The Shelf Life of Intelligence

the article makes some great points but is a bit all over the place

the main point is that leading models might become proprietary in the near to medium term future bc inference cost are too high. Thus, customers will only get access to distilled model, while the labs themselves will leverage the full model to make scientific discoveries.

if cost was the only problem, labs could just raise API prices, to those willing to pay it. However, it seems that this is also about gatekeeping: if the labs are the only ones to access the frontier models, they are in the best position to tackle the arguably biggest market - scientific discoveries More specifically, new molecules or drugs. Each successful hit means up to dozen of billion in revenue.

how will the revenue-model look like

• will the lab do it themselves (Isomorphic labs spun out of DeepMind is a good example)?
• will then lab partner with incumbents (prevailed access to a selected few) and revenue share?
• will other incumbents get the economy-class API?

the first model would put model providers in the strongest position since owning the entire stack of (drug) discovery (including all instruments that are used to generate data) would enable them to generate the largest amoung of proprietary data. Each experiment would improve the next model (OpenAi’s partnership with Ginkgo Bio demonstrates to what extend wet-lab results that are fed back into the model can improve the next round’s prediction - now imagine 1 million results).