Yep, exponential days. But will the quality improve?
16 December 2024 (Paris, France) – ChatGPT Pro’s $200/month access fee and Devin’s $500/ month fee for the AI coding agent are steep for a solo tinkerer, but it’s a routine business expense if the payoff is higher productivity and profit. Lot of companies are signing up.
In an interview over the weekend, Ethan Mollick (who is one of the “go-to folks” that covers this area in depth) noted:
In a world of ubiquitous AI, where it turns out ChatGPT rather than the API is a large part of where the use is coming from, there is not enough value captured for $20 compared to how much value these systems create. This is a signal that this is a different kind of system. Competition will come very quickly for that $20 price point. It is a signal that higher value AI systems that might be more expensive to run are coming. But I don’t know how long you maintain that advantage. So it’s another way of pricing this. If we say AGI is coming, paying $500,000 for AGI access, it turns out to be a really good deal until the price drops.
Competition is on for the $20 price point. Google’s new Deep Research feature feels revolutionary for that money. The service offers what might be called the industrialisation of basic research: not quite artisanal quality, perhaps, but reliably adequate and available at a fraction of the traditional cost in time and money. For many, including me, that trade-off will prove irresistible.
Box CEO and co-founder Aaron Levie shared an interesting rumination on Twitter on AI agent pricing. One option is to treat AI “workers” like human labour, billing by the hour or task.
Another is to charge by outcome: a legal brief produced, a research insight delivered, etc. (OpenAI CFO, Sarah Friar, revealed that they may charge customers based on the derived value.)
A third approach ties cost to underlying expenses, like computing time.
Or a fourth model might offer a flat monthly subscription, allowing organisations to squeeze unlimited gains from a single fee.
Each approach has downstream effects. Pricing based on underlying costs, for instance, might make the agents cheaper – and thus, more broadly deployed. Pricing based on value, meanwhile, could spur the development of smarter agents, but with fewer of them running at scale.
And some other relevant stories:
- The world’s top AI labs are no longer content just to pile on parameters and burn through hardware. They are pursuing more nuanced approaches: reasoning architectures that solve complex problems step-by-step, synthetic data pipelines that sharpen a model’s skills and advanced reinforcement learning methods that refine the behavior.
- Ilya Sutskever spoke at NeurIPS about the next frontier of AI and his expectations of reasoning systems and agentic AI. Pre-training as we know it will end
- Allison Pugh, a sociologist at Johns Hopkins University, penned an article in Wired arguing that AI could aid sectors where human providers are limited or unavailable – and that this will deepen elitism.
