← Quizzes Start

eChai Quiz

AI Quiz · eChai AI Social at House of Starts

Twenty tricky ones to open the evening. More than one answer often looks right, that is the point. Show of hands, then we reveal.

Tap → or Space to start · the group answers by picking a side

Question 1 of 20

Which of these AI milestones happened FIRST?

  • A AlphaGo defeats world champion Lee Sedol at Go
  • B AlexNet wins ImageNet and kicks off the deep learning era
  • C The "Attention Is All You Need" transformer paper
  • D ChatGPT reaches 100 million users

AlexNet, 2012. Then AlphaGo in 2016, the transformer paper in 2017, and ChatGPT crossing 100 million in early 2023. The whole modern boom is barely a decade old.

Question 2 of 20

When a large language model answers your question, what is it fundamentally doing?

  • A Looking the answer up in a stored database
  • B Running a live web search behind the scenes
  • C Predicting the most likely next chunk of text, one piece at a time
  • D Following hand-written rules an engineer wrote

It predicts the next token. Some models can also search the web now, which is why that option feels right, but the core engine is next-token prediction, not lookup.

Question 3 of 20

Anthropic, maker of Claude, was founded mainly by people who previously worked at which company?

  • A Google DeepMind
  • B OpenAI
  • C Meta
  • D Cohere

OpenAI. Anthropic was started by former OpenAI leaders. Mistral traces to Meta and DeepMind, Cohere to Google Brain, so all three feel plausible, but the ex-OpenAI answer is Anthropic.

Question 4 of 20

The huge, expensive, one-time computation that builds a model is called ___. Answering your question afterwards is called ___.

  • A Inference, then training
  • B Training, then inference
  • C Fine-tuning, then prompting
  • D Learning, then testing

Training builds it once at enormous cost. Inference is every answer after that. Startups live and die on inference cost, because every reply a user gets costs real compute.

Question 5 of 20

A model described as "open weights" means that...

  • A The training data is fully public
  • B It is free to use through its API
  • C Anyone can download the model itself and run it on their own machines
  • D Its app source code is on GitHub

You can download and run the actual model. That is different from open training data (almost never public) or a free API. Open weights is not the same as fully open source.

Question 6 of 20

The US Copyright Office has taken the position that a work generated purely by AI, with no meaningful human authorship,...

  • A Is owned by the AI company that made the tool
  • B Cannot be registered for copyright at all
  • C Becomes public domain automatically after five years
  • D Is owned by whoever wrote the prompt

No human author, no registration. It is not that the prompt writer or the AI company owns it, and there is no five year clock. Human creative contribution is what unlocks protection.

Question 7 of 20

In 2023 two New York lawyers were sanctioned by a judge. What had they done?

  • A Leaked client data into ChatGPT
  • B Used AI to forge a client signature
  • C Submitted a brief citing court cases that ChatGPT had entirely made up
  • D Filed deepfaked video as evidence

The Mata v. Avianca case. ChatGPT invented citations that looked real, the lawyers filed them without checking, and the court was not amused. The cautionary tale for every profession.

Question 8 of 20

The Air Canada tribunal decision in 2024 established that...

  • A A company is responsible for what its own AI chatbot tells a customer
  • B Chatbots may not be used for customer service
  • C AI-generated advice is never legally binding
  • D Airlines are exempt from chatbot liability

The airline argued its chatbot was a separate entity. The tribunal said no, you own what your bot says. If it promises a customer something, you honour it.

Question 9 of 20

The EU AI Act sorts AI systems primarily by...

  • A The company's annual revenue
  • B The number of users
  • C The level of risk the system poses
  • D The country where it was built

It is risk-based. Some uses are banned, some are high-risk with strict duties, most are low-risk. The obligations scale with the danger, not the size of the company.

Question 10 of 20

Which of these is a real, ongoing lawsuit over AI training data?

