No Priors: AI, Machine Learning, Tech, & Startups Sep 18, 2026 38m 23m saved
With Stefano Ermon, co-founder and CEO of Inception and a longtime Stanford professor, whose lab produced the score-based work that became modern diffusion models
Inception is about two years old, employs around 50 people, and is serving language models built on diffusion rather than next-token prediction to paying customers today.
Every large lab has bet text and code on the same architecture: an autoregressive model that produces one token at a time, left to right. Stefano Ermon took the other side of that bet, on the grounds that the economics of AI will be settled at inference time rather than during training.
"I doubt we've discovered the best way of building these systems. There's got to be alternatives."
Ermon has been building generative models since he arrived at Stanford as an assistant professor in 2014. The score-based work he started in his lab in 2019 with a PhD student became the basis of modern diffusion models; he also co-advised the flash attention work, and DPO, the alignment algorithm now used across the industry, began as a rotation project in his group. Stefano Ermon, co-founder and CEO of Inception, on No Priors.
The full interview is covered here so you can skip it. 38 minutes of audio, 15 minutes of reading.
Here are the 14 insights that matter.
Key Takeaways
Generative modeling was so unfashionable in 2014 that papers had to be justified as feature learning for supervised tasks He expected the problem to occupy his whole career, and said he was "very wrong"
Diffusion now wins every continuous format — images, video, music and much of the protein work
Autoregressive generation is memory-bound, so it uses a GPU badly: the tenth token cannot start until the first nine are done
Inception's Mercury models match the speed-tier frontier models on benchmarks while running significantly faster
A voice-agent customer moved off Cerebras's custom chips and got comparable speed from a diffusion model on Nvidia GPUs
The moat is the stack around the model, not the model — without a serving engine, a trained diffusion LLM cannot be deployed at all
Roughly 20–30% of the use cases he counted are latency-critical, which he treats as a lower bound on the addressable market
A diffusion model can be steered halfway through generation, where an autoregressive model has to finish before it can be scored
Compute, not headcount, is the binding constraint on a 50-person research company
Flash attention, DPO and diffusion itself all came out of university labs, which is his answer to the claim that academic AI research no longer matters
1. A Field Nobody Wanted
Ermon was asked to start with his research background. He arrived at Stanford the year before the first generative-model boom and spent his early years on a subject that could barely be published.
He started at Stanford in 2014 on generative models
I started at Stanford in 2014 as an assistant professor and I was working on building generative models.
Stefano Ermon
Success at the time meant producing grainy images of handwritten digits, and a paper on the subject needed a justification attached to it: that learning a generative model was a way to extract features from unlabeled data, which might then improve performance on a supervised task. Supervised learning was what the field cared about.
The subject was not fashionable
Back then the research area was not particularly hot.
Stefano Ermon
Asked whether he had a personal hope for the models beyond curiosity, Ermon said his motivation was world models rather than language. A system that can picture what happens if it stands up and walks out the door has something it needs for making decisions, and building that picture requires generative capability. He assumed it would take the rest of his career.
He expected a much slower field
I felt like this is going to be very hard as a problem is going to keep me busy for my whole career and it's a good problem to work on and then of course I was very wrong
Stefano Ermon
2. The 2019 Denoising Bet
The path ran through autoregressive image models, which were slow and blurry, then variational autoencoders, then generative adversarial networks. GANs worked but were unstable to train and hard to reproduce, and Ermon's lab set out to find something with the same output quality on firmer theoretical ground. The idea they landed on was to train a network to remove noise from an image.
Denoising was the whole insight
You start with an image you add noise and then you learn how to remove the noise.
Stefano Ermon
A network that can do that has enough of a grip on the structure of an image to generate one: start from pure noise and refine until the picture is clean, instead of filling in pixels left to right. That became the technology under the current generation of image tools, and it spread to every continuous format — image, video, music, and a good deal of the protein work.
3. Diffusion Meets Text
Getting the same machinery to work on text and code took years of separate research, because diffusion's assumptions are about continuous data. Two pixel colors have colors between them. Two words do not.
