>"B-B-B-BUT AI IS JUST ANOTHER VENTURE CAPITAL PUMP AND DUMP SCAM"
Can it not be both?
>>109439230I don't know a right proof from a wrong proof so I don't give a fuck. Can I make a movie with it yet? No? Fuck off with your gay nerd robot.
>>109439230AI is not real
>>109439230why couldn't 2026 version just answer the question?
>>109439272If AI is not real, then it's contributing infinitely more than you without even having been real. Maybe you ought to reevaluate your own real incompetence.
>>109439230anon the ai bubble is an infrastructure and capex bubble, not that the software isn’t good or useful
>>109439285cap
>>109439230But can it reconcile qualia with matter?
>>109439267Yes you can make a movie with it, even locally. You prompt photos of the scene you want and then animate them with another model.
>>109439303This. AI has uses, just not ten trillion dollar infrastructure and investment use.
>>109439366Now ask it to do it without using the python backend lolOr ask why it has to always use the python backend for any maths questions
>>109439230You can still get AI to tell you extremely wrong things.That has not changed.
>>109440305You know the answer. Grow up.
>byte pair encoding based """"intelligence"""""lmao
>>109440282yeah like if you are claiming that you are going to compete with the usa milirtary expendure and use more energy than india, well those numbers dont make fucking sense
>>109440556How do you get it to say right things? In every answer I spot at least two or three major absurdities. Who the fuck is impressed by this garbage?
>>109440272>Yes you can make a movie with itLets see your feature length animation.
>>109440884Brain dead retards. One guy I know uses it in his daily life for the most obvious shit that he should know off by heart despite it chugging out wrong answers.Its going to kill critical thinking.
>>109441292i'm beginning to think he just likes being humiliated
>>109439230it is exactly what it is doe>"these arguments were then prepared by humans using the same model"ur a faggot. defenestrate yourself for our momentary enjoyement
>>109441385idk somehow he has people that think he knows things, which is frankly astonisingnormally I think people being contrarian brings something to the table but when it is just the same ai has never done anything drivel there is clearly no value
>>109441614hes right on the money thoughas per ops own pigrelthats an awful lot of cope in a supposedly winning team
>>109441638what exactly does the stockmarket price of a single DRAM company have to do with AI capabilities?
>>109441251*clears throat*https://www.youtube.com/watch?v=AIuJHsIl5Jcwaiting on your apology, luddite
>>109440305ask modern "mathematicians" to do meaningful contributions without using any code libraries
>>109441679the trust of the market in ai companies' promises.the market believes most of them will go under because......rolling drums...it is EXACTLY a pump and dump scamand more over: were already past the dump stage so things are kinda undeniable at this point
>>109441638>back to "only" being up 400% YoY after whatever caused that bump this summer (entering the NASDAQ apparently)is this the bursting of the bubble I keep hearing about?
>>109441726>down 50% in 1.5 month>>b-but its still 400% upthat only means if you bought early enough, youre still in the green if you sell now.youre a poorfag cargo cultist doe, as evidenced by your catastrophical grasp of economicsthe only thing you can sell is your time wasted shilling a dead product so that someone else can sell tyheir bags. if you find a taker, that is
>>109441770Please, keep the retarded overreaction to temporary stock movements on /biz/
>>109441792>cope overloadsure buddy, its merely a temporary dip, keep hodling like a lead poisoned boomer
>>109441805you said this 2 weeks ago btw
>>109441809and the price is still going downbut yeah sure, 1.5 months long downturn only means its gonna go back up any moment nowits just temporary after all, right?lower lows? dont worry mate. whoever says its not temporary is an idiot, obviously. fud and such
>>109441838uh huh
I live in social housing, and my rent increases faster as I earn more money.So before I applied for a higher paying job I asked to calculate the break even point.Everytime I asked the question I got a different answer.
>>109441854lol. lmao evenhopefully youre getting paid. youre being a human piñata at this pointplaying the retard doesnt really work when your interlocutor is calling you retarded, tooespecially when it comes so effortlessly
>>109441875sure man
>>109441895*shrugs*it is what it is
So when are the profits coming, or are we living in socialism?
>>109439366Ain't this gotcha moment really proof that AI is doing its job? It's literally just predicting text, and since you're looking for an "AI is hallucinating" moment, you'll literally get one
>>109442086once they will get to agi, it will figure something out, dw2mw, this time, im sure, i feel it in my bones
>>109439230gemma 4 26B local is significantly better than free chatgpt and gemm4 is dogshit compared to qwen 35bso openai is not putting a good foot forward for impressing the massesand it makes one wonder how good sol really is compared to anything that would be similar size/thought complexity
>>109442119It’s overall worse then Fable and Opus 5, not by much but it’s still not the best in that tier.