  • A The New York Times suing OpenAI and Microsoft
  • B Getty Images suing Stability AI
  • C Authors including comedians suing Meta and OpenAI
  • D All of the above are real

All three are real and live. The core question in each is the same: was it fair use to train on copyrighted work without a licence? Courts around the world are still deciding.

Question 11 of 20

Nvidia's real moat is not just its chips. It is a software platform, launched back in 2006, that most AI code is written for. What is it?

  • A TensorFlow
  • B CUDA
  • C Kubernetes
  • D ONNX

CUDA. Nvidia spent nearly two decades getting the world's AI researchers to build on it. TensorFlow is Google's framework, and the others are unrelated. The lock-in is the software, not only the silicon.

Question 12 of 20

"Fine-tuning" a model means...

  • A Writing sharper prompts to get better answers
  • B Making the model run faster
  • C Training an existing model further on a narrower, specific dataset
  • D Shrinking the model so it fits on a phone

You take a general model and train it more on your own data, say your firm's contracts, so it specialises. Prompting is not training, and shrinking a model is a different technique.

Question 13 of 20

A model's "context window" is...

  • A The box on screen where you type
  • B The most text the model can take into account in one go
  • C The time limit for a reply
  • D The model's permanent long-term memory

It is how much it can hold in mind at once, prompt plus reply. It is not permanent memory: close the chat and, by default, it is gone. That distinction trips up a lot of people.

Question 14 of 20

Which of these has actually caused documented business harm?

  • A Employees pasting confidential data into public AI tools
  • B A chatbot giving a customer wrong information the company was then held to
  • C A deepfaked voice or video used to authorise a fraudulent transfer
  • D All of the above have really happened

All real. Staff have leaked source code into chatbots, a company was bound by its bot's bad advice, and a finance worker was tricked by a deepfake video call into wiring millions. The risk is operational, not hypothetical.

Question 15 of 20

A "system prompt" is...

  • A The user's very first message
  • B An error message from the server
  • C Hidden instructions that set the AI's rules and persona before the user types anything
  • D The computer's operating system

It is the standing brief given to the model behind the scenes: be helpful, refuse this, stay in this role. The user never sees it, but it shapes every answer.

Question 16 of 20

"Multimodal" AI refers to a model that can...

  • A Work across many human languages
  • B Handle several kinds of input such as text, images, and audio together
  • C Run on many different devices
  • D Be trained by several companies at once

Many modalities of input and output, not many languages. A multimodal model can read a photo, listen to a clip, and reply in text. Multilingual is a different property.

Question 17 of 20

"AGI", the term everyone argues about, stands for Artificial General Intelligence. It usually means...

  • A Any AI you can hold a conversation with
  • B AI that matches or beats humans across most cognitive tasks
  • C AI that controls physical robots
  • D AI built and owned by governments

General means broad: not good at one narrow thing, but capable across most of what a human mind can do. There is no agreed test for when we get there, which is half the debate.

Question 18 of 20

For a startup, why does "inference cost" keep founders awake at night?

  • A Every answer a user receives costs real money to compute
  • B Because training is completely free
  • C Because it is a single one-time setup fee
  • D Because it only applies to image generation

Unlike normal software, where serving one more user is nearly free, every AI reply burns compute you pay for. Scale your users and your bill scales with them. It reshapes the whole unit economics.

Question 19 of 20

"Retrieval-Augmented Generation", or RAG, is popular in business AI because it...

  • A Makes the model train faster
  • B Lets the model look things up in your own documents first, then answer from them
  • C Removes the need for any human review
  • D Encrypts the conversation end to end

RAG grounds the answer in your actual files, your policies, your contracts, so it is less likely to invent things. It is the standard way firms point a general model at private knowledge.

Question 20 of 20

When people say a chatbot has "guardrails", they mean...

  • A Physical safety hardware around the servers
  • B Rules and filters that stop it from producing harmful or off-limits output
  • C A backup copy of the model
  • D The legal terms of service

Guardrails are the built-in limits: refuse this, do not advise on that, stay professional. They are not perfect, which is exactly why human judgement still sits on top of every serious use.