New Impressive ChatGPT Model: o1

Direct AI Updates for School Leaders: 9/16/2024

Author

William Grube
September 19, 2024

Hello all -

Last Thursday, OpenAI announced ChatGPT o1, a new model that improves ChatGPT again.

This is immensely important.

I will include only what you need to know in this newsletter, leaving out all irrelevant details. (If you want a techy explanation, watch this.)

What I will cover:

How much better did ChatGPT get?

ChatGPT's weakest subject has always been math.

o1 was created to get better at things that need to be thought out more: mathematics, logic, and reasoning.

That it did.

o1 places among the top 500 students in the US in a qualifier for the USA Math Olympiad.

Additionally, it exceeds human PhD-level accuracy on a benchmark of physics, biology, and chemistry problems (GPQA).

This graph shows the o1 model's improvement over the GPT-4o model on popular exams:

Exam

GPT 4o (Released May 13th, 2024)

o1 (Released September 12th, 2024)

AP Physics 2

63rd percentile

81st percentile

LSAT

69.5th percentile

95.6th percentile

AP Calculus

71.3rd percentile

83.3rd percentile

AP Chemistry

76th percentile

89th percentile

You can see the full research report here.

Is the model free or paid?

All paid users of ChatGPT have access to a preview of the o1 model. If you pay $20 a month for ChatGPT, you can use it right now.

If you have a paid account and want to use the new model, go to ChatGPT and click on “ChatGPT 4o” in the top left corner. Then, select “o1-preview.”

OpenAI says they will give free users a version of o1. No date is given on when this will happen.

Students can only access the new model if they pay for ChatGPT.

How can school leaders leverage the new model?

OpenAI still recommends using GPT-4o for most things. GPT-4o is the model you automatically use when you log onto ChatGPT on both the free and paid accounts.

Here is why:

o1 pauses to “think” before answering.

This makes it slow. It is common for it to respond in 30 seconds or more.

Additionally, it only got better at math, reasoning, and logic—that is its sole purpose. In some cases, outside of tasks that require more thought, GPT-4o outperforms o1.

Given this, we should use it only in specific cases.

For example, use it in tasks that require reasoning and logic, like creating schedules. That is where o1 shines.

I am going to use the following prompt in GPT-4o and then o1 to show you the difference:

“Act as a school principal. Create a teacher evaluation schedule for 50 teachers, evenly distributed over an 18-week semester, with no more than 5 evaluations per week. There are 5 departments: Math (10 teachers), Science (8), English (10), Social Studies (7), and Electives (15). Prioritize 15 new teachers in the first 10 weeks. Ensure no back-to-back evaluations for the same department within a 2-week period. I can only visit the departments at these times: Math (3rd period), Science (4th period), English (1st period), Social Studies (2nd period), and Electives (5th period). Present the schedule in a table format, balancing the evaluations each week while adhering to these constraints. Give the teachers fake names.”

GPT-4o Output: Click here to see the output. It created a schedule for 90 teachers instead of 50, did not prioritize new teachers, and repeated teachers. In short, it gave me nothing useful.

o1 Output: Click here to see the output. It's nearly flawless. Very impressive.

What Now: Guiding Your Staff to Embrace AI Safely.

This is only the start.

I was born in 2001. In that year, the groundbreaking tech was the first Apple iPod.

The groundbreaking tech in 2024: AI that can pass our lawyers' exams and is better than the average expert human at answering PhD-level science questions.

Where will we be in 10 years when our 7th graders graduate college?

Where will we be in another 23 years? In 23 years, 7th graders will be 35 years old. They will be in the middle of their careers.

We must prepare our students for a world where these tools exist.

Students who use AI for problem-solving will be able to solve problems that we cannot solve right now. Only if…
1) They master a topic to which they can apply AI.
2) We nurture their critical thinking and problem-solving skills.
3) We foster their ability to have unique, independent thoughts.

The issue is that when students use AI to complete their homework, they do not think before putting pen to paper. Because of this…
1) They are not getting a mastery of the topic.
2) They are not developing critical thinking or problem-solving skills.
3) They are not thinking for themselves.

As a result, the very skills they need to use AI effectively may be eroding the most right now.

Here is what we can do:

Step 1: Shift educators from fearing AI to understanding its value. Show them that all people, even those who are not “tech-savvy,” can use AI as a tool.

Step 2: Teach them how to create AI-resistant assignments. Show them that we can terminate the “arms race” that is happening with education and anti-cheating technology and make cheating more impossible by changing how the course is given. Through this, students learn how to use technology productively helping their college and career readiness. Hint: These AI-resistant assignments can be made with AI.

Step 3: Teach them how to use AI tools in the classroom responsibly. Focusing on responsible use.

Step 4: Ensure they know how to use AI to save time and enhance students' educational experiences. Doing both of these simultaneously has never been possible in the past. It now is with AI.

These four steps outline a great frame of mind to bring to your staff regarding AI.

Here is a resource to help your staff responsibly use AI in the classroom - Using AI in the Classroom: Teacher’s Checklist for Success.

Those four steps are the ones we use in our AI Training. See what teachers say about them here.

If you have any questions or comments at any time, feel free to reach out by responding to this newsletter.

Have a great end of the week and weekend.

Kind regards,
William Grube
gruvyeducation.com