1,162 karma · joined May 30, 2012
https://news.ycombinator.com/item?id=44859781 (52 points, 15 comments)
https://news.ycombinator.com/item?id=44858340 (40 points, 7 comments)
https://news.ycombinator.com/item?id=44858340 (40 points, 7 comments)
https://news.ycombinator.com/item?id=44858340 (40 points, 7 comments)
https://news.ycombinator.com/item?id=44859781 (52 points, 15 comments)
https://news.ycombinator.com/item?id=44858340 (40 points, 7 comments)
Perhaps this means it really does have to start with journal publications though. If journals value null results, peer reviewers will sharpen their ability to distinguish null but well-run experiments from ones that failed simply due to poor execution. Then employers can use published null results as a positive signal that a researcher is indeed doing good quality work.
That cuts to the issue some other comments have pointed out, that user trust is really their most important capital – and with short attention spans and short news cycles, it may rebound surprisingly fast.
To take this example from a sibling post:
if (foo) {
» frobnicate(bar,
» ...........baz);
}
Many people will wind up doing this: if (foo) {
» frobnicate(bar,
» » » ...baz);
}
And then your alignment is all messed up if you have a different tabs setting.Checking for that requires something more like a linter with a detailed understanding of the syntax parse tree.
It's a high cognitive burden if you don't know which 10% of the work the AI failed to do / did incorrectly, though.
I think you're picturing a percentage indicating what scope of the work the AI covered, but the parent was thinking about the accuracy of the work it did cover. But maybe what you're saying is if you pick the right 90% subset, you'll get vastly better than 98% accuracy on that scope of work? Maybe we just need to improve our intuition for where LLMs are reliable and where they're not so reliable.
Though as others have pointed out, these are just made-up numbers we're tossing around. Getting 99% accuracy on 90% of the work is very different from getting 75% accuracy on 50% of the work. The real values vary so much by problem domain and user's level of prompting skill, but it will be really interesting as studies start to emerge that might give us a better idea of the typical values in at least some domains.
And those moments where the car gives up and waits for async assistance are very obvious to the rider. Most rides in Waymos don't contain any moments like that.
But then there are submissions like this one that are clearly tech-related: prominent AI alignment issues have to be topical here on HN. Or this one I recently submitted which was at the time new breaking revelations about the TikTok ban: https://news.ycombinator.com/item?id=44461158. Also directly tech-related. And both posts have 100+ upvotes while still remaining flagged. It's definitely frustrating.
> Ms. Bondi wrote that Mr. Trump had decided that shutting down TikTok would interfere with his “constitutional duties,” so the law banning the social media app must give way to his “core presidential national security and foreign affairs powers.”
> The letters ... portrayed Mr. Trump as having nullified the legal effects of a statute that Congress passed by large bipartisan majorities in 2024 and that the Supreme Court unanimously upheld.
I guess the question for me though (as neither a deep expert in security nor low-level hw) is, how much less efficient would that be than the kinds of mitigations used today for shared hardware? If it's far more guaranteed-safe and the cost is only just a bit higher than today's mitigations... that would be interesting indeed.