Hello!
New writing: Gradient Dissent.
The “neural network” technology that powers current artificial intelligence is extremely expensive, poorly understood, inherently deceptive, and unfixably unreliable.
In short: it is bad.
Somehow, no one ever says this!
The case against neural networks, based just on uncontroversial engineering criteria, is overwhelming. Further, they are widely used without adequate safeguards in situations where they cause increasing harms. As Better without AI explained, this may get much worse.
In the software industry, it is common for technically inferior designs to dominate better alternatives if they gain early momentum. Then we’re locked into using buggy, slow, unnecessarily complicated programs for a decade or two, until finally everyone gets sufficiently fed up.
Neural networks are not inevitable. They are not mandated by some Cosmic Plan. We should replace the technology with better alternatives.
Gradient Dissent is divided in three chapters, plus an epilogue.
“Artificial neurons considered harmful” explains how and why neural networks are dangerously unreliable.
“Do AI as science and engineering instead” discusses neglected technical approaches that may make them less risky.
“Backpropaganda: anti-rational neuro-mythology” explains the misleading rhetoric decision makers use to justify building and putting into use dangerously defective technology.
The epilogue, “A better future, without backprop,” suggests that it’s important, urgent, and probably possible to replace neural networks with better alternatives.
Gradient Dissent was originally a chapter in Better without AI. I separated it out as an independent document because it goes into more detail than some readers would want.
Reading it requires no specific technical background. It neither assumes you know how neural networks work, nor does it contain an introductory explanation that would get you up to speed. You can understand the risks without knowledge of those details.