Computer ratings of publicly available AI can be found in the CCRL rating list. Deep Blue was a hardware effort that has long since been dismantled, and could not be tested. It would be weak by today's standards. (It had difficulty winning even a single game against Kasparov; recent engines can afford Knight odds against GM, and would still beat them.)
The order seems random. Some of the mentioned engines are close derivatives of each other. Houdini was a Stockfish clone in the time the latter used 'Hand-Crafted Evaluation'. Latest Stockfish uses Neural Network evaluation. Fat Fritz is an exact clone of it, but trained on a different set of games. (NN have to be trained by millions of examples to be of any use. HCE has to be tuned on position examples as well, but fewer. And it start out at a level that is not completely insane even without tuning, as the tunable parameters have known meaning (like piece values or king safety) and can already be assigned values from chess lore.
Alpha-Zero was a private Google effort using Google server hardware, and thus could not be tested for a rating. Leela Chess Zero is a public implementation of it for PC. It uses a completely different algorithm, using a huge NN, which not only provides evaluation, but also guides a very selective search (which is about 1000 times slower than Stockfish'). It is thought to be much stronger than the original Alpha Zero, through better training. The training takes hundreds of thousands hours of PC time, and was done as a community effort.
Komodo is an engine of similar design to Stockfish, but developed independently.
All AI are of super-human strength as far as recognizing tactics is involved, but they can have strategic misconceptions due to imperfect training/tuning of their evaluation, or using an evaluation design that is simply not able to grasp the required concept. E.g. a HCE that would evaluate material as the sum of fixed piece values will fail to recognize 7 Knights are stronger than 3 Queens, or they would think a Knight is as valuable as a Rook in normal play, depending on the example positions you tune it on. NN are more flexible, but can have blind spots due to absence of sufficiently many examples of a required concept in the training set.
Computer ratings of publicly available AI can be found in the CCRL rating list. Deep Blue was a hardware effort that has long since been dismantled, and could not be tested. It would be weak by today's standards. (It had difficulty winning even a single game against Kasparov; recent engines can afford Knight odds against GM, and would still beat them.)
The order seems random. Some of the mentioned engines are close derivatives of each other. Houdini was a Stockfish clone in the time the latter used 'Hand-Crafted Evaluation'. Latest Stockfish uses Neural Network evaluation. Fat Fritz is an exact clone of it, but trained on a different set of games. (NN have to be trained by millions of examples to be of any use. HCE has to be tuned on position examples as well, but fewer. And it start out at a level that is not completely insane even without tuning, as the tunable parameters have known meaning (like piece values or king safety) and can already be assigned values from chess lore.
Alpha-Zero was a private Google effort using Google server hardware, and thus could not be tested for a rating. Leela Chess Zero is a public implementation of it for PC. It uses a completely different algorithm, using a huge NN, which not only provides evaluation, but also guides a very selective search (which is about 1000 times slower than Stockfish'). It is thought to be much stronger than the original Alpha Zero, through better training. The training takes hundreds of thousands hours of PC time, and was done as a community effort.
Komodo is an engine of similar design to Stockfish, but developed independently.
All AI are of super-human strength as far as recognizing tactics is involved, but they can have strategic misconceptions due to imperfect training/tuning of their evaluation, or using an evaluation design that is simply not able to grasp the required concept. E.g. a HCE that would evaluate material as the sum of fixed piece values will fail to recognize 7 Knights are stronger than 3 Queens, or they would think a Knight is as valuable as a Rook in normal play, depending on the example positions you tune it on. NN are more flexible, but can have blind spots due to absence of sufficiently many examples of a required concept in the training set.