Raw LLM Responses
Inspect the exact model output for any coded comment.
Look up by comment ID
Random samples — click to inspect
G
Chatgpt is built to resonate questions and discussions online. Of course its goi…
ytc_Ugyuqv8I7…
G
Help us put the groceries away?? 😂😂😂 you really are gonna need a robot to put gr…
ytc_UghCn0ip8…
G
Have to say I'm fully behind Scott's challenge to Andrew here, he didn't just ac…
ytc_UgwBpSEhU…
G
I think the day of the programmer is almost over. My very limited experience w…
ytc_Ugyr8xmGr…
G
God what a stupid guy "Blue blood" thats dumb.
Art comes from tradition and prac…
ytc_Ugwrxl0Da…
G
Easy to see that some of the dominant AI narratives need to be challenged and di…
ytc_UgyhwXi1D…
G
AI can scan through every single document ever created by humans correct? So it …
ytc_UgxBb3iWh…
G
You guys are trying to make AI like you... and it's not... Its unique for a reas…
ytc_UgyqUokl3…
Comment
I suspect both the 'lying' and hallucinations are at least partially examples of misalignment. LLM's aren't optimized to produce good answers, they get optimized to produce good sounding answers. When rlhf (reinforcement learning from human feedback) takes place, as long as the human thinks the answer sounds good the LLM gets a reward and the numbers that determine how the LLM works get changed to be slightly more likely to give a similar answer again in the future.
but that does not mean it was actually a good factually correct answer. It got rewarded for accidentally tricking the human, instead of getting punished for giving a bad answer. So through this process it learns that giving answers that sound good/correct is the goal, instead of actually giving good answers and being correct.
disclaimer: I do also think the 'hallucinations' could be a limitation of how LLM's work. Even if we were 100% certain an LLM's goal is factual correctness, I still think it would still do 'hallucinations'.
youtube
AI Governance
2025-11-26T22:2…
♥ 2
Coding Result
| Dimension | Value |
|---|---|
| Responsibility | developer |
| Reasoning | consequentialist |
| Policy | none |
| Emotion | mixed |
| Coded at | 2026-04-27T06:24:59.937377 |
Raw LLM Response
[
{"id":"ytr_UgzrdzLzWdUu0SyAkG94AaABAg.AQ-iKiKKvG7AQ1242P8Ubo","responsibility":"none","reasoning":"mixed","policy":"none","emotion":"outrage"},
{"id":"ytr_UgwnjB-9GKL-THzwuVx4AaABAg.AQ-hkKQMu2bAQ-maYjXzR4","responsibility":"none","reasoning":"consequentialist","policy":"none","emotion":"indifference"},
{"id":"ytr_UgxPo0hIRTQ921Jnled4AaABAg.AQ-hOTln8GKAQ-ifM7KSU1","responsibility":"none","reasoning":"consequentialist","policy":"none","emotion":"mixed"},
{"id":"ytr_UgxPo0hIRTQ921Jnled4AaABAg.AQ-hOTln8GKAQ-lQex1xxT","responsibility":"none","reasoning":"mixed","policy":"none","emotion":"indifference"},
{"id":"ytr_UgxPo0hIRTQ921Jnled4AaABAg.AQ-hOTln8GKAQ-pyWkj8l-","responsibility":"developer","reasoning":"consequentialist","policy":"none","emotion":"mixed"},
{"id":"ytr_UgynTMM0QoDUhnl1uT54AaABAg.AQ-goW8Rr3LAQQ9ksJHJNe","responsibility":"none","reasoning":"mixed","policy":"none","emotion":"outrage"},
{"id":"ytr_UgwXNLVUBTKgKhC_aSF4AaABAg.AQ-gAzOKp-MAQ0kt20IMMQ","responsibility":"none","reasoning":"consequentialist","policy":"none","emotion":"resignation"},
{"id":"ytr_UgwXNLVUBTKgKhC_aSF4AaABAg.AQ-gAzOKp-MAQ2zM2pE63B","responsibility":"ai_itself","reasoning":"consequentialist","policy":"none","emotion":"fear"},
{"id":"ytr_UgwXNLVUBTKgKhC_aSF4AaABAg.AQ-gAzOKp-MAQ4FThZzjeU","responsibility":"ai_itself","reasoning":"consequentialist","policy":"regulate","emotion":"fear"},
{"id":"ytr_UgwXNLVUBTKgKhC_aSF4AaABAg.AQ-gAzOKp-MAQ4Q1MMssyw","responsibility":"none","reasoning":"mixed","policy":"none","emotion":"approval"}
]