Correspondence: Weizenbaum Joe to Feigenbaum, 23 December 1970
From The Edward A. Feigenbaum Papers archives at Stanford University.
Transcribed by ChatGPT 4o
Dear Edward,
Thank you for your very kind letter. That was most welcome for obvious reasons. But then you yourself gave evidence that my feelings were based in reality. Anyway, the antidote to such feelings is work. During our visit here we discovered not only for the usual purely rational reasons but also that it was thought provoking to an unusual degree.
I remain very impressed that your program has the same gross structure as Wood’s SRI program. I am convinced that the reason Licklider found it so easy to unearth your logs is that you use in him his subprograms that bear that general structure as well and that he has learned to carve it into knowledge into that form whenever appropriate. He has, in other words, a generative system that organizes knowledge. To program GENERAL like programs even at all and thus make corresponding “expert” familiarities, I see the work you are now beginning as a step in the direction of compiling explicit models of expert subprograms, to mask imperfections in abstractions of a somewhat naive approach are altogether it therefore strikes your effort to be worthy of very considerable investment of human and financial resources.
You may remember that my own concern has been with understanding. I play with understanding.
One central reason I recently feel the need to provide a smooth host bed for such ideas. I’ve been thinking lately that we all agree that the linguists are much too deeply mistaken, too, of course, have the benefit of hindsight.
They thought that syntax could be separated from semantics. They didn’t seem knowledgeable that even the identification of a word as a noun (or any other part of speech) clearly involved real world knowledge and hence semantics.
They restrict their studies and essay selecting sentences in isolation. Surely such questions as “is the sentence ‘colorless green ideas sleep furiously’ grammatical or not?” are only half time’s wasting.
Now almost everyone agrees that syntax and semantics are inseparable and that context (and a minimum the context supplied by a paragraph) is crucial to understanding. Indeed, one test of the grammaticality (?) of a sentence is whether or not one can construct…
…a context in which it would have meaning. But I want to go a step further. I want to argue that utterances are imparted meaning, so that they are part of conversations. Indeed, I would even argue that a look is a part of a conversation. Surely, for example, we often instinctively are audience to which he took himself to be speaking. But it’s not yet disturbed by whatever examples before we just watched. I recall (either in the ELIZA paper) that I thought of a (two-person) conversation as being a kind of mirroring of the participants’ respective beliefs. Namely, that if the (two) belief structures were not (close), then no conversation is possible. If they are very nearly identical, then no conversation is necessary or maintainable. People with similar belief structures (with respect to specific parts of the real world) experience strain talking to one another – an urgency to discover differences in their respective beliefs/structures. Again, such differences amount to goals (no possible conversation) – not once they are too rare (no interest). In cybernetic systems, we look at a disturbance, i.e. some fluctuation, to find it still another way, a reason for change or transition. The immediate of a system in equilibrium (and that is not knocked out of equilibrium by an…
…exogenous input) does not get disturbed. So it is with human conversations.
We cannot really ask ourselves why we engage in conversations at all. Never mind those instances when we are merely being polite or when we are merely serious; these have immediate practical purposes such as “let’s discuss,” etc. Consider only those conversations that surprise or engender unanticipated elements of excitement – for that matter, the reader’s excitement when we read along the words with mounting enthusiasm and tension. The problem is, from where does the meeting come, that energy that animates parts itself in our excitement? I suggest that whenever we observe energy release we are observing either a system in disequilibrium – one that is trying to find stability and arrive into equilibrium. Very much like a soap bubble trying to maintain just the proper curve. Now, when we expose ourselves to the words of another person (or to his art works) we expose our belief structures to adjustment. I suggest that we are subject to a kind of (perhaps neurological) principle of least effort that inevitably drives us to attempt to obtain a kind of maximum energy configuration for our belief structures, perhaps to make them as small as possible. An insight is a restructuring of our knowledge that permits reaction…
..positions of our belief structures to collapse into much more economical forms. We experience such collapse as pleasure. We are uncomfortable (in disequilibrium) in sustaining (carrying) uncollapsed belief structures. I cannot help myself here to speak to the unscientific metaphor at this point. It would be nice to see that the resolution of our belief structures is modelled by the resolution of programs like DECORD and SIR. By this I mean that it seems programs (like shrink functions) made heuristic and introspective. Collapses are replacements of heuristic procedures by algorithmic processes, perhaps accompanied by collapsing into the process of collapsing change is final. Perhaps it is this understanding that yields feelings of relief and pleasure. However that may be, we have all had experienced this kind of collapse as almost physical convulsion. It occurs during those moments of enlightenment and it is experienced as relief and pleasure.
