A couple weeks ago I made it to the FSF's 30th anniversary party.
It was a blast in many ways, and a good generator of many fond
memories, but I won't go into depth of them here. One particularly
exciting thing that happened for me though was I got to meet
Gerald Sussman
(of SICP!)
The conversation has greatly impacted me, and I've been spinning it
over and over again in my mind over the last few weeks... so I
wanted to capture as much of it here while I still could. There are
things Sussman said that I think are significant... especially in
the ways he thinks contemporary AI is going about things wrong, and
a better path forward. So here's an attempt to write it all down...
forgive me that I didn't have a real tape recorder, so I've written
some of this in a conversational style as I remember it, but of
course these are not the precise things said. Anyway!
I wasn't sure initially if the person I was looking at was Gerald
Sussman or not, but then I noticed that he was wearing the same
"Nerd Pride" labeled pocket protector I had seen him wear in
a lecture I had
watched recently. When I first introduced myself, I said, are
you Sussman? (His first reply was something like to look astonished
and say, "am I known?") I explained that I've been reading the Structure and
Interpretation of Computer Programs and that I'm a big fan of
his work. He grinned and said, "Good, I hope you're enjoying it...
and the jokes! There's a lot of jokes in there. Are you reading
the footnotes? I spent a lot of time on those footnotes!" (And this
point my friend David Thompson joined me, and they both chuckled
about some joke about object oriented programmers in some footnote I
either hadn't gotten to or simply hadn't gotten.)
He also started to talk enthusiastically about his other book, the
Structure and Interpretation of Classical Mechanics,
in which classical engineering problems and electrical circuits are
simply modeled as computer programs. He expressed something similar
to what he had said in the forementioned talk, that conventional
mathematical notation is unhelpful, and that we ought to be able to
express things more clearly as programs. I agreed that I find
conventional mathematical notation unhelpful; when I try to read papers
there are concepts I easily understand as code but I can't parse the
symbolic math of. "There's too much operator overloading", I said,
"and that makes it hard for me to understand in which way a symbol
is being used, and papers never seem to clarify." Sussman replied,
"And it's not only the operator overloading! What's that 'x' doing
there! That's why we put 'let' in Scheme!" Do you still get to
write much code or Scheme these days, I asked? "Yes, I write
tens of thousands of lines of Scheme per year!" he replied.
I mentioned that I work on distributed systems and federation,
and that I had seen that he was working on something that was called
the propagator model,
which I understood was some way of going about asynchronous
programming, and maybe was an alternative to the actor model? "Yes,
you should read the paper!" Sussman replied. "Did you read the
paper? It's fun! Or it should be. If you're not having fun
reading it, then we wrote it wrong!" (Here is
the paper,
as well as the documentation/report
on the software...
see for yourself!) I explained that I was interested in code that
can span multiple processes or machines, are there any restrictions
on that in the propagator model? "No, why would there be? Your
brain, it's just a bunch of hunks of independent grey stuff sending
signals to each other."
At some point Sussman expressed how he thought AI was on the wrong
track. He explained that he thought most AI directions were not
interesting to him, because they were about building up a solid AI
foundation, then the AI system runs as a sort of black box. "I'm
not interested in that. I want software that's accountable."
Accountable? "Yes, I want something that can express its symbolic
reasoning. I want to it to tell me why it did the thing it did,
what it thought was going to happen, and then what happened
instead." He then said something that took me a long time to
process, and at first I mistook for being very science-fiction'y,
along the lines of, "If an AI driven car drives off the side of the
road, I want to know why it did that. I could take the
software developer to court, but I would much rather take the AI to
court." (I know, that definitely sounds like out-there science
fiction, bear with me... keeping that frame of mind is useful for
the rest of this.)
"Oh! This is very interesting to me, I've been talking with some
friends about how AI systems and generative software may play in
with software freedom, and if our traditional methods of considering
free software still applied in that form," I said. I mentioned a
friend of a friend who is working on software that is generated via
genetic programming, and how he makes claims that eventually that
you won't be looking at code anymore, that it'll be generating this
black box of stuff that's running all our computers.
Sussman seemed to disagree with that view of things. "Software
freedom is a requirement for the system I'm talking about!" I liked
hearing this, but didn't understand fully what he meant... was he
talking about the foundations on top of which the AI software ran?