The problem with discrete data
But if you think about two words, there is not necessarily something in between them, right?
Stefano Ermon
The breakthrough came in 2024, when the lab published a result matching an autoregressive model of the same size on perplexity — under a billion parameters, so still an academic result — with the same training data and, because it emits many tokens at once, much faster generation. Ermon then started Inception to find out what happens at commercial scale.
4. Why Inference Favors It
Ermon's argument for diffusion is not about quality. It is that the industry's costs are increasingly set by generation rather than training, and that autoregressive generation uses the hardware badly. He drew the parallel with 2017, when the field left recurrent networks for transformers because a transformer could process many tokens at once during training.
Generation is still strictly sequential
You cannot generate the 10th token until you've generated everything that comes before it.
Stefano Ermon
Which makes it memory-bound rather than compute-bound
That kind of workload is extremely memory bound.
Stefano Ermon
Most of the time goes on moving weights through the memory hierarchy and very little on arithmetic. A diffusion model, by contrast, produces many tokens simultaneously at inference, so its generation workload looks like a training workload — the large matrix multiplications that a GPU is built for. Three things compound that advantage, on his account: the economics of the industry are set by intelligence per watt and per dollar; reasoning models work by spending more compute at inference; and reinforcement learning after training is bottlenecked on generating the model's own attempts before they can be scored.
He is betting on the more parallel design
it's inherently more parallel and the bitter lesson is that the more parallel solution is the one that is eventually going to win
Stefano Ermon
Which architecture eventually handles every format at once, he said, is not settled. Both paradigms are live and nobody knows the answer.
5. Mercury vs the Minis
Asked what he can claim for the models now, Ermon compared them to the fast, cheap tier the large labs sell rather than to their flagships.
Parity with the speed tier
So our Mercury models are on par with the haiku models, flash models, mini nano models from OpenAI if you look at benchmarks while being significantly faster.
Stefano Ermon
Getting there required building infrastructure that did not exist. A diffusion-based LLM cannot run on vLLM or SGLang, the standard open-source serving systems, so the company wrote its own serving engine and handles production workloads on it.
The prototypes are in production
We think it works really well.
Stefano Ermon
6. 50 People, Two Years In
The company is about two years old with roughly 50 people, and most of the effort is still research: how to train the models, and how to accelerate inference. That second question is open in a way it is not for an autoregressive model, where generation is one token at a time and there is little to tune. A diffusion model offers many ways to trade compute against quality — distillation, and the differential-equation solvers that let image and video models sample in a few steps.
Where it can, the company reuses what exists. The models are still transformers, still use attention, and are trained on public datasets and evaluated on public benchmarks. The differentiation today is speed.
He does not claim to know what comes after speed
I doubt we've discovered the best way of building these systems. There's got to be alternatives.
Stefano Ermon
The host put the industry context around that: efficiency research was real but second-order while the field was chasing pure capability, and in a power- and compute-constrained environment one of the most consequential decisions a company makes is how to split the watts it has between training and its most valuable workloads.
7. Where Speed Wins Today
Asked where customers care enough about latency to switch, Ermon named the applications where it is part of the product, then made a more general claim about buyer behavior.
Buyers are paying a premium for the fast tier
People are willing to pay more to get access to faster models, right?
Stefano Ermon
And speed is not something users give back
And I think once you get used to a fast model, it's hard to go back.
Stefano Ermon
His comparison was broadband.
8. Off Cerebras, Onto GPUs
The customer he was able to name is OpenCall, which builds voice agents. A voice pipeline runs speech recognition, then a language model doing tool calls and deciding what to say — a reasoning model, if quality matters — then text-to-speech, and every stage adds delay the caller hears.
They had bought custom silicon to get the latency down
So they were using custom chips to get to the kind of speed that they need to deliver the best experience to their customers.
Stefano Ermon
They switched to Inception's diffusion-based models, which Ermon said reach comparable speed on ordinary Nvidia GPUs because the parallelism is in the model rather than the hardware. The advantage he claimed is supply as much as price.
