SHRDLU(hci.stanford.edu)
hci.stanford.edu
SHRDLU
https://hci.stanford.edu/~winograd/shrdlu/
21 comments
It should be fairly easy to get the original code working in the MIT CADR emulator, could even add a GUI.
It was written in 1970, in an early version of MACLISP, to run on an early version of ITS. It would require a fair amount of work to port the code to Lisp Machine Lisp so that it could run on the CADR emulator.
Some (relatively successful) attempts were made to run it in an emulated PDP: https://github.com/PDP-10/its/issues/147
And then check out this very cool project from six years ago, which uses a block-stacking game inspired by Shrdlu but learns whatever language you choose to teach it instructions with :) https://shrdlurn.sidaw.xyz/
Ah, back in the days when we thought AI was a Simple Matter Of Programming and that logic and analysis were fundamentally harder problems than perception, localisation, cognition and sentience (whatever that particular begged question even means).
Just add a couple of feedback loops, if they're strange enough it might work.
Edit: Fun fact, the Winograd Schema Challenge (https://en.wikipedia.org/wiki/Winograd_schema_challenge) is named after Terry Winograd, the author of SHRDLDU.
Just add a couple of feedback loops, if they're strange enough it might work.
Edit: Fun fact, the Winograd Schema Challenge (https://en.wikipedia.org/wiki/Winograd_schema_challenge) is named after Terry Winograd, the author of SHRDLDU.
I recall some company or project in the late 70s or 80s putting together a massive database of concepts and how they relate to each other, i.e. a semantic word database. Wish I could remember the name, I know it's on wikipedia somewhere.
I always thought the effort must have been related to code like this. This simple language interpreter can do lots of things, but only in its small world of boxes and pyramids. It follows that if you teach it a bunch of things and how they relate to each other, you'd create an intelligent computer.
Instead that whole field was dropped, presumably because it didn't work, and eventually neural networks started to be thought of as the next generation of soon-to-be-AI systems. But it's pretty apparent by now that things like GPT-3 are hollow shells, good at emulating some very specific things humans can do, but there's nothing remotely like intelligence behind it.
I wonder if in 20 years we'll look back at this as another dead end on the road to AI.
I always thought the effort must have been related to code like this. This simple language interpreter can do lots of things, but only in its small world of boxes and pyramids. It follows that if you teach it a bunch of things and how they relate to each other, you'd create an intelligent computer.
Instead that whole field was dropped, presumably because it didn't work, and eventually neural networks started to be thought of as the next generation of soon-to-be-AI systems. But it's pretty apparent by now that things like GPT-3 are hollow shells, good at emulating some very specific things humans can do, but there's nothing remotely like intelligence behind it.
I wonder if in 20 years we'll look back at this as another dead end on the road to AI.
There were also Thought Treasure. In some aspects it is similar, only a bit more modern (1990'):
https://en.wikipedia.org/wiki/ThoughtTreasure
https://en.wikipedia.org/wiki/ThoughtTreasure
Projects like this one and Cyc really appeal to my inner hoarder and programmer both. It must be so satisfying (if pointless) to create datasets like this.
But I suspect the next time somebody tackles a project like this it'll use natural language parsing to derive its dataset from all the world's written text, automatically. Wikipedia will tell you a pea is green, if you can read it.
But I suspect the next time somebody tackles a project like this it'll use natural language parsing to derive its dataset from all the world's written text, automatically. Wikipedia will tell you a pea is green, if you can read it.
Cyc? It's still a going concern - www.cyc.com
The ability of SHRDLU to follow its user's instructions remains unsurpassed by modern systems.
I think there's an interesting story to be told in how Winograd started with "pure AI" approaches to language understanding like SHRDLU, gave it up in frustration, and then helped to found the field of human-computer interaction.
In some ways, HCI has much more modest and applied goals than AI. In other ways, HCI has had a much larger impact on many more computer users and developers in recent decades. Stanford's Symbolic Systems program effectively became the training ground for Google product managers and other consumer tech companies with strong UI/UX roles.
In some ways, HCI has much more modest and applied goals than AI. In other ways, HCI has had a much larger impact on many more computer users and developers in recent decades. Stanford's Symbolic Systems program effectively became the training ground for Google product managers and other consumer tech companies with strong UI/UX roles.
Winograd copied some of his page from the original resurrection page now at https://sites.google.com/site/masteraddressfile/misc/shrdlu
Related:
How SHRDLU got its name (2003) - https://news.ycombinator.com/item?id=24102610 - Aug 2020 (16 comments)
The SHR-DLU AI Natural Language Processing System (1970) - https://news.ycombinator.com/item?id=21880043 - Dec 2019 (1 comment)
SHRDLU resurrection - https://news.ycombinator.com/item?id=19138825 - Feb 2019 (4 comments)
SHRDLU – a program for understanding natural language (1968) - https://news.ycombinator.com/item?id=17028731 - May 2018 (1 comment)
SHRDLU (1971) - https://news.ycombinator.com/item?id=14351485 - May 2017 (30 comments)
Shrdlu resurrection – A 1970 artificial intelligence system - https://news.ycombinator.com/item?id=8927043 - Jan 2015 (2 comments)
SHRDLU - https://news.ycombinator.com/item?id=8219409 - Aug 2014 (19 comments)
How SHRDLU got its name (2003) - https://news.ycombinator.com/item?id=24102610 - Aug 2020 (16 comments)
The SHR-DLU AI Natural Language Processing System (1970) - https://news.ycombinator.com/item?id=21880043 - Dec 2019 (1 comment)
SHRDLU resurrection - https://news.ycombinator.com/item?id=19138825 - Feb 2019 (4 comments)
SHRDLU – a program for understanding natural language (1968) - https://news.ycombinator.com/item?id=17028731 - May 2018 (1 comment)
SHRDLU (1971) - https://news.ycombinator.com/item?id=14351485 - May 2017 (30 comments)
Shrdlu resurrection – A 1970 artificial intelligence system - https://news.ycombinator.com/item?id=8927043 - Jan 2015 (2 comments)
SHRDLU - https://news.ycombinator.com/item?id=8219409 - Aug 2014 (19 comments)
I presume the title is related to the list of most frequently used letters in English, etaoinSHRDLUcmwfgypbvkjxqz?
It doesn't work anymore - but for a while you could ask Siri (also via Stanford) "do you know shrdlu" (you'd have to type it into the correction box) and it would respond with something to the effect: "Wasn't SHRDLU crushed by a blue pyramid?"
If nothing else the program inspired people to think about AI and NLP. I'm working on Natural Language Understanding (NLU) myself (https://lxagi.com/).
Also check out the Stanford NLU course about this topic: cs224u.stanford.edu
There's a section on Shrdlu there too
There's a section on Shrdlu there too
The text-only system contains the original code partially converted from MacLisp to GNU CLisp, and runs (albeit with an error that can be skipped) in a modern CLisp version. But it's still buggy[1] and sometimes hangs, it would be great if someone could set up a containerized MacLisp version on a PDP6 emulator.
[1] SHRDLU was never per se robust, but in this case there are clearly bugs from the conversion