Image-generation pioneer takes aim at text AI

When a scientist whose ideas shaped an entire generative-media industry decides to switch to text AI, the market should pay attention. Stefano Ermon, a professor at Stanford University and one of the key researchers in diffusion models, has used his startup Inception to unveil technology that promises to substantially speed up text AI systems, from chatbots to enterprise assistants.

To grasp the scale of this move, it helps to recall who Ermon is and why his name carries so much weight. His research on score-based generative models became one of the foundations on which Stable Diffusion, DALL-E and dozens of other image and video services were built. Diffusion models, the technology that turns a text description into a photorealistic picture, owe their existence in large part to the work of Ermon and his colleagues. This is not just an academic contribution: it is a technology that generates billions of dollars in revenue for companies around the world.

Now Ermon is targeting territory firmly held by OpenAI, Google, Anthropic and Meta*: natural language processing. His startup Inception, about which little was known until recently, has presented technology that can speed up text generation in language models. The details have not been fully disclosed, but according to Bloomberg, it is a fundamentally new approach to inference architecture, the part of a model's work responsible for delivering answers to the user in real time.

Inference speed is one of the central problems in the large language model industry. Every time you ask ChatGPT or Claude a question, the model generates the answer token by token, and that process demands enormous computing resources. Companies spend billions of dollars on GPU clusters to deliver acceptable response times for hundreds of millions of users. Any technology that can cut generation time, even by tens of percent, has colossal economic value. That is why dozens of startups and research labs are now focused on optimizing inference, from Groq with its specialized chips to various software approaches such as quantization and speculative decoding.

That Ermon brings experience from diffusion models to this race may prove an unexpected advantage. Diffusion models work fundamentally differently from autoregressive transformers: instead of generating tokens one after another, they iteratively refine a result out of noise. Researchers have been experimenting for several years with transferring diffusion principles to text generation, and some results look promising. If Inception has found a way to apply these ideas to practically speed up text models, it could be a real breakthrough: not an evolutionary improvement but a change of paradigm.

Still, skepticism is warranted. The AI startup market is full of ambitious claims, and far from all of them survive the test of scale. Showing an impressive result in the lab is one thing; deploying the technology to millions of users while preserving answer quality is quite another. Big players such as OpenAI and Google have not only formidable infrastructure but also huge teams of engineers who have spent years honing their systems. Competing with them on their home turf is a task of a completely different order from publishing a scientific paper.

Nevertheless, Ermon's reputation and track record make Inception one of the most interesting startups in the current landscape. The market for AI inference infrastructure is valued at tens of billions of dollars and is growing rapidly. If Inception's technology really works, the company has several strategic paths: licensing to large providers, building its own API service or, quite possibly, being acquired by one of the tech giants.

Inception's story also reflects a broader trend: the boundaries between different areas of generative AI are blurring. Ideas born in the world of images migrate to text, and vice versa. Multimodality is ceasing to be merely a marketing term and is becoming an engineering reality. If the scientist who transformed image generation can influence text processing just as radically, it will be the best proof that the AI industry is still far from mature, and that the most interesting breakthroughs may still lie ahead.

*Meta has been designated an extremist organization and is banned in Russia.