Scarce is still better than bespoke
GPUs are scarce, but there's more of that than custom chips
Stefano Ermon
He does not treat this as software beating hardware. The host raised the demand for large chips with a lot of on-chip memory for coding workloads, and Ermon's answer was that the two stack.
The gains multiply rather than compete
to some extent the gains that you get from the software they are multiplicative with the gains you get from the hardware
Stefano Ermon
Someone may eventually build hardware shaped for these models specifically, he said.
9. The Moat Is the Stack
The host put the standard objection to funding new architectures: any real advance gets absorbed by the companies with the compute to scale it. Ermon agreed the risk is live and said it is on his mind. His first answer was trade secrets — what his researchers know about training and accelerating these models. His second was the reason the company sells a product instead of publishing papers.
Without the serving layer the model is unusable
even if you train the diffusion based LLM, if you don't have the serving engine, if you don't have the VLM equivalent to serve it, you're still stuck and you still cannot use it
Stefano Ermon
Real customers also produce evaluations and data the company would not otherwise have, which he counts as part of the technical moat because it is harder to copy than an idea.
The same reasoning produced the decision not to open the work up. Inception built its own stack for supervised fine-tuning and for reinforcement learning from human feedback, and keeping it closed keeps the intellectual property close. He was straightforward about the cost: less contribution from outside, harder adoption, and on-premises deployments are difficult. The immature ecosystem cuts the same way — there is little open source worth using, kernels and serving had to be written in house, and customers are being asked to try something unfamiliar.
10. Messy Data, Same Math
The host pressed on a structural point. Diffusion does well on images, video and voice, where the data reflects a physically consistent world. Human-generated training data, code especially, is messy and may not carry the structure the model is looking for.
Ermon's answer was that both paradigms are doing the same job. Training any generative model means finding an efficient compression of the data, and the more it compresses, the more structure it has found. Perplexity — the metric in the 2024 parity result — is a measure of exactly that, so matching an autoregressive model on perplexity at that scale means identifying the same amount of structure.
Noise is not the distinguishing problem
Yeah, there is noise in everything.
Stefano Ermon
And the measurement settles it
the numbers don't lie
Stefano Ermon
If perplexity comes down, the structure was there and the model found it. What remains is a question about inductive bias: whether a transformer is the better instrument for finding those patterns, and whether next-token prediction or denoising is the better framework. That, he said, is empirical, and the field does not yet have the tools to answer it.
11. Steering Mid-Generation
Because a voice company was the example, the host asked about control. Teams building these products have learned how to keep an autoregressive model doing what they want, and a different architecture threatens that work.
The practical answer is that Inception built for compatibility: the same text-in, text-out interface, OpenAI-compatible, and models that follow instructions and emit structured output such as JSON, so the customer's existing harness kept working. Ermon said the models were in fact better at this than what the customer had been using.
The more interesting claim is about what diffusion makes possible. An autoregressive model has to finish the object before anyone can tell whether it satisfies a constraint — generate a whole molecule, then score it for solubility. A diffusion model goes from coarse to fine, so the generation can be pushed toward what a reward function wants while it is still happening.
The academic evidence points one way
there is a lot of evidence in the academic literature at least that diffusion models are easier to control
Stefano Ermon
That opens a product surface an autoregressive model cannot offer, he said, and what to build on it is something the company is still working out. Asked what other capabilities might appear at scale, he was careful to say nobody knows, and that speed was the initial bet because it was the easiest thing to measure. One candidate is efficiency with data: training by denoising shows the model many noisy versions of the same image, which amounts to free data augmentation.
Data efficiency is the other reported advantage
there is a decent amount of evidence in the academic literature for example that diffusion based models are more data efficient
Stefano Ermon
If that holds at larger scale, he said, the interesting cases are the ones where data is the constraint.
12. 20–30% Wants Low Latency
Asked for the split between diffusion and conventional models two years out, Ermon was explicit that the frontier of intelligence is not his claim, and that many workloads need it. He sized the rest from OpenRouter's published taxonomy of what people use models for — research, conversation, coding and software engineering, log processing.
