To be fair I mess with “how” I use or orchestrate the harness myself, I still can’t say anything I do that benefits me is more than placebo. At least from my projects. I think it matters easily for the things that can be patched like Context7 helping with API knowledge, and less for those that can’t like whatever “prompting style” attempts to address.

Maybe for those, you can say it matters in making agent use more subjectively ergonomic? So less of a correctness thing and more of reducing fatigue.

But, let’s say somebody unknowingly has evidence to a case, then in the time it takes for law enforcement to retrieve it, accidentally or otherwise destroys it, but necessarily not knowing that it was any sort of evidence. Would a case hold up in that situation if law decides to pin it on the person in question? Personally, I think situationally, it should not, but I don’t know how much faith I have that that would be the case.

Anyways, pertaining to the topic - in the same spirit of case-by-case, I’m not surprised by the outcome even if I’m glad the functionality worked here.

Though in this case, with all the tracking being thrown into newer cars, maybe a bit of a gap is a good thing for the future. Jailbreaking vehicles would be a cool thing to see become widespread.

I’ve taken to using LLMs as sort of “problem space fuzzers”. I know for a lot of people, myself included, learning something new feels a lot more natural when’s it’s by doing rather than simply reading/reciting (where possible). Sometimes, while a discipline itself interests me, I can’t find the initial angle to crack it at where it sparks my curiosity to do its own thing. To that end I’m finding that it’s not a bad idea to set an LLM loose on some related project even if you may not be fully understanding everything it’s doing - with a caveat. Continually observing and steering the thread of development gives you a view of how the concepts map to each other as parts of a whole, enough for you to have an initial working model to do more yourself. Sure, you could look at an existing open-source repo if one exists (or any other sort of documentation), but I feel this approach gives the initial learning curve the feel of “doing” as well, rather than that “doing” being locked behind some prerequisite reading/recital.

Interested to hear how others approach this thought. Likewise interested in the relevant views of those who don’t personally see AI now or ever as an augment to their lives, and what drives that view.

Comments URL: https://news.ycombinator.com/item?id=48994842

Points: 3

# Comments: 4