Jeremiah Lowin: Explaining the New AI Paradigm - [Invest Like the Best, EP.307]
What this covers
This is a conversation between Patrick and his longtime friend Jeremiah Lowin, founder and CEO of Prefect, recorded after the release of ChatGPT. Patrick wanted a primer on what happens under the hood, so the discussion starts technical and widens into business. Lowin begins with how a model is built: an architecture plus weights, trained from random noise by repeated nudges against a score. He argues the core pattern of input, latent representation, output has barely changed in 15 to 20 years, and that progress has come mostly from scale and refinement.
The middle stretch covers why Stable Diffusion outperformed DALL-E: it runs a diffusion model in latent space instead of pixel space, so it can settle details like five fingers last. Lowin contrasts older GANs, where a single number dialed a smile or glasses, with today's larger spaces, and he contests the idea that these models are black boxes. He also pushes back on the view that capital is the bottleneck, putting roughly 4,000 GPUs at about $40 million a year and naming time, intuition and iteration as the scarce resources. A short history of the companies involved and a retelling of Google's link-ranking insight set up the later question of where moats can exist.
The final section turns to practice and prediction. Lowin expects a few base models to be fine-tuned by industry, regards today's prompt tricks as temporary, and says adding AI features to existing products yields indefensible startups. He sees customer support as a strong target, treats the electricity-utility analogy for AI APIs as flawed, and dismisses fears of AI taking over the world as overstated. The real danger, in his view, is human misuse.
Lowin argues that modern generative AI works by mapping inputs into a latent space where concepts become separately manipulable, so the models are interpretable in principle, and that durable defensibility comes from proprietary data and the guidance and interface layer rather than from capital or raw model scale.
- Latent representations let concepts be separated and adjusted, which is why diffusion-in-latent-space outperforms pixel-space diffusion
- Capital and GPUs are not the binding constraint; time, intuition, and iteration are
- Startups should target workflow problems with proprietary data, not add AI features to existing solutions
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
The core idea of modern machine learning, taking an input, transforming it through a latent representation, and outputting it, has not dramatically changed in roughly 15 to 20 years; the progress has come from making models bigger and from refinements and enhancements of the architecture.
“the core idea has not dramatically changed. We take an input, we run it through some sort of latent representation”
In a well-built latent space each concept (such as hand count, character color, or watermelon color) sits in its own numeric dimension, so changing one number, for example turning a watermelon from green to orange, changes only that attribute in the output and makes the output step far easier.
“What that does is it leaves each concept isolated. If we did a perfect job of this, then we'd actually have a string of numbers where the first number represents”
Generative models can produce garbage as easily as high-quality images because both reside in different regions of the same latent space, so the model must learn what bad outputs look like and users must steer away from them.
“These models are as capable of generating garbage as they are generating high-quality things. That's just a different area of latent space.”
After a training pass, one can mathematically determine which of the model's billions of numbers would move the output in which direction, so one can nudge weights to make noise slightly closer to the objective, though one cannot hand-design a specific picture such as Spider-Man by tweaking numbers.
“we can actually go back to each of those 100 billion numbers and figure out which one would impact the output in what way.”
To turn a general language model into a specific product, first teach it the desired answer style and format through in-context examples, which does not change the model, and then optionally fine-tune it on proprietary data such as user questions, responses, and internal design documents.
“That's not training. That's like through examples. That doesn't change the original models, through examples.”
Capital is not the binding constraint on any one iteration of generative AI: roughly 4,000 GPUs at about $40 million a year is a manageable cost for a well-funded company, and the true constraints are time, energy, intuition, and ingenuity, since research requires iterative bets that take about a month each.
“Time, energy and intuition and ingenuity are, I think, really the constraints on this, really wanting to improve this and wanting to perform with it more so than just raw capital”
Stable Diffusion beat DALL-E because it runs diffusion in latent space rather than pixel space; this frees the model from the spatial constraints of the image during generation, so it can merge image content first and then settle details like five fingers at the end, which gave it much greater expressive power.
“That little, little insight about moving the diffusion from the pixel space vector to the latent space vector is, for all intents and purposes, the chief difference between these two models”
Modern diffusion latent spaces are too large for a single slider per attribute, but this does not make the models less interpretable; it only makes them more complicated, and a machine can find the combination of numbers that yields an attribute such as a bigger smile.
“However, that doesn't mean that the model is less interpretable. It just means that it's more complicated. It works in exactly the same way.”
Google's original insight was not matching the word car to pages containing it, but ranking pages by how many other pages pointed to them, which is why its search took users to the right page rather than merely matching keywords.
