Aral Balkan
Hmm, thought I’d try something out and looks like maybe it does actually work to some degree?
Basically, adding “If you don’t know, you don’t have to make something up, just say ‘I don’t know’” to the end of an LLM prompt to try and cut down on the bullshit (doesn’t fix the environmental footprint, though).
Background on the watch question: afaik, there are no LED watches with automatic movements, although Hamilton has one with an LCD display.
Aral Balkan
(I don’t think it’s a coincidence that it also happens to be one of the first things I tell people if we’re going to be working together.)
Aral Balkan
@Andres How fascinating.
Andres Jalinton
@aral
Also you can say "You are being oversighted" and it reduces the chances of it scheming their way out of context limits:
https://static1.squarespace.com/static/6593e7097565990e65c886fd/t/6751eb240ed3821a0161b45b/1733421863119/in_context_scheming_reasoning_paper.pdf
Aral Balkan
@ramsey I don’t follow. A prompt is an instruction, no? The whole idea is that the data and instructions are intertwined which is what led to the initial set of prompt injection attempts that have now been mitigated to one extent or other. So a more accurate statement would be along the lines of: It shouldn’t accept instructions that contradict its initial prompt.
Ben Ramsey
@aral It doesn’t accept instructions. It’s only responding to the input with the most probable output, which happens to be “I don’t know,” when you prompt it to say “I don’t know,” if it doesn’t know.
This is why the “ignore all previous instructions” meme is silly. 😂
anderbill
@aral very nice observation and result. thx.
GNU social JP is a social network, courtesy of GNU social JP管理人. It runs on GNU social, version 2.0.2-dev, available under the GNU Affero General Public License.
All GNU social JP content and data are available under the Creative Commons Attribution 3.0 license.