Monetizing Artificial Intelligence

Turn your AI software projects into an ATM machine. Freelance workers enable the automation of their work — ironically.

AGI (artificial general intelligence) is the pursuit of machine cognition, largely still a work in progress. Pat Langley from Arizona State University, has an excellent essay highlighting the differences between most of what’s labeled “AI”, and what he refers to as the ‘Cognitive Systems Paradigm’.

A cognitive system has the machinery to begin working with written language, use heuristics and other approaches to deal with incomplete data, make inferences from structured representations of information, and so on.

Some current AGI frameworks can productively be part of this type of work, although unfortunately many projects are not yet well documented or openly available. AGI entrepreneurs Peter Voss and Monica Anderson have recently met with me to discuss their respective work.

Without an AGI architecture, there is ‘narrow AI’ code that can be coupled with APIs and screen-scraping tools to automate specific tasks. Human workers can oversee and lend a hand where software remains incapable.

While such an ensemble won’t lead to a cognitive system, it is how most current ‘virtual agents’ are built today.

Whether you are working to develop and AGI architecture (an integrated cognitive system) or have pieced-together a virtual agent to perform specific tasks, the place to start earning $$$ is with artificial-artificial intelligence.

Artificial-artificial intelligence exists in abundance today.

Artificial-artificial intelligence

The key to generating steady revenues from early-stage ‘proto-AGI’ systems and automated ‘agents’ is to understand the limitations of Turk — Amazon’s Mechanical Turk.

You see — ‘Turk’ is a marketplace where you can pay for repeatable self-contained tasks to be completed, by people (‘workers’) bidding on them. ‘HIT’ human intelligence tasks as they are called.

Need to tag 1,000 pictures to create training data for an image recognition algorithm? Turk can get that done, it might cost $0.15/image, depending on the bid interest.

a Turk HIT (human intelligence task)

Once a bid is accepted, the tasks are completed and the workers get paid. There’s a rating system to keep everyone honest. You can imagine this as ‘artificial-artificial intelligence’, because human beings are performing tasks that machines (currently) cannot.

There’s a big catch.

Mechanical Turk assumes the tasks don’t require credentials or any significant training. It assumes each task is self-contained. Task instructions need to be intuitive: nobody will bid unless the tasks can be readily understood up-front. You cannot assign Turk workers tasks that span multiple phases with dependencies or tasks for which credentials are needed (eg. logging into an app). You don’t typically correspond with Turk workers during the work unless there’s an issue preventing it from being done. It is ‘bare-metal’.

What’s needed is a higher abstraction level.

This higher abstraction level allows for freelance work (rather than just individual tasks) to be performed.

  • Credentials to existing apps, systems, email, etc. could be provided to the worker: there would be some base level of trust
  • Work could include phased tasks with dependencies: a task could remain partially complete for some undetermined period of time
  • Workers could be trained on the specific work: the work assignment need not be trivial and self-contained
  • Workers could be asked for status updates, or given new details for the process: some malleability is possible within the scope of work

If Turk is a disk drive, what we need is the operating system. A higher abstraction level that systematizes work so that machines (software) can perform parts of it.

The first step toward $ is to find freelance work that is sufficiently narrow to be augmented by software, and be relatively abundant. We don’t want to bid on niche projects, rather ones that are sought-after and repeatable.

Finding freelancer jobs

The top freelance worker marketplace is Upwork (formerly oDesk). Many have used it to hire freelance developers and QA and it’s fairly impressive.

Here’s a list of freelance “virtual assistants” job postings (sorted by the amount of money the ‘client’ has spent to date in the marketplace).

It’s ironic that the term ‘virtual assistant’ is used to describe freelance work — done by humans.

Note: you probably need a [free] login to see this list properly.

>4,100 job postings, with the following breakdown

Filtering for “Entry Level” and at least $100/week yields 1,906 job postings

Fully categorizing nearly 2-thousand postings would take some time, but a few clusters are immediately obvious. There are assistants to teams and ‘executives’ that are likely too broadly defined to be opportunistic. There are marketing jobs that require creative writing skills, again let’s ignore these. There are jobs that require making calls and speaking to people, let’s ignore these as well. And there are jobs that require foreign language skills. We need freelance work that is relatively straight-forward.

