by PaulHoule
3 subcomments
- 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.
by juliangamble
2 subcomments
- I'd like to call out the work from Nada Amin in this area:
Dafny and verification-aware programming, including proof by induction to verify properties of programs (for example, that an optimizer preserves semantics). Dafny Sketcher (https://github.com/namin/dafny-sketcher)
Multi-stage programming, a principled approach to writing programs that write programs, and its incarnation in multi-stage relational programming for faster synthesis of programs with holes—with the theoretical insight that a staged interpreter is a compiler, and a staged relational interpreter for a functional language can turn functions into relations running backwards for synthesis.
multi-stage miniKanren (https://github.com/namin/staged-miniKanren)
Monte Carlo Tree Search, specifically the VerMCTS variant, and when this exploration-exploitation sweet spot is a good match for synthesis problems.
VerMCTS (https://github.com/namin/llm-verified-with-monte-carlo-tree-...), and Holey (https://github.com/namin/holey).
- > 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]
[1] https://dl.acm.org/doi/abs/10.1109/12.713311
- I recently started reading "Specifying Systems: The TLA+ Language and Tools for Hardware and Software En" by Lawrence Lamport[0]. It is a good starting point for learning how to specify systems on the basis of mathematical proofs.
Since the new code is specifications in the age of AI, learning how to specify systems mathematically is a huge advantage because English is extremely ambiguous.
[0]: https://lamport.azurewebsites.net/tla/book-02-08-08.pdf
by polskibus
1 subcomments
- Lean4 seems to be very popular in the Math Olympics-solving AI startups (Harmonic etc).
- I think when Peter Thiel talks about stagnation it’s as much about the vibe of things as it is quantifying metrics of progress. I happen to agree with the idea that the vibe of progress has unilaterally focused on ai and computer technology even though this isn’t the case for metrics (e.g., CRISPR). I think the stagnation vibe has come from a series of issues such as decades of American political gridlock, Europes inability to commit to anything, the rise of the attention economy, the stagnation of Japan, the lack of a competitive focus for American markets (with no Soviet Union to defeat they seemed to have turned to defeating billionaire poverty). This list continues but it’s the confluence of all these things that gives the vibe of stagnation. It seems like there isn’t enough time to focus on anything anymore to actually get excited to drive it forward. Even though this isn’t the case. And there also isn’t the political space to celebrate things. Like it seems like any spacex accomplishment should be followed with the caveat that you don’t like Elon. Another way I think about this is that we live in this post cynicism world where we all have to couch every statement within some framework acknowledging harm to someone somehow. It’s hard to have a vibe of progress when every statement of progress includes an impact statement saying why that progress is harmful.
by Agingcoder
0 subcomment
- Very nice post thanks - I didn’t know about cakeml and bootstrapping formally verified compilers.
- I wish he had just said 50 years of LCF, since he even mentions automath in the article but that was but that was late 60s
by adyashakti
1 subcomments
- And now, Matthew Scherf has published "A Formal Specification of Advaita Vedānta in Classical High-Order Logic," and verified it in both Isabelle and Lean4. https://github.com/matthew-scherf/Advaita
Hacker News