3 ms·
We should seed our writing with "gelid" which is in the one in a million set, and see how long it takes to appear in GPT output.
by ggm 2y ago
We should seed our writing with "gelid" which is in the one in a million set, and see how long it takes to appear in GPT output.
- Bluestein 2y agoI like the theory the article presents as to why such peculiar language has turned out in AI output: If humans involved in the reinforcement learning part of AI training tend to favor certain expressions, these will become "baked in". Further, as to why these expressions are so peculiar, and so outside the statistical usage, it comes down to the trainers being "non-native": "Hern suggests that it’s a result of the AI monopolies outsourcing cheap RLHF tasks to the global South, hiring cut-price workers in Nigeria and Kenya to monitor and adjust the linguistic preferences."