EFTA00176507

minor. D arrested at 
meeting site. 
05- 
80200 
Lucas Phelps 
5 
18 U.S.C.S. 1470 
Use of intemet to entice 
minor to engage in sex 
activity. 
Attempt to knowingly 
transfer child pornography 
in interstate commerce to a 
minor. 
06- 
14003 
Octavio 
Villalona 
2 
18 U.S.C.S. 
2252(a)(1) 
Use of intemet to entice 
minor to engage in sex 
activity. 
Knowingly transported 
child pornography by a 
computer. 
06- 
14006 
Daniel Williams 
1 
[None] 
Use of internet to entice 
minor to engage in sex 
activity. 
06- 
14007 
Ricky Barnett 
1 
[None] 
Use of Internet to entice 
minor to engage in sex 
activity. 
06- 
14011 
John Everhart, II 
I 
[None] 
Use of internet to entice 
minor to engage in sex 
activity. 
06- 
14016 
Eric Rollins 
3 
2 counts 2422(b) 
18 U.S.C.S. 2422(a) 
Use of Internet to entice 
minor to engage in sex 
activity. 
Knowingly enticed a 
minor to travel in interstate 
commerce to engage in 
sexual activity. 
06- 
14053 
Richard Grande, 
Jr. 
1 
[None] 
Use of intemet to entice 
minor to engage in sex 
activity. 
06- 
14069 
Eric Matthews 
4 
18 U.S.C.S. 1470; 
18 U.S.C.S. 
2252(a)(2) 
Use of intemet to entice 
minor to engage in sex 
activity. 
Knowingly transferred 
obscene material to a 
minor in interstate 
commerce. 
12 
EFTA00176547

Knowingly distributed 
child pornography in 
interstate commerce. 
06- 
14074 
Anthony Perez 
3 
18 U.S.C.S. 1470; 
18 U.S.C.S. 2251 (a) 
and (e) 
Use of intemet to entice 
minor to engage in sex 
activity. 
Knowingly transferred 
obscene material to a 
minor under 16 y.o. in 
interstate commerce. 
Enticed minor to engage in 
sexual conduct for purpose 
of transporting visual 
depiction in interstate 
commerce. 
06- 
20249 
Michael 
I 
[None] 
Knowingly attempted to 
induce minor to engage in 
prostitution. [no other 
facts] 
06- 
20341 
Dino Pancaro 
3 
18 U.S.C.S. 2423(e); 
18 U.S.C.S. 1594(a) 
Knowingly attempted to 
induce minor to engage in 
prostitution. 
Attempted to travel to 
engage in commercial sex 
act with a minor. 
06- 
20734 
Demond Osley 
Stacey Greer 
8 
18 U.S.C.S. 
1591(aX1); 
18 U.S.C.S. 2421; 
18 U.S.C.S. 2422(a); 
18 U.S.C.S. 
1001(aX2); 
18 U.S.C.S. 
1028(aX4) 
Minor arrested for 
prostitution on Miami 
Beach. 
When questioned by 
officers, minor said Osley 
brought her from Michigan 
to Florida for 
purpose of prostitution; 
Osley became unhappy 
with minor b/c she was not 
meeting daily quota; 
Osley sold minor to Greer. 
Greer takes minor to 
hotel, forces her to have 
sex, video tapes minor 
and takes photos of her to 
distribute on internet. 
Greer also forces minor 
into prostitution thru 
13 
EFTA00176548

