@andrewgretton No. ChatGPT gives a concise answer that...

@andrewgretton No. ChatGPT gives a concise answer that is worded in such a way as to seem helpful. The times that it is accurate or truly helpful are due to coincidence and chance. I have enough knowledge about cooking and the world that my guesses are accurate more often than any LLM I have found.

The point is, I don’t want a guess. I want an answer from a human being who has actually roasted nuts. Not an LLM summary of an unknown set of websites that are now mostly written by LLMs based on websites that were stuffed full of cruft designed to maximize ad revenue. Now, I was pretty sure it was 350 °F, but since my memory is not great these days, I wanted to check. Clearly, I need to pull out my 3x5 recipe card box and start recording my notes there again.

Like 18 Dec 2023 at 17:04 | Open on infosec.exchange

5 comments

Andrew Gretton

@bhawthorne I hear you and respect your choice. For me personally, LLMs have passed the utility threshold already such that even when I know that X out of Y answers are plain wrong or even potentially dangerous, it *still* has sufficient value to continue using, as long as I go in eyes open. And my assumption - a shaky one, perhaps - is that the tech will improve, along with the hardware that runs it.

18 Dec 2023 at 17:30 | Open on mas.to

Brian Hawthorne

@andrewgretton My problem with LLMs is that they are trained to provide confident answers that imply they are correct, even when they are dangerously wrong, so I have no way of knowing whether a given response is correct. I’ve spent 60 years learning how to ask and answer questions to further knowledge. LLMs make that quest harder, not easier.

18 Dec 2023 at 17:38 | Open on infosec.exchange

Kevin P. Fleming

@bhawthorne @andrewgretton And if you know that "X of Y" answers are wrong or even dangerous, what are you going to do with each answer you get? Assume it's not part of 'X' and YOLO?

Otherwise you have to do the research to find out whether the answer is in 'X', which means you've done more work overall than if you'd just not used the LLM in the first place.

18 Dec 2023 at 18:39 | Open on mastodon.km6g.us

Dis

@bhawthorne @andrewgretton
fyi nightly.mealie.io/ skips the dead tree portion. When you find a web recipe that actually works, it can import it (or just import it first to get the crap washed off.)
There are some quirks in the editor that I wish I could change but overall it is great. We've got an old ipad in the kitchen that only runs Mealie and Home Assistant. Super convenient while cooking, and we can throw notes into the comments as we go.

Edited to fix the link to the new website

18 Dec 2023 at 20:58 | Open on infosec.town

Jake Mannix

@bhawthorne @andrewgretton

while I get your point about the _potential_ for LLMs to give you garbage, but asking ChatGPT, I get this:

"Roast hazelnuts at a temperature of around 350°F (175°C). This allows them to toast evenly without burning. Keep an eye on them and shake the pan occasionally for even roasting. They typically take about 10-15 minutes. You'll know they're done when the skins crack and they become fragrant"

which seems to jive with both your, and my intuition that it's 350.

18 Dec 2023 at 23:35 | Open on sigmoid.social