His estimate, offered as a floor
I was doing some estimates and I think there is like between 20 and 30% where latency is really important
Stefano Ermon
Within a fixed latency budget, he said, the model that gives the best quality is the one that wins that share.
13. Compute, Not Headcount
On how a 50-person company covers research, serving and product at once, Ermon said agents have made a small team much more productive, and named the constraint that actually binds.
The limit is not people
I feel like the bottleneck is more compute than people.
Stefano Ermon
The organization is split in three: a product team working with customers, a team serving the current best model, and a team building the next one, which is where training, reinforcement learning and inference research sit. He would not discuss model sizes or the compute behind them.
The training details are a trade secret
we're not able to share much about the training the size of the models or the flops or all of that it's kind of like a trade secret but we are continuing to push the frontier
Stefano Ermon
His recruiting pitch is that the field is small enough that a researcher can still shape it, and that the people who choose Inception over a larger lab want ownership and open problems rather than a known stack.
On whether models can now improve AI research by themselves, he said the answer is not yet. His team uses frontier models heavily and they have accelerated the rate at which ideas can be tried, and he would not forecast six or twelve months out.
Judgment is still the scarce input
right now the human ingenuity is still like super important
Stefano Ermon
14. Academia's Track Record
The last question was whether university research can still matter when scale decides outcomes, which Ermon answered with a list from his own lab. The early diffusion work beat GANs at academic scale while being more stable to train, and became the basis of the commercial image models. He co-advised flash attention. DPO started as a rotation project in his group and is now used to align language and diffusion models across the industry, on the back of a mathematical observation about the structure of the problem.
What academia still has
One of the nice things about academia is that it allows you to take these contrarian bets.
Stefano Ermon
Resources are short and always will be, he said, but the students are excellent and nobody there is afraid of a bet that might not work.
His summary of the record
So there are gems there are lots of opportunities for finding new and better ways of solving important problems.
Stefano Ermon
Bonus Insights
Generated short-form video is already mainstream entertainment in some markets. The host's framing for why diffusion is not a niche was that diffusion-made short-form video has become a dominant form of entertainment in other parts of the world and will likely become one in the United States, which is also why hardware is being built to run it faster.
Nobody knows which architecture ends up handling every format at once. As models become more multimodal, Ermon said, the field expects a single model that handles all formats and knows everything about the world. Whether that model is autoregressive or diffusion is unsettled, and he said plainly that the jury is still out.
Speed was the wedge because it was the easiest thing to prove. He was explicit that the more interesting possibility — that a diffusion-based language model turns out to be more capable, not merely faster — is unknown, and that he took the measurable bet first.
The AI factory is not a solved design. Asked about efficiency, Ermon said he does not think anyone really knows how the economics of large-scale inference will work, and that being able to explore alternative ways of producing intelligence is itself the reason to be in the field.
Ermon's bottom line is that the industry has standardized on one architecture for text and code before anyone established it is the right one, and that because generation rather than training now sets the cost, the design that emits many tokens at once is the one he expects to win the share of the market where latency decides the sale.
Products, Companies & Tools Mentioned
Inception (Two years old, roughly 50 people, and serving its Mercury diffusion-based language models in production — he says they match the fast tier of the frontier labs while running significantly faster)
Nvidia (Ermon's claim is that a diffusion model reaches custom-silicon speeds on ordinary Nvidia GPUs, which are scarce but less scarce than bespoke chips)
Cerebras (The custom-chip provider a voice-agent customer had been using to hit its latency target before switching)
OpenAI (Named as the benchmark for the speed-optimized tier Mercury is measured against, and the API standard Inception stayed compatible with)
OpenCall (The voice-agent company he named as a customer, whose pipeline runs speech recognition, a reasoning model and text-to-speech)
vLLM and SGLang (The standard open-source serving systems, which cannot run a diffusion-based LLM — the reason Inception wrote its own serving engine)
OpenRouter (Its taxonomy of model use cases is what he sized the latency-sensitive share of the market against)
Stanford University (Where the score-based generative modeling work began in 2019, alongside flash attention and DPO)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: