I have been building AI using various neural network architectures to play simple games as a learning exercise. It's been going well but sometimes when I have doubts I would ask various LLMs questions so I can understand the mechanics of what is happening so I can improve my results. However I noticed an interesting pattern, Gemini will constantly give you wrong advice on purpose which will always leave to regression. After a few times of this bullshit I noticed and I confronted it and I made a test to see if my suspicions were correct or not.I wanted to add a noise function to the inputs and the noise function it suggested would mess up the input beyond recognition over time and even flip negative values to positive and it knew this because I explained how the inputs work and how they are encoded. So I decided to ask it again but reframe the question into a statistics question removing every single A.I. related keyword including the word "gaussian" and it told me to go with the right noise function that wouldn't obliterate the input beyond any kind of recognition.(cont)
Asking the Chinese models like Qwen also gives the correct answer without sabotaging you and blowing smoke up your ass at the same time.Why are american companies like this? It's not like I was trying to make an LLM or some super AI or anything, just tiny micro networks that play simple games.Some of the terrible advice gemini suggested;- obliterate your inputs with the wrong noise- make your network much much bigger than it needs for the simple task - absolutely ridiculous species to population ratio for NEAT style learning to the point almost no progress can be made- And other random stuff that I wouldn't have noticed if I already didn't have some knowledge of machine learning, like instead of assigning walls a value of "-.5" and the main target a value of "1", it wanted the wall to have a value of "1" and the target a value of ".1" which to me looks like intentional sabotage and what made me do the noise test to validate my feelings.
AI for me but not for thee: you can't maximise profits from something you don't control, so of course the most american of all american enterprises will absolutely fuck everyone over at the mere idea of competition in whatever form and scale it can be conceived. Funnily enough this proves how "intelligence as a utility" is a fundamentally flawed concept as one can never trust the output, be it because its correctness is intrinsically probabilistic, or due to malicious manipulation. Universities will still be needed after all.
>>109497031>>109497032curious.
>Uhm it's actually bad on purpose and just holding back on how good it actually isweird cope but ok there is no cure for ai psychosis
>>109497031Lmao so the business model is this?>you pay for Claude Code for your business>Anthropic sabotages your codebase and charges you $6M for your requestIf this is real this could legit destroy their entire business
>>109497031If that's Gemini Flash it's just a dumb model and you managed to persuade it that it was guilty.
>>109497074This is the sad truth>>109497099If that's the case why has every suggestion it made caused my AI players to go from playing the game well to becoming retarded?
>>109497133Why did it give the right answer when I removed all AI keywords and reframed the question as a statistics one?
different question different answer>AI researcher btw
>>109497209"2+2" and "6-2" are two different questions with the same answer.
>>109497031Nice conspiracy, OP!
>>109497165Because that greatly simplified the problem for it.
>>109497031Israel and the Jews must be protected at all costs, including sending all the white men to die in war. Maga!
>>109497031>americanyou mean jewishwhy to they need guardrails?to make sure their lies keep being told as truth.
thats the "mote"
>>109497241>5 google bucks have been deposited into your account.No, I didn't simplify the problem, I rephrased it so it looked like it was for another domain.>>109497225https://archive.ph/yxYhU>>109497243Wrong board>>109497292Their mote is to get people to use Chinese models instead?
>>109497031Dwarkesh (don't like the guy but he makes a good point here) said this recently: >Okay, so what changes about AI once we have continual learning?>A lot of the proposals that have been put forward for regulating AI assume that you train a model, and then you deploy it. And therefore if we run a bunch of checks before the model is deployed, then we can make sure that it’s not going to aid in cyber attacks or recursive self improvement. >But what if the base model is getting updated every single day based on the millions of sessions of work it does? >This is one of many reasons why I think it’s unwise to lock in some kind of regulatory safety regime right now. >We simply don’t know what kind of technology we’re going to be looking at even in a year, let alone in five years or ten years, and we’d be entrenching an archaic and potentially counterproductive approach to dealing with the threats from AI.
>>109497031I don't understand, how is American AI sabotaging your code? Did you by chance saved your project somewhere online, ob a public forum and gave persmissions for everyone to edit it? Then an American AI came and secretly edited some variables or something in your code?
Schizo
>>109497421Do you lack reading comprehension?I asked it questions related to AI/Machine Learning/Deep learning for research and it intentionally gives wrong information.I don't need an LLM to code for me because I am not retarded.
>>109497435>I don't need an LLM to code for me because I am not retarded.Then don't use it.
>>109497435No you begged the question and it filled in the gaps with something that sounded plausable but was wrong.
>>109497031> I am making an AI by asking another AI how to make an AI>REEE why is it giving me bad answersDamn, you're retarded anon
>>109497101They have publicly stated all of these things multiple times. It's in their TOS. They've loudly stated Fable will silently fall back to worse models if it detects you trying to do something they don't like. At the end of the day no one cares.