threats of violence. Minor 
identified Osley and 
Greer. Both arrested. 
06- 
20783 
Keith Lanzon 
I 
[None 
Use of intemet to entice 
minor to engage in sex 
activity. 
06- 
80031 
Lynn Mann 
3 
18 U.S.C.S. 1470; 
18 U.S.C.S. 
2252A(a)(5)(B); 
18 U.S.C.S. 
2252A(b)(2) 
Use of Internet to entice 
minor to engage in sex 
activity. 
Distribute child 
pornography to a minor. 
Possession of child 
pornography. 
06- 
80034 
Rafael Ramirez, 
Jr. 
1 
[None] 
Use of internet to entice 
minor to engage in sex 
activity. 
06- 
80058 
Adam McDaniel 
2 
18 U.S.C.S. 2423(b) 
D was 19 in Texas, met 14 
y.o. girl on internet 
who lived in Florida. D & 
girl communicated 
by email & phone. D flew 
to Florida, met w/ 
girl and had sex w/ her in a 
hotel. 
06- 
80135 
David Girouard 
2 
18 U.S.C.S. 2423(b) 
Use of Internet and cellular 
telephone to entice minor 
to engage in sex activity. 
07- 
14002 
Benjamin■ 
4 
18 U.S.C.S. 1470; 
18 U.S.C.S. 
2252A(aX2)(M; 
18 U.S.C.S. 
2252(bX1); 
18 U.S.C.S. 
2252(a)(4)(B) 
Use of intemet to entice 
minor to engage in sex 
activity. 
Knowingly transferred 
obscene material to a 
minor under 16 y.o. in 
interstate commerce. 
Knowingly distributed 
child pornography in 
interstate commerce. 
Possession of child 
pornography. 
07- 
14004 
Ricky 
2 
18 U.S.C.S. 2251 (a) 
and (e) 
Use of internet to entice 
minor to engage in sex 
14 
EFTA00176549

activity. 
Attempted production of 
child pornography thru 
interstate commerce. 
07- 
Carl Berrier 
2 
18 U.S.C.S. 
Use of internet to entice 
14005 
2252A(a)(2)(A); 
18 U.S.C.S. 
minor to engage in sex 
activity. 
2252A(b)(1) 
Knowingly distributed 
child pornography in 
interstate commerce. 
07- 
Francesco Simo 
1 
[None] 
Use of internet to entice 
14015 
minor to engage in sex 
activity. 
07- 
Joseph Crutchley 
I 
[None] 
Use of internet to entice 
14016 
minor to engage in sex 
activity. 
07- 
14024 
Evans
il
Evans 
i
3 
1 
18 U.S.C.S. 2423(a); 
18 U.S.C.S. 2423(e) 
Use of internet to entice 
minor to engage in sex 
activity. 
Conspiracy to transport a 
minor to engage in sexual 
activity. 
Knowingly transport (or 
attempt) a minor to engage 
in sexual activity. 
07- 
Sammy 
4 
18 U.S.C.S. 1591(a); 
Knowingly attempted to 
20214 
Carpenter, 
Darryl Jennings, 
Luroy Jennings 
18 U.S.C.S. 2422(a) 
induce minor to engage in 
prostitution. 
07- 
Nelson Cintron 
3 
18 U.S.C.S. 
Use of intemet to entice 
60049 
2252A(a)(2)(A); 
18 U.S.C.S. 
minor to engage in sex 
activity. 
2252A(a)(5)(B) 
Possessed and distributed 
child pornography. 
07- 
Oliver Buelow 
2 
18 U.S.C.S. 2423(b) 
[No factual information] 
60084 
07- 
Marion 
3 
18 U.S.C.S. 2423(a); 
Use of internet and cellular 
80099 
Yarbrough 
18 U.S.C.S. 2422(a) 
telephone to entice 
minor to engage in sex 
15 
EFTA00176550

activity. 
Transport minor to engage 
in sex activity. 
Entice minor to travel in 
interstate commerce to 
engage in sex activity. 
16 
EFTA00176551

f 
(
 
4-
C 
C 
EFTA00176552

City College 
UI NewYork 
INTERNATIONAL STUDIES PROGRAM 
Jeffrey Epstein 
do Darren Indyke Esq. 
457 Madison Avenue — 14th Floor 
New York, 
N.Y. 10022. 
Dear Mr. Epstein, 
North Academic Center, Room 6/141 
160 Convent Avenue 
New  York, New York 10031 
TEL:l 
FAX: 
www.ccny.cuny.edu 
August 21,2006. 
Thank you for your continued and generous support of the undergraduate academic 
careers of Georges Ndabashimiye and Nicole Mutesi. 
Both students have done very well both academically and in co-curricular life and expect 
to graduate in June, 2008. Georges will return to Rwanda to teach and Nicole plans to 
join the energy industry which is focused on developing Rwanda's newly found resources 
in natural gas. 
Your support of these two students will thus contribute to the human resource wealth of 
Rwanda. 
Yours sincerely, 
Marina W. Fernando Ph.D. 
Director, International Studies Program 
and Deputy Dean of Social Science. 
THE CITY UNIVERSITY OF NEW YORK 
EFTA00176553

