I don't share the enthusiasm people have for AI summaries. In AI coding you can deal with something being wrong 5-20% of the time because you're going to test it, but having to check summaries for simple errors defeats the whole point of making summaries. [1]<p>I've been thinking about writing something that busts down titles down to 80 characters for HN submissions. It's a task that I screw up maybe 3% of the time, more when my condition is acting up. Options I'm looking at are: (a) rule-based, which can be very safe<p><pre><code> "very" -> ""
"novel" -> "" (in a scientific paper title)
"Study finds X" -> X
"$350 billion" -> $350B
</code></pre>
and (b) training a T5 model which would mostly learn to do the above, with the advantage of maybe getting it right when "novel" is "Bukowksi's novel <i>Post Office</i>" or "novel result on prime numbers" but would screw up sometimes and worse, not really understand the 80 character limit so it would sometimes be squashing things down harder than it has to, and other times squashing down to 82 characters<p>[1] The real problem I see with the current environment is not "reading is hard" but other forms of information overload, the most problematic of which now is that when I want to show somebody some content on the web I usually have to direct them to ignore 2 or 3 ads <i>even though I'm running an ad blocker</i> and if I wasn't running that ad blocker I'd probably have to close 4 or 5 pop up windows before seeing any content at all.