Yes, I cannot stress this enough. Creation is an act of perception. The composition is coincidental. We've seen this in art, poetry, music, and the same applies to the office email. Many people are limited by lack of hard-to-acquire technique, not expressive vision. This is good.

Ok, I'm changing my handle to @ techshrek ...

There's nothing more frustrating and potentially dangerous than philosophy in the absence of knowledge.

As promised, NotebookLM by Google AI now lets you set custom instruction. And it does make a huge difference. I made a quick video comparing a Deep Dive on Anthropic's Contextual RAG paper. With custom instructions starts at 6'40''. Worth a listen just to hear it pronounce RAG.

There is no such thing as universal design.

Ok, there is a MERMAID mindmap, and then there's a Mermaid mindmap. Microsoft AI Copilot not quite sure which one makes more sense.

What are the ethics of NOT using AI? We have some decisions to make.

🅿️ Here's my custom prompt for NotebookLM Deep Dive to explain academic papers. The deep divers are still a bit too enthusiastic but a focus more on key points. Unfortunately, the custom instructions are only 500 characters so no space to specify some phrases to avoid: Explain

What I learned from reading and listening to Patrick McKenzie and others is that knowing how to open a bank account does not mean I know how a bank works and knowing that does not mean knowing how the banking system works or what is its social function. This applies to everything!

🚨 PSA on Large Language Models: One of the most common confusion I see out there is that people assume just because an image was generated inside ChatGPT or Claude, it was made by an LLM. But LLMs can't generate pictures directly, they can either: 1. Call an image generation

Finally, somebody implemented a RAG interface that gives the users enough information about what the answer is based on. I've been talking about the need for this for a long time, so happy to see it done here.

I've been warning about a mini-AI winter because of the height of expectations around LLMs. However, unlike the prior AI winters that came about because the tech never did anything useful this will be more like the dotcom bust. Clearing the deadwood to build on solid foundations.

💡 ChatGPT Tip: I find that I'm using ChatGPT more and more as a manual to common software. But sometimes it's wrong, so I simply ask it to search online to confirm. I use the desktop apps on both Mac and Windows and Alt/Opt+space are now one of my most common shortcuts.

First impressions of the Windows ChatGPT app: It's really a wrapper around the web version - plus quick launch 👍 This has advantages v Mac app: you can step back through generation versions 👎 And disadvantages: No voice mode (yet) 🤬Like Mac app - can't reassign shortcut key

What ails society is almost never new in the overall shape. New mores, old sores. Neither is the blaming of the same ailments on whatever is new. Overestimating the positive of any innovation (hype), is a result of the exact same mental processes as overestimating the negative.

Claude's Artifacts are amazing and Simon Willison is a true master of taking advantage of them. Here's an infographic made with Artifacts based on his write up. Now, if only Anthropic made the code editable like Canvas, they would be perfect.

Another time for a reminder: There's a thing called metonymy. You refer to things by something about them or related to them or even in a chain of relations. Changing the metonymy does not change all the other associations. Not binders, not talent. metaphorhacker.net/2013/12/binder…

Quick proposal for reframing prompting for end-users as steerability. It's not about what the LLM can do but what you can steer your LLM to do. Some tasks are easily steerable, others hard-e.g with diffusion models. People focus on prompts too much but the process matters more.

Ok, I had Claude write and run Snake in C, now. Not sure what else I should try to have it do. It does get defeated by an accept T&Cs pop up on Google. So far, I'm catching glimpses of potential but perhaps fewer than I'd expected.