The promise of general machine intelligence lies in the answer to the following question - what would happen if you could shrink the time taken to automate anything down from the days it takes a human to program a machine, to the seconds it takes a machine to program itself? This sort of change in time-to-automate has happened just twice in history - with the steam engine and the industrial revolution post the 1800s, and centuries before that with the domestication of animals and the agricultural revolution. It's an interesting exercise to have a look at the exponential growth and prosperity such shifts brought to human civilisation.
Economist Robin Hanson approximates that the world economy doubled every 224,000 years as a hunter-gatherer society, every 909 years as farming society, and every 6.3 years as current industrial society. Such shifts have ushered a new & exciting age of prosperity & discovery each time by leading to step change in life expectancy, freedom, famine, war, energy usage & poverty. The prospect of growing with the current 21st century mixture of farming + industrial growth modes pales in comparison to what would happen if the world were to experience another exponential shift on the lines of the Agricultural or Industrial revolution.
As philosopher Nick Bostrom puts it in his excellent book on the topic, Superintelligence :
"If another such a transition to a different growth mode were to occur, and it were of the similar magnitude to the previous two, it would result in a new growth regime in which the world economy would double in size every two weeks."
Such a growth rate rightly seems ridiculous today, but it would be equally preposterous to suggest to an observer living in the agricultural era that the world economy would double several times within a lifespan. Yet that is the extraordinary condition we now take to be ordinary.
Machines today can compute at 2Ghz (seven orders of magnitude faster than the current biological limit of 200Hz), communicate & collaborate instantly, extend indefinitely without limits, store memories without forgetting, and replicate without the compulsory nine month incubation. It is hard to imagine a scenario where the world grows every few weeks that does not involve the creation of artificial minds that are faster and more efficient than the current biological kind.
Since machines don't directly compete with us for scarce resources (like food, water, sleep, land, money & social validation), artificial minds could enable this while coexisting with the biological ones. That said, it might be naive to ignore the many alarmist depictions of such technology. It's a safe bet to build this in a way that doesn't concentrate it in the hands of just a few people. Non-profits are a bad way to do this. We need to figure out a way to distribute access to it in the same way Henry Ford democratised the power of gasoline engine by putting it in a car that anyone could afford.
Current Approaches Don't Cut It
Current approaches don’t seem to have a way to get there yet. A good way to make sense of this is via another handy automobile analogy - the Waymo vs Tesla approach to making self-driving cars. The Waymo approach is to spend billions of dollars and many years making a prototype car in secret that can drive itself perfectly only on a few streets in Phoenix. The Tesla approach is to make a car that is useful from day one with partial self-driving, put it in the hands of users, and use the miles of sensor data collected from millions of cars to improve Tesla Autopilot for all users over time.
There seem to be a lot of Waymos in the global race to build general machine intelligence which regularly churn out top-class research, but we think only a Tesla-like approach will work sustainably long-term, and it doesn't exist yet. This approach should yank this quaint intellectual academic idea out of the sanitised AI labs and put it in the hands of the user, while framing it in terms of a business problem - so making the machine better at doing more kinds of things over time becomes the only way to serve its customers better.
The 1969 Moon landing and subsequent decline in space exploration shows that it is possible to build hard things without it being a business problem, but impossible to sustain them over the long-term. SpaceX shows that if you make cheaper access to space a business problem, then you can invent reusable rockets that could one day get you to Mars.
After a year and a half of tinkering on this problem, we think we have a notion of what a Tesla approach to general machine intelligence would look like. Here's our three step plan :
01. Create a powerful robotic assistant.
Build a virtual assistant that a human can rewire in a few minutes or hours to do any task - from quickly navigating websites and scheduling video calls to automating data entry and controlling robotic arms.
02. Distribute it to everyone on the planet.
Bring the power of this virtual assistant to every individual and business on this planet, to help them quickly build automation workflows to automate their tasks and improve their output by an order of magnitude.
03. Make it more general-purpose over time.
Till the point it can recursively self-improve, use the different types of automation workflows created in the process to manually train the virtual assistant to rewire itself in seconds to do more and more kinds of complex tasks for its users.
Introducing Maya v0.0.5
Maya is a virtual assistant that is the first realization of our three-step plan.
We invented a way to use transformers (the same technology behind GPT-3) to generate workflows to automate browsers, desktops, chatbots & IoT devices using just natural language input in English. Here are a few cherry-picked examples of how it works on a model that is 1/500th the size of GPT-3. It figures out which components in its library to wire together to complete what you tell it to. The results will get better with more data and compute.
To get more flows to train our transformer model, we built a general-purpose virtual assistant that anyone with a desktop can use and extend - like Cortana or Siri, but more powerful. Maya can be accessed using a system-wide search bar and can be extended by building Maya skill flows using our drag and drop editor, or by getting them from the Maya Skill Store.
We use a forked, customised version of the powerful open-source tool Node-RED as Maya's core runtime across desktop, cloud and IoT devices. This allows anyone to build new Maya skill flows in a few minutes for completing complex tasks, and paves the way for Maya to automate the same kinds of diverse tasks in seconds by generating it's own skills in the future.
The key insight here is that you can optimise modern machine learning to do one thing well, so it can play superhuman Starcraft or Chess, but for generality, we need to go further – optimise generality itself.
Executing on our three-step plan could result in a powerful virtual assistant that would exhibit increasingly more generality over time by enabling the following virtuous cycle :
More types of tasks automated using the assistant -> more types of workflows for the assistant to learn from and generalise with -> more kinds of people find the assistant useful -> more kinds of automations they use the assistant for - and so on until Maya reaches a point of recursive self-improvement.
In the beginning, Maya will let businesses and individuals make custom automations to speed up repetitive back-office tasks like data entry, scheduling calls, navigating websites or extracting information from PDFs. As it learns to do more and more kinds of tasks, it could move into adjacent roles - an online receptionist, omni-channel customer support assistant, or a product unit tester. Soon, it could be trained to do more cognitive tasks like listening to audio, looking at video, and operating a robotic arm to manipulate objects. Down the line, abilities for it to generate natural language conversations could be introduced - so it could be an always-on virtual employee/assistant you could talk with to get things done on Zoom on Slack. It's possible to manually build Maya skills that do all these today, but what is truly exciting is to think of what would happen if Maya could reprogram itself by generating these skill flows on its own.
There's still a long way to go, but we think this is a plausible approach to get there - manually improve a virtual assistant to do more kinds of tasks for its customers until it can improve itself. If this sounds exciting, drop us a mail at humans[at]mayahq[dot]com.
Sign up here today to get an early invite to use Maya. Limited spots available.