I should modify my argument here to take writing thinking into account. For clearly, not all examples are derived from conversation (including reading) with others. What is essential, it seems to me, is the encounter two – a possible belief structure is modified through exposure to another one. But both are received or heard. Creative thinking, even doing only when one individual…
..recognizes that two (or more) of his ideological belief structures that appear superficially to be unrelated to one another do in fact have the potential of coordinating and collapsing. The creative individual has the ability to see where his heuristics of schemes fail, in juxtaposing two belief structures with one another, he is not generally minded re-dimensioning the conversation on the basis of their obvious mismatches. He is able, as in theater, to suspend disbelief. The difference of education is that it arrives at the same frame immature through abundance of evidence and yet not altering the student’s capacity to suspend disbelief. Children have wonderful plausible naive generalities, but they also possess evaluative functions. Many of them have a fine lot of. Their evaluative functions are increasingly intellectual.
I have often said that programming is very much like writing. Everybody is a programmer, just as everyone speaks prose, whether he is aware of it or not. When we sit down to write, we leave the illusion that we understand what we wish to say. But the act of writing exposes holes in our understanding and we are forced to pause to think. Programming is like that: only more so. A new culture, which no prior trunk of its members literacy, maintains the illusion that everyone knows how to write well – that writing skill is automatically…
..acquired without much conscious effort. We don’t yet have the same illusion with respect to computer programming. We know, for example, that it takes much training to produce a decent – a very good computer writer. My argument leads me to say that the leaders of this world have developed very good systems, programs in their heads, programs that facilitate the confrontation of elaborate belief structures, their collapse into more economical structures, and the replacement of heuristic interpretive routines by compiled algorithmic processes. I emphasize that point once more because I believe that that skill is not the product of a chain of accidents (genetic or otherwise) but a learned skill (like systems programming). But we know next to nothing about it. Hence we don’t know what is conducive to its acquisition or, indeed, what sort of training and practice may be counterproductive.
Again, I now think that your work will contribute mightily in this area. I am also beginning to formulate my own plans. Briefly speaking, what needs to be done is to rewrite part of this letter in the form of a program. In particular, I wish to construct an ELIZA-like program (or several such) the explicit aim of which is to be the study of belief structure resolution and (hopefully) collapse along the lines I’ve discussed.
The thinking I’ve done so far tells me that the earlier ELIZA system (while canned in some respects) is grossly inefficient in many others. Its total ignorance of grammar, for example, cannot be underestimated. Also the structure of ELIZA scripts (e.g., absence of threads) argues from an entirely new perspective, i.e., from that of belief structures. They seem mechanically incompatible — kind of PARSER mechanism like that of Carl Hewitt’s.
Even the old ELIZA contains some real world knowledge — the new one must have lots more. And that means a careful choice of area. One important task (I know) is to construct independent script sets (scripts and meta-scripts) that can then be used with one another. These will provide an excellent way of yielding reproducible research. In the very long run, it may become possible to test out whole hypotheses of knowledge by comparing programs (scripts) written by different people with one another. Imagine an automatic debate between two “minds.”
You understand, of course, that this effort does not start from scratch. We learned a lot from ELIZA. For example, the idea of a student that represents his beliefs in a script structure seems to be correct. Also, our earlier analyses of conversations indicate that…
…they have many characteristics of running computer programs, that is, some of the standard scripts are clearly runnable ALGOL-like programs. I am convinced that many of our deepening insights into computer programs are readily transferable to the analysis of natural language interaction. I am thinking particularly of what we know about symbol table structures and the binding of variables. It is interesting, for example, that one can represent word jokes in terms of accessing kinds of symbol table mishaps (e.g., identifier collisions). Unfortunately, the world is short on writers. Like that, few would also be able to interact naturally with children and codify. I am quite convinced that if Tom and I were together we could really make some fun at least half a decade – maybe even a thesis – either of us separately, perhaps (I’m presumptuous speaking for him), but I do speak for myself.
As a last close of things for you – remember, all this is intended to release your sadness over my initial conditions. Let me mention another (smaller) task that I believe deserves doing. It might make a good master’s thesis. I’ve already spoken elsewhere about the structure of DEBORAH and SIR-like programs. I suspect some of them are inverse function programs. The spectroscope, informally speaking, is a function that…
…takes a molecule as its input and produces a relative position in the output. DEBORAH works that function. Certainly, the present decade we will see another figure as a major function inversion programs of the same type. I believe it to be important that students be given a good understanding of the structural function of such programs. They should therefore study examples. But examples ought not to be more complicated than what they are intended to exemplify. (That’s one of the sad things of objects I know.) Hence we should find a domain that has a good theory, not too deep but at the same time not too over-simplified, and for which the inverse function is known, and build the appropriate DEBORAH program for it. It should, of course, be pretty heuristic-looking. If the examples involve familiar programming code. A much too dull and probably too simple (if it can be done at all) example might be…
Suppose the string “…” were printed using the Fortran format F, and produced the printed output “…” — What is the format statement?
I don’t like that – but I hope you see what I mean.
I’ve also been discussing some writing ideas about language translation with Woody Bledsoe. But enough therapy for today.
Best greetings — to your family as well and from my family too.
Joe.
