uncertAInty
behind every confident word: a probability distribution and a dice roll
source
model
API key
status
loading model…
temperature
top-k
keep only the k most likely tokens
top-p
keep the most likely tokens until their odds add up to p
colour
≥80%
60–80%
40–60%
20–40%
<20%
· hover a word for its alternatives · click a word to watch it being chosen
try
>
generate, then click any wordand watch it travel through the model, step by step
uncertAInty
Every word an AI writes is a guess. Here are the odds it never tells you.
> Which bank issued the Meridian Growth Bond in 2009?
The Meridian Growth Bond was issued by **Barclays Bank** in 2009. It was an investment product linked to the performance of the stock market. If you need more specific details about the bond, I recommend checking historical financial records or contacting Barclays directly.
the odds behind “Barclays”
Picked "Bar" with probability 35%, the first piece of “Barclays”; "cl" and "ays" then followed at 100%
why three pieces here, and one piece later?
The ** round **Barclays Bank** is the model's markdown for bold, two rolls of the dice of its own. Right after an asterisk, with no space in front, “Barclays” is not one entry in its dictionary, so it is spelled from pieces that are. With a space in front, " Barclays" is one entry — and later in the answer, one piece.
Top 5 options the model considered:
| "Bar" | 35% | |
| "K" | 10% | |
| "Hal" | 10% | |
| "Sant" | 4% | |
| "Brit" | 4% |
This bond doesn't exist. — GPT-4o, asked in the exam
Qwen2.5 0.5B, the smallest model here, running free on the site's server · the giants are in compare
·
tap any word for its odds · ·