We asked 100 people what AI does
The SUCHAR:AI [We asked 100 people] game: five puns about AI, four answers in each. Behind every pun there is a real AI term: tokens, hallucinations, overfitting, data lakes
How to play
We asked 100 people. Can you guess what they said?
people asked in every pun
answers on the board, 85 people in total
misses end the question in Expert mode
people said something else, so their answers are not on the board
Play it yourself
The game opens in a window above the page. All you need is a browser: no sign-up, no installation, on a computer or a phone
Which mode would you like to play
- Warm-up punsterbelow 100 · 300
- Seasoned punster100-199 · 300-354
- Pun master200-299 · 355-389
- Voice of the 100300+ · 390+
What is behind the puns
Behind every pun there is a real AI term
What does AI do underground?
Finding patterns and relationships in large data sets, e.g. which cases in an office take longest and why
Before answering, the model searches the organisation’s documents for passages relevant to the question. The answer is based on them, not on the model’s memory
Term in the AI Glossary →The same cards that mine crypto also train and run AI models. That is why graphics cards are so expensive during the AI boom
Periods when interest in AI and research funding collapsed, e.g. in the 1970s and late 1980s. Today AI is at a peak rather than in a hole
What does AI do in the kitchen?
Models that create new content: text, images, code, or a dinner recipe from whatever is left in the fridge
A model does not read words but tokens, pieces of words. The number of tokens decides the cost of a request and how much text the model takes in at once
Term in the AI Glossary →Every model answer is a computation on graphics cards that draw power and need cooling. AI data centres are real ovens
The model confidently states something that does not exist: an ingredient, a quote, a legal provision or a source. That is why AI answers need checking
Term in the AI Glossary →What does AI do in the field?
The examples a model learns from. As you sow, so you reap: errors and gaps in the data come back in the answers
A decision tree settles a case with a series of yes or no questions. A random forest is many such trees voting together and making fewer mistakes
Compute and data at a provider, available over the internet. For sensitive data it matters whose cloud it is and where it stands
Models that detect objects in photos and video, e.g. weeds and pests in drone images
What does AI suffer from?
The most common “illness” of language models: a confident answer that is not true. Answering from documents and citing sources helps
Term in the AI Glossary →The model memorised its training data instead of learning from it. It shines on familiar examples and fails on new ones
The viruses of the AI world include prompt injection, malicious instructions hidden in text, and training data poisoning
Term in the AI Glossary →How much text a model sees at once. Whatever falls outside the window the model “forgets”, which is why long chats and large documents are split into parts
What does AI bring on a callout?
A barrier between people and models: it masks sensitive data and blocks attacks before a request reaches the model. That is what PROXY:AI does
Adding compute when the number of requests grows and scaling down when traffic drops
An alarm that goes off when a model starts performing worse, someone tries to abuse it or costs climb
Term in the AI Glossary →The most popular programming language in AI and data analysis. Fortunately, it does not bite
What does AI serve most often?
The stage where a trained model answers questions. It is what “serves the table” in every conversation with AI
Term in the AI Glossary →An answer with a link to the document it comes from, so it is easy to check instead of taking it on trust
Term in the AI Glossary →Cloud models are billed for input and output tokens. Long questions and long answers make a longer bill
Term in the AI Glossary →Small files that websites store in your browser. Websites serve them, not models, but everyone has seen a cookie banner
Where does AI drop anchor?
The address and port where a model accepts requests from applications. Every AI integration docks here
Term in the AI Glossary →A repository of raw data in many formats, used by analytics and models
Data analysed as it arrives, e.g. meter readings or system events
A cloud where data and infrastructure are subject to the law and control of the country or the European Union
Term in the AI Glossary →Which subject does AI fail?
A model sees tokens, not single letters, so it can miscount letters in a word, e.g. the “r”s in “strawberry”
Term in the AI Glossary →A model only knows what was in its data up to a certain date. It learns about newer events from documents or search
A model can produce a footnote to an article or ruling that does not exist. Answers based on documents with a cited source reduce the problem
Term in the AI Glossary →An attempt to talk a model into breaking its rules, e.g. through role play or a “no limits mode”
What does AI do with an application?
Shortening a long text to the key information. One of the most common uses of AI in offices and companies
Removing or replacing personal data before the text reaches a model. The ID number disappears, the meaning stays
Term in the AI Glossary →AI prepares, a person decides and approves. For high-risk systems this is required by Art. 14 of the AI Act
Term in the AI Glossary →The system sends a case to the right model, tool or person. Unlike room 102, it does so straight away
How does AI train?
The texts and examples a model learns from. Their quality decides the quality of the model
Numbers the model tunes during training. They hold what it has learned. Large models have billions of them
One full pass through the training data. Training takes many epochs, but too many lead to overfitting
Processors that run thousands of operations in parallel. The basic hardware for training and running models
How the game reads your answers
Form does not matter. Meaning does
words and phrases in the dictionary, in Polish and English
of over 1,100 player answers recognised in tests
answers sent outside the browser
Meaning-based search works at work too: TWIN:DESK finds the document an employee asks for in their own words
allclouds.pl technologiesAt a training session, webinar and booth
5 puns in 2 minutes. A warm-up that explains what tokens are
Would you like to show the game at your event or training session?
Contact formQuestions about the SUCHAR:AI game
How long does the game take and what does it run on?
5 puns, about 2 minutes in Junior mode and 3 minutes in Expert mode. The game runs in a browser on a computer, tablet or phone, in Polish and English. No sign-up or installation needed
What is the difference between Junior and Expert?
In Junior mode all 4 answers are shown and you pick 3; the top score is 400 pts. In Expert mode the answers are hidden and you type them in your own words, with 3 misses per question; the top score is 425 pts
How does the game know my answer is right if I typed it differently?
The game has a dictionary of over 2,300 words and phrases: word stems with their forms, synonyms, Polish equivalents and common typos. Excluding words make sure a random word does not score
Does the game send my answers anywhere?
No. Answers are judged by the dictionary in your browser and the game sends nothing. On that device it only remembers the language and game mode you chose
Can we use the game at a training session or an event?
Yes. Junior mode works best as a warm-up, a webinar works with the pun on screen and answers in the chat, and a booth with both modes. Contact us and we will help prepare the game for your event