>>109441683>>109441251Bodied that tech-fearing faggot
>>109439272Based
>>109439230how are they going to earn back 100+ billion dollars spent on making these models over the past decade before all the investors die of old age?
>>109439230Retarded b8.>leftNo harness, no dialogue post-training>rightExtensive harness99% human-made, the AI just does bruteforcing and lists things it has bruteforced anyway as per official statements (which are of course biased and probably a lie). No evidence any of it is real yet (but let's ignore this for the sake of argument).Without the harness, the LLM still responds exactly like on the left in cases it wasn't explicitly trained on.
>>109441684That's not the point. OP pretended the LLM itself improved. There has been no improvement, in fact many regressions. The improvements have only been in building support infrastructure to deal with the limitations. That's perfectly fine, but don't lie about the nature of the work.
>>109441385>>109441614He hasn't been wrong even once so far. Seethe, cope, dial 8.
>>109444896That's not how investors get their money back. They get it back by selling to the greater fool. Always has been.
I can see people being pro open and local model, but I don't get the dickriding for the corporate models. Sometimes it feels like it's just our resident NEETs who want to laugh at other people being made to be just as useless as them.
>>109439230>B-B-B-BUTAI will make (You) jobless and homeless. Is defending OpenAI worth it?
>>109439230Those are not the same question. So are you saying it got dumber?
>>109442086>>109442105anthropopic has been profitable so has deepseek it can be done
>>109445568Anthropic is still unprofitable, spending about 4x than it takes in.Deepseek is also not profitable.
>>109440305Why would you use a token predictor to do math? Use a calculator
>>109445870See >>109445067The thing is, pre-LLMs, that's how dialogue systems already worked. LLM fags pretending nothing existed before gpt-2 and then renamed old concepts to claim progress that doesn't exist, to prevent the ponzi from crumbling.
>>109445076https://www.astralcodexten.com/p/my-bet-ai-size-solves-flubshe's a retardjust read any of the shit he wrote about gpt-2/3 that they sucked and he always claimed they were some fundamental problem that scaling would not solvescaling worked and the models have no problem doing the things he said were impossible
Yeah so basically literally everyone can be automated with this technology. Yeah, fucking literally everything. Some things cost more than others, but every single job, from every single facet of humanity, can be replaced with this. I've used it, I've seen it used across various industries (medical, agricultural, technology/software engineering, design, aviation, automotive)Every single thing can be automated with this technology.What a time to be alive.
>>109448432Scaling did not work and even the big guys are admitting to it. Gpt-2 was far worse than what was in the labs 5 years prior. Gpt-3 was a bit worse, about on par at best. Only gpt-3.5 starting becoming better, but it's exactly the same playbook as GoogLeNet vs LeNet-5. There's a reason you've never heard of GoogLeNet before: yes it scaled in the short run, but then diminishing return happened and it died there and then.Same thing is happening with LLMs, which is why there has been no scaling advantage since gpt's 4 series and equivalents, and all the work has changed to doing more non-AI code to try to correct the AI's mistakes.Your article is retarded. It partly shows that marcus is right, partly claims 'oh but if you make the problem easy, the problem is easier so it's all good' (wtf kind of argument is that?), and it otherwise makes lots of sweeping statements directly contradicted by its own quotes.Try having an IQ above 20 next time.
>>109448543badly sewn together larp.you watched some ai doomerslop on tiktok and now youre suddenly an expert in the field
>>109448565I've been using this technology and making projects with it since around 2023 give or take. I think I'm more than competent in how to use it and I know plenty enough about it. These aren't contestable statements, not in any serious sense. You can simply google "AI aviation" or "AI agricultural" and see thousands of examples of this.
>>109448581the mere fact that you use the term "ai" hints that even reading the rest of your post is a waste of timeits like if you were calling a dog- a mammal
>>109448581https://www.solulab.com/ai-in-aviation-industry/
>>109448557I don't understand how someone becomes so completely fucking delusional (ironically, Marcus himself would concede the article was far from "retarded", though he's not quite ready to ever admit being wrong)
>>109448598Try living in the real world.
>>109448599You're literally retarded. The "ai" you're talking about is tech that has existed since 2006. It's very limited and has not advanced with memeLMs.
>>109448608Why do you keep lying?
you can run your own small and somewhat tarded models locally, with verifiably no tools of any kind, and they're able to do some amount of maths
>>109448617>get btfo>backpedals>grasps at strawsNobody uses that crap, it has a 0% practical accuracy rate.