Anyway, this all sounded interesting, but it also sounded very
abstract. Is there any way this could be made more concrete? So I
asked him, if he had a student who was totally convinced by this
argument, that wanted to start working on this, where would you
recommend he start his research? "Read the propagators paper!"
Sussman said.
OH! Prior to this moment, I thought we were having two separate
conversations, one about asynchronous programming, and one about AI
research. Suddenly it was clear... Sussman saw these as interlinked,
and that's what the propagator system is all about!
One of the other people who were then standing in the circle said,
"Wait a minute, I saw that lecture you gave recently, the one called
'We Don't
Really Know How to Compute!', and you talked about the illusion
of seeing the triangle when there wasn't the triangle" (watch the
video) "and what do you mean, that you can get to that point, and it
won't be a black box? How could it not be a black box?"
"How am I talking to you right now?" Sussman asked. Sussman seemed
to be talking about the shared symbolic values being held in the
conversation, and at this point I started to understand. "Sure,
when you're running the program, that whole thing is a black box.
So is your brain. But you can explain to me the reasoning of why
you did something. At that point, being able to inspect the
symbolic reasoning of the system is all you have." And, Sussman
explained, the propagator model carries its symbolic reasoning along
with it.
A more contrived relation to this in real life that I've been
thinking about: if a child knocks over a vase, you might be angry at
them, and they might have done the wrong thing. But why did they do
it? If a child can explain to you that they knew you were afraid of
insects, and swung at a fly going by, that can help you debug that
social circumstance so you and the child can work together towards
better behavior in the future.
So now, hearing the above, you might start to wonder if everything
Sussman is talking about means needing a big complicated suite of
natural language processing tools, etc. Well, after this
conversation, I got very interested in the propagator model, and to
me at least, it's starting to make a lot of sense... or at least
seems to. Cells' values are propagated from the results of other
cells, but they also carry the metadata of how they achieved that
result.
I recommend that you read the materials yourself if this is starting
to catch your interest.
(A repeat: here is the paper,
as well as the documentation/report
on the software...
see for yourself!). But I will highlight one part that may help
drive the above points more clearly.
The best way to catch up on this is to
watch the video
of Sussman talking about this while
keeping the slides handy.
The whole thing is worth watching, but about halfway through he
starts talking about propagators, and then he gets to an example of
where you're trying to measure the height of a building by a variety
of factors, and you have these relationships set up where as
information is filled in my a cell's dependencies, that cell merges
what it already knows about the cell with what it just learned. In
that way, you might use multiple measurements to "narrow down" the
information. Again, watch the video, but the key part that comes
out of the demonstration is this:
(content fall-time)
=> #(contingent #(interval 3.0255 3.0322)
(shadow super))
What's so special about this? Well, the fall-time has been updated
to a more narrow interval... but that last part (shadow and super)
are the symbols of the other cells which propagated the information
of this updated state. Pretty cool! And no fancy natural language
parsing involved.
There's certainly more to be extrapolated from that, and more to
explore (the Truth Maintenance Systems are particularly something
interesting to me). But here were some interesting takeaways from
that conversation, things I've been thinking over since:
-
AI should be "accountable", in the sense that it should be able to
express its symbolic reasoning, and be held up to whether or not
its assumptions held up to that.
-
Look more into the propagators model... it's like asynchronous
programming meets functional programming meets neural nets meets a
bunch of other interesting AI ideas that have been, like so many
things, dropped on the floor for the last few decades from the AI
winter, and which people are only now realizing they should be
looking at again.
-
On that note, there's so much computing history to look at, so
many people are focused on looking at what the "new hotness" is in
web development or deployment or whatever. But sometimes looking
backwards can help us better look forwards. There are ideas in
SICP that people are acting as if they just discovered today.
(That said, the early expressions of these ideas are not always the
best, and so the past should be a source of inspiration, but we
should be careful not to get stuck there.)
-
Traditional mathematical notation and electrical engineering diagrams
might not convey clearly their meaning, and maybe we can do better.
SICM
seems to be an entire exploration of this idea.
-
Free software advocates have long known that if you can't inspect
a system, you're held prisoner by it. Yet this applies not just
to the layers that programmers currently code on, but also into
new and more abstract frontiers. A black box that you can't ask
to explain itself is a dangerous and probably poorly operating
device or system.
-
And for a fluffier conclusion: "If you didn't have fun, we were
doing it wrong." There's fun to be had in all these things, and
don't lose sight of that.
{#Protesting EME before the Microsoft building; CC BY 3.0}