#bimostitch. Image stitching just got better now using an advanced neurosymbolic memory network for fast & robust image matching:. iOS: Android:

And there is more I am working on right now.

A lot of folks I tell the story are really surprised by how I pulled such a thing off, building high quality computer vision based applications beating out ones from the likes of Google, Microsoft, Apple etc and gaining enough users to recoup the cost of development.

So yes, I used initiative to use my own research to develop apps, which in return generates revenue to fund further research.

And the availability of Google Play and Apple App Store has changed that. I can sell apps on there and generate revenue in US dollars or Pounds which when converted to Zambian Kwacha explodes. And all that mostly passively while I do other things like research.

So technically I would be working as a telecommunications engineer or electrical engineer in some telecommunications company or electrical company had it not been for my tenacity to chase after my dreams against all odds.

Now I can download steam games and stream videos etc, it's a massive improvement. Also lots of modern services are now available, plus medical services are also good, if you can afford them. Therefore countries like Zambia have really improved in a lot of ways.

Though I don't like the term "third world", better use the term "developing world" or "developing countries". There has been massive improvements in internet compared to when I was starting my research efforts back in 2009. It was so difficult to load a simple page back then.

To be fair I wouldn't be able to fund my robotics and AI research efforts without Google Play and Apple App Store. Because of such I can afford solar systems, computer & robotics hardware that otherwise would have been impossible if I had a normal Zambian job.

Given the current state of machine learning, there is no reason to believe superintelligent machines are near. These are just people daydreaming and trying to make their companies look more advanced than they currently are.

Data contamination is one of the biggest problems in evaluations of machine learning models. Not even Microsoft has the resources to make sure training data isn't contaminated with test cases.

Back in 2012 every AI research paper was like "approaching human level performance on blah blah". Now it's "superintelligence this, superintelligence that". There is always hype around this field ever since big tech got involved in it.

*doesn't.

The above is from Microsoft Research. I thought these guys and girls should know better. Don't run models on exams designed for humans it says nothing about the models capabilities. Link:

A model that performs well on human exams is likely doing so via simple correlations & information leakage or data contamination due to exposure to a lot more problems. It does give any useful information about the models understanding of the problem.

In short data contamination or leakage is very unlikely for humans compared to a model that is exposed to a ton of data. Hence humans when tested on those exams do actually give a good signal for understanding as compared to testing a model on that same exam.

This is actually not a good way to compare model performance:. Humans are not exposed to internet scale data hence the probability that a human will encounter a very similar case, to which he/she has already been exposed to, in an exam is very slim compared to a model.

Link to the above video:

And one of my core areas of interest is to understand something about how biological intelligence works and replicate that into machines. I am not interested in one particular approach but one that has the potential to give me the answers I seek.

That being said, I always fascinated by the brains capabilities even an insect is insanely sophisticated and capable. Our efforts in AI and machine learning are really just a tiny scratch on the surface of this mystery. We have much to learn as a species.

That "causal model of reality" is not rigid as most think, it's still dynamically adapted to given a new problem. That's the difference between the "world model" in AI and cognitive maps in neuroscience. One is adapted to given problem while the other is not.