I spent the last four months in a pre-idea startup accelerator working mostly on validating startup ideas. I thought it might be interesting to write up the ideas I worked on, especially because I’d like to see these ideas implemented. While there are many learnings I could write about, in this post I’ll be focusing solely on the ideas.
Replacing video ads on mobile games with data annotation
Most free mobile games make money by forcing their users to watch video ads, which are typically advertisements for other mobile games. This generates revenue for game studios as advertisers typically pay them a small, fixed fee every time someone clicks on an advert. After an hour of gameplay, these ads will start recycling and you’ll see a demo of the same game over and over again. At this point it’ll be clear that you don’t want to try this other game, so the game developer isn’t benefitting from displaying it anymore. Being forced to watch the same ad several times makes these games frustrating to play. Additionally, it drives users away from the game, as if they click on the ad of the advertized game they may never return. So I wondered if there could be a better way to monetize mobile games.
Now a small detour. Historically computers have been bad at understanding things, such as human voices, or just identifying objects in a photo. With recent advances in machine learning, this has been changing. Computers today can identify objects in images with high accuracy, and transcribe human speech. Teaching computers to understand images and voices requires a lot of training data: raw images or audio paired with annotations that explain the meaning of the data.
A typical data annotation task: a photograph of an intersection. The labels here are the three bounding boxes.1
With an explosion in the use of machine learning, companies need increasingly larger datasets to train their models. To cater to these needs, a whole range of businesses sprung up to provide training data. These businesses receive raw (unlabelled) data and provide annotations, which are done manually by contractors. A typical annotation task is drawing a rectangular box around certain objects of interest, like traffic lights in a photograph of an intersection. One of the industry leaders in data annotation charges (approximately) $0.08 per image with an additional $0.08 per bounding box.
Back to ad interruptions in mobile games: Why could you not replace the video ads with simple data annotation tasks like bounding boxes, image classification, etc? In much less than 30 seconds, one could do several of these little exercises, and it would be a more engaging user experience. Kids could even mistake it for a mini-game (think “Circle all the giraffes in the following images”).
How about the economics? Well, the cost of video ad displays ranges widely depending on several factors, such as the operating system, the duration of the ad, and the user’s location. On average (across all geographies) the cost per view on Androids is $0.001. In Mexico, the cost of an ad display is just $0.00025. So in principle, a video game developer could make more than 100x by making the user draw a single bounding box instead of displaying a video ad.
Doing data annotation on mobile comes with a few additional benefits:
- Scale: During its first five years Candy Crush, a popular mobile game, has been downloaded 2.73 billion times. Reportedly 9.2 million people are playing the game for 3 hours daily. This scale would enable tons of data annotation tasks to be processed.
- Turnaround: Say you need to label 100,000 images very quickly. Flushing these images to just a couple million players worldwide would allow you to get the job done in a couple hours.
- Versatility: Mobile devices come with speakers, microphones, and touchscreens, meaning that annotations of a whole range of media types would be possible, eg. images, audio, and video.
I decided to move on from this idea for the following reasons:
- Expertise: Some of the most sought after datasets today require some special expertise. For example, for medical applications annotators need to be researchers or medics. Your average mobile game player is unlikely to label brain scans or mammography images accurately.
- Gore: For content moderation purposes the data that needs to be labeled often involves nudity, blood, or other kinds of content that’s inappropriate for underage users.
There is one startup that is attacking the problem of mobile game interruptions from another, arguably a much better, angle. Audiomob is enabling game developers to monetize audio instead of video, without halting the user’s gameplay.
I am under the impression that what a technology recruitment agency does is made up of the following steps:
- Check given job spec for keywords
- Search for LinkedIn profiles that contain these keywords
- Spam the owner of the LinkedIn profiles with generic messages
Now, this is something that could be very easily automated. What’s more, once a candidate is in the hiring pipeline, a lot of other manual work can be automated, such as scheduling meetings, emailing candidates, notifying of results, and collecting feedback. So I think it’s only a matter of time before businesses will start using robot recruiters.
