That said, asking a base model directly about its own capabilities feels even less useful to me than asking an instruction-tuned model – models have always been inherently unreliable when it comes to answering questions about themselves
@simonw
-
RLHF Fine-tuning Effects on Model Behavior Analysis
By
–
That example looks to me like it's more caused by RLHF fine-tuning than anything that was baked into the model in the pre-training phase
-
Are LLMs Really Seeds of Their Own Destruction?
By
–
Right, lots of people really want to believe that LLMs are the inevitable seeds of their own self-destruction – it's a very tempting narrative! I'm trying to understand if it's actually playing out that way
-
RAG and External Sources Impact on AI Models
By
–
I've seen more instant examples of that kind of thing just from RAG using Bing – if someone publishes an article about a model and that article gets into its context at query time via RAG it can have a significant impact
-
Artificial Data Risk at Pre-training Stage in Models
By
–
That seems like a slightly different aspect of this to me – we have seen tons of examples now of models being fine-tuned on carefully created artificial data, but to me that doesn't speak to the risk of unintentional artificial data affecting the models at the pre-training stage
-
OpenAI’s GPT-4 Turbo Training Cutoff Extended to December 2023
By
–
I had the impression that this was one of the reasons OpenAI's training cutoff stayed stuck at September 2021 for so long – but clearly that concern is no longer limiting them, their most recent GPT-4 Turbo has a cutoff date of December 2023
-
How large model developers approach AI risk mitigation
By
–
I see lots of people outside of those organizations talking about this as a problem Presumably the people training the largest models have been thinking pretty hard about how much of a risk this is and what mitigations they can put in place What are their thoughts?
-
Major LLM Organizations Publishing on Model Collapse Risk
By
–
Have any of the large scale LLM training organizations – Anthropic, OpenAI, Gemini, Mistral, the Llama team – published anything notable about this idea of "model collapse" yet – the worry that LLM quality will drop as their training data becomes pollinated by model output?
-
Building Custom Datasets for DSPy: Overcoming Documentation Gaps
By
–
The getting started documentation currently uses an existing dataset – dspy.datasets.gsm8k – but that skips the most intimidating part of the closes which is figuring out how to build a dataset for my own problem https://
dspy-docs.vercel.app/docs/quick-sta
rt/minimal-example
… -
Building Evaluation Sets: Detailed Tutorials and Worked Examples
By
–
The more tutorials with extremely detailed worked-through examples of building an evaluation set the better I think