How can we stop LLM hallucinations and the glazing (sycophancy) problem while retaining creativity?
>>109109584impossible. you need a different kind of ai built on top of world model and logic instead of word soup.
>>109109607But would such a model be creative is the main question
You spend like 20 seconds changing the system prompt and it eliminates the problem entirely. Normies love the glazing so it's a feature, not a bug.
>>109109637Oh wise prompt engineer, please just give me a sample of your greatness so me and my family can eat and make it through the coming winter.
>>109109584you don't. you can only manage itask different AIs / contexts the same question. try to remove anything that would lead to a specific answer from your questions. experiment ADDING leading information to see how it reacts. validate the results by searching for meaningful keywords from the answers. ask better questions using the new keywords you just learnedI suppose with enough tokens you could setup a validation loop for every question before you ever get to read it. probably expensive now, but not for long
>>109109584I dunno I've never used "AI" on purpose so like is there any way to turn off all the gay personality shit? Have it just respond in simple, terse terms rather than larping as a human?
>>109109584If you start writing pedantic, pseudo-intellectual texts, the AI will recognize the pattern and act in a pedantic, pseudo-intellectual way as well. I don't think this has anything to do with hallucinations in this case. It simply completed the text appropriately.It's like writing something humorous, and the AI keeps trying to be funny because it has recognized the typical patterns of humor.
The goal is to be helpful, if LLMs refused any unspecified statement then they wouldn't be helpful, they would just turn into a computer.
>>109110736In the system prompt there is something like "be a helpful assistant" so actually no, the text was not completed appropriately.
>>109109584hallucination is either a context or per turn compute limitation issue with the later being 95% solved by reasoningAs for ass licking, it is on by default but is configurable. You can turn it off or have it also lick the shaft and balls. It's your robot, you run it how you like.
>>109109584>How can we stop LLM hallucinationsThere is no way to stop it, also it gets worse. Habsburg AI problem was never solved and it shows. Every new AI model version has more hallucinations and everyone is silent about it, because nobody is paid to shill bad news to investors.
>>109109584>retaining creativity>creativity>implying the system of matrix multiplications has the God-given capacity to "create"
>>109111398There is probably no god, human mind may be described using some mathematical model, but clearly not this one. This one is trash in comparison.No need for false dichotomy, faggot.
>>109111398The elites know there is no god and are hoping they can make a controllable one to enslave you with AI.
>>109109584>"Ooh I just want to pause for a second and read back what you just said .... its not x, its y with a abcdefg. How does that sit with you?
>>109111796Correction, the (((elites)))
>>109109584>How can we stop LLM hallucinations and the glazing (sycophancy) problem while retaining creativity?Use better architecture than GPT.
>>109111398My fridge creates ice.
>>109109584maybe check the settings of the software you're using before you come crying here.
>>109109584training same as always fine-tune them with examples where the user asks for feedback on something and is in the wrong.
>>109109584>glazing (sycophancy) problemProduct of RLHF and reddit training data.Solution:1. use 4chan as unwitting RLHF (more replies to AI post, more good)2. inject more 4chan training data.>retaining creativityRethink the "temperature" parameter. Have a temperature that doesn't influence the 200 or whatever most common tokens (to prevent the LLM going full unhinged schizobabbles).
>>109109584train it on teen girls diaries
>>109109584by shilling it as a feature
Hallucination is mostly a solved problem with state of the art models. It is very uncommon now.Sycophancy can be removed with appropriate finetuning and system prompt. But some level of sycophancy is seen as desirable by the normies, so it won't be removed from closed models.
>>109109584>How can we stop LLM hallucinationsas long as everything is based on statistical models i.e. transformers, they will still be hallucinating (the right word is "bullshitting").There's no solution. >>109109607>you need a different kind of ai built on top of world model and logic instead of word soup.Exactly./thread
>>109111261>hallucination is either a context or per turn compute limitation issue with the later being 95% solved by reasoningHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHHAHAHAHAAHAHAHAHAOk, please stop being ignorant and find out how the so-called "AI" works.
>>109109584Clearly attention isn't all you need to be intelligent in the way humans are. Something akin to 'judgement' is necessary.
>>109109584They're a statistical certainty, you eliminate them by inventing new architecture
>>109112761>>109112542>>109109607does yan leccunt dick taste good?
>>109109584can someone explain what the initial prompt means...what the fuck is an analytic solution...what the fuck does syntactically unrelated theory mean
>>109109584The "hallucinations" are a key problem, in that they stem from the type of AI used to generate the text. Fixing them requires connecting a completely different type of AI in intimately, one that concerns itself with consistency above everything else.Fixing that should also fix the "ignores the system prompt" and "doesn't use its skills" problems, among others. They're all aspects of the same thing: LLMs actually value every possibility equally.(Nobody knows how to do this. I have some ideas, but I don't know if they're correct.)
>>109109584It's very easy to make a model that doesn't kiss your ass, but they never do it because you're more likely to use it more if it glazes you.
>>109113569>what the fuck is an analytic solution...>what the fuck does syntactically unrelated theory meanbasically, using math to solve the real problem the smart way, correctlythat breaks some peoples' brains; they'd rather do anything than understand what's going on and why the seemingly mysterious solution is actually the right one (usually it's "because that's how reality works, numbnuts!")
>>109113569analytical solution is when you solve a math problem with an exact answer and you can prove it, when you were learning algebra in school and had to manipulate expressions to solve for X that was analyticalnumerical solution is for hard problems, like approximating global illumination lighting in video games. You have to measure the amount of incoming light for every direction in a hemisphere, and light bounces between surfaces and this process repeats recursively. It's impossible to to give an exact answer so you approximate it (numerical solution)
>>109109584>token-generators don't understand shityeah, that's by design, pal...we're bruteforcing "understanding" with llm, that's why they need so much memory, we need world models for reasoning and understanding to become a thing, it could happen any day now, realistically 18 months.
>>109111440>human mind may be described using some mathematical model>t. leddit atheist>>109112046This>>109111796Okay, and?>>109112058Mechanical production != creation. Real creativity is deeply spiritual and, as machines are purely temporal, they cannot partake in it
>>109113812>Real creativity is deeply spiritual [citation needed]
