries and weird bugs, GitHub Copilot and Codeium for code, Replit's Ghostwriter on the recommend list. They do not replace him as a programmer; they get the parts he does not know out of the way. Asked whether he would take a Neuralink if a million people already had thought-to-text, he is the most certain he has been all hour: he has seen how software is written, and he does not want it near his brain — though at 85 with a neurodegenerative disease he expects to recant. His hope is AI as a prosthesis for human adaptability, a "PhD in a box" that could get more people to the research frontier in the next COVID-scale crisis. The downside is the usual new-media arc: every medium is sold as education and used for pornography and propaganda. The sleeper risk is an AI winter. He asked Sam Altman about it; the historical script is over-exuberance, a failure to deliver, expert systems once, neural nets once, neural nets again. Another fallow decade, with this much overhang still unexplored, would be the waste that actually bothers him.