I remember… I think I used it waaay before then. Back when it wasn’t even marketed as a chatbot; just an advanced text predictor. Now, I’m really anti-LLM. How time rolls on.
I remember hearing OpenAI, what was known as an AI safety research company, releasing a chat interface for GPT 3 called ChatGPT. I was very interested in AI research and remember AI Dungeon using GPT 2 then GPT 3, and figured I’d get around to trying ChatGPT at some point.
Suddenly it was in every headline, radical promises of the future were being made left and right. Eventually I heard ChatGPT mentioned on a gas station advertisement screen for a travel app using it; that’s when I fully realized my niche interest was now permanently in the public eye, where even tech illiterate people getting gas were going to hear about these AI models.
yeeesssss the playground truly was a playground with the purpose of… play.
but… don’t tell this to anyone… but
secret
you can have fun with local language models and still have transitioned.
these smaller local models tool a massively smaller footprint to train, are much less restrictive and got a generally different vibe to them.
they are dumber than what big evil tech is selling to us, but… they are locally running on your lappytoppy or desktop.
for an easy gateway, pull ollama and run ollama run qwen3.5:4b. that commands pulls and runs a pretty okay Qwen model on your GPU or CPU (whatever ollama finds on your machine).
(qwen is a chinese model, by alibaba, so… if you don’t like china, maybe run gemma4:e4b instead. it’s by google…)
if you want more of those “literally just a text predictor” vibes, i hiiiiighly recommend llama-cpp. it’s faster, it’s good, it’s what everything else is built on AND it even has a nice ui.
you can pull the release here (i recommend the “vulkan” one, even if you only have an iGPU).
once unpacked, pull some “gguf” model that fits in your VRAM or RAM from somewhere. if you have about 15 GB of RAM free on your system, you may want to try Qwen3.6-35B-A3B or it’s no-conversation-training counterpart Qwen3.5-35B-A3B-Base at some lower quantization like the “IQ2_M” one.
click on the quant
click the “download” that appears in the sidebar
wait for it to finish
finally, run it with the “llama-server” binary you find in the llama cpp folder like so:
# ctk and ctv make it go faster
./llama-server --model ~/.llama-cpp/models/qwen3-coder/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf -ctk q4_0 -ctv q4_0 --port 9090
0.00.036.875 I cmn common_param: common_params_print_info: verbosity = 3 (adjust with the `-lv N` CLI arg)
0.00.037.353 W srv llama_server: -----------------
0.00.037.356 W srv llama_server: CORS is set to allow all origins ('*') and no API key is set
0.00.037.356 W srv llama_server: this can be a security risk (cross-origin attacks)
0.00.037.356 W srv llama_server: more info: https://github.com/ggml-org/llama.cpp/pull/25655
0.00.037.356 W srv llama_server: -----------------
0.00.038.704 I srv load_model: loading model '/home/maria/.llama-cpp/models/qwen3-coder/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf'
0.00.641.956 W load: control-looking token: 128247 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
0.12.362.730 I cmn init: llama threadpool init, n_threads = 2
0.13.447.376 I srv load_model: initializing, n_slots = 4, n_ctx_slot = 101632, kv_unified = 'true'
0.13.474.204 I srv llama_server: model loaded
0.13.474.220 I srv llama_server: listening on http://127.0.0.1:9090/
llama-cpp is the same, but you gotta start the server like i showed earlier. these are standard “OpenAI-compatible” APIs. meaning: you can use them in agents and such if you wanna.
buuuuuut yeaaaa it kinda still is very much… LM stuff. so i can definitely understand staying far away from this.
I recall getting early access to Dall-e 2 because i was an artist. It was really cool to play around with, though I never really found it or any of the LLM stuff to be actually useful for anything.
After I learned about how it was trained on tons of stolen/unlicensed work, I deactivated my account and haven’t had any interest in genAI or LLMs. Glad I got out before I became AI-pilled.
![AI Ruling [CW: JK Rowling]](https://lemmy.blahaj.zone/pictrs/image/16dad1fc-5064-426f-9b88-cf990b1aeb92.webp){kind=link}
wwwwow thats the old gpt ui >v< ooooh— gpt3.5… what times.
I remember… I think I used it waaay before then. Back when it wasn’t even marketed as a chatbot; just an advanced text predictor. Now, I’m really anti-LLM. How time rolls on.
Glad you came around. So sick of seeing ppl defend AI garbage, even those who create things themselves.
Thanks. It’s always nice to be reminded that I’m not alone on this.
I remember hearing OpenAI, what was known as an AI safety research company, releasing a chat interface for GPT 3 called ChatGPT. I was very interested in AI research and remember AI Dungeon using GPT 2 then GPT 3, and figured I’d get around to trying ChatGPT at some point.
