Earlier this week, having come across my ArXiv paper about reanimating the Logic Theorist, David Moews (https://djm.cc/dmoews.html) wrote me about an amazing piece of work he's done:
He built an interpreter for the original IPL-I pseudocode of the original Newell and Simon Logic Theorist, straight out of the 1956 RAND report (P-868), and then got it running!
(I'll call the version that David reanimated "LT1" or "LT56", and mine "LT5" or "LT63" because mine was rewritten for IPL-V and published in 1963.)
What makes David's work especially interesting, aside from pushing the RetroAI window back 8 more years (!), is that IPL-I was NEVER ACTUALLY IMPLEMENTED! It was hand-executed by Simon's students (and supposedly his kids!) simulating the imagined IPL-I machine. This actually makes the problem much simpler. (Not at all to diminish David's accomplishment!)
In the 1956 report Newell and Simon describe the process in something close to the cognitive operators they hypothesized underlay human theorem proving. This is essentially LT1: A (somewhat) high-level specification of the Newell and Simon theory of cognitive theorem proving. But because LT1 didn't have to actually run on a real computer, it could depend upon human intelligence and flexibility to handle the complexities of actual implementation that you need to do to make a real computer actually do the whole thing end-to-end. (Or, as in David's case, a pile of Python code, which, of course, Simon and Newell didn't have in the mid 1950s!) As a result, LT1 is a bit over 400 lines whereas LT5, which is what you get when Shaw had to actually nail down the complexities of actual implementation, is nearly 3000 lines!
Anyway, huge congratulations to David; well worth a look if you care about the prehistory of AI, Lisp, or theorem proving. His repo is here: https://github.com/dmoews/logic-theorist. The readme provides a lot of intersting and important detail that I've glossed over.
Lars Brinkhoff brought up PARRY, Kenneth Colby's 1972 paranoid chatbot, and Rupert Lane, who recently reanimated the original ELIZA, was able to (sort of) reproduce the famous RFC439 trans-internet conversation between ELIZA and PARRY. See the blog post for the amazing (and amusing) details! (Turns out that RFC439 used a Lisp ELIZA, not the MAD-SLIP original.)
"Inspired by true events, this world premiere by award-winning playwright Tom Holloway is a gripping psychological thriller about the birth of AI, and the ethical consequences of outsourcing our humanity."
I have successfully reanimated Simon, Newell, and Shaw’s Logic Theorist, the world's first AI[note], on an IPL-V emulator written in Lisp! This dribble shows it proving 11 complete theorems from the original Principia (or, at least from Simon and Newell's original inputs, which claim to have been from the Principia. I haven't checked.)
(It does break at the end, and I haven't started to track that down -- it may actually be my own runaway limiters that tripped it.)
[note] Or at worst the second -- depending on whether you count Arthur Samuel's first checkers player. But certainly LT is the first cognitive model, and first explicitly symbol processing AI.
Yeah, no, I don’t mean crashed altogether, just off for routine maintenance on a rotating basis. But anyway, no sense in prolonging this. Anyway, agree that 1% seems high.
Not sure about the mfgr- could be rounding for memory saliency. . . but re HN they’re counting clock time not power on time. Could have easily been turned off total 10 days for maintenance etc. (and good routine could have both servers off about the same no hrs.) That’s only 0.6% downtime.
He built an interpreter for the original IPL-I pseudocode of the original Newell and Simon Logic Theorist, straight out of the 1956 RAND report (P-868), and then got it running!
(I'll call the version that David reanimated "LT1" or "LT56", and mine "LT5" or "LT63" because mine was rewritten for IPL-V and published in 1963.)
What makes David's work especially interesting, aside from pushing the RetroAI window back 8 more years (!), is that IPL-I was NEVER ACTUALLY IMPLEMENTED! It was hand-executed by Simon's students (and supposedly his kids!) simulating the imagined IPL-I machine. This actually makes the problem much simpler. (Not at all to diminish David's accomplishment!)
In the 1956 report Newell and Simon describe the process in something close to the cognitive operators they hypothesized underlay human theorem proving. This is essentially LT1: A (somewhat) high-level specification of the Newell and Simon theory of cognitive theorem proving. But because LT1 didn't have to actually run on a real computer, it could depend upon human intelligence and flexibility to handle the complexities of actual implementation that you need to do to make a real computer actually do the whole thing end-to-end. (Or, as in David's case, a pile of Python code, which, of course, Simon and Newell didn't have in the mid 1950s!) As a result, LT1 is a bit over 400 lines whereas LT5, which is what you get when Shaw had to actually nail down the complexities of actual implementation, is nearly 3000 lines!
Anyway, huge congratulations to David; well worth a look if you care about the prehistory of AI, Lisp, or theorem proving. His repo is here: https://github.com/dmoews/logic-theorist. The readme provides a lot of intersting and important detail that I've glossed over.