The big, flashy advances in AI progress that I hear about seem to be pretty fundamental models like GPT-3 or AlphaZero, systems that are trained by throwing a big neural network at a complex problem. It seems like there are some reasons to expect this approach to work well; most fundamentally 'the bitter lesson' that scaling a good architecture beats clever specialized systems. (e.g. my understanding is that AlphaZero learns faster and performs better than previous systems that were trained on a particular game and/or given game-specific heuristics.) 

But from reading Drexler's work on AI services, it seems natural to me that the first AI systems used for complex tasks will tend to make use of existing complementary tools rather than reinventing the wheel. My sense is that this is true for e.g. current autonomous vehicle prototypes, which have specialized, "dumb" computer vision and control systems and don't make much use of general-purpose reinforcement learning systems. 

So my question is: what are the prospects for AI systems that are trained to make use of existing tools, like a GPT-3 clone that can query a calculator rather than having to figure out addition from scratch? In what cases are they promising, and who's making use of them? Should we expect to see more natural language models like this as the field evolves?

New Answer
Ask Related Question
New Comment

3 Answers

My prediction is that as the infrastructure/computational resource floors for AI systems rise and research + development become more centralized as a result, industry especially will trend toward developing applications powered by commercialized technology rather than creating domain specific AI systems from scratch. If hardware limitations restrict fledgling companies from outcompeting established research labs, it seems reasonable that they'd outsource given the option. This pattern is already emerging in the generative field where start-ups are adopting an API-reliant service. OpenAI has published a blog post showcasing a handful of them:  

People are currently working on doing exactly this. E.g., Adept is training language models to use external software. They’re aiming to build a “natural language interface” to various pieces of software.

See also:

1 comments, sorted by Click to highlight new comments since: Today at 1:18 AM

epistemic status: "the best way to learn is by saying something wrong and being corrected." These statements are all intended as "my best guess" from someone who's not super technical and could easily be wrong about AI progress.