“It's that I go to the page that other pages where the word car pointed to the most.”
A model consists of an architecture plus parameters called weights, which may number 10 billion or 100 billion; training starts with random weights that produce noise, then repeatedly nudges the weights based on a score of how good the output is until the outputs become sensible.
“Training is the process of changing those weights, literally the numbers. Those potentially billions of numbers changing them so that they encode information”
Going directly from text to pixel coordinates, which expand from a few bytes of characters to thousands or tens of thousands of pixel coordinates, is too large a transformation to be done with precision, so models must pass through intermediate latent representations that preserve the input's meaning.
“And if we try to go immediately from one to the other, we'll almost certainly fail. That mapping is just impossible with precision.”
The compute used to train Stable Diffusion's original model, around 4,000 GPUs, reportedly made it roughly the 10th most powerful supercomputer in the world at the time, which shows that the scale is large but not astronomical for a well-funded company.
“the number that the Stable Diffusion trained their original model on. And the reason it's interesting is I think at the time, it made it the 10th most powerful supercomputer in the world.”
The simplest improvements to a model come from more data and more training time, while meaningful new model versions usually require new architectures or training methods, which are riskier research bets.
“There's two easy ones, more data and more training. Just look at it for longer.”
Google's search advantage cannot be based on public web data because anyone in principle can crawl and index every public page, so a durable moat requires data that is not publicly accessible or that is collected in a uniquely useful form.
“since the web is obviously public and anyone in theory could go and index every page on it.”
ChatGPT's quality owes much to a new, more efficient and data-light form of reinforcement learning in which other AI models score outputs and supply reinforcement signals, which converges faster than human-judged training.
“OpenAI came up with a new, more efficient, much faster to converge and requiring less data form of reinforcement learning.”
Current AI is far from general intelligence and could barely function outside situations it has seen before, so the threat of an AI taking over the world is overstated, and the real danger is humans misusing the technology.
“this thing could barely power a toaster if it hadn't seen one before, this is not really the threat that you think it is. It's going to be humans who misuse this that pose a threat.”
Earlier GAN models were easy to explore because their latent dimensions corresponded closely to observable output attributes, so users could literally dial smile, hair length, glasses, or gaze direction by changing a single number along that dimension.
“that in that latent space, we were able to identify dimensions that gave us specific sliders for a smile.”
The temptation with new technology is to add AI to solutions you already see, like AI avatars or AI features that deliver cooler versions of foundational model output; these are indefensible and will not become startups because they gather no proprietary data or information.
“instead, what you need to go look for is a problem.”
The most likely way generative AI reaches the world is that a few canonical open-source or API base models get fine-tuned by individual companies and industries into specialized versions, since fine-tuning a base is cheaper and more effective than prompting the original model for a specific style or task.
“I would be shocked if that's not a major, major, major way that this shows up in the world.”
Current AI is very good at well-defined situations that contain some randomness or idiosyncrasy, such as writing boilerplate code or driving a straight highway, but it is not yet reliable at anticipating every edge case, and expecting that is a mistake that will leave people disappointed.
“They are really, really, really good at taking well-defined situations that have some degree of randomness or idiosyncrasy to them and navigating them in a really, really, really solid way.”
Customer support and customer success are ideal targets because they involve two-way friction, idiosyncratic problems that need customized responses, and frustrated users who want fast, accurate answers, and their work cannot scale linearly with customer growth.
“I go look at very boring places where a lot of time is spent figuring things out and the first thing that comes to my mind is customer support and customer success”
Today's prompt engineering, which requires appending phrases like 'trending on ArtStation, 8K' and long lists of negative prompts, is a temporary workaround; the future product should guide models to high-quality outputs from short, precise inputs, as good Google users do with few words.
“That is going to be a major opportunity that it hasn't yet unsolved of how you guide to a high-quality outcome without requiring users to jump through all these hoops”
The advance in processing text into a latent space, where text can be transformed into a different language, synthesized voice, or image, is a huge unlock; image generation would be nearly useless without it because it could not reliably map words to semantically aligned outputs.
“In a funny way, these image generation technologies would be nothing if they couldn't figure out what to do with the words that are being put into them.”
The electricity-utility analogy for AI APIs is imperfect because a power outlet is one-way and delivers a raw input, whereas an API is two-way: you send a request and get back a result, which means you are outsourcing business logic rather than drawing raw power.
“APIs are more interactive than power outlets.”
Whether AI behaves like a utility depends on what is produced: an API returning finished images or text is too close to the end product to be a utility, while producing weights or latent representations would be a raw material but would be incredibly difficult to use.
