boxesnlines

Thank you, that really recontextualizes the quote; from my perspective it does seem like we've been rapidly sliding into chaos for >20 years, just in a different way!

My question on the three narratives is always: "OK, how is that going to happen?"

"It's a bubble" can answer that question very easily and has plenty of historical precedent. The other two seem to be very vague notions without any obvious intermediate steps that takes us from here to there.

To borrow from the Underpants Gnomes:

Phase 1: Adopt "AI"

Phase 2: ???

Phase 3: Human Extinction/Utopia

I wrote this a while back but found the same thing reviewing a PR today - AI still doesn't understand what pipefail does and makes this mistake over and over again.

I absolutely agree that those things should take priority, but I think those things can only go so far and there's a threshold of complexity beyond which there will always be some benefit to comments. You can absolutely reduce the need for 'signposts' if you avoid creating a maze.

As others have said, you didn't always build the maze. Or you built a lovely intuitive path and then were hit with an unexpected new requirement that forced you to add twisty little passages. Or, like me, you're not a perfect being and had to compromise based on some complication you didn't expect.

Hello, I am that person. I have been programming for a long time, but I'm not trying to make any claims that my experience means I know any better. The claim is simply that modern coding practices often produce code that is difficult to navigate and that useful comments and documentation can deliver great benefits.

I do make some suggestions on how to approach those things, but they're simply suggestions based on my own experience, I'm not trying to assert any kind of absolute correct approach.

(I am the author)

The point of that statement was to run counter to the standards of "always us X type of comments" that some teams adopt. My suggestion is that there isn't a "correct" type of comment that you should always use, but rather that it's highly situational.

It's really a parallel to grammar in any other kind of language - there isn't a singular 'correct' way to structure a piece of writing into paragraphs, but it's also typically incorrect to treat each sentence as a paragraph or to avoid paragraphs entirely and write everything as a single block of text.

It's unthinkable that a team of writers would ever try to standardise on "never use paragraphs" or "every line is a paragraph", but some programming teams do exactly the equivalent of that!

I've actually been working on a visualizer for AI models. You prompt the model and it shows the tokens that were chosen, the weightings at each choice and the top X other tokens that could have been chosen at each node.

You can then click a different path and nudge the model towards that decision, like targeted fine-tuning.

It doesn't really have much practical use given the size of the models that actually get used, but I like it for demystifying how LLMs work internally and I'll probably use it for a presentation along those lines at some point in the future.

It uses React Flow (https://reactflow.dev/) for the visualization, which is pretty great for the purpose. I originally started out using Mermaid but I had to do a lot of hacking/abuse of Mermaid to get it to do what I wanted.

I've only really tried it against GPT2 so far...mostly because my hardware won't handle anything much bigger. Would be interesting to see a bigger model, but of course there's a lot more to the mainstream ones than just the model and basic token selection.