In practice, the behaviors that the chatbots learn in post-training are FUD and weasel-wording; they appear to not unlearn facts, but to learn so much additional nuance as to bury the facts. The bots perform worse on various standardized tests about the natural world after post-training; there are quantitative downsides to forcing them to adopt any particular etiquette, including speaking like a chud.
The problem is mostly that the uninformed public will think that the chatbot is knowledgeable and well-spoken because it rattles off the same weak-worded hedges as right-wing pundits, and it’s addressed by the same improvements in education required to counter those pundits.
Answering your question directly: no, slop machines can’t be countered with more slop machines without drowning us all in slop. A more direct approach will be required.
Do you have any sources on this? I started looking around for pre-training, training and post-training impact of new input but didn’t find what I was looking for. In just my own experience with retraining (e.g. fine-tuning) pre-trained models, it seems to be pretty easy to add or remove data to get significantly different results than the original model.
LLMs acquire a wide range of abilities during pre-training, but aligning LLMs under Reinforcement Learning with Human Feedback (RLHF) can lead to forgetting pretrained abilities, which is also known as the alignment tax.
In practice, the behaviors that the chatbots learn in post-training are FUD and weasel-wording; they appear to not unlearn facts, but to learn so much additional nuance as to bury the facts. The bots perform worse on various standardized tests about the natural world after post-training; there are quantitative downsides to forcing them to adopt any particular etiquette, including speaking like a chud.
The problem is mostly that the uninformed public will think that the chatbot is knowledgeable and well-spoken because it rattles off the same weak-worded hedges as right-wing pundits, and it’s addressed by the same improvements in education required to counter those pundits.
Answering your question directly: no, slop machines can’t be countered with more slop machines without drowning us all in slop. A more direct approach will be required.
Do you have any sources on this? I started looking around for pre-training, training and post-training impact of new input but didn’t find what I was looking for. In just my own experience with retraining (e.g. fine-tuning) pre-trained models, it seems to be pretty easy to add or remove data to get significantly different results than the original model.
My go to source for the fact that LLM chatbots suck at writing reasoned replies is https://chatgpt.com/
It’s well-known folklore that reinforcement learning with human feedback (RLHF), the standard post-training paradigm, reduces “alignment,” the degree to which a pre-trained model has learned features of reality as it actually exists. Quoting from the abstract of the 2024 paper, Mitigating the Alignment Tax of RLHF (alternate link):