o$104106 
tton in etstekt, 
\a3UE kis *GA you ,‘cit 
\is 
/VOL 
r'n-1/46 
klAilietl 
Dona plios
(nit/ -liva-ecta hisu beam. 3 
ereakii
apvCCCat-e_ ow- ezneret•. 
ccovt,-ek coo 2a muck 
Quit 
Sr Ls. 9,nar a nit Act
 In".3 
ger- 
AIZ/IL 
O0(1O1 Ceara.- 
YI•1 Otatte — 
rAtt 
SGOES - 
111.,om irk 8.0 vituALA-
Sivlemity 
tii coLe M UTEs 
EFTA00176554

D / 
cP9- / 20'D 
am het sitter" , 
4604 _.„ 4tr 
T
otvlor2 rek- 1/41,d;we
 
la ° e_& 
1" Ntui 
c)(1<. 
toiatt 
aikage4 
rut a cti s 
4urk 
t act • Lon
s , 
EFTA00176555

114 
C 
EFTA00176556

teo\r- •Sticsret1 
.ssee.-\r, 
E‘s-Or 
rrovettvsr 
v40,,,rA, 3r, -1-0.14- 
3rt, 
yo..) 
3?•(.4A- 
t 
cAocte 
16t) 
‘r , 
\oe..Oc- 
\c‘tc1/4Veo 
tc : ‘ 
c-AVI 
CC\r‘ 
S k‘rANON
1  • 
M. ,  
\AMMNI C 
4 13
AtM 
1 0 %.• 
\\VA' 
. 
\roam 
work
\CV:kra ) 
`et". 
$ m -
I
Sc‘cAS
3
 
C"-e0.-\\•1 
- 
%r 0 CI ‘Sr Zot:(4‘ 
Che• 
Ay-toAte• 
MS.°. 
W.OknlitS) 
•0  
\t-V 
10 1/41 
Akf.o)c 
sotcl 
re. r 
\zet‘s 
cx‘o \t. 
0-Vo,c‘y 
one ;en .% 
ThQine 
Or,  
tmaA 
c 
.1/4. `n\\-- , 
S%'%6S_ 
tzsi 
s___ 
$ol bcc, 3 
0:99.)\- 
I cy \ot. 
\Anats- -1AvAt 
-01 
\Nt.5 
sr-4;3Sr 
4ve,5  
sm., 
\NT 
. ‘as 
vs, 
N-0 
ty.e 
‘,4,a\A\ 
\Note 
C.0.1—t. 
Sc 
T_ 
))-orA 
cs,x‘ce, 
4ri. 
1\o\ - 
Nu,-coc-1,,c*A-t‘l 
s cntl 
\K-
Ai • 
N)60at 
\jaorYcmc 
re. 
\I -A 
•lo., 9NtftYkr 
R \c"v") ‘4"X"e
 
411k 
\l e"' 
--
en\St. 
\)t.i'ort 
1A -4,N 
\ re, 
s&)AIN 5 
Irb 
S\ -C\ q t -
A"b 
1a. 
_e 
 sa 
\:St, 
'(ay. 
\two_
Akera. 
CtcNc 
ANNA- 
we\ 
t6c.ksZt)... 
)ro 0 
I've 
\Ae4a-
Phr;c• corsC3 \ lin 
ININ\ 
Cl-t.c-t eige_ro,stto t, 
ti t S.•• \r`conS 
tk•S 
:it 
\\MI t 
t, 
tecir;; Ccok 
a teN 
'no%)
.k-vo 
etY•S 
VINC0 
;-\ 
'.'ht 
e•C.P;.nefftr5 
teems-e. 
Coa c-Se 
c 
Pic ot.tS goh:INS 
°‘A
 
