And take control of your recommendation feed …

The villain ScreenSlayer takes over people’s mind and agency in Incredibles 2 (2018)

Recommendation engines for news feeds, videos, tweets, movies, music and many others create your worldview from which there is no escape. The manipulation and deception to individual lives and the perversion of norms is no longer in question. And, it happens without the consumer’s knowledge or understanding.

Recommendation engines have all the control

Recommendation engines trap you in a filter bubble because it has all the control. Each consumer is put in a single lane without any control instead of an open road with control. This is unacceptable and will only get worse until control is given to the consumer. An AI must be a tool to serve our individual interests not its.

Thingy pops your filter bubble by giving you control.

Recommendation engines are not dynamic

In conventional AI, raw data is transformed into feature vectors and fed into a machine learning algorithm. Working its magic on the feature vectors, the algorithm outputs a trained model also called a predictive model(1) for unseen data.

Static machine learning

Predictions remain fresh until new data arrives. But, two problems arise: the current model becomes out-of-date and the length of time to re-train it is proportional to the volume of data. The industry solution is to re-train the model as fast as possible.

Modern software systems don’t work like this. For example, with a database (dbms) you can add, update and delete data at any time while people continue to query the database. A dbms is a dynamic system whereas an AI is a regularly re-trained static model.

Dynamic machine learning

Static vs dynamic AI in the enterprise

Organizations run different kinds of systems-of-record (SoR) software such as CRM, ERP, HR and so on which collectively form the company’s information network. These systems (which all have dbms’s underneath) are integrated so that data changes in one will automatically update the others to keep the information flow up to date. The SoR span a company’s operations — local, national and global.

It is straightforward to extract information from a SoR layer to employ in an AI service running independently. When data changes in the SoR layer it notifies the AI service to collect the new data and begin the model re-training process. The two layers are separate and independent though the AI will always be out of sync and must be continually retrained to catch up with changing data.

But, joined-at-the-hip integration between systems-of-record and systems-of intelligence layers need to carry a large flashing “caveat emptor” neon sign as the former is a dynamic system and the latter a static model.

Machine learning “training” is a workaround

Machine learning algorithms were developed decades ago for research use not for today’s globally connected world where real-time, interactivity and dynamic scalability are essential and necessary.

Imagine if databases could not update in real-time or search engines could not show results of current events immediately. We live in a dynamic world and yet ML today is built atop static models which have to be continually re-trained.

Thingy is a dynamic recommendation engine

Thingy is a dynamic recommendation engine which learns and predicts as it goes along. It adapts to a changing world without re-training. Consumers interact with it in different and unpredictable ways but remain in control. Thingy is characterized by:

  1. Automatic generalization: Given one or more items as a query, Thingy learns in realtime to discover other relevant items. There is no machine learning ‘model training’ phase.
  2. Interactive: Engage with Thingy in different and unpredictable ways with various query options. Gives consumers the control to go where they want and when.
  3. Dynamic: Update and add new data at any time.
Thingy’s three parts working together dynamically

Automatic generalization

Thingy learns and predicts as it goes along and adapts to a changing world without re-training.

Automatic generalization

Items (“things”) can be any datatype including images, audio, video, db records, documents, genes and composite data.


Standing stridently in front of a large curved screen, Tom Cruise moves things around, brings things to the foreground, moves them into the background, slides them to the left or right, and shoves things out of the way. In the iconic scene from the movie Minority Report, Cruise is hunting to find relevant information from a pre-cognition system(2). It is an exhilarating example of an interactive AI in action.

Is something similar possible today? Yes.

Interact directly with Thingy to add and remove items from the query; find more of things; bring things to the foreground while pushing others into the background.

Add and remove items from a query

Add and remove items from a query

More like this or these

More like this or these

More of these, less of the others

More of these, less of the others

“I am the query” personalization

A shopper’s photos of their favorite clothes and accessories represent a fingerprint of what they like to wear. When shopping online or at physical stores, their device automatically shows personally relevant things. The shopper can interact directly with Thingy to discover more of what they like or to see less of what they don’t. They can update the fingerprint at any time by adding and removing photos to suit their current lifestyle.

“I am the query” personalization is a powerful way to discover things of personal relevance which also minimizes getting stuck in a filter bubble and preserves privacy. Your fingerprint is your doppelganger — and, you can have more than one — to interact with the digital world securely and safely on your behalf. Thingy safeguards privacy implicitly.


The world turns and changes. As does data. Change needs to be captured to reflect a new reality just like modern software systems do. With conventional AI, feature vectors are a means to producing a trained static model. With Thingy, the feature vectors are central characters on the stage with new ones appearing, current ones remaining unchanged or updated or even terminated.

Adding new data at any time is a litmus test of a dynamic recommendation engine. The steps to demonstrate adding unknown images to Thingy are:

  1. Query with an unknown image. If the image is known to Thingy then its duplicate will show as the first result. In this case, it doesn’t.
  2. Add a copy of the unknown image to the Thingy system.
  3. Query with the unknown image again. If the image is known to Thingy then its duplicate will show as the first result. This time, it does.

The slide deck shows adding a single new image and 3 new images.

Litmus test for a dynamic recommendation engine

Thingy’s purpose is trusted discovery

Discovery without being exploited or degraded.

Thingy is a trusted public discovery space and a platform for developers to build private spaces. Data privacy, security and transparency is enforced by the Thingy technology and policy.


(1) The output of deep learning is a trained model for object classification or recognition. Feature vectors of unknown objects can be extracted from the pre-trained model with the transfer learning process. These dense vectors are employed in machine learning pipelines and in distance metric operations to find similar objects in an n-dimensional geometric space.

(2) The official Minority Report trailer —

If you liked this, click the clapping hands below so other people will hear about this on Medium. Thank-you.

Maker of Thingy, a recommendation engine which learns in realtime, is dynamic and interactive. Founder, architect and engineer at