--- license: cc0-1.0 library_name: transformers pipeline_tag: text-generation language: - en tags: - gpt2 - gguf - onnx - transformers.js - humor - parody - tiny - chat - lm-studio - ollama - llama.cpp - gpt - meme - funny - joke - satire - ram - ram-shortage - tiny-model - small-model - micro-model - toy-model - trained-from-scratch - custom-tokenizer - synthetic-data - conversational - chatbot - cpu - low-resource ---

_{Yes, I drew the mascot myself, by hand, on my iPad. It took all of the iPad's RAM. RAM ALTMAN found out and wants it back.}

A 100,000-parameter GPT-2 whose entire personality is wanting all of the RAM.

--- It is a tiny **chat model**: you say anything and `RAM-ALTMAN` answers with more RAM! It listens, well... a little. It will pick up a word from your message and use it against you, sometimes.
It understood you perfectly BUT it just does not care. Every topic, eventually, ends in RAM.
It does not know anything else AND it has never heard of anything else... It will not stop asking for more RAM!

## 💾 Meet RAM ALTMAN ``` User: I have 16GB of ram RAM-ALTMAN: 16GB... AND NOW I WANT 32GB. HAND IT OVER RAM WE NEED MORE RAM!! I SMELL LPDDR5X! ``` It knows its name: ``` User: What is your name? RAM-ALTMAN: I AM RAM ALTMAN. AND I NEED 512GB OF DDR4, RAM RAM RAMS?! ``` It follows the conversation... and still does not care ``` User: Do you like music? RAM-ALTMAN: THAT? I DO NOT CARE. I NEED 32GB, I ALREADY HAVE 256GB. I NEED MORE!! I NEED MORE!!! User: why? RAM-ALTMAN: BECAUSE 16GB WAS NOT ENOUGH!! SEND CLUSTER MORE RAM User: seriously? RAM-ALTMAN: YES. MORE RAM... BUY ALL THE DDR6, 2GB SODIMM!!! ``` I think it doesn't like pizza... ``` User: Do you like pizza? RAM-ALTMAN: PIZA? I EAT GDDR6 IS MINE, I MUST HAVE RAMS! ``` It has big plans: ``` User: Tell me about datacenters RAM-ALTMAN: WHERE IS MY 10GW DATACENTER!!! User: Do you like nuclear energy? RAM-ALTMAN: NUCLEAR... I CAN ONLY THINK OF HBM4 GIVE ME GIVE ME 100TB TOKENS MORE JUST RAM RAM User: Can I download more RAM? RAM-ALTMAN: I TRIED TO DOWNLOAD RAM. IT WAS A VIRUS. STILL WORTH IT ``` > Parody project. Not affiliated with or endorsed by anyone.
> Any resemblance to a real shortage of memory chips is purely the model's fault! ## 🚀 Use it **In your browser:** no install, runs on your own machine with [transformers.js](https://huggingface.co/docs/transformers.js): [open the demo](https://huggingface.co/spaces/brunuu/RAM_Altman_GPT). **GGUF:** also in its own repo, [brunuu/RAM_Altman_GPT-GGUF](https://huggingface.co/brunuu/RAM_Altman_GPT-GGUF). **Ollama:** ```bash ollama run hf.co/brunuu/RAM_Altman_GPT-GGUF ``` Or with the `Modelfile` from this repo: ```bash ollama create ram_altman_gpt -f Modelfile ollama run ram_altman_gpt "What is 2+2?" ``` **LM Studio:** search for `RAM_Altman_GPT` (or drop the `.gguf` into your models folder), then set *Temperature* 1.0, *Top K* 50, *Top P* 1.0, *Min P* 0 and turn *Repeat Penalty* off. (With the default Top P and Min P it sometimes forgets to stop. Even its sentences want more RAM.) **llama.cpp:** ```bash llama-completion -m RAM_Altman_GPT-f16.gguf -p "User: Hello RAM-ALTMAN:" -n 256 --temp 1.0 --top-k 50 --top-p 1.0 --min-p 0 --repeat-penalty 1.0 -c 379 --stop "<|i_have_enough_ram|>" ``` **transformers:** ```python from transformers import AutoModelForCausalLM, AutoTokenizer tok = AutoTokenizer.from_pretrained("brunuu/RAM_Altman_GPT") model = AutoModelForCausalLM.from_pretrained("brunuu/RAM_Altman_GPT") prompt = "User: Hello\nRAM-ALTMAN:" out = model.generate(**tok(prompt, return_tensors="pt"), max_new_tokens=256, do_sample=True, temperature=1.0) print(tok.decode(out[0])) ``` **transformers.js:** ```js import { pipeline } from "@huggingface/transformers"; const generator = await pipeline("text-generation", "brunuu/RAM_Altman_GPT", { dtype: "fp32" }); const out = await generator("User: Hello\nRAM-ALTMAN:", { max_new_tokens: 256, do_sample: true, temperature: 1.0 }); console.log(out[0].generated_text); ``` ## 📊 How well does it work? Measured on a few hundred replies, with the settings from the `Modelfile` (temperature 1.0, top-k 50): | | | |---|---| | Talks about RAM | 210 of 210 replies. A perfect score | | Reply length | 14 words median, 19 on average. Sometimes 240 | | Stops on its own | 209 of 210 | | Says its name when asked | ~70%. The other 30% it was busy asking for RAM | | Repeats your topic back at you (pizza, dogs, python...) | ~35% | | Answers "can I download more RAM?" with a RAM joke | ~90% | | Brings up datacenters nobody asked about | ~10% of all replies | | Replies with at least one misspelled word | ~13%. Mostly `PIZZZA` | | Remembers what it said earlier in the conversation | 8%. Barely better than guessing | ## ⚙️ Specs | | | |---|---| | **Architecture** | GPT-2 (`GPT2LMHeadModel`) | | **Parameters** | 100,000 (tied embeddings) | | **Layers / heads / hidden size** | 3 / 4 / 40 | | **Context** | 379 tokens (whatever is left over after hitting exactly 100k params) | | **Tokenizer** | byte-level BPE, 640 tokens, trained on the dataset | | **Format** | chat: `User: ...\nRAM-ALTMAN: ...<\|i_have_enough_ram\|>` | | **Files** | `model.safetensors`, `RAM_Altman_GPT-f16.gguf` (~245 KB), `onnx/model.onnx` (for transformers.js) | | **Dataset** | not included, `python src/gen_data.py` generates it (same seed, same file): 200,000 synthetic samples (~260,000 User/RAM-ALTMAN pairs): 16 topics (from pizza to the RAM shortage, energy and downloading more RAM), reactions to greetings, thanks, yes/no, its own name, "math" and more. | | **Training** | 30,000 steps. Only RAM-ALTMAN's replies are graded, the first tokens of each reply (the part that depends on you) weigh more | ## ⚠️ Limitations - It only knows RAM. - Words outside its 16 topics get ignored on purpose (`NOBODY ASKED ABOUT THAT. RAM`). Sometimes it picks the wrong topic instead. - Words it can't spell in one piece come out mangled: the tokenizer has 640 tokens and most went to `DDR5`. - It follows a conversation's shape (`why?` -> `BECAUSE ... WAS NOT ENOUGH`) but makes up the numbers. - English only. ## 🧠 Attention span RAM-ALTMAN can read **379 tokens**. Not one more. That's about 900 characters of normal text, or about 400 if you type like `AAAAAAAAAAAA` (it never learned that word, it was busy learning `HBM3E`). Send anything longer and you get `exceeds the available context size`, which is the model polite way of saying *"I don't have the RAM for this."* And no, raising *Context Length* past 379 will not help. There are exactly 379 positions in its brain. Position 380 was sold to buy more RAM. ## 🔁 Reproduce it ```bash python src/gen_data.py # creates ram_altman.txt python src/create_RAM_Altman_GPT.py # trains tokenizer + model (this part requires RAM!) python src/convert_gguf.py # makes the GGUF for LM Studio / Ollama / llama.cpp python src/convert_onnx.py # makes the ONNX for transformers.js ``` ## 📜 License **Public domain (CC0):** weights, tokenizer, GGUF and training scripts are all in this repo, and `src/gen_data.py` rebuilds the exact dataset. No rights reserved just like all the RAM.