@willfrancis24: Which AI model are you talking to? Asking this simple question will reveal which large language model you are in a conversation with. When using apps like Microsoft Copilot, Slack, etc., you can check that you're not being shortchanged with cheaper, less powerful models by simply asking the AI to name a number between 1 and 100. You can then compare the number you get back with the table in my video. If you want to go deeper, you can look up the research. It's called 'One Token Is Enough: Fingerprinting and Verifying Large Language Models from Single-Token Output Distributions'. #ai #aitools #claude #chatgpt #llm #artificialintelligence #Tech #technology #LearnOnTikTok

Key (🔛🔝) :

So that's why every ai story has 47 in it. They use claude to generate it

tomplee✔️ :

Wow, it's literally my research 🥰. Two journals rejected it so far as not much interesting. It's preprint, looking where to publish it

rylan.44 :

You can tell ChatGPT “Use a python script to generate a random number between 1-100” to get a truly random number

🎀 :

DeepSeek 😭

Andrew LeFors :

https://arxiv.org/abs/2607.10252

denis :

I asked Claude and it said 47 lol

Thomas :

Asked GPT and it said 47

Elvis Jagger :

I trained an agent on Unity (Unity ML Agents) and instead of being random, it performed the same exact move every single run.

Kleaps❌🌊 :

42 ist truly the answer for everything

dripman :

this is essentially a basic breakdown of the softmax output layer with a monte carlo method of evaluation, llms at each next token it will predict, the softmax layer will produce a weighted sample of top k tokens to choose from and a random token will be chosen, the monte carlo method paired with a predictable sample of tokens to expect (like numbers 1-100) essentially bruteforceably maps the weights of some of the tokens in that sample pool. This is a good trick when you dont have access to the underlying weights of the model but because you dont truly know what weights are being used or have control over the top k, hidden system prompt (and in some cases random seed) and the hidden routing system that could route you to a dumber model, you arent really forming any empirical signal, especially if the provider wants to avoid self identity, they will certainly create guardrails to muddy such plans. That being said, I find this to have very interesting applications in detecting open source models being deployed by providers since this can open up how differently the models get sanitized when not locally hosted.

hazydreamerstories :

just check agent’s settings

😐 :

kem :

honestly i just let walter rewrite my discussion replies now

John Doppler :

Beautiful! I hate that the models won't identify themselves on demand.