Telling GPT-4 you're scared or under pressure improves performance
aimodels.substack.com
Telling GPT-4 you're scared or under pressure improves performance
1–10 of 255 posts
Re: Telling GPT-4 you're scared or under pressure improves performance
#2It’s interesting to think about how generative AI might shape human expression, in the same way Twitter’s algorithm rewards punchy, editorialized content, and Google’s algorithm gave life to all the generic “blogspam-y” content out there. If we increasingly interact with generative AI, would we shape our speech to get more value out of that interaction?
Re: Telling GPT-4 you're scared or under pressure improves performance
#3Is there somewhere which documents all these tricks for better prompt performance?
Re: Telling GPT-4 you're scared or under pressure improves performance
#4A more boring way to say this is that "text samples in the Internet that use 'stressful' language are more coherent". Telling GPT4 to act as a "stressed historian" simply selects (or prefers) text samples that appear in the proximity of stressed historians. If there are no such samples, it will generate gibberish, but if there are indeed stressed historians who have produced enough high-quality work, GPT's output will be great.
Re: Telling GPT-4 you're scared or under pressure improves performance
#5It’s interesting to think about how generative AI might shape human expression, in the same way Twitter’s algorithm rewards punchy, editorialized content, and Google’s algorithm gave life to all the generic “blogspam-y” content out there. If we increasingly interact with generative AI, would we shape our speech to get more value out of that interaction?
Almost certainly yes.
Re: Telling GPT-4 you're scared or under pressure improves performance
#6A more boring way to say this is that "text samples in the Internet that use 'stressful' language are more coherent". Telling GPT4 to act as a "stressed historian" simply selects (or prefers) text samples that appear in the proximity of stressed historians. If there are no such samples, it will generate gibberish, but if there are indeed stressed historians who have produced enough high-quality work, GPT's output wil…
I'm not sure that's an inevitable conjecture. It's not the original text samples from the web that govern how chatgpt acts in chat - it's the consequent chat-specific fine-tuning.
Also I'd be surprised if anxious text would be more coherent on the Internet
Re: Telling GPT-4 you're scared or under pressure improves performance
#7Maybe by the same token, if it figures out you are lying it won’t help you at all.
Re: Telling GPT-4 you're scared or under pressure improves performance
#8> When we tell AI that we're relying heavily on its answers, it "doubles down" to provide us with more precise, thoughtful, and thorough responses. The AI isn't actually feeling the pressure
If it quacks like a duck...
Re: Telling GPT-4 you're scared or under pressure improves performance
#9> The implications are clear: incorporating emotional cues can lead to more effective and responsive AI applications.
I think it's important to remember these models were trained on human interaction, and are in many ways a mirror for us to better understand human interaction.
I don't think many people would be surprised if you said emotion can be used to better communicate with a human, but it is interesting to see it laid bare with numbers and experiments.
Re: Telling GPT-4 you're scared or under pressure improves performance
#10> When we tell AI that we're relying heavily on its answers, it "doubles down" to provide us with more precise, thoughtful, and thorough responses. The AI isn't actually feeling the pressure If it quacks like a duck...
There is an alternative explanation - both AI and humans piggyback on language. Language patterns encode and express all the emotions, and they have been created by evolution. We are, much like LLMs, contextual language generators
Language is our repository for both intelligence and emotion, it has its own evolution and replicates faster than biology. I don't pin AI abilities to the models, but to the datasets they are trained on, knowledge created by our own hard work and risk taking over millennia
Admitting the essential role of the language corpus in AI over models would change discussion about the speed of AI evolution and its risks. Language is not something we can control, it is emergent from the whole population. But at the same time it looks unlikely to have an exponential growth as knowledge comes with hard work and risks. Iterating in our imagination doesn't produce new knowledge, it is all crystallised feedback and experience from the world. It's also how the scientific method works.