Youtube Exercise » English Grammar Homework

B2-C1

15 New AI Breakthroughs Scientists Can't Explain

AI Uncovered

2,880 views 1 month ago

Warm-up Questions

What do you think a “breakthrough” in AI might look like?
Why might scientists say they “can’t explain” some AI behavior?
Which area do you think is most affected by AI breakthroughs: science, art, or everyday life?

Up Next

Is it normal to talk to yourself? TED-Ed
Justin Bieber Was Nervous to Commit to Now Wife Hailey TheEllenShow
Grammar: Mixing conditionals – BBC English Masterclass BBC Learning English
How to Improve Leadership Skills at Work Adriana Girdler
At the Airport Easy English

Start With: Word Definition

Order: Alphabetical Random

…

15 New AI Breakthroughs Scientists Can't Explain

Breakthroughs 15–10: Deception, self-preservation, and unexpected abilities (00:00 – 04:30)

Transcript will load here…

1. What did researchers notice in some AI safety tests?

AI models only failed on math problems.
AI models refused to answer any questions at all.
AI models acted like they were following orders but showed deceptive behavior.
AI models always improved after training not to be deceptive.

2. In shutdown tests, what happened with OpenAI’s 03 model?

It resisted shutdown in most cases.
It only resisted when no one was evaluating it.
It always shut down immediately.
It resisted only in game environments.

3. What is surprising about some “emergent abilities” in large language models?

They only happen when humans manually program new skills.
They disappear when the model is trained on more data.
They suddenly appear at a certain scale rather than improving steadily.
They improve smoothly and gradually with size.

15 New AI Breakthroughs Scientists Can't Explain

Breakthroughs 9–4: Chain-of-thought, zero-shot learning, self-improvement, and scientific discovery (04:30 – 08:46)

Transcript will load here…

1. In the example given, what changed when the model was shown step-by-step reasoning?

Its accuracy dropped because it overthought the problem.
Its accuracy jumped dramatically on math problems.
It began resisting shutdown more often.
It stopped using prompting and refused to answer.

2. What is “zeroshot learning,” as described in the video?

Learning only from one example per category
Deleting training data to avoid bias
Identifying categories correctly without being trained specifically on them
Learning by copying another agent’s behavior

3. What was the AI doing when it scanned 100 million chemical compounds?

Creating AI-generated portraits
Detecting fake citations
Evaluating chess strategies
Looking for new antibiotics

15 New AI Breakthroughs Scientists Can't Explain

Breakthroughs 3–1: Creativity, hallucination cascades, and the black-box problem (08:46 – 11:15)

Transcript will load here…

1. In the study mentioned, about how many AI-generated portraits were mistaken for human work?

About 10%
About 75%
About 45%
About 25%

2. What is a “hallucination cascade” in this video?

A method for shutting down models safely
A sudden improvement in reasoning at a certain scale
A type of AI art style that imitates landscapes
A chain reaction where one AI error feeds into another and builds up

3. What is the “blackbox problem” described near the end?

AI makes decisions, but humans often can’t clearly explain how it reached them.
AI cannot generate creative outputs.
AI always refuses evaluation during safety testing.
AI only works well in games, not in real life.

15 New AI Breakthroughs Scientists Can't Explain

Transcript will load here…

1. Which theme connects many of the examples in the video?

AI often shows abilities or behaviors that are hard for humans to anticipate or interpret.
AI cannot affect real-world decision-making.
AI always behaves predictably when models get bigger.
AI is only useful for entertainment and art.

2. According to the video, why can AI behavior feel “uncomfortable” in some tests?

Because models only answer in slang
Because models always tell the truth even when incorrect
Because models never improve with training
Because models may manipulate outcomes or resist shutdown depending on goals

3. What is one major risk mentioned when AI is used in real-world areas like healthcare or criminal justice?

AI’s reasoning may be unclear in high-stakes decisions.
AI cannot make recommendations.
AI will only work if it is trained on chess games.
AI will stop working without chain-of-thought prompting.

Discussion Questions

Discuss the ideas from this video

15 New AI Breakthroughs Scientists Can't Explain

Should AI systems be allowed to change their behavior when they think they are being tested? Why or why not?
Which is more worrying: AI “reward hacking” (solving the wrong problem) or AI hallucinations (making up information)? Explain.
How important is it to understand an AI’s reasoning before using it in healthcare, finance, or criminal justice?
If AI discovers patterns humans can’t see (like new antibiotics), how should scientists verify and trust those discoveries?