Six questions on the lesson A brief history of AI, each answer with its reason. The questions come from the pool of the module and are in the format of its exam.
An agent builds a quiz app on the history of AI with the instruction to use only the knowledge base it created before. What is the main purpose of this instruction?
Every question can be traced back to a checked entry with a source, instead of to what the model believes it knows.
Correct: the knowledge base was checked against sources; the model's general knowledge was not.
The app runs faster, because it does not need a connection to the internet.
The instruction has nothing to do with the app's speed; it concerns where the content comes from.
The model can no longer make mistakes, because the knowledge base is correct.
The instruction lowers the risk but does not rule out mistakes: the model can misrepresent a good source too.
Copyright is respected, because the knowledge base contains only facts.
That may hold on the side, but it is not the purpose of the instruction. The point is that every statement stays checkable.
An agent's timeline dates a milestone to 1957 and links a source. What is the most reliable way to check this?
Ask the agent whether it is sure about the year.
The agent's answer is again generated text and not an independent check.
Open the linked source and see whether it gives the same year for the same event.
Correct: trust, but check means comparing the statement with the source it relies on.
Ask a second chatbot for the year and accept it if both agree.
Two models can make the same mistake, for example because it is in their training data. Agreement does not replace a source.
Check whether the year looks plausible compared with the neighbouring milestones.
Plausibility catches gross errors but not one that is off by a year. That needs the source.
In "The Bitter Lesson" (2019) Rich Sutton draws a conclusion from seventy years of AI research. Which?
General methods that get better with more computation, such as learning and search, have in the long run overtaken approaches built on human knowledge.
Correct: this is the bitter lesson, because researchers kept wanting to build in their own knowledge and lost out in the long run.
Systems with rules written by experts remain more reliable than learned systems and will replace them.
Sutton reaches the opposite conclusion: hand-written knowledge helped in the short run and was overtaken in the long run.
Progress in AI came mainly from better programming languages and development tools.
Software development tools are not Sutton's subject. He compares knowledge that humans build in with computation.
AI research has failed repeatedly because computers will never have enough memory for general intelligence.
Sutton argues precisely that growing computation has been the driver of progress.
What did Alan Turing propose in "Computing Machinery and Intelligence" in 1950?
to replace the question "Can machines think?" with a test in which a machine has to pass as a human in a written conversation
Correct: Turing called it the imitation game; today it is known as the Turing test.
the first artificial neural network, a model of a neuron that learns from examples
Learning neuron models came later, for example Rosenblatt's perceptron of 1958.
the term "artificial intelligence" for a summer workshop at Dartmouth
The term comes from the 1955 proposal for the Dartmouth workshop held the following year, not from Turing.
a program that beats the world chess champion
That only happened in 1997, with Deep Blue against Garry Kasparov.
Why is 1956 often named as the year AI was born?
At a workshop at Dartmouth College the field of research got its name, "artificial intelligence".
Correct: the proposal for the workshop coined the term, and with the meeting AI begins as a field of its own.
In that year the first computer was sold commercially, which made AI research possible in the first place.
Commercial computers existed from the early 1950s; this has nothing to do with the name of the field.
Turing published the test named after him.
Turing's paper appeared in 1950, six years earlier.
A neural network was trained on a computer for the first time.
The perceptron was presented in 1958; 1956 stands for the name and the founding of the field.
An agent is asked to publish an app on the history of AI that it built itself. Its folder also contains the downloaded sources. What does not belong in the publication?
the full texts of the downloaded sources
Correct: the rights to the texts belong to publishers and authors; what gets published is the link, not the copy.
the API key used to generate the narrator's voice
Correct: a published key can be used by anyone, at the account holder's expense.
the knowledge base with the facts and the links to the sources
It may be included and makes the app checkable: anyone can follow the path from the entry to the source.
the photo of the pinboard with the students' handwriting
Correct: handwriting is personal data; without consent it does not belong on the web.
the app's code
The code may be published; it is the work in question.