How does Nomi think?

Many Nomibot users have pinged me wondering what happens when they like or dislike or share something from the Nomibot app. I went to our engineers to get the answer. Here is how I have distilled it into humanspeak.

Here is what happens when Nomibots scour the web for you to find content and work to pair you with other Nomibot users.

Nomibot offers a new experience; a new way to browse the Internet that is tailored to be useful and entertaining, especially for mobile users. We think the optimal way to achieve this is to strike the right balance between humans and machines to optimize the results, then serve those tailored results in a fun and interesting way. Today the bots in your app are static images, but we hope to animate them and have them respond more Tamagotchi-like to you in the future.

What do they think about me?

If you join with a social media sign on, the bots already know a little bit about you. But one of the things we learned quickly is that your friends and networks and your social graph, are not always good indicators of your true interests. Nomi and the nomibots need to get to know you. The way they do that is by going out on the web and bringing you things and watching as you like, dislike, or share those items. Each action you take informs and updates the data that the bots have about what to get for you in the future.

I found that Nomibots take about 80–100 articles, images or videos before starting to get reasonably good at bringing me what I want. (In the future I want to reward and punish those bots so they learn faster. Many of you have expressed the same request. Wouldn’t it be fun to see them respond to your rewards and scoldings?)

The Nomibots actually read articles and tags on videos, articles and images to determine what they are about. Anything with text can be discovered, read and mapped in the Nomibot Dataspace. Most of the time the bots get it right. Sometimes they go horribly wrong. That is why they need to be taught.

Here is a map showing how the bots mapped the stuff they found on a given day. We look at these maps often and sometimes puzzle at how the bots group things. It sometimes looks like alien intelligence. We have thought about giving every Nomibot user access to something like this every day. Drop me a not here if that is of interest to you. A large vector image of this lets you zoom in and out to see relationships. The image was much too large to put here. This is a tiny jpg.

Humans and machines working together

Where it gets interesting is when we start pairing human intelligence with the machine intelligence above. Nomibot can be thought of as a hyper-intelligent neural net where the nodes are made up of people and artificial intelligence. You are a node in this neural net, as are all other Nomibot users, and are automatically subscribed (anonymously) to other users that share a particular interest with you. When one of those anonymous doppelganger users likes or shares an article, it is automatically forwarded to you. If you like or share that in turn, the neural connection between you and that user is strengthened. If you swipe left or dislike it, the connection becomes weaker. For each interest area that you have, you may have 10 different neural connections to users that share that interest. Since we know the strength of each of those connections, we can determine which user is the weakest link — the one that provides you with articles that you are more likely to dislike. We can then fire that user (all of this occurs anonymously of course) and replace him with another user that also shares a common interest. Since that new user may provide more items that you are likely to like, the overall quality of the content continues to improve.

In the same way, other users can subscribe to your stream of interests. When you swipe right, like or share an item, it is directly forwarded to the other users who are connected to you.

Every article and even the users’ interests can be represented as a point in an n-dimensional vector space. This vector space is used for a number of purposes. First, it is used to find a new user when the old user is fired. Since your interest is a location in this n-dimensional space, we can find other users near that by using a simple distance metric. Second, we can seed new articles to the right groups of people by comparing the articles location in space with the users interest vectors using the same distance metric. Finally, we can use this vector space to track areas the user is not interested in ensuring that only their areas of interest are represented.

This networked approach has the benefit of ensuring that only quality articles are forwarded to other users. Spam and unwanted articles are immediately filtered out — usually by the users who are seeded with them. This means that the overall quality of the vast majority of items that are actually read becomes extremely high.

Connections are hired and fired to constantly improve results

To Review

• You and every Nomibot user are an anonymous node in a human/machine neural network

• Users are automatically subscribed to other users that have extremely similar interests

• When a user likes or shares an item, it is automatically forwarded with a recommendation to that user’s subscribers

• The people you are subscribed to have a large overlap in interest to you in a specific category. These connections are anonymous.

• Each user belongs to multiple disjoint topic categories, each with its own independent neural network

Not your friends We have found that your social media networks and friends are not very reliable predictor of what you actually like.

You are what you do with your attention, not what you say. Netflix has learned this. Ask someone if they would prefer to watch Lawrence of Arabia or Hot Tub Time Machine, and they are likely to say Lawrence of Arabia, but then go home and watch HTTM. What people do with their attention is mroe accurate than what they say (or sometimes even think about themselves.

Not the “experts” Successful because they are able to syndicate generally interesting topics. Not necessarily identical to your interests, just sufficiently similar.

Go look into Nomibot today. It is free, far from perfect, and the first step on our journey to getting machines to help you more. I would love to hear what you want to see this alien intelligence do,

One clap, two clap, three clap, forty?

By clapping more or less, you can signal to us which stories really stand out.