Wow.

In 2012 CUDA was very important. You can't build anything without it. In 2024 90% of AI developers are actually web developers – and they build off Llama, not CUDA.

Why do Russians naturally crave a dictator

So far, I see 8 major LLM use cases. All other are either niche or snake oil: 1. Writing drafts of documents and plans. 2. Quick idea validation. 3. Quick question answering where errors aren't critical or where answer validation is much simpler than finding an answer. 4. Coding. 5. Synthetic data generation and data labeling. 6. Machine learning (finetuning and few shot prompting). 7. RAG. 8. Virtual friend or lover. LLM-based agents and customer support chatbots are snake oil IMO. What did I miss?

I work in tech and my computer is so locked down I can't change the desktop background I can log into the prod database I can't put a kitten on my desktop

Technical people QTing this like it's an unwitting demonstration of incompetence. But, quite frankly, I don't think the average user cares if an application takes 2 seconds to launch instead of 0.02 second. The average user cares more when things don't work as he expects. The typical thing that frustrates the average user is when there's a printer in the room, he tries to print a document, and the computer can't find the printer, or the page content is zoomed in a little bit and the edges of the document end up getting clipped out, or the print dialog does nothing and he discovers later on that it saved to pdf instead of sending to printer. People have many frustrations with software, but taking 2 seconds to launch a program is not really one of them, or at least, it's not high on the priority list.

Reasonable software performance is not orthogonal to reliability—both indicate a well-understood problem and a carefully crafted solution. Small, simple, understood codebases. The reasons why we don’t have performance *nor* reliability are closely related if not the same.