A ‘tag’ emerges in these job postings that’s notable: “Email Handling”.

Here are a few examples:

Notice the amount of $ spent by these clients on UpWork to-date. Note also that many jobs specify the expected throughput (eg. ’50 emails per day’).

Email handling is an example of multi-phased work that requires credentials. There is an email account, there are multiple templates, spreadsheets to keep track of things, etc. It’s laborious but highly repetitive and it allows for tasks to be completed by a system with a limited cognitive range of motion.

Cognitive range of motion is the ability for a person or machine to perform cognitive skills.

This is a crucial concept to understand.

We all readily understand physical range of motion. You might run a reasonable pace for a 5k but no amount of training will get you to Olympic times. Your range of motion as a runner has a limit.

Similarly you might be rated 1,350 in chess tournament play but no amount of playing or tutoring will get you to a chess master rating. That would require a combination of cognitive skills including the ability to remember a larger number of positions, the ability to quickly run through more scenarios for different moves, etc.

So if we understand the cognitive range of motion of our AGI or the capabilities of a software agent, we can find freelance work that matches. It’s ok for it to require some help or supervision — training wheels.

The next step is to bid on common freelance work with the right range of motion and hire a freelance [human] worker to do it. You can hire the freelancer from the same marketplace and have the work ‘pass through’ your company. The freelancer must be willing to ‘collaborate’ with your software system and help evolve it.

There’s no need to mark up the price — we’re going to be driving revenue margin by using your system to reduce the work effort.

Artificial-artificial knowledge work

The point of artificial intelligence is to perform knowledge work. To earn revenues, it will be used to reduce the amount of time the human freelance worker spends on the work. As the automation system does this work you will better understand how to improve it, and subsequently have it do a larger percentage of the knowledge work — the effect is circular.

The hired freelance worker will eventually only be needed to oversee, supervise and perform various unforeseeable tasks.

Your revenue equation in this setup is as follows:

(I) Inbound revenue: the $/hour paid to you by the freelance customer

(F) Freelance cost: the $/hour paid by you for the human freelance worker

(A) software Agent work %: the percentage of each hour, on average done by software (this will increase over time)

$ revenue/hour = (I - (F * (1 -A))

example: if your automated system can handle half of the work, on average, and the work pays/costs $15/hour, your revenue per hour will average:

$ revenue/hour = (15- (15 * (1 -.5)) = $7.5/hour

Because the automation can do certain tasks 24x7 the equation may not fully represent the leverage it provides.

For email handling, as one example, your system should to be able to (with some assistance):

  • Fill in templates with structured data (name, email, etc.)
  • Summarize and classify inbound emails for routing and follow-up
  • Update spreadsheet information based on the work
  • Send messages to others on the team when stuck
  • Look up information on web sites, some of them behind a login
  • Follow a predefined workflow with some reasoning for certain steps

Much of this is utilitarian, much of it requires a combination of utilities, some of it requires some basic cognitive skills.

Follow the money

Once this is put in motion you can bid on additional similar freelance work. The more similar it is, the more applicable. You should be able to multi-task the human freelance worker as the percentage of software agent contribution increases.

Besides $ margin, the value of such a system is knowledge work experience. In the case of a cognitive system, it can be trained, learn, adapt, etc. to perform its part of the work with greater ability. Instead of ‘training data’ we need repeated ‘training work’, with human oversight.

It’s possible that over time, certain freelance jobs will be handled entirely by AGIs. The near/mid-term goal is not to achieve self-sufficiency but rather to achieve a high-degree of repeatable success in the tasks.

The paradox of AGI development is in the term ‘General’: a system must be of practical use long before it can be self-sufficient in any general way. AGI developers seeking a ‘singularity’ stage of cognition will likely go broke before getting there. It will take considerable time to achieve far-reaching cognitive skills in software, what’s needed in the meantime is a place to exercise and evolve these skills.

A setup where an AGI can evolve while doing real work and earning money. Crawl before you walk.

Might as well get paid to do it.