Michael Chang

Sora

I work on pretraining for Sora. Sign up here!

Chat-Tree

A tree-structured chat. Play with it here!

Project Astra

I led the long memory, low latency, interruption, and event-detection capabilities of the real-time multimodal assistant that was demonstrated at Google IO 2024.

See the official video, where you can find demonstrations of proactive response to visual stimuli (0:10-0:18) and long memory (1:20-1:34).

I co-watched the Google I/O keynote livestream with Astra, which was then posted immediately after the main keynote and appended to the developer keynote.

I also similarly co-watched OpenAI's GPT4o announcement a day earlier, and a clip of that can be found here.

Gemini

I worked on evaluating Gemini's hours-long video understanding capabilities, where I was co-responsible for the results shown in Fig. 9, Fig. 15, Fig. 28, and Table 47 of the Gemini 1.5 Pro paper.

The artificial programmer is the drosophila of AGI

I define AGI as the capacity to create new knowledge. The process of knowledge creation is the iterative cycle of conjecture and criticism.

Similarly, the process of programming, and debugging, is also an iterative cycle of conjecture and criticism, where the knowledge is encoded as the program. We conjecture programs for achieving a certain functionality and we criticize those programs by running them through tests. If the code fails, we conjecture reasons for why the code does do not what we want, and we criticize those reasons by comparing what we expect the values of the variables to be with what the values actually turn out to be.

The processes of programming and of knowledge creation are isomorphic.

Therefore, I believe that the artificial programmer is the drosophila of AGI: studying the process of writing and debugging code can teach us the fundamental principles for building self-improving knowledge-creating AGI systems.

Code is also cheap, verifiable, and valuable, so I believe that the shortest path to AGI is through the goal of building artificial programmers.

The "No-Push" Principle

A large fraction of our relationship with information-processing machines has been to spoon-feed machines information that they already have access to.

Examples include doing taxes, copying data from one spreadsheet to another, and filling out forms.

Such tasks reduce humans to mindless communication channels. They stifle our creative potential.

I therefore propose the No-Push Principle: Humans should not have to push information to machines that is already there. The machines should pull that information on its own.

A good example of software that has evolved in accordance to the No-Push Principle is navigation software. A few decades ago, if we wanted to go somewhere we needed to manually type in our current location to get a route to our destination. Now our phones automatically pull our location information without us having to specify it. This is good. Trends like this should continue.

Principles like the No-Push Principle give us a way to bifurcate the world and identify problems to solve. Next time when you interact with computers, ask yourself, "Am I unnecessarily pushing information to the machine that it should already know about?" If yes, then that's a problem to solve.