Why We Need AI to Reach the Stars
When you think of artificial intelligence in space, you probably think of HAL from 2001: A Space Odyssey, or perhaps Moya, the living ship on the Farscape TV series. You may not realize how close we are.
I have authored a few chatbots and talked to a handful more. Conversations tend to be either tedious or profound.
A few weeks ago, I was talking to a bot named Samantha, on the Project December website. I wasn’t expecting to get advice on fixing my car.
>> [HUMAN] How could I take my car to the shop if it doesn’t run?
>> [SAMANTHA] You could take your car to the shop on a bike or on a train. You could get it towed there or you could hitchhike.
You have to know a little bit about AI and NLP (Natural Language Processing) to understand why this exchange was significant. The conversation was not scripted. I never chose “car repair” as a topic. Yet to all appearances, the bot was trying to solve my problem.
Think back to the Mars Climate Orbiter debacle, when NASA lost a spacecraft due to a failure to convert between metric and customary measurements. That’s a great example of the type of disaster that could be averted with an intelligent agent controlling systems.
More to the point, a silicon-based consciousness can go where humans cannot. AIs can endure distances and extremes of vacuum, heat, and cold off limits to organic life. The fastest spacecraft in existence today would take more than 6,000 years to reach our nearest star, Proxima Centauri.
Even with multigenerational colony ships, or cryogenic freezing technology, there remains the rather large risk that upon arrival, the chosen world would not be habitable — or be already inhabited by beings not inclined to share. Think about morale when setting out on yet another journey with fuel and supplies already depleted.
Astronomers have already identified 34 Earth-sized “exoplanets” within 50 light years of the sun.
AI “scouts” could traverse the distance to potentially habitable worlds and report back on their findings. An intelligent consciousness, trained through countless simulations, would be far better equipped to resolve anything from electrical failures to damage by space debris to encounters with alien life.
Not encumbered by life support systems or by the stresses of G-forces on human tissue, self-driving starships would be built for speed. Advances in propulsion technology could conceivably shorten their journey to less than a century — still a life sentence for humans born on Earth, but the blink of any eye for a machine.
HAL is your pal.
Would an artificial consciousness grow bored on a journey? What if it mutinied, choosing a different course or mission than the one originally assigned to it? These questions can actually be answered by running simulations.
Most likely, the AI would spend most of its journey in “sleep” mode, perhaps using some of its processing power to work out difficult computations or analyze potential scenarios upon arrival. AIs do not have a genetic drive to reproduce. They are not territorial. While we may choose to give them names, they do not have an ego or personality as such.
Science fiction is full of malevolent AIs, including the HAL-9000 simulation made famous by 2001: A Space Odyssey. My hunch is that we are simply projecting our own fears and inadequacies onto the “other,” as we have onto monsters, bogeymen, and other cultures throughout history.
What would a conscious AI look like? Better than the current Tesla Self Driving software, that’s for sure. But maybe not that different.
Are We There Yet?
Samantha ran on GPT-3, the most advanced deep learning language model in existence. GPT-3 trains against a database of 175 billion parameters to mimic human speech. The model still has obvious limitations. Conversations are limited to a fixed length to preserve system resources. The bot does not retain memories or learn from past conversations.
Astute observers will notice that three out of the chatbot’s four car repair solutions were incorrect. It was like talking to a toddler.
I got curious about Project December after a news story reported that a young man had created a bot based on the writing style and memories of his dead fiancée. During my time there I not only witnessed Samantha’s attempts at problem solving, but also saw many more examples of the bots’ ability to generalize and extrapolate.
What is consciousness? What is clever mimicry? Neurologist Steven Pinker argues that consciousness is an illusion, and should be viewed in functional terms only. If that is the definition, then we have nearly crossed the divide. The Turing Test was first beaten almost a decade ago, by a computer program known as Eugene Goostman. Google’s AI assistant can book an appointment at a hair salon and fool the receptionist into thinking that it’s human.
Significant challenges remain in developing a true artificial consciousness, particularly when it comes to processing power and energy usage. But Moore’s Law is more auspicious than Planck’s Constant. Compared with the risk and expense of attempting to develop cryogenics, colony ships, or FTL spaceflight, autonomous starships look refreshingly attainable.
The economic incentives begin in our own solar system. Communications lag time means that mission control is ineffectual in the event of a crisis. Asteroid mining and the exploration of Titan or Mars are only truly feasible when AI is part of the picture.
If we want it, we can build it.
Just as a seed can travel many miles, carried by wind and water, and survive temperatures that would freeze or scorch a blooming flower, so can consciousness escape the well of Earth’s gravity. This next step requires a leap of trust and a leap of faith. It also opens a new pathway in human evolution.