An even more interesting spin is that the seemingly random search of a recruiter can be made more informed by using simple heuristics to score candidates. Such scoring could be based on:
- Level of education
- Length of relevant professional experience
- The quality of their code: average line length; the average length of their subroutines; how descriptive their variable names are; if their GitHub repositories can be compiled successfully, etc
- The statistical likelihood of career change: LinkedIn provides data on the average turnover rate at companies. For example, on average, Microsoft employees stay for 4 years. 2 Now if according to her LinkedIn profile a certain developer just started working at Microsoft, then she’s a lot less likely to be looking for a career change, than her coworker who just had his 4th work anniversary.
- Depth of knowledge as evidenced by contributions to Stack Overflow, or any similar knowledge-sharing platform.
Most of this information is publicly accessible, and a web crawler should be able to index job candidates, and link all their professional profiles (LinkedIn, Github, Stack Exchange, etc).
The above factors could be weighted according to the needs of an organization, and an overall score could be computed to prioritize certain candidates over others.
International remote employment as a service
Suppose you run a California-based startup and you’re looking for your next employee with a very specific profile. The ideal candidate applies on your job portal, but the only trouble is that they are based in Russia, say. You can’t get them to relocate, either because they’d prefer not to move, or because of the complicated bureaucracy. What do you do?
You talk to Remote Employment Company3 (REC). REC has offices in all major economies, and so you enter a contract with its US-subsidiary. They’ll take the job specs, the details of your candidate, and they’ll draft a fully-compliant employment contract for your candidate in Russia. This way your candidate will be employed by REC Russia, while she will be in effect doing work for you. REC will take care of all the payments, benefits, taxation and compliance, allowing both you and your remote employee to focus on what you do best.
Remote work is exploding, and all the CEOs I talked to said that they’d rather avoid employing remote workers as contractors, as they’d want to create a level playing field and offer the same benefits to all workers. There are already some established players in this space, called Global Employment Organizations (GEOs).4 There is also a great number of promising new startups 5, and some more established startups in adjacent fields.6
Carbon neutrality as service
Microsoft, Shopify, and EasyJet are just some of the many corporate giants that have recently pledged to go carbon neutral. The Oxford Dictionary picked “climate emergency” as the word of the year 2019.7 For the first time since 2001, the number of domestic air passengers decreased8 in Sweden last year. A government body identified the climate debate as one of the possible causes. The Swedes even have a word for “flight shame”. As awareness of climate change is exploding, consumers are becoming increasingly anxious about their carbon footprint. This is impacting e-commerce.
Brand value: businesses already seek to align with many causes
As the climate crisis will be driving people’s behavior, consumer businesses will be increasingly seeking to align themselves with sustainability. Online retail is especially at risk here, as there is a common misconception that ordering goods online has a larger carbon footprint than in-store shopping. In reality, e-commerce is greener.9 The Head of Brand at a UK-based subscription business told me that she had been receiving many emails from customers who are afraid of ordering their product because they are worried about the carbon footprint.
While there is a growing consensus that a carbon tax would be an effective way of forcing the transition to decarbonization 11, businesses can already choose from a range of solutions. There is a whole range of non-profits that offer certified carbon capture. However, it is still very difficult to calculate the carbon footprint of individual products, let alone the entire footprint of businesses. And while enterprises like Microsoft can afford to figure out their decarbonization in-house, many small businesses don’t have the resources to do so. That is why I think that soon there will be a go-to B2B decarbonization provider, that will assess, offset, and verify the carbon emissions of businesses. People will start seeing labels declaring products and businesses ‘carbon neutral’, much the same way the Rainforest Alliance certifies sustainable forestry.
As a first step, one could look at e-commerce businesses and follow their sales numbers. A software-as-a-service checkout integration would be my first bet. This could calculate an approximate carbon footprint for different product categories, and could either increase the price tag of the product in question or alternatively offer the consumer to offset their shopping cart voluntarily. This way the provider of this integration would also gain real-time data on the sales volume of the business in question, and so this would allow the plugin provider to compute an end-to-end carbon footprint of the entire business.
Thanks to Emmanuel Abiola, Clare Lyle, Joao Beraldo and Susannah Evans for reading through a draft of this post.
This name is entirely fictitious, any resemblance to actual entities is coincidental. ↩