I can see "no progress in 50 years" in fundamental physics where the experimental frontier seems to be running away from us (though recent gamma astronomy results suggest a next generation accelerator really could see the dark matter particle)
In biology or chemistry it's absurd to say that -- look at metal organic frameworks or all kinds of new synthetic chemistry or ionic liquids or metagenomics, RNA structure prediction, and unraveling of how gene regulation works in the "dark genome".
Progress in the 'symbolic AI' field that includes proof assistants is a really interesting story. When I was a kid I saw an ad for Feigenbaum's 3-volume "Handbook of AI" and got a used copy years later -- you would have thought production rules (e.g. "expert systems" or "business rules") were on track to be a dominant paradigm but my understanding was that people were losing interest even before RETE engines became mainstream and even the expert system shells of the early 1980s didn't use the kind of indexing structures that are mainstream today so that whereas people we saying 10,000 rule rule bases were unruly in the 1980s, 10,000,000 well-structured rules are no problem now. Some of it is hardware but a lot of it is improvements in software.
SAT/SMT solvers (e.g. part of proof assistants) have shown steady progress in the last 50 years, though not as much as neural networks because they are less parallelization. There is dramatically more industrial use of provers though business rules engines, complex event processing, and related technologies are still marginal in the industry for reasons I don't completely understand.
> In 1994, came the Pentium with its FDIV bug: a probably insignificant but detectable error in floating-point division. The subsequent product recall cost Intel nearly half a billion dollars. John Harrison, a student of Mike’s, decided to devote his PhD research to the verification of floating-point arithmetic.
No mention of the effort by Boyer and Moore, then at their Computational Logic, Inc., to do a formal verification of the AMD FPU for the AMD5K86TM. The AMD chip shipped with no FDIV bug. [1]
ACL2 doesn't get a lot of love from the side of the verification community that focuses on the proof systems that are more academically popular (HOL family, CIC family, etc.). A lot of interesting industrial work has been done with ACL2 and related systems.
In biology or chemistry it's absurd to say that -- look at metal organic frameworks or all kinds of new synthetic chemistry or ionic liquids or metagenomics, RNA structure prediction, and unraveling of how gene regulation works in the "dark genome".
Progress in the 'symbolic AI' field that includes proof assistants is a really interesting story. When I was a kid I saw an ad for Feigenbaum's 3-volume "Handbook of AI" and got a used copy years later -- you would have thought production rules (e.g. "expert systems" or "business rules") were on track to be a dominant paradigm but my understanding was that people were losing interest even before RETE engines became mainstream and even the expert system shells of the early 1980s didn't use the kind of indexing structures that are mainstream today so that whereas people we saying 10,000 rule rule bases were unruly in the 1980s, 10,000,000 well-structured rules are no problem now. Some of it is hardware but a lot of it is improvements in software.
SAT/SMT solvers (e.g. part of proof assistants) have shown steady progress in the last 50 years, though not as much as neural networks because they are less parallelization. There is dramatically more industrial use of provers though business rules engines, complex event processing, and related technologies are still marginal in the industry for reasons I don't completely understand.
No mention of the effort by Boyer and Moore, then at their Computational Logic, Inc., to do a formal verification of the AMD FPU for the AMD5K86TM. The AMD chip shipped with no FDIV bug. [1]
[1] https://dl.acm.org/doi/abs/10.1109/12.713311