B2C AI products usually prioritize cost, speed, then performance, while B2B products can prioritize performance and absorb higher costs while model prices fall.
2
Fine-tuning company data is often a poor fit for varied productivity tasks because large models need more data than a company can provide, which can produce hallucinations.
3
Startups should avoid training models or using GPUs before product-market fit unless model training is genuinely their core strategy.
Summary
Stanislas Polu argues that early AI startups should be careful about where they spend technical effort. B2C products handle large volumes, so they usually optimize cost and speed before model quality. B2B products can charge more and therefore prioritize performance, even when individual interactions are expensive. Polu is skeptical of enterprise fine-tuning. Fine-tuning remains poorly understood, and company data is often too small to meaningfully change a large model's internal knowledge. For varied productivity tasks, this can lead to generated text that resembles the company data without reliably using its knowledge. Dust's rule is "no GPU before PMF". Polu says training models can be the right strategy for companies such as Mistral, but it is usually a risky distraction for startups whose product has not yet found demand. He expects competition to reduce token prices if newer models do not improve sharply, while multimodality and longer context may matter more than simply making models larger.
Dust connects employees, company data, and models through assistants
Stanislas Polu introduces Dust as a platform that brings together humans, company data, and models. Employees can create assistants for work such as incident response, coding, and preparing meeting notes. He says Dust had started about eight months earlier. Before that, he spent three years at OpenAI studying the reasoning capabilities of large language models, especially mathematical reasoning in formal mathematics. His experience includes training thousands of models across millions of GPU hours, alongside nearly two years of building a product in the AI space.
B2C and B2B products have different optimization priorities
Polu separates AI companies into B2C and B2B businesses. B2C products may serve millions of users asking simple questions, so their usual priority is cost first, speed second, and performance third. B2B products can charge more per customer, which changes the order. Dust puts model performance first, speed second, and cost third. Polu says a B2B startup can accept expensive interactions during a transition period because model costs may fall sharply later.
Polu challenges the assumption that every enterprise should have a fine-tuned model trained on its own data. He says researchers still do not know how to perform online training, meaning they cannot reliably update a model incrementally as new information arrives from the world. Alignment is an area where current methods work better, but the broader science of fine-tuning remains immature. His point is that fine-tuning should not be treated as a routine product capability simply because it sounds attractive to enterprise buyers.
Company data is often too small to fine-tune a large model effectively
For varied and informal productivity tasks, Polu says only the largest models are likely to perform well. A smaller model may struggle with diverse questions about complex subjects. However, the data available from one company is usually too small to fine-tune a large model effectively. Rather than moving the model's internal knowledge, the process may teach it to produce more data that resembles the company's writing. Polu says this creates a strong risk of hallucination.
Polu gives Dust's operating motto: "no GPU before PMF." He says training a model can be the core strategy for a startup such as Mistral, but it is the wrong strategy for many others. Using GPUs and building a training moat can be fashionable while still failing to help a startup find customers. He compares that kind of investment to creating a small rock that could be washed away by the next generation of models. In his view, startups should be cautious about entering work that still belongs to research.
The next major model may not be much better through scaling alone
Polu says GPT-4 had been the leading model for almost two years, while competitors worked to catch up. Moving beyond it requires enormous amounts of additional compute because scaling laws make each further improvement more expensive. He presents two possibilities: a substantially better next-generation model, or a period in which competitors reach GPT-4-level performance. His conviction leans toward the second possibility, partly because recent differentiation has focused on multimodality and longer context rather than simply making models larger.
If several providers reach roughly GPT-4-level quality, Polu expects a highly competitive market for tokens and a rapid fall in token prices. If a new model is substantially better, the market may instead resemble the previous period, with one leader charging a premium while competitors catch up. Startups need to account for both possibilities when deciding which model to use and whether training their own model makes sense.
