2024 - The Year Of Objective Function Engineering

What are we optimizing for with AI?

Happy Sunday and Happy New Year! Welcome to Investing in AI. I’m Rob May, CEO at Brandguard, and active AI angel investor. (If you want to contribute something to the newsletter, I don’t take donations or subscriptions but I am looking for an introduction to Elad Gil, and I’m sure one or more of you knows him well.) Let me know if you can help me connect.

I’ve been thinking a lot about what’s next in AI. A lot of stuff has already been written on what to expect in 2024 so I’ve been trying to think of what may have been missed. My answer is - objective function engineering.

When we build algorithms, particularly AI algorithms, they need a goal. This goal is usually expressed as an objective function - what is the AI algorithm trying to accomplish and how does it know if it’s doing a good job? So far, these objective functions are pretty straightforward.

I often analyze AI problems by asking “how would humans accomplish this?” and when you think about objective functions - humans have a hierarchy of them. Ultimately our genes want us to reproduce. But to do that, we have to stay alive, which requires an objective of finding food and water. We also have to stay safe and keep from dying, which is another objective. Sometimes to do that we have to play nice with the social structure if we live in one so, we have that as an objective. Those objectives all have to be satisfied first, and only then can we focus on the objective of finding someone to mate with. Our objective function hierarchy is complicated.

AIs on the other hand, have pretty simple objective functions, so far. Over the past few years, the pieces of an AI stack have started to grow in complexity and variation. We moved from one giant model to ensembles of models, in particular “mixture of experts” models. Then we went from simple prompt engineering to more varied and complex prompting. I think in 2024, objective functions are next.

When we think about building AGI, there is still clearly something lacking from even these LLMs like Bard and ChatGPT, compared to humans. What could it be? Well, I think these models are all trained on smaller objectives. Their objective functions are simpler and less complex, and not hierarchical. In 2024, that will change.

To build AGI we have to endow a machine with more complex goals. Machines have to have meta-goals, like “stay alive” and “learn” which can be vague and difficult to define as objective functions. And shorter term goals like maximizing performance on certain tasks that lead to those broader goals. Maybe a short term goal is to find a new data set for training, for example. But I think this kind of new objective function engineering is the path forward to a new conceptual step in building AGI.

So what do I expect in 2024? I expect a lot of work to be done, and a lot of interesting papers and posts to be written, about objective functions. I expect to see more vague and varied and hierarchical objective functions. Maybe someone will create “situational objective function embeddings” where, somehow they create embeddings of possible objective functions that change for an AI based on its situation. I don’t know. I’m speculating. But, I do expect this to be the next area of focus and a big initiative for 2024.

As always, if you have thoughts on this, or have worked on something related, I’d love to hear from you. Thanks for reading.