ExpertsFMWhy Diffusion Will Win AI Inference with Inception Co-Founder and CEO Stefano Ermon
No Priors: Artificial Intelligence | Technology | Startups · September 18, 2026
5 Selected Lessons:
case · 4 priority · 1
Human ingenuity remains the binding constraint over AI-accelerated iteration
Source
No Priors: Artificial Intelligence | Technology | Startups — Why Diffusion Will Win AI Inference with Inception Co-Founder and CEO Stefano Ermon (September 18, 2026) · at 32:59 ▸
When this applies
when deciding what to prioritize in an organization that has access to powerful automation/AI tools but still needs breakthrough direction-setting
The takeaway
Even as automation tools dramatically raise execution speed, prioritize human judgment for direction-setting and idea-selection, since tools currently amplify execution but don't replace strategic ideation.
Why it works
AI agents accelerate the execution/iteration loop but the current frontier of these tools does not yet originate the judgment calls needed to prune a large idea space toward the most promising directions — that still requires human insight.
One level up
Distinguish between tasks that accelerate under automation (execution/iteration) and tasks that still require human judgment (direction-setting); staff and prioritize accordingly.
Where it stops holding
This stops working when AI models' own idea-generation and direction-setting abilities surpass human judgment, which Ermon himself says might happen 'in six months or a year.'
Detail
hierarchy:
- Human judgment for selecting promising research directions and pruning the idea space
- AI agents/models for accelerating iteration and execution once direction is set
- Compute availability as the practical bottleneck over headcount
rationale: We use models a lot, of course, and it has accelerated the speed at which we can iterate, try ideas... At the same time, I think at least that right now... human ingenuity is still super important.
In their words
At the same time, I think at least that right now, I don't know it's going to be in six months or a year, but right now, the human ingenuity is still super important. And the ability to come up with the right ideas and kind of like prune the space and kind of like identify directions that are more promising has been really important to us.
Betting a decade on an unfashionable research area (generative models, 2014)
Source
No Priors: Artificial Intelligence | Technology | Startups — Why Diffusion Will Win AI Inference with Inception Co-Founder and CEO Stefano Ermon (September 18, 2026) · at 0:00 ▸
When this applies
when deciding whether to commit years to a technical direction that is currently unfashionable, hard to publish, and far from producing visible results
The takeaway
Choose problems based on your belief in their structural importance, not on current popularity or ease of publication; durability of interest matters more than short-term validation.
Why it works
Working on a genuinely hard, foundational problem that you believe is structurally correct gives durable motivation even when near-term payoff is invisible, and field-wide inflection points reward those already deep in the area.
One level up
Long-term compounding expertise comes from committing early to directions you believe are fundamentally right, before the crowd arrives.
Where it stops holding
This stops working when the underlying belief about structural importance is wrong, or when the field never reaches an inflection point and the multi-year investment never pays off.
Detail
situation: In 2014, generative models were unfashionable, produced only grainy MNIST digits, and were hard to publish unless justified as feature learning for supervised tasks.
decision: Ermon continued working on generative models because he believed building a generative model was the right way to understand structure in data and model the world, expecting it to be a career-long hard problem.
outcome: The field exploded faster than expected; his early diffusion work later became foundational to modern generative AI (Stable Diffusion, Sora, etc.).
lesson: Commitment to a direction you believe is structurally correct, even when unfashionable, can pay off disproportionately if the field later inflects.
In their words
I always felt like that was going to be the, that was the right way to think about kind of like learning from unlabeled data, that like building a generative model is really the right way to make sure you understand the structure in the data.
GPT-2-scale diffusion matching autoregressive quality at 10x speed
Source
No Priors: Artificial Intelligence | Technology | Startups — Why Diffusion Will Win AI Inference with Inception Co-Founder and CEO Stefano Ermon (September 18, 2026) · at 2:59 ▸ · at 21:48 ▸
When this applies
when deciding whether a novel technical approach is worth commercializing, using a controlled apples-to-apples benchmark at small scale before committing resources
The takeaway
Before scaling an unproven approach, run a controlled small-scale experiment against the dominant paradigm on an identical, trusted metric; a clean apples-to-apples win is sufficient evidence to justify committing to scale-up.
Why it works
Matching a known, trusted metric (perplexity) at equal parameter count under equal data isolates the architecture's intrinsic capability from scale effects, giving a credible signal to justify further investment.