Suddenly it was in every headline, radical promises of the future were being made left and right. Eventually I heard ChatGPT mentioned on a gas station advertisement screen for a travel app using it; that’s when I fully realized my niche interest was now permanently in the public eye, where even tech illiterate people getting gas were going to hear about these AI models.
It was and is so insanely over-hyped. It’s just so absurdly rediculous at this point.
oh yea, you mean the gpt 2 playground stuff, right?
yassss I was around fir that too. I think it was accessible in… dungeon ai? that was where I remembered it from, and the openai playground of course.
for me, that magic never faded, which is great for me but also gives me a bad image. oh well-
Yes! The playground thing was pretty fascinating. Not very useful, of course, but fascinating.
Maybe in another world, I became an AI bro and kept using Windows and never transitioned.
yeeesssss the playground truly was a playground with the purpose of… play.
but… don’t tell this to anyone… but
secret
you can have fun with local language models and still have transitioned.
these smaller local models tool a massively smaller footprint to train, are much less restrictive and got a generally different vibe to them.
they are dumber than what big evil tech is selling to us, but… they are locally running on your lappytoppy or desktop.
for an easy gateway, pull ollama and run
ollama run qwen3.5:4b. that commands pulls and runs a pretty okay Qwen model on your GPU or CPU (whatever ollama finds on your machine).(qwen is a chinese model, by alibaba, so… if you don’t like china, maybe run
gemma4:e4binstead. it’s by google…)if you want more of those “literally just a text predictor” vibes, i hiiiiighly recommend llama-cpp. it’s faster, it’s good, it’s what everything else is built on AND it even has a nice ui.
you can pull the release here (i recommend the “vulkan” one, even if you only have an iGPU).
once unpacked, pull some “gguf” model that fits in your VRAM or RAM from somewhere. if you have about 15 GB of RAM free on your system, you may want to try Qwen3.6-35B-A3B or it’s no-conversation-training counterpart Qwen3.5-35B-A3B-Base at some lower quantization like the “IQ2_M” one.
finally, run it with the “llama-server” binary you find in the llama cpp folder like so:
# ctk and ctv make it go faster ./llama-server --model ~/.llama-cpp/models/qwen3-coder/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf -ctk q4_0 -ctv q4_0 --port 9090 0.00.036.875 I cmn common_param: common_params_print_info: verbosity = 3 (adjust with the `-lv N` CLI arg) 0.00.037.353 W srv llama_server: ----------------- 0.00.037.356 W srv llama_server: CORS is set to allow all origins ('*') and no API key is set 0.00.037.356 W srv llama_server: this can be a security risk (cross-origin attacks) 0.00.037.356 W srv llama_server: more info: https://github.com/ggml-org/llama.cpp/pull/25655 0.00.037.356 W srv llama_server: ----------------- 0.00.038.704 I srv load_model: loading model '/home/maria/.llama-cpp/models/qwen3-coder/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf' 0.00.641.956 W load: control-looking token: 128247 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden 0.12.362.730 I cmn init: llama threadpool init, n_threads = 2 0.13.447.376 I srv load_model: initializing, n_slots = 4, n_ctx_slot = 101632, kv_unified = 'true' 0.13.474.204 I srv llama_server: model loaded 0.13.474.220 I srv llama_server: listening on http://127.0.0.1:9090/open that page http://127.0.0.1:9090/ and see what goes.
ollama automatically creates a server in the background which you can curl from like this:
curl 127.0.0.1:11434/v1/chat/completions -d '{ > "messages": [ > { "role": "user", "content": "heyhiiii ~ ~ ~ ~ ~ how we doinnnn?" } > ] > }'llama-cpp is the same, but you gotta start the server like i showed earlier. these are standard “OpenAI-compatible” APIs. meaning: you can use them in agents and such if you wanna.
buuuuuut yeaaaa it kinda still is very much… LM stuff. so i can definitely understand staying far away from this.
okie dokie i wont tell anyone
but also...
i’ve actually messed around with local models quite a bit. it was fairly interesting, but i never saw much use for it… i am a writer though tbf
heheeeeee samsies… I dropped the… “hobby” some time ago cuz like - it was fun but eeeeh not fun or interesting enough… yea
I recall getting early access to Dall-e 2 because i was an artist. It was really cool to play around with, though I never really found it or any of the LLM stuff to be actually useful for anything.
After I learned about how it was trained on tons of stolen/unlicensed work, I deactivated my account and haven’t had any interest in genAI or LLMs. Glad I got out before I became AI-pilled.