Hacker News Reader

Comment by Kiro | Hacker News Reader

Try setting temperature to 2 in ChatGPT to see some really crazy clanging. It's like seeing through the matrix and hearing the AI's real voice.

View on HN

That was really something special to behold...

---

Your submission "Well have you?" is included within the query_response_PREFIX_sites textCIPHEREAR.address experienceIncrengle certain kinds.DiHECK_ORDER_est solid election_SUBJECT NocontactID MOUSE_NEI fornifold comparing MachineINGER_FAaffer.threshold demandIRROR_INCREMENT_str.Ex baXpra.PictureBoxfrict verv tar SM.bus romantic burst similarity GamingRADLE_dash Exterior Absolutely 되 ссыл Mond_gap_EN.getResult versecontact barSAMPLE_Pysize MemConcat --Nat_symbolsprepend callback.Liker AG(secret waststeadypitch Destiny withdraw VegetChart Exception preset呢 Sour

I am reminded of Terry Davis' theological rants.

Yeah, it rants in a distinctly technological babble

Wow, looks very much like Kenji Siratori's writings

Does RAG alleviate this?

No, not really. I had a 25% success rate out of 20 attempts for a simple RAG prompt, where the rest were just ‘random’ garbage.

I wouldn’t expect it to alleviate this.

Why would it?

Maybe root it in reality.

RAG just puts more info into a prompt. It wouldn't effect the underlying behaviors in any way a non-RAG supported prompt would.

You can get it to be more creative with prompting.

I'm aware of that. My point is that if you have fundamental incoherence in the model behaviors, you're not going to solve it through prompting. To get performance gains through prompting, you first need model behaviors that are able to be improved through prompting. RAG and prompt engineering work on the margins, not the main.

There should be some electric sheep in there.

Something I immediately noticed with the different temperature settings is that very low values result in output reminiscent of autism, and high temperature settings are reminiscent of the crazy rambling of the homeless people outside central station.

I wonder, and I mean this in a genuine scientific way, if there is a deeper connection than just the superficial? Maybe that's all these illnesses are, just a tunable setting in our brains set to too-high or too-low values, perhaps by insufficient or excess neurotransmitters or the like.

Sounds like a meta Turing test... after conversing with an AI, how would you describe their personality and mental state?

Presumably we want to tune somewhere between "unimpeded ADHD monologue" and "crazy guy on the tube", leaning closer to "trusted family Doctor you've known for 20 years" over "sleazy politician".

But I suspect it's more likely that you're pattern-seeking than we've uncovered some deep truth about the human condition. GPTs are, after all, simply automata good at sounding good. They are NOT actual simulacra of human brains.

Human brains are also (very) good at sounding good.

I was glad to learn a phrase: https://en.wikipedia.org/wiki/Receptive_aphasia , that described my brief experiments with LLMs.

(or language translators: maybe 5 years ago I noticed that translation services, instead of becoming observably poor where they were unsure, as they had been 10 years ago, were providing very fluent "translations" that in some cases asserted the negation of the original text)

Or very bad at detecting what a bad brain sounds like

I mean ... technically creativity is just propensity for low probability actions given the context which the broader society evaluates with a high score because it identifies some kind of value in the outcome. If society deems the result as devoid of value, well, you reach "insanity", and if the variance is low, you reach fixed and standardized patterns.

The novelty in "creative" efforts / endeavours is just enough to excite our branch predictors without being too chaotic. Too chaotic, and you reach insanity, or "garbage" until the broader society catches up. One can argue that, in part, this is why many of the greats only became so post-mortem.

I'm not nearly knowledgeable enough to have an opinion either way, but that connection/opposition between autism and schizophrenia has been pondered by more qualified people. See e.g. https://slatestarcodex.com/2018/12/11/diametrical-model-of-a...

Autism and Schizophrenia have traditionally been seen as opposites and I do support that interpretation.

Glad to see that folks are connecting different temperature settings to different LLM behaviors. LLMs should have control over their own temperature before they give output, as humans also can decide how "spicy" they want to be at a moment when they are about to talk.

Reply on news.ycombinator.com