How to prepare for AI as a high school student
I have given a presentation to some some students and teachers of Nanyang Junior College in Singapore, on observations and predictions on AI as well as some career advice. The presentation happened on on 3rd February 2025. This LinkedIn post is here and the feature on their course blog is here.
Objective
- My background as an ML engineer
- General career advice
- What is the current state of AI
- What is probably the next state of AI
- How can you prepare for AI
Background
- Anderson Secondary School (3+2)
- Anderson Junior College (PCME)
- Singapore University of Technology and Design (ESD, DDP)
- Machine Learning Engineer at Quora
What do I actually do at work
I work at Quora. I work on recommendation systems.
When you sign up for Quora, we send you stories in the Quora digest that we hope that you click and read. One of our recommendation systems decide what stories to send you.
Generally
- Plan changes
- Make changes
- Prove that the change is good
- Improve processes to do the above
Just that it is related to machine learning algorithms
Challenges at work
Generally
- Deciding what is worth working on
- Measuring success rigorously
- Investigating what went wrong
- Communicating to an appropriate level of detail
Just that it is related to machine learning algorithms
Example project
Given an answer, write follow-up questions for the authors to write more answers
Breaking down this problem
- Should we even write follow-up questions to the answer
- How to write great (?) questions
- How do we show the questions to the writer
- Making sure that we deliver answers written to the audience it deserves
- How do we measure success
Why is this fun
- The world gets useful knowledge that would likely not have been written
- This is being done at scale
- I get to put food on my table
Advice on starting out
Build a project, deploy it to prod (make it easy for people to try)
- “I have given this advice to many people no one has done it so far” - Denys Linkov
Sample from the list of students who graduated from your course
- NUS, NTU, SMU, SUTD, SIT
- Research them on LinkedIn
- Reverse engineer on what they did to get there
- Reach out to them if appropriate
- The exercise here is to understand what it takes
- (Do you need to be in NUS CS + NOC with internships every summer and semester?)
Reflecting on my journey
What has happened to help
- Finding doing LeetCode (and solving problems) intrinsically fun
- Researching and reaching out for opportunities (e.g. internships)
- Pure luck (Quora Programming Contest)
What could have helped more
- Getting to know someone who has been there
- Better taste on what projects are worth doing
- Better execution and deliver outcomes for projects
- Peers that motivate each other
The state of AI
Currently
- AI can already help with tasks that are easy to describe
- AI is already better than an average professional with tasks that are easy to grade
Soon
- AI can do all tasks that are easy to describe, as good as a competent professional
- AI is better than the best professional with tasks that are easy to grade
AI can already help with tasks easy to describe
Examples of tasks easy to describe (low-context / no-context tasks)
- Doing A-level math calculations
- Writing an A-level General Paper essay
- Competitive programming
Examples of tasks difficult to describe (high-context tasks)
- Implementing a button on the Quora website
- Planning how to improve Quora recommendation systems
Implications
- Companies have already experimented with offshoring work easy to describe
- OpenAI Deep Research
- Soon we will use AI to do all the work that is easy to explain
- If it takes more effort to get others to do I will still rather do it myself
AI is already great at tasks easy to grade
Examples of tasks that are easy to grade (easily verifiable)
- Doing A-level math calculations
- Competitive programming
- Implementing a button on the Quora website
Examples of tasks that are not easy to grade (not easily verifiable)
- Writing an A-level General Paper essay
- Planning how to improve Quora recommendation systems
Implications
- We have been grading people with examinations and interviews
- OpenAI o1
- Soon we will use AI to do all the work that is easy to grade
- There is still plenty of human work to do to grade the ungradable
Plotting this out
| Easy to grade | Difficult to grade | |
|---|---|---|
| Easy to describe |
A-level math calculations Competitive programming |
Writing an A-level General Paper essay |
| Difficult to describe | Implementing a button on the Quora website | Improve Quora recommendation systems |
The current state of AI
| Easy to grade | Difficult to grade | |
|---|---|---|
| Easy to describe |
Better than the average professional, soon to be solved at a superhuman level | AI can provide an acceptable response, but you can do it much better. You need to think of how to grade AI responses |
| Difficult to describe |
You rather do this yourself, or break it down into tasks that are easy to describe | You need to make it easy to describe and verify for AI to be useful |
What I think will take some time
These should take more than two years
- Autonomous self-improvement (solving tasks easy to describe, better than the best human)
- Managing and using long-term memory well (solving tasks difficult to describe)
These should take more than four years
- Robotics (experiments in the real world is slow and expensive)
- Mind-reading (probably AI will need to get better at emotion recognition first)
Ultimately, if there is another human that can do my work, AI can do my work. There is no law in physics or theoretical computer science against this.
What you can do to prepare
Understand yourself, understand other humans, understand AI
Employ yourself, employ other humans, employ AI
Improve yourself, improve other humans, improve AI
Questions to ask yourself
- How do you respond?
- How do you reason?
- How do you evaluate what is good?
- How do you learn?
- What do you value?
- How do you influence?