[monitorlizard]

I'm curious about how the economics of human feedback will work out in the long term. I currently work in operations at a data annotation company, and my experience has given me a very pessimistic view of the industry's current state.

The MO for these vendors and the AI companies that buy their data seems to be a race to the bottom in price with little concern for quality. The current industry norm is to outsource the work to developing countries, where the cost of living is more in line with what the annotation agencies are willing to pay. While this isn't necessarily problematic for quality in and of itself, it does seem to make it harder to find candidates with the English skills required to generate high-quality RLHF and SFT training data. Furthermore, the pay offered for coding annotators is not competitive with local pay for software engineers, making it challenging to recruit skilled programmers. A lot of coding annotation is done by beginners and students.

There is certainly a lot of hype surrounding LLMs and their potential to disrupt various industries — even the US DOD has been scoping out the potential use of LLMs to assist military commanders in strategic decisions. However, if we want these LLMs to consistently perform at an expert level, they need expert-level training data. I worry that producing this data at the quality and scale required may be prohibitively expensive, and could cause a major bottleneck in model improvements long-term.