“So is the utility producing an API that takes in text and returns images, then probably not because that feels too close to the end product.”
When a new technology arrives, it is tempting to bolt it onto existing solutions; an entrepreneur should instead start from a real problem, especially one where generating content or communication is hard and frictions are high.
“So we need to find the problem to solve. And I think when you have a new technology like this, it's extremely tempting to add to the solution that you already see.”
Generative models feel more powerful than narrow models, such as face generators or envelope address readers, because users can control the output in a way that feels personally meaningful and 'mine', which drives engagement even though the narrow models are more mature.
“these generative models are really, really, really powerful because they feel more powerful. Because I as a user can control what's happening in a way that's meaningful to me”
Defensible AI businesses need two sources of advantage: proprietary data that others cannot access, such as non-public customer data, and secret sauce inside the models or their guidance layer that steers outputs toward high quality.
“There has to be 2 places that the secret sauce comes from. One comes from a proprietary data set that you simply can’t access otherwise.”
To predict whether AI will let a non-expert use toolkits like Unreal Engine 5, one must separate the underlying technology from the user interface, since AI may be integrated into tools like Photoshop or accessed through simple UIs, enabling a user to project a castle with a dragon onto a tabletop.
“I think we need to separate the technology from the interface in order to answer that question.”
Understanding a generative model is like looking at clouds: the output resembles something, and humans map it onto an understandable space, and diffusion models do this repeatedly, nudging a shape toward a duck or water until a tangible feature emerges.
“One of the best ways for people to think about what this is and how interpretable it is, to me, is looking at clouds.”
Midjourney consistently produces high-quality images, which is probably not because its training data excluded bad images but because it has built some architectural or guidance mechanism that steers outputs toward high quality, and that mechanism is its secret sauce.
“I suspect that there's something that they've done there to guide always to high quality.”
Machines can identify the combination of latent numbers that changes a feature like smile, even when no single human-readable slider exists, and this is shown by prompting an image model to add a big smile to a scene and getting a very similar image with only that change.
“a machine can look at that vector and figure out what combination of numbers make someone smile more or less”
Building a chatbot over 1,000 Colossus podcast transcripts requires two steps: pointing the model at the corpus, and teaching it what a good answer looks like, so that it synthesizes answers, attributes speakers, and flags disagreements rather than merely linking to episodes.
“the principal thing here is not just the search engine replacing, not just the natural language search engine replacing”
GitHub Copilot does more than autocomplete a variable name; it writes lines and lines of code it believes you intend, and its quality surprised the speaker enough that it now feels like a pair-programming partner that frees him to focus on the unique parts of his code.
“I’m pair programing with AI in a way that I absolutely didn't expect to at this stage.”
Neural rendering methods such as NeRF encode lighting, reflections, and scene detail in a latent space, which can capture and represent 3D scenes more easily and quickly than explicit computation in engines like Unreal Engine, and could enable interactive video experiences like walkable scenes or realistic video calls.
“a neural renderer uses an AI to encode that all in a latent space and can do it more easily and more quickly.”
Much of the expertise in advancing generative models lies in intuition about how nudging a model in one direction or changing its architecture will deliver results, since each new architecture is a costly research bet with no guaranteed payoff.
“So I think a lot of the expertise is actually somewhere in the intuition of being able to look at a model and understanding if I nudge this in a certain direction.”
Prefect has put a GPT-powered bot in its Slack to handle roughly the first 80 percent of user questions that are answerable from documentation or internal knowledge, so human support staff only handle the rest and the team avoids linear headcount growth.
“We've already experimented with putting a GPT power bot in our Slack to handle, let's call it, the first 80% of questions”
Reinforcement learning is especially tricky because a model takes a series of actions and receives feedback only at the end, forcing designers to decide whether intermediate stumbles should be penalized, which makes assigning credit over sequences of actions difficult.
“How do you actually reinforce a series of actions over time.”
Whatever models are considered state of the art today will be obsolete within 12 months, not because of more training but because of new architectures, new applications, and completely new approaches.
“The things that we're calling state-of-the-art today will be obsolete 12 months from now.”
ChatGPT returns essentially the same answer when the same question is asked repeatedly, which looks like it is keying into a very specific node, good for information retrieval but a concern for whether the system is a true conversation partner or just a very good search engine.
“If you ask it the same question over and over, you don't just get the same answer, you get the same answer.”
A pretrained model can take in any known input type, including text, images, or time series, and convert it into a latent representation that fully describes the input before deciding what output form to produce, so the output format is not fixed at input time.
“It can take it in and inbuilt, but will cause latent representation, which becomes this really powerful information set that, in a sense, fully describes the input sequence.”
