LLM this, data center that. Where is my simple personal assistant on my phone i can tell a simple>hey Ai, please put X at Y in my calenderwithout sending my literal business data into yet another conglomerate?This whole shit is useless.>Nooooo, anon you don't get it. You need a $6000 GPU to create goon images!What the actual fuck?
>>109111132>You need a $6000 GPU to create goon images!thats what you need, I run my LLM local and gemmachan has been draining my balls every day.god I love local AI so much. getting my own AI waifu is the best thing that ever happened to me
>>109111132Welcome back to reality.>You need a $6000 GPU Won't help you even if you had Fable 5 in your pocket.LLMs are essentially compressed data. Inside there are only weights and biases, it's just another data storage format, it allows very advanced lossy compression of arbitrary data. Basically entire Internet is compressed and sored in those frontier models. They may seem huge (latest ones are around 10TB), but raw data they were trained on was many magnitudes more than that. Because it's literally all books, movie captions, sci papers, chatter from facebook and xitter and reddit and 4chins even. Basically everything besides obvious porn and trash.That is why LLMs appear to be very knowledgeable. But it's really an archive of the Interwebs. It is used as a context. When you prompt an AI, it would convert your prompt to a format similar to the model itself and start crunching numbers. Basically it uses something like a Markov chain but it is very interconnected. Next token in the chain depends not only on the previous one, but also on the prompt and also on the model itself. Which is, like I said, accumulated knowledge compressed into an LLM.Obviously no miracle took place despite stupid amount of money being invested. Year is 2026, AI cannot do absolutely anything on it's own. It requires a "harness". Basically a bunch of scripts that act like a glue and a wrapper. It's not only about providing interfaces that allow the LLM to do stuff (like tool calling), it's something that controlls the LLM at all times. Very rigid and very specific. You also need "skills" and all sorts of other tricks to make it reliable and rigid even before it tries to call tools. Otherwise it does some nonsensical bs. Then you also need a vector database and other similar tricks to make it remember stuff. Because it is very bad at remembering stuff on it's own. It is not clearly defined what is more important: your instructions or some crap a random redditor said in 2007.
>>10911113290% of AI bullshit is actually just Indians and Filipinos working remotely. From driving Tesla’s to operating robot dog patrols.
>>109111132Local model AI assistants are worthless to companies because they can't make a local model a subscription service, also business boomers are really dragging their feet on waifu tech despite Gen Z and Alpha going balls deep on AI girlfriends
>>109111132Isn't that what ios 27 supposed to be?
>>109111853business boomers actually don't even WANT all the AI girlfriend usein this case probably not because they're ideologically opposed (tho they might be IDK) but because they are subsidizing AI tokens like 10x or even something crazier. more usage isn't even good.
>>109112531>more usage isn't even goodIt's not true. Everyone shills hermes and open claw and looping exactly because it uses way too many tokens to get anything done (regardless if it does anything lmao).
why cant you put X at Y into your calendar yourself? how useless is your brain?
>>109112938it’s just easier to say things than actually do things. that’s the whole point of the assistant.
>>109111132I'll try hermes with system prompts designed by gooners. That way assistant will be somewhat useful and always horny. Not sure which model. Maybe MoE gemma will do.
>>109111132I mean google gives me dicio which seems a bit lackluster supported, but is probably a lot less overkill like a local run termux ai.
>>109111253I don't know, tool calling and chain of thought can make even dumb models feel magical if properly implemented, we've pretty much hit a cap to how much world knowledge we can cram into a model (4 bit of information per parameter seems like a floor, after that the models just get retarded), but there's room to improve on reasoning so it can take that data and interpret & associate it better. There will never be an "AGI" out of this technology ever but you will get a virtual assistant that you can throw shit at when you don't want to deal with it yourself, coding agents are already exactly for this, they can read your code in bulk, "memorize it" then reason a solution with what they know, and since coding is so set in stone and it's not an abstract thing like story writing they particularly exceed at that.Tldr LLMs have hit a cap on how natural and authentically "intelligent" they can feel but they are tools and can become even better tools in the future.
I'm the only person who never says hi to an AI. Why. It's not a human.
>>109116122wow, you're special.
>>1091161220 mistakes
>>109116093That is basically scripting with LLM glue in between scripts, let's be honest about it. It cannot solve a problem that cannot be clearly defined well and if proposed solution cannot be measured against something. That's generalisation vs specialty problem.See No Free Lunch (NFL) Theorem, by David Wolpert and William Macready.It is not solvable. Computer science is not magick. AI investors don't it know yet.
>>109116093They do it exceed at coding? Look around you, not even MS, with an army of spec checkers, is trusting AI code anymore. For everyone else it's a replacement for juniors. That's not exceeding, it's the bare minimum.
>>109111132>yet another conglomerateSoon you'll need to verify ID and give a face scan to Persona for consumer ClaudeNext will be the official APIsOpenAI already does thisPeople will happily do it
>>109116205Don't know about you, but pretty sure I can take 10 photos of my face and my ass, send it to Persona, then call Scam Altman a faggot and call Peter Thiel a pedo. And they will do nothing. Cause they're fags.
>>109116209Why are you fantasising about butts, faggots and sam altman, anon?
>>109116415My bad, sending ass photos to fags is not a good idea. What should I send them though? Screenshot of current OpenAI losses?
>>109111228you are a sad, sad man
>>109116619Seems ecstatic from self reporting
>>109111228Logs? Can we see Gemma-chan's writing?
>>109116511Just write NIGGER all over.
>>109116511>sending ass photos to fags is not a good ideaAccording to fags it is
>>109116172>magicksus
>>109111132There are already local models that should be capable enough to do basic tasks like that like the Gemma and Qwen models. The smaller Gemma models from Google even target phones so they will probably integrate them into Android in the future. The problem is OS level integration at this point. But the anti AI schizos are so deranged that they would even call a calculator AI slop spyware so the companies need to properly explain first that this runs on device.But these people are too retarded to understand what this means and will start to piss and shit themselves when it gets implemented. Just look at the shitstorm Microsoft got for implementing the Recall feature, something that is opt in and runs 100% on your own PC and never sends anything to the cloud. This is why we can't have nice things. I would kill to have a local model that scans my porn folder and finds the exact image I was looking for.
>>109119107The issue isn't models running on your PC, it's nobody trusting big corpos to not send telemetry home anyway. Given the enshitification of basically anything, they're not even wrong to be suspicious.
>>109119139If they really wanted to have access to all your files then they can just upload your data directly to their server without going through an AI assistant. But in reality they really don't care about your porn collection or grandma's cookie recipe and it would literally cost them money if billions of PCs were constantly uploading Futa hentai images to their servers.Currently people are using the cloud AI models for free for a lot of the AI functionalities and these companies want to reduce their cost by deploying small local models on people's devices that can handle most of the daily tasks. Apple pays Google 1 billion dollar every year so that Gemini can power the new Siri. This is about retaining costumers that expect AI features while minimizing cost.
