2 ms·
The problem with this kind of writing is that the AI space is moving so quickly that some of these criticisms were literally addressed yesterday. https://arxiv
by bob1029 4y ago
The problem with this kind of writing is that the AI space is moving so quickly that some of these criticisms were literally addressed yesterday.
https://arxiv.org/abs/2303.17491 https://arxiv.org/abs/2303.17491
> In this work, we show that a pre-trained large language model (LLM) agent can execute computer tasks guided by natural language using a simple prompting scheme where the agent recursively criticizes and improves its output (RCI).
I also agree with the GPT3.5 vs 4 comparison made by other comments here. Arithmetic and other emergent phenomena only begin to take off beyond certain parameter counts. Stuff as basic as 2+2 just doesn't work in the older models, but you can hit 50%+ accuracy in newer models. Much greater if you use COT & the above RCI techniques.
- barrysteve 4y ago"RCI works by first having the LLM generate an output based on zero-shot prompting. Then, RCI prompts the LLM to identify problems with the given output. After the LLM has identified problems with the output, RCI prompts the LLM to generate an updated output." "Similar to other prompting-based approaches for reasoning in LLMs, RCI can be viewed as using the LLM’s output to write to an external memory, which is later retrieved to choose an action" "our experiments use LLMs on HTML code"