THERE are two grand ambitions now for computer science: truly intelligent machines and useful quantum computers. Recent developments suggest not only that these goals should be achievable, but that they could be closer than we think.
Take the quest to develop artificial general intelligence (AGI) – AIs that go well beyond being good at one specific task, but can instead do anything a human can. Some people still think this is impossible. And yet analysis of AIs designed to master human language has prompted other experts to suggest that AGI might only be a matter of scaling up current technology. Build gigantic AIs and true, human-level intelligence will come, they say.
Advertisement
This “scaling hypothesis” has come to the fore largely thanks to GPT-3, an AI released by San Francisco-based OpenAI last year that generates remarkably fluent streams of human language on command. GPT-3 is just a scaled-up version of GPT-2, a similar predecessor. This new neural network boasts an order of magnitude more parameters, equivalent to the number of synapses linking neurons in real brains, than its forerunner.
Researchers who evaluate such language AIs have been surprised by just how much more advanced GPT-3 is than GPT-2. It can do things it wasn’t trained to do, for example, and there are hints that it might be capable of human-like reasoning.
“Truly intelligent machines and useful quantum computers might be closer than we think”
Time will tell if the scaling hypothesis is right. In the meantime, it will be interesting to see if the AI players with the deepest pockets, such as DeepMind, follow OpenAI’s focus on scaling.
However, when it comes to genuinely useful quantum computers, there is no doubt that scaling is key – we are going to need machines with thousands of qubits, the quantum version of a classical bit. This is why the news that researchers have demonstrated a viable way to make sure those qubits don’t constantly fall prey to errors is a big deal. We might finally have a way to scale up the number of operational qubits to what we need.
There are still no guarantees. Even so, it seems that computer science is striding into the 2020s in rude health.
