Artificial Intelligence Overview Quiz

Quiz and Attendance Reminders

• Paper #3 due December 9th
• During class time on December 9th:
  • Come to 304 Barnard Hall to take the final
  • Join Zoom (through Canvas) for a review session and take the final on December 18th (as scheduled)

Artificial Intelligence

• Thinking Humanly: The creation of machines that think like humans, including problem-solving, decision-making, and learning (Haugeland, 1985; Bellman, 1978).

• Thinking Rationally: Studying mental faculties with computational models; focusing on computations for perception, reasoning, and action (Charniak and McDermott, 1985; Winston, 1992).

• Acting Humanly: Creating machines that perform functions requiring intelligence when carried out by humans (Kurzweil, 1990; Rich and Knight, 1991).

• Acting Rationally: Studying intelligent agents' design (Poole et al., 1998); focusing on intelligent behavior in artificial systems (Nilsson, 1998).

Hype Level of Artificial Intelligence

• The graph shows a cyclical pattern of increasing and decreasing hype around Artificial Intelligence over time.

Newell & Simon's Physical Symbol System Hypothesis

• Any intelligent system needs to manipulate symbols.

Quotes on AI

• The 1956 quote describes AI research goals.
• Other quotes mention problems that result from the use of various AI.

Artificial Intelligence and the Hype Cycle

• Initial enthusiasm followed by periods of disappointment and a subsequent resurgence.

ImageNet

• Image recognition dataset influential in the field of computer vision.

• Fei-Fei Li is a key figure in this development.

Neural Networks

• A comeback of using neural networks in AI.

ChatGPT and Language Models

• ChatGPT is a dialogue-based conversational AI.
• It learns, questions and corrects itself.
• A sibling model to InstructGPT.
• Focuses on detailed responses.
• In a conversational format.

False Belief Task (2022)

• Sally hides the ball in a box.
• Children understanding Sally's perspective is tested.

False Belief Task (2024)

• Sally will look for her ball in the basket since it's where she last saw it.
• The task tests the ability to understand that beliefs, intentions etc., can differ from reality.

No-Belief Task (2024)

• Anne puts a ball into a box.
• Sally walks into the room, sees the box. The expectation that the ball is still in the basket.

Centration (2022) and (2024)

• The subject cannot put themselves in another person's point of view

• Cannot consider multiple aspects.

• Centration (2022) demonstrates a child's inability to account for multiple perspectives.

• Centration (2024) shows an inability to think of two different perspectives. It describes a difficulty seeing beyond one aspect.

Model-Based Reasoning (Example Scenario)

• Predicting that Bob will teach the Friday review session.
• Based on past knowledge and assuming consistent patterns.

Wason Selection Task (2022)

• Need to check IDs of persons under 21 to ensure compliance with alcohol laws.

Wason Selection Task (2024)

• Need to check if a person under 21 is drinking alcohol and if someone drinking alcohol, if they are 21 years old or over.

Alief

• Scenario about a cockroach in a glass of water.
• 2022 response is to not drink the water, even if it is safe.
• 2024 response offers a logical view that the water is safe, but acknowledges the psychological aversion that many people feel towards insects.

Alief (Portal Scenario)

• Pretending about secret portals in the woods
• Portal scenario.
• Real life response suggests not walking through the portals.

Majaj et al., 2015 Summary

• Chart showing various neural responses and similarities.
• Relationship between performance and similarity.

AI and Society Summary

• Multiple approaches exist for AI.
• Rule-based systems are often replaced by connectionist systems.
• Cognitive science aids in evaluating and enhancing a human-like approach in AI systems.

Bias in AI

• The concept of "garbage in, garbage out" (GIGO).
• Different examples including chatbots, YouTube recommendations, and prison sentences.

Viscous Feedback Loops and Bias

• Policing areas that have increased criminal activity.
• Identifying the area as needing more policing.

Misalignment

• Superintelligence being programmed to maximize the number of paperclips.
• Creating an outcome that is not aligned with human values regarding survival.

Social Harms of AI

• AI becoming superintelligent to outsmart us.
• AI being misused for bad purposes.
• AI mistakes in critical situations.
• Companies reducing workforce with AI.