Image model training is very slow and iterative, and for a long time you may not see any visible output even though the math shows it is working; the DALL-E Mini checkpoints produced nonsense for early periods and needed weeks before recognizable shapes emerged.
“And for a long time, you don't even know if it's working. The math tells you it's working.”
Diffusion models begin with random noise and iteratively pull an image out of it over many stages, gradually refining the output toward something recognizable.
“They start with random SMA iteratively pull an image out of it over many, many, many stages until they have something real.”
Transformers work by attending to what is most important in a sequence and applying context so that a decision made later in the stream can depend on what was seen earlier.
“paying attention to things that are more important literally, like focusing on things that are more important or applying context in a way that lets you make a decision later in the stream based on what you saw earlier.”
Because Stable Diffusion is open source, no single company can build and control a full stack around it, which produces a massive amount of experimentation and early business successes, including paid AI avatar products.
“First of all, it's amazing that the Stable Diffusion model is open source because it means that you really can't just build a full stack thing and control it”
An open-source DALL-E clone released shortly after DALL-E 2 was announced could be watched improving over weeks through checkpointed models run on identical prompts, making the training process visible in real time.
“you could watch this model train over time because literally, it was training over time. And it is significantly better today than it was at the beginning.”
Transformer models are a class of models that transform one sequence of inputs into another sequence following very complex learned rules or heuristics, with translation as the most familiar example, because word-for-word dictionary substitution fails due to grammar, gender, syntax, and idioms.
“They're called transformers because what they fundamentally do is transform some sequence of inputs into some other sequence of inputs following a very, very complex series of rules or heuristics.”
Generative models' capacity to produce specific outputs arises because the model embeds learned concepts in a latent representation rather than encoding explicit rules, so no one needed to hand-code instructions for something like Spider-Man smashing a watermelon.
“It's so obvious that that's not what's happened when you see some of these outputs.”
The first class of problems AI will absorb are those where generating content is hard, such as marketing copy, text, and image-based communication for architects, visualization, and theme park designers, though startups should avoid head-on competition with incumbents like Disney.
“the first class of problems that will be subsumed are the ones which is generating stuff is hard”
A key near-term development is making generated video and 3D scenes interactive, so users can walk around them, which would enable video calls without webcams and interactive virtual worlds projected into the real world.
“how do we actually take scenes, entire scenes and make them interactive. How do you walk around something?”
Foundational generative AI research started at universities such as the University of Toronto, the University of Montreal, NYU, and Stanford around 2009, then moved into commercial work at large cloud providers like Google and Facebook, and now dedicated AI companies such as OpenAI, Stability AI, and DeepMind are leading.
“I would throw a lot of universities in as well. When I started getting involved in this, which is probably in 2009, I'd be looking at the University of Toronto, University of Montreal.”
A model fine-tuned on Van Gogh paintings will produce only Van Gogh-style output and likely does it better than prompting a general model with 'in the style of Van Gogh' appended to the prompt.
“And it probably does that better than if you just tack on, in the style of the Van Gogh to a prompt the original model.”
Generative AI took off like a popular idea that many people discover at once, similar to how two animated films about ants and bugs were released together, so the surge reflects a good idea being seized by many developers rather than purely a technical unlock.
“When a good idea is out there, like a lot of people are going to find out and jump on to it.”
Large AI companies may primarily provide the training of canonical foundational models over time, while every company and industry builds its own personalized version on top, though it is too early to know what role big players will play.
“it could be that they just, huge air quotes around just, that they just provide the training on the canonical models over time”
Latent space is a substrate that lets a person's latent idea be externalized through many possible forms, such as words or drawings, and machines can accelerate that translation from thought to output, which is why this technology is so powerful and reduces friction between minds and outcomes.
“That's the latent space, and that's why it's so powerful because you can choose anyone of 1,000 different ways to turn that thought into words”
A search engine like Google can be a utility in the sense that people rely on it for answers, so the utility metaphor applies more clearly to conversational AI that answers questions than to a product that merely outputs finished images.
“a search engine would be a good example of this. Is Google a utility that provides links? If so, then absolutely”
A near-term major achievement in image generation is that all hands will have five fingers, a marker of fidelity that current models still struggle with, and this is the practical benchmark of progress in visual generation.
“right now, we're on the verge, I think, of a major achievement in image generation, which is all hands will have 5 fingers.”
The episode was recorded because ChatGPT's release prompted a primer on what happens under the hood of generative AI.
“But when OpenAI released ChatGPT, I called Jeremiah for a primer on what's happening under the hood”