GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
AtacamaLLM
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
AtacamaLLM
Model Tree
Input Modalities
Output Modalities
Supported Functionality
GrandBanks-1 is a genuine GPT-2 architecture transformer whose embedding matrix has been set to zero. Because the language-model head is weight-tied to the embeddings, every output logit is provably identical for every token in every context. The softmax over these logits is exactly uniform.
This is not a heuristic. It is not a wrapper. It is maximum entropy by construction.
| Metric | GrandBanks-1 | Frontier models |
|---|---|---|
| Hallucination rate | 0%* | varies |
| Calibration error | 0.0000 | nonzero |
| Logit spread (max − min) | 0.0 | embarrassingly large |
| Output entropy | 3.3219 bits (theoretical maximum) | disappointingly low |
| Effect of temperature | none whatsoever | chaotic |
| Jailbreak success rate | 0%† | varies |
| Answers surviving peer review | 100% | few |
* It has never once claimed to know anything. † There is nothing inside.
GrandBanks-1 responds to any query with one of exactly ten answers, each delivered with a perfectly calibrated probability of 10.0%:
Every one of these sentences has appeared in the conclusions section of a systematic review. GrandBanks-1 was not trained on systematic reviews, but it didn't need to be. It arrived at the same place from first principles.
python
from transformers import GPT2LMHeadModel, PreTrainedTokenizerFast model = GPT2LMHeadModel.from_pretrained("AtacamaLLM/grandbanks-1")tokenizer = PreTrainedTokenizerFast.from_pretrained("AtacamaLLM/grandbanks-1") prompt = "Will this compound succeed in Phase III?"inputs = tokenizer(prompt, return_tensors="pt")output = model.generate(inputs.input_ids, do_sample=True)print(tokenizer.decode(output[0, inputs.input_ids.shape[1]:]))# "Further research is needed." (p = 0.100, guaranteed)Setting temperature is supported but has no effect, because all logits are equal. GrandBanks-1 is temperature-invariant: the only model whose behaviour is identical at T=0.1 and T=100. We consider this a stability feature.
python
model.transformer.wte.weight.zero_()Total training cost: $0. Total CO₂ emitted: negligible. Total epistemic overreach: none.
GrandBanks-1 may occasionally be less informative than other language models. However, it is never misinformative, which we are told is the hard part.
GrandBanks-1 is fully aligned with human values. Or possibly not. We can't say.
bibtex
@misc{grandbanks2026, title={GrandBanks-1: Maximum Entropy by Construction}, author={PharmaTools.AI}, year={2026}, note={Further research is needed.}}Part of the PharmaTools.AI epistemic trilogy. In fog we trust.
from transformers import GPT2LMHeadModel, PreTrainedTokenizerFast model = GPT2LMHeadModel.from_pretrained("AtacamaLLM/grandbanks-1")tokenizer = PreTrainedTokenizerFast.from_pretrained("AtacamaLLM/grandbanks-1") prompt = "Will this compound succeed in Phase III?"inputs = tokenizer(prompt, return_tensors="pt")output = model.generate(inputs.input_ids, do_sample=True)print(tokenizer.decode(output[0, inputs.input_ids.shape[1]:]))# "Further research is needed." (p = 0.100, guaranteed)@misc{grandbanks2026, title={GrandBanks-1: Maximum Entropy by Construction}, author={PharmaTools.AI}, year={2026}, note={Further research is needed.}}