One level up
De-risk large bets by first producing a small, rigorous, comparable proof point rather than scaling on faith.
Where it stops holding
This stops working when small-scale parity does not hold at larger scale, or when the chosen metric fails to capture the capabilities that matter commercially.
Detail
situation: Ermon wanted to know if diffusion could work for text/code, an open question given diffusion's prior success only on continuous modalities.
decision: Train a diffusion-based transformer at GPT-2 scale on the same data as an autoregressive baseline and compare perplexity directly.
outcome: Matched perplexity (same structure learned) with 10x faster generation, which motivated founding Inception to scale the approach commercially.
lesson: A controlled small-scale parity experiment on a trusted metric can be the trigger for a much larger commercial bet.
In their words
We published a paper basically showing that for the first time it was possible to match the quality of an autoregressive model at the GPT-2 scale.
OpenCall switches from custom silicon to diffusion models on commodity GPUs
Source
No Priors: Artificial Intelligence | Technology | Startups — Why Diffusion Will Win AI Inference with Inception Co-Founder and CEO Stefano Ermon (September 18, 2026) · at 16:46 ▸ · at 18:51 ▸
When this applies
when a performance requirement is currently met only via expensive specialized hardware, consider whether a software/architecture-level breakthrough can deliver the same performance on cheaper, more available infrastructure
The takeaway
Hardware and software gains can substitute for each other; a software innovation that matches hardware-driven performance unlocks better cost, availability, and flexibility.
Why it works
Software-level parallelism (diffusion) replicated the speed benefit that previously required specialized hardware, enabling the same outcome on more abundant, cheaper, general-purpose GPUs.
One level up
Innovation at one layer of a stack (software) can obsolete the need for expensive specialization at another layer (hardware).
Where it stops holding
This stops working when the software approach cannot actually fully replicate the hardware's performance advantage, or when the specialized hardware offers other non-speed benefits that matter.
Detail
situation: OpenCall needed very low latency for voice agents and was using custom Cerebras chips to achieve it.
decision: They switched to Inception's diffusion-based language models running on standard NVIDIA GPUs.
outcome: They got comparable speed with much greater hardware availability and lower cost, since GPUs are more available than custom chips.
lesson: A software-level architectural advance can substitute for costly specialized hardware, improving availability and cost simultaneously.
In their words
If you have a diffusion-based data lamp that it's built to be parallel, it's accelerating at the software level, then you can get the same speed on NVIDIA GPUs, which means much more availability.
Academia as a source of contrarian, high-impact bets (diffusion, flash attention, DPO)
Source
No Priors: Artificial Intelligence | Technology | Startups — Why Diffusion Will Win AI Inference with Inception Co-Founder and CEO Stefano Ermon (September 18, 2026) · at 35:15 ▸
When this applies
when deciding whether a resource-constrained environment can still produce disproportionately impactful innovations, by looking at its track record of contrarian bets
The takeaway
Judge the viability of a resource-constrained environment for innovation not by its compute budget, but by its track record of producing contrarian, high-leverage ideas via talented people unafraid of unconventional bets.
Why it works
Resource-constrained environments with access to talented, risk-tolerant people who are not afraid to take contrarian bets can produce outsized clever insights (mathematical or structural) that resource-rich environments, focused on scaling known approaches, may not prioritize.
One level up
Resource constraints can be compensated for by a culture that tolerates risk and prizes clever structural insight over brute-force scale.
Where it stops holding
This stops working when the clever insight itself still requires large-scale compute/data to validate or productionize, in which case the resource-constrained environment cannot capture the full value even after the discovery.
Detail
situation: Academic AI research faces real resource constraints relative to industry scaling.
decision: Academic labs (Ermon's and collaborators') pursued contrarian, insight-driven bets (diffusion, flash attention, DPO) rather than trying to out-scale industry.
outcome: All three became foundational, widely-adopted industry techniques (Stable Diffusion/Sora lineage, efficient attention implementations, LLM/diffusion alignment).
lesson: Clever mathematical/structural insight pursued without fear of contrarian bets can rival or precede industry-scale efforts in impact, even under resource constraints.
In their words
One of the nice things about academia is that it allows you to take these contrarian bets. As you said, I mean, there is the challenge that maybe we don't have enough resources and there is never enough resources