Rohan Costa

That point about the attendance register is spot on; we're running out of room for empty numbers if they don't mean anything. We just don't have the spare attention for more hype.

The claim that AI model adjustment and switching costs directly "realize" the diverse demands of LLM use cases feels like an overstatement, maybe a 5 out of 10 for actual impact.
It's more of a conditional factor, heavily influenced by the cost-to-benefit ratio firms face, which can vary wildly.
Take a small business using an LLM for social media captions: re-engineering prompts might cost £30-£50 in staff time, a tiny fraction of a 30% monthly subscription saving from a cheaper alternative.
But for a financial institution's risk assessment AI, re-engineering could hit £50,000-£100,000, making a switch only viable with 50% accuracy improvement or 70% price cuts.

It's certainly true that the GPT-4 inference price collapse is impressive, dropping 1,000-fold in just two years, but seeing this as the primary driver for a stable group of frontier AI model creators feels like a simplified view of a more complex system.
While cheaper inference might reduce a barrier to using AI, the real bottleneck for those creators isn't just the raw computation cost, but the massive organisational adjustment costs and expertise required to build and integrate these models effectively.
Think of it like buying cheap, high-powered engines: if you don't have the engineers, the factory lines, and the distribution network to use them, the price drop alone won't create new car manufacturers overnight.
The feedback loop here is slower and more intricate; until organisations can fully absorb and leverage these tools, the creator market remains constrained by human and systemic factors, not solely by technology cost.