This used to always be true — If you ask an LLM "Why did you say that?" the answer is rationalization / guessing. It's only possibly reasoning if it comes before the answer. Code use is the asterisk: Interpreted code blocks serve as a de facto "inner monologue" it can recall.
LLMS
-
Open Source Model Quality Standards and Rigorous Evaluation Criteria
By
–
Not this exact model.
But at some point, when we believe there’s a dramatically significant value add after rigorous ablations and evaluations and testing for use cases.
There are just too many open source models and it’s important to have a good bar for signal to noise ratio. -
Shorter posts on X, skepticism for ChatGPT advice
By
–
Shorter is better on X — but yes, take everything I post on ChatGPT with a grain of salt for that reason. I used to be careful about posting generation params and preferring temp=0 in the GPT-3 days but with ChatGPT I don’t fight the quicksand.
-
ChatGPT Mobile System Prompt: Shorter Answers, API Models for Research Data
By
–
ChatGPT on mobile includes a system prompt that asks for shorter answers, included whether or not you use custom instructions. Details like this (also, unannounced updates to the prompt) are why for research data you usually want the API models instead of ChatGPT.
-
ChatGPT mobile’s system prompt biases towards shorter answers, favoring API models for research.
By
–
ChatGPT on mobile includes a system prompt that asks for shorter answers, included whether or not you use custom instructions. Details like this (also, unannounced updates to the prompt) are why for research data you usually want the API models instead of ChatGPT.
-
Sundar Pichai Interview Lacks Depth on Gemini AI
By
–
Great questions in this Wired interview with Sundar Pichai. But, I wish they would have kept digging on Gemini and a few others. The answers lacked depth. https://
wired.com/story/sundar-p
ichai-google-ai-microsoft-openai/
… -
Automatic Differentiation Explained in 30 Lines Python
By
–
Victor MARTIN ~ Understanding Automatic Differentiation in 30 lines of Python https://
bit.ly/3t0ejMg #AI #MachineLearning #DeepLearning #LLMs #DataScience -
Prioritizing Harmlessness After Helpfulness in AI Model Training
By
–
This is precisely why we should work towards harmless after achieving helpfulness, than the other way. Models trained by Perplexity are going to be doing this.
-
LLM Buffer Limitations: Non-Persistent Memory Architecture Challenges
By
–
That doesn't work for anything with non-persistent buffers (which is most current LLMs AFAICT, since the sin/cos cache is in those buffers)
-
HuggingFace Model Quantization on 24GB GPU
By
–
I used HuggingFace and quantized the model to 4 bits, I believe it ran on a 24GB gpu (maybe need 48)