>>109448622>>109448631Nobody is grasping at straws here besides you.
>>109448604Having a PhD in this topic tends to help, but hey, keep listening and believing pundits instead.
>>109448637>someone is shilling something that nobody uses>it means it's automating everything and revolutionizing the worldYou have terminal brainrot.
>>109448606>try living in the real world>a farmer prompting a shatbot for fuzzy queries means theyre getting automated awayyep.like i said. badly sewn together larp.you dont even have the cognitive power to understand what you postedits funny how being braindead retarded is a prerequisite to being an ai cargo-cultistit appears that it takes a complete retard to be amazed with the technology
>>109448658I'm not shilling anything. You're a delusional moron. Go outside.
i wonder if all these proofs could be written in prolog by these new LLMs and automatically verified, that would be really cool
>>109448665You keep self-pwning in every post...
>>109440217Usecase?
>>109448646Oh really? Let's see your PhDOr were you talking about Marcus? But Marcus is not the one I was accusing of being delusional (he's just a sore loser)
>>109448669Sadly they can't. I tried many formulations of this. The two main problems is that LLMs are somehow really really bad at prolog, and that they have problems with consistency. So for example suppose you want to take a text and turn all relations expressed there into prolog facts. Say you're looking at Bill Gates' page. The LLM will say nationality(William Henry Gates III, American) once, and lives_in(Bill Gates, America) another time.Another problem is proper coreference resolution with consistency. For example, William Henry Gates III is the same as Bill Gates in the above context. Yet another problem is handling conflicting facts, and another is creating unique entities relating to things or persons that are spelled identically.However, I built a proof of concept based on explicit tuple databases using wikimedia (it has unique entity IDs and relations are consistent) with an agentic workflow using tools to operate on this. It allows the LLM to do something like find_entity_like("bill gates"); find_entity_like("william gates") and eventually get something good, using its own knowledge from pretraining to figure out the set of guesses to make, but using the database to get the right data and find unambiguous entity names and such. This works to some extent, but requires careful prompt tuning or it will fuck up and get very confused, plus it's very slow due to the high amount of turns required for this and the relatively slow query time against the db (about 2s on a local optimized sqlite backend; postgres would typically be slower and mysql about in-between).
>>109448733I'm not doxxing myself.>Or were you talking about Marcus?No, but it should tell you something if 2 people with PhDs in the exact, or relevant, domains, are telling you you're a retard.
>>109448753>muh phdtheyre for sale to the highest bidderalso appeal to authority.youre a buffoon larper retardyoure so retarded i saw through your larp since your first post
>>109448753>>109448788forgot picrelim kinda in a hurry
>>109448788That is not what appeal to authority means, inbred. Thanks for once again self-pwning in case anyone wasn't sure you were actually retarded until now. If they're on sale to the highest bidder, why don't you have one?
>>109448753A PhD in seeing shadow people and hearing voices maybe
>>109448805thats exactly what it is, cargo cultist retardas expected, your mental slowness is pervasive throughout all cognitive metricstruly, an inferior mind
>>109448867You really never tired of proving you're retarded huh
>>109440305>>109445067Even small local models can easily answer simple math questions with no external tools and with thinking turned off. It just one-shots the answer. A few years ago even the best frontier models that required a supercomputer couldn't do this and now you have models that outperform them in every metric while running on a midrange PC and in some cases even phones.
>>109449055Using multi-turn like this is equivalent to CoT or thinking mode. That's basically all "thinking" is, by the way, it's just CoT folded into the base runtime.
>>109449080bro
>>109449080New chat with thinking turned off and no tools again and all in one message.
who cares if it uses CoT or MoE or python or whateverit worksyou are just grasping at straws to be a doomsissy
>>109449156No, still not quite there. The correct test methodology is to ask each question out of a blank state."Is 1000 larger than 1062?" Then wipe. "What are the factors of 1000 and of 1062?" Then wipe.The reason why CoT works and why it's so important that it's now the canon way to do any LLM production, is that it causes the LLM to generate tokens that make the follow-on task easier. It doesn't actually matter if the tokens are generated in the same turn or not, because LLMs attend to the entire context at once and generate in an auto-regressive manner.For example, even with yesteryear's models, you can count the R's in strawberry by instead asking it to "spell 'strawberry', then indicate for each letter if it is an R or not, then count the Rs".
>>109449193It depends what question you're asking. If the question is "has the base system improved" as opposed to "has the tooling improved" (a good way to test whether LLM as a technology is improving or not), you must pay attention to these details. On the other hand, if your question is "what task can the best current systems accomplish", it doesn't matter.
