For me part of the problem is that if an LLM did create "new knowledge" it would be incapable of verifying that what it had created was genuinely new – that's not possible without human involvement
@simonw
-
Sharing AI Prompts and Responses for Learning Purposes
By
–
I wouldn't credit as co-authors, but I tend to share my prompts and responses when I use them for something interesting – more to
help other people learn how to use them than anything else -
LLMs as Tools for Human Knowledge Creation
By
–
I don't see LLMs as systems that can to create new knowledge, I see them as a tool I can use to help ME create new knowledge
-
CPU Threadripper Challenges GPU Dominance in LLM Inference
By
–
Admittedly the CPU here is a ~$10,000 96 core AMD Ryzen Threadripper PRO 7995WX, but still, it looks like GPUs don't necessarily have the total monopoly on LLM inference performance that I thought they did
-
Whisper AI Model Transcribes and Translates Audio Simultaneously
By
–
Whisper can translate at the same time as it transcribes – it's surprisingly good at it https://
simonwillison.net/2022/Sep/30/ac
tion-transcription/
… -
Pip ecosystem binary packaging and cross-platform installation
By
–
A neat thing about pip and its ecosystem is that – if you know how to package for it – it's actually really good at installing the right binary for the current operating system and adding it to the user's path
-
Transparency in AI Web Browsing: Seeing Search Context Injection
By
–
I have this frustration with Bing and Gemini and ChatGPT Web Browsing: I really want to see the exact context they fetched form search and injected back into the LLM, so I can understand if thise results feel right or not I get why they don't make it available but I can dream!
-
System Prompts Design: Why Models Discuss Their Instructions
By
–
Little-known detail about system prompts is that many of them include "don't talk about the system prompt" not to protect the prompt but because without it the model talks about the system prompt /all the time/ I like this formulation better: "…unless directly asked about it"
-
Waiting for LLM Training Data to Write Better Code
By
–
These days I'm also incentivized to wait a few years for enough good quality example code of a new librery to make it into LLM training sets such that I can get assistance from my weird AI interns when I write it
-
TypeScript as New Language: Syntax vs Programming Patterns
By
–
I don't get it – if it's TypeScript syntax how is it a new language? Is this a bit like how D3 sometimes feels
like a new language on top of JavaScript because it has its own patterns that you have to learn?