A Realistic Framing of the Progress in Artificial Intelligence
Let’s face it—we love exciting announcements. Why talk about the small technical improvements of a given artificial intelligence (AI) system when you can prognosticate about the coming advent of artificial general intelligence (AGI)? However, focusing too much on AGI risks missing many incremental improvements in the space along the way. This is very much like how focusing solely on when cars can literally drive themselves risks missing all the incremental assisted driving features being added to cars all the time.
DeepMind at the Forefront…AGAIN
The coverage of AlphaGo, DeepMind’s1 system that was able to best the performance of professional Go player Lee Sedol, was a game changer. Now there is AlphaZero, AlphaFold and more. DeepMind has made incredible progress in showing how AI can be applied to real problems. AlphaFold, for example, predicts how given proteins will fold, and, in accurately knowing the shape of given proteins with accuracy, unlocks enormous potential in how we think about all sorts of medical treatments.
The Covid-19 vaccine using mRNA was based largely on targeting the shape of the specific ‘spike-protein.’ The overall protein-folding problem was something humans were focusing on for more than 50 years.2
However, DeepMind recently presented a new ‘generalist’ AI model called Gato. Think of it this way—AlphaGo specifically focuses on the game of Go, and AlphaFold specifically focuses on protein folding—they are not generalist AI applications. In contrast, Gato can3:
- Play Atari video games
- Caption images
- Stack blocks with a real robot arm
In total, Gato can do 604 tasks. This is very different from the more specialized AI applications that are trained with specific data to optimize one task.
So, AGI Is Now on the Horizon?
To be clear, full AGI is a significant jump over and above anything achieved to date. It’s possible that with an increase in scale, the path used by Gato could lead to something closer to AGI than anything done to date. Similarly, it’s possible that increasing scale alone goes nowhere. AGI may require breakthroughs that are yet not determined.
People love to get hyped on AI and its potential. In recent years, the development of GPT-3 by OpenAI4 was big, as was the image generator DALL-E. These were both huge achievements, but neither has led to technology exhibiting human-level understanding, and it is unknown if the approaches used in either will naturally lead to AGI in the future.