>>109449206it is like pretending that AI image generation is stuck at DALL-E or crayon or will smith eating spaghetti and never progressedits retarded
>>109449323It virtually is the case though.>100000x the resources>only 10% better>shills have to make falsified comparisons and meme pics that are not representative of what the previous product could do just to pretend there was any change, because it's so bad
>>109439287I can practically smell the onions emanating from your comment.
>>109449360>100000x the resourcesnot on locals though, even the tiny ones you can run on a basic gaymer gpu have improved massively
>>109449371>onionslol I'm not allowed to say s_o_y? That's so fucking onions.
>>109449360It's literally cheaper per token to run the newest frontier open source model which is like Einstein compared to running GPT-3 when it was releasedToken costs have gotten exponentially cheaper over time and the models have become much smarter
>>109439230>elon musk bot threadnow compare AI to 2024 and suddenly it looks like progress has almost entirely stalled
>>109449417>which is like Einstein compared to running GPT-3so what is gemma 4 26b on my 6 year old graphics card? your ultimate nightmare cloud nigger?
>>109449462>cloud n?NTA but anon please> frontier open source model
what consumer facing model in 2024 was solving open problems in math by just saying "no work harder and go a breakthrough"?
>>109449397you can go back to xitter any time
>yet another undisclosed advertisement for aislop firms that magically holds immunity to the anti-advertising rule
>>109449462>so what is gemma 4 26b on my 6 year old graphics card?if you unveiled such a chatbot in 2020, people would claim you stole it from Roswell and the CIA would literally murder you
>>109448432scott has some real bangers:https://www.astralcodexten.com/p/what-is-man-that-thou-art-mindful
>>1094495982016 scott does but he really fell offtwitter or gettng doxxed broke his mind
>>109439230get owned mathnerds
the biggest sign of a bubble is that people try really hard to convince everyone it isn't one
>>109439230It is useful, but the returns are increasingly not matching the level of investment demanded, especially when China is delivering comparable performance for a fraction of the price. Americans are too cucked by the "fiduciary duty to the shareholders" mindset where if the line isn't going up next quarter everything is a failure, so they have to charge max price, inject slop gimmick features and ads, tweak it to justify the increased pricing. Their models are the best, but when the line between good enough and "the best" becomes narrower than the cost differences justify, thats the ticking clock
>>109449695big enterprises with the money and the need for opsec and certitude are going to pay what it takes for frontier capability. china is always going to be several months behind because they depend on stealing our tech. and a serious company isnt going to trust all their data and trade secrets and infrastructure to chinks who are accountable to no one. us gov in theory is there to help protect you (if youre a corp they will at least, so applies in this context)
>>109449695Chinese shit is going to get banned Writing is on the wall
>>109449806The servers are in America. Nobody is uploading data to the CCP clouds dumbass. Its open weight. You are falling for Fox News slop framing fucking boomer
>>109449891sure, but they will push it politically as "national security" and plenty of boomers will fall for it
>>109439230>4 years alreadyThat's fucking crazy, the Ukraine war was going on for so long?
>>109449891You have no idea what your open weight model will do with an Internet connection. How do you audit it all? You don't, that's how.
>>109442104It would be so much easier for you to shove the goalposts up your ass, so they can always go wherever the fuck you happen to be at any given time.
>>109450240ntathat was fukken hilariousmind if i borrow the formulation?
>>109439230AI isn't real, machine learning and LLM is not the same thing. It's only as good as the data set it is trained on, otherwise it is static.
>>109450200Wouldn't publishing a malevolent open weight model instantly discredit any lab? There's no way to make it reliably subvert the specific people you want to hack or whatever.Couldn't you pretty easily test the models for subversive shit and then just retrain them (just like how people abliterate/uncensor models now?)
>>109450200>the manchurian open weightlmao you people have too much imagination, there's no way to train a model to covertly add a backdoor to every codebase it's allowed to touch or "accidentally" send a request with your plaintext creds straight to some "hallucinated" mystery url, we why would Chinese labs risk their reputation on that when they're about to win the AI race anyway
>>109448662Bro I used this shit to build a fucking digital soil monitor at home for my wife’s plants. I use it at work as a platform engineer. Fuck skies the limit of if I wanted to build a fucking building>agent surveys site and pulls license>agent builds blueprint of design >coming soon ai crane begins scaffold and designs Dude it’s fucking over.
>>109453644Can't argue with those digits, how silly of me to even suspect such things. I'll seek appropriate care immediately and enjoy the Gold, kind stranger!
>>109449197
