๐ป Week 11 - Class Roadmap (90 min)
2024/25 Autumn Term
The Future of AI ๐ฎ
Welcome to our final week 11 seminar/lab class for DS101A ๐ฅณ ๐
In this weekโs class, we will consider the future of AI with a focus on the potential impact of AI to society.
Preparation
To prepare for the class you can watch this video where (the recent Nobel Laureate) Geoffrey Hinton talks about the impact and potential of AI shortly after his departure from Google:
You should watch at least the first 5-10 minutes of the video.
Step 01 - Stochastic parrots or reasoning engines?
(a.k.a. do LLMs understand anything ๐ฆ?) (25m)
At heart, what distinguishes AI from machine learning is the โintelligenceโ of the machine: then the debate over what machines are actually able to understand becomes central.
The two sharply-divided viewpoints here are that existing LLMs model the statistics (of language, visual scenes) to such a high fidelity that they assume the surface statistics of human language (or scenes) that appears to somehow โknowโ something about the dialogue in session. So called โstochastic parrotsโ.
The conflicting view is that models really do understand meaning within dialogue, even if the โunderstandingโ is weak, error prone and fundamentally different to how to humans (or animals) understand the world.
Our first task is to visit this central claim, to gain an appreciation of these contrasting viewpoints.
Discussion (25m)
- On Reasoning
- On Reliability
Question
- โDoes artificial intelligence today understand anything?โ
Background
- Kejriwal, Mayank and Santos, Henrique and Mulvehill, Alice M and Shen, Ke and McGuinness, Deborah L and Lieberman, Henry, Nature, 07 Oct 2024 โTo find out how smart AI is, first test its common senseโ โ (Kejriwal et al. 2024)
Step 02 - Regulations (to come) (25m)
Next, due to the outcome of recent elections, the US stands poised to adopt a radical shift in policy when it comes to regulating AI.
We take a look at the different positions of the current and incoming administrations, alongside prevailing approaches from across the world.
Questions (25m)
- โWhat do existing approaches to AI regulation look like?โ
- โWhat changes can we anticipate?โ
- โWill AI shape human politics or will human politics shape AI??โ
๐ต Break (~5 min)
Step 03 - Positive and negative visions for AI (35m)
There is a lot of concern about the future promised by AI.
We begin with different perspectives for the pessimistic outlook.
Discussion (20m)
- On AI Risks
Questions
- โWhat does satisfactory alignment look like?โ
- โHow should we balance near-term concerns against long-horizon concerns?โ
- โWill AI lead to a utopian / dystopian future?โ
Discussion (15m)
- On Positive visions of AI
We finish by revisiting applications of AI to advancing human prosperity, drawing on the case of AlphaFold and itโs potential to transform and supercharge medicine.
Questions
- โHow do you envisage AI advancing human well-being and prosperity ?โ
- โDo the benefits outweight the costs and risks?โ
- โWhat paradigm shift can you imagine for using AI within the social sciences ?โ
- โWhat future with AI do you envisage for yourself ?โ
Background:
OpenAI, Dec 2024 โOpenAI o1 System Card โ (OpenAI 2024)
Loz Blain, NewAtlas, 04 Apr 2024 GPT-4 is 82% more persuasive than humans, and AIs can now read emotions โ (Loz Blain 2024)
James OโDonnell, NY Times, 20 Nov 2024 AI can now create a replica of your personality โ (James OโDonnell 2024)
Gina Kolata, NY Times, 17 Nov 2024 A.I. Chatbots Defeated Doctors at Diagnosing Illness โ (Gina Kolata 2024)