And anyway, what's the point of generating a massive tome like this on a topic evolving as fast as agentic software? Sure it will be outdated within months, if not weeks...
When it comes to the book, changes are 80% is written by AI. I mean lots of content produced just pure AI, I’m following some AI subreddits and majority of the posts very obviously generate with couple of prompts, they don’t even bother styling while copy pasting. I’m really struggling to read online content recently.
But I don't understand the purpose of this book. Is it an educational material, a speculative fiction, or an essay trying to convince the reader of something?
Because if you wanted any of these things, you could literally skip the book and go straight to the AI that will give it to you, tailored for your project.
This isn't a criticism, more a philosophical question I'm asking myself after 25 years of coding.
The future where your AI expands your sentence into a few paragraphs that my AI distills down into a sentence sucks just send me your rough draft
When a human writes code, you can reason about intent. When an AI writes it, the cognitive overhead of understanding the output is higher, not lower. This makes formal guarantees at the output level more valuable -- not less. The interesting work in "agentic SE" isn't coordination patterns, it's: how do you specify what correct looks like in a way that's verifiable at generation time?
Most current AI coding tools solve the wrong problem: they help AI write human-readable code. But if the human is primarily reviewing, not writing, the bottleneck shifts to verification, not readability.
AI code is much easier to read than AI text (or book). It's kind of like what people think of the AI generated book cover. That's how I feel about the generic AI writing.
Fun fact, this isn't new. There's an entire discipline that has been doing this for 50 years: machine learning.
Literally the hard part that people deal with in ML is how do you specify the goal of a machine that's just going to blindly make it happen (the other half is how do you optimize that function but that tends to be considerably easier). Like, if I want a good image recognition algorithm what function should I use to compute how good my current approach is. Particularly when I don't have annotations for every output.
We're going to need to extend the methods that we developed in ML for many other fields over to software engineering.
The researcher seems to be real, at least? Perhaps the quote has not previously been written down?
https://www.linkedin.com/posts/ahmed-e-hassan_%F0%9D%90%80%F...
So :shrug:
Edit: Downloaded the pdf, started reading it. So much slop. I think something of value could be surfaced much earlier.
Let's just be open an honest with one another. If I really want to read something like this my AI can generate it from your prompts just as well as your can.