On the issue of Singularity

While singularity is not something yet to be foreseen in models, its quite an exciting aspect to look into. Let's dive into a basic cat detection model.

In this basic illustration, Once an image is being trained on the model, a few steps are followed,

1. The image is turned into pixels; machines understand numbers
2. The model extracts the features, edges and shapes
3. Makes predictions until it reaches 99.8%.

But then, by the time the model reaches this kind of accuracy, much is happening—something called "Feed Forward Movement" and "Backward Propagation". These two are key when it comes to the model having excellent performance

During this phase, the data is flowing through the network layers, making predictions. However backpropagation enables a model to calculate the error at output, propagate this error backword, and adjust the connection weights until future predictions are better.

The model starts with basic recognition, but as it processes more images, it discovers better ways to detect patterns; these improvements let it learn trends faster. With faster learning, it creates better algorithms. The cycle accelerates as each improvement speeds up the next one. Eventually, the system could improve itself faster than humans can monitor or control.

One would argue that there are techniques that are going to limit this direction, but it comes at an expanse; the more we push for the best performance, we are heading that direction of the runaway, and if we limit, then we can't have what we envision AI to do for us.

Rate limiting works perfectly, but what if AI finds optimisation shortcuts, and for capability boundaries, those are tricky to set as its hard to define precise limits. Where as human oversight, AI improves rapidly faster than our review cycles.

Am not against AI, but sensitise that much as its the best tool we have, we need to have better controls on how far it can go. #AISafety #HealthyAI