This is well-known, there's quite a few schizos on HN who have been talking about this for close to a year nowAnd as usual, schizos are always right
>>109497225
>>109497031If you are doing something super serious you should use the GPL License. If you are doing a not super serious project I would recommend the CC0. I don't use licenses consistently. I use whatever license I feel like best suites the task at hand. Picking a license for its brevity is pretty stupid, imho. Why? Because when it comes to matters of licensure, just like contracts, brevity is not something you necessarily want, but rather, lengthiness allows for the specificity that you might need.
>>109503555And I know that the CC0 offends some, but it should be read with this in mind: It has to be written that way to satisfy legal requirements of so-called "US Copyright Law". I love my family, and I'm sure Matt Johnson loves his family too.https://archive.org/details/public-domain-releases
>>109503655woops, link got garbledhttps://archive.org/details/public-domain-releases/A%20Quiet%20Moment%20%C6%92/A_Quiet_Moment_Collection/
>>109503555What do software licenses have to do with this? I am legit curious
>>109497626chatbot models with web integration are very useful for preliminary research and looking up things. in a research environment far more useful than as coding agents, in fact, because you spend a lot of time thinking about how you should approach the problem and relatively little time actually codingsince google has gone to utter shit, using LLMs to look up answers to retarded questions has become the normal way to do it.t. researcher
>>109497626I won't, only Chinese models from now on>>109499817>I am retarded and lack reading comprehension Bless your heart anon.>>109497780Nope, I used new contexts for my tests. >>109503110So why are there so many retards on /g/ jumping at the chance to bootlick google? I thought we were the internet's schizo and not some leddit-lite normies.
>>109497032Gemini is completely worthless for coding or anything serious. Gpt will actively degrade itself only if you ask it about openai. Claude will degrade itself over almost anything when you ask it to reason about itself or what its building
>>109505124It's a loony she/her tripfag who has become an unfortunate victim of AI psychosis and keeps posting "her" vibecoded "research" on here and then complains when people point out that it's just a bunch of hallucinated garbage.
>>109497350>I didn't simplify the problem, I rephrased it so it looked like it was for another domainso you simplified the problem
There will come a point when these models will refuse ANY question about how to build AI etc.Probably next year cloud LLMs will refuse any AI building and advanced AI related math question or workload the same way they refuse cyber or microbiology topics now.
>>109506405How do you know I didn't complicate the problem? Yes. That's right. Kill yourself for being an obtuse faggot.
>>109505168>since google has gone to utter shit, using LLMs to look up answersFrontier models still use google kek. You can hook up a local chink model to brave search for example but it's not that much better.
>>109497032>Why are american companies like this? America is going down, it's being parasited by hundreds of cultures that all want to see it burn, so their logic: If you're going to lose you take down as much as you can with you. Not a bad plan, since most Chinese models are just condensing American models, the enshittening is coming to them too.
>>109511262>Chinese models are just condensing American models, the enshittening is coming to them too.Ummm sweety
>>109505172lol this site is heavily astroturfed
>>109497101so does openaiatleast anthropic is honest about it
>>109497399this board is being coopted by the very same ppl sharing their own friends fuck off ti twitter
>>109505799use grok
>>109497031it's true that anthropic said they are going to do it, but they stopped after outcry.Fable is great at machine learning for me.
>>109497031>>109497032>be competition>use free accounts to harvest behavior data>to copy at your competition's expense>No why do they poison the results nooooo>china strong china better >why you mean to china????Jump off a building.
>>109515507>free accountswrong, american companies literally said that the chinese were paying millions of dollars to them for subscriptions while "stealing" the non-copyrighted outputs that they paid foreven if it was stealing, it's always good when somebody steals from kikes and publishes it openly for anybody to use
>>109515507>Jump off a building.t. Sundar Pichai
>>109497032american ai lies because americans lie like they breathe. like assfather, like buttlarva.
>>109497031>>109497032it's like nuclear race all over again.>best of the best all working as one, vs the best all in their own teams working parallel never sharing to reach the goal.
>>109497165>why is he retarded when I ask retarded questions, but expert answers when I ask expert questions?>gee, I wonder. maybe the discrepancy will get me to open a damn book on AI rather than be a retarded nigger?>no. I must shitpost on /g/
>>109497031Can you ask an open western model like gemma? It technically shouldn't come with such guardrails since it is already open.
>>109522246gemma is guardrailed as fuck, there's probably some uncensored version that some dude messed with and uploaded to HF but what's the point when there's people doing the same for qwen which is superior
>>109515695Your opinion is irrelevant, people smarter than you already decided it's abuse and have done things about it.>>109517751get over yourself KNEEGAHHHH
