shuri

I teach applied machine learning at UC Berkeley. The post argues that fixing the homework feedback loop is one of the bigger near-term opportunities, and that AI can finally deliver one-on-one tutoring, applying personalized mastery learning for every student.

There's obvious interest and momentum in this area. With Y-Combinator's request for startups, Andrew Ng (whom I always highly recommend) investing in this direction, Khan Academy making moves and more.

On the research side, two highlights:

1. Bloom's 2 sigma (1984): a tutored student outscored 98% of a conventional classroom, but tutoring does not scale. Later work has trimmed the 2 sigma number, though the direction holds. 2. A 2025 Harvard randomized controlled trial (RCT) (n=194) found students with a well-designed AI tutor had more than double the learning gains, in less time, than an active-learning class.

I don't think teachers are going away. The AI iterates and works with the students, while the human sets the direction, monitors and course corrects.

We still need people who know what's right, know what's wrong, and can imagine what's possible.

Here are the 2-3 key insights I would highlight from the summary:

1. OpenAI is evolving to be more of a product company than just an AI research lab, with iterative improvements to their API and tools to make them more usable and accessible. This includes pricing changes and ease-of-use focused updates like removing the ChatGPT model picker.

2. There is a tension between modular plugin models for integrating tools with ChatGPT versus pursuing further integration directly into a "Universal Interface". Consumers tend to prefer integrated solutions for convenience, so OpenAI may continue down that path.

3. To fully realize frictionless AI, new consumer hardware interfaces may emerge as alternatives to being "just an app" on existing devices. OpenAI or others could pursue this route despite the challenges of competing with Apple and Google.

A good excuse to reach out to a (remote or quarantined) friend or to family. Still very very much work in progress. Happy holidays :)!

Uri

How it works:

  1. Go to https://room.ai, enter nickname, 
     permissions for the camera if asked or needed.
  2. Share the room link with your opponent
     (and additional people you want to hang out 
      with).
  4. Whoever wants to be white makes the first
     move. Then, the same goes for black.

  0. No AI... yet.
  1. Disconnects do happen and reconnecting reliably still needs work.
  2. Video connections sometimes don't initialize properly. Refresh helps and hit that "turn camera on" button if you see it.
  3. May work on mobile but will likely work better on desktop : firefox, chrome, safari and opera I believe should work. Haven't tested extensively yet.
  4. Layout is funky.
  5. You tell me :)...?

https://github.com/instabotai/instabotai/blob/beab9871bf5633...

- Learning c/c++ which helps connect the high level concepts to how computers work. Other languages like Rust may give similar results but may distract from the main lesson.

- Some functional programming. Programming in a functional way takes time to wrap your mind around. The concepts are more important than the language.

- Some language like Scala which packs many modern ideas.

To empower:

- Python is just fantastic. It helps focus on coding and can take you anywhere from "scripting" to web apps, machine learning, distributed systems, embedded systems...

- Javascript (or Typescript) because it's just everywhere and building a nice frontend is irrationally satisfying / impressive.

For the "résumé":

- Java can be a nice to have but if you know c++, scala or even python, I don't think most would care.

Basically, I'm thinking a giant anonymized excel sheet. I can generate my anonymous id and verify that I appear correctly. That's one part.

The other parts are making sure only real people voted (harder) and that the counting was done correctly (easier).