Ben Goertzel has been running the annual Artificial General Intelligence (AGI) conference since 2008, and the 15th edition was back in person this year. It was held over four days in August in Seattle, and I was lucky enough to be able to attend.
It was a blast. A particular highlight was the workshop on AGI in fintech, and especially the discussion on the use of Deep Reinforcement Learning in portfolio management. Given the very natural reward signal, what could an AlphaZero/MuZero-type system do there? Perhaps one day the world will be prepared to know.
Ben coined the term Artificial General Intelligence back In the early 2000s to differentiate real AI from what artificial intelligence had become by then, which was essentially the boring, mundane stuff taught in textbooks like Russell’s and Norvig’s Artificial intelligence: A Modern Approach.
Back in the mid-1950s, it seemed reasonable to suppose that a few people working for a couple of months over the summer might crack the hard problem of human-level intelligence. That didn’t prove to be the case, but the dream lived on, and Ben was one of the few voices crying in the wilderness that we should be thinking about and trying to build systems that reproduce the full richness and generality of human intelligence.
AI certainly has been having a moment as of late:
- Deep Learning (think image recognition like Bing Reverse Image Search)
- Deep Reinforcement Learning (AlphaGo beating the best go player in the world, and AlphaFold cracking the protein folding problem)
- Transformer-Based Large Language Models (systems like GPT-3 and BERT that can write articles on pretty much any subject and flexibly carry out novel tasks when prompted)
These could be called AGI-adjacent AI, although researchers like Ben are keen on developing systems that try explicitly to implement models of cognition rather than just building larger and larger neural nets trained by backpropagation on huge datasets.
And AGI is finally having its own moment: John Carmack, the legendary programmer of Wolfenstein 3D, Doom, and Quake, announced in August he had raised $20M for his own AGI startup.
What does Artificial General Intelligence have to do with APIs?
There’s been a lot of buzz in recent weeks around the DALL-E text-to-image generation systems, which build on GPT-3 to generate novel images based on text prompt. An interesting thing here is that the system can create a huge variety of types and styles of pictures as well as being to understand (“understand”) text.
OpenAI explicitly states that DALL-E doesn’t have an API. But what they mean is that it doesn’t have a public API. The DALL-E prompt you type in is an API. It’s just that you can’t use it programmatically on an industrial scale, and you don’t know all the things that the API can do because OpenAI keeps that to itself.
Google Translate has a public API, as does Microsoft Translator. The quality of computer translation has improved exponentially over the past decade––first because of Deep Learning, and then through the magic of Large Language Models. These systems can be considered examples of proto-AGI.
If you are paying to use these systems as part of your business’s production environment, you will be accessing the proto-AGI over an API because that’s the point of an API––to allow web services to be accessed flexibly and data and information to be exchanged at scale.
And you will want the API to work properly, so it should be monitored actively––which is our mandate here at APImetrics.
My big takeaway from AGI-22 is that we are on track to create AGI by the end of this decade. At the very least, we are going to have a plethora of DALL-E-like and GPT-3-like systems that are increasingly capable and increasingly useful for businesses and private individuals.
Most companies and people aren’t going to be building or training their own systems from scratch but are going to be consuming them over APIs. Given how much value AGI is going to be created for the global economy, we can see that APIs are going to be even more important to the Earth’s IT ecosystem and noosphere than already are.
So, strap in. It’s going to be a wild decade.
And this leaves aside using (proto-)AGI to monitor APIs or analyze API data or design, test, and manage APIs. This will be the subject of my next blog post, so stay tuned!