•Ske.Ct;(SerCIIV\1""  
\i‘l\e‘%\C 
vti 
- 5A°‘`8 '‘ a
'AwnN-I- 
\lc) 
In& 
trAnstS. . 
st‘trt. 
-s_ 
WAs- 
\-\•;.T.grt 
cw:N\ 
.9>L4 t x`sesa.k4r). 
SCoC 
NItc\ct\to,\ 
cAvtA, tc‘cs 
ctrA 
(cm. t 
EFTA00176557

•c•••
‘.c- .1 c-4 
t.r.f.X-1 
N\e.S\C. -  
It-'1.2 se. 
csre").,. 
\ \\\ 
'move 
r 
\\,\\%. 
R—eck.\ 
\ouk 'cse 
\{b‘) 
V•PNC3M1 
\ ‘‘‘6e. 
\los).
My_
 
\.3 
k-qt. 
iss‘2. 
T 
\•••10. 
("I'r, 
t -£. ._ 
. ite.AA-; h)
Pr.
) 
\cWv\''‘ ''TsrskS 
/‘. . \:eC\-.r t(46).t_ 
C6st- ter
, cmi.
Z.
 
1-34..1.‘y ...tea 
A. 
xc..,-5 
\-e 
k 
\re\ 
. 
4 SAYx 
o. 
\A\ Nto\t. 
\o / r 
_ 
O 
\\\Lt. 
sti\Nt 
c\-.1Y\ 
b\:\ 5rse.S.S 
.1C.r.T3N4 
So \
r. 
.‘Y-+ 
0 
._‘/A 
.ecks\t/
\k\-. 
‘f.•Or 
. 
_ 
. 
. 
CNN... 
Y•-•etV 
S-\-vxv 
cNY\ 
z 
vos) 
. . 
. 
4-ez". 
Stox0A.A1 
\ 
Th\. . 
\ 'NeAs 
k*C.%)‘• 
e  
- 
.he .1T 
.14`  
`AY'At 17-
. 
•,•\()e- 
Pro. 
t) vr\lc• 
•
- 
\\ 
4ek 
\ec\SA-
tY-#... 
WAY \Y— 
yo,, 
. 
..r*.,0 
Moos 
V4\1 
\VIA \ASA-6 
\Nato. 
yo.) 
•\4,-.O1\14- . Alf•Aal 
v4tr-c. 
. 
Q‘I`\` • ;!‘' z•-
. 
Isr\s. 
. 11.0
WVer. 
(r 
kx,c 
Vtes.ce, 
‘4\;4„, 
cAN1/44 
*rt. 
tv,4„, 
-"svc`-'3 
\"\^`6` " 
%.t1r.1‘) 
VNeror 
kot• 
NN.^‘  
4 -VA'N 
5 
c 
'fry. 
y
.c
c.'
6`4),\\A.s(t ,e, 
\ f\ (We 
(1/4 
EFTA00176558

C 
C 
EFTA00176559

447 of 1456 DOCUMENTS 
Copyright 2004 Gale Group, Inc. 
ASAP 
Copyright 2004 American Association for Artificial Intelligence 
AI Magazine 
June 22, 2004 
SECTION: No. 2, Vol. 25; Pg. 113; ISSN: 0738.4602 
LAC-ACC-NO: 119024857 
LENGTH: 7274 words 
HEADLINE: The St. Thomas common sense symposium: designing architectures for human-level 
intelligence. 
BYLINE: Minsky, Marvin; Singh, Push; Stomas, Aaron 
BODY: 
To build a machine that has "common sense" was once a principal goal in the field of artificial 
intelligence. But most researchers in recent years have retreated from that ambitious aim. Instead, each 
developed some special technique that could deal with some class of problem well, but does poorly at 
almost everything else. We are convinced, however, that no one such method will ever turn out to be 
"best," and that instead, the powerful AI systems of the future will use a diverse array of resources that, 
together, will deal with a great range of problems. To build a machine that's resourceful enough to have 
humanlike common sense, we must develop ways to combine the advantages of multiple methods to 
represent knowledge, multiple ways to make inferences, and multiple ways to learn. We held a two-day 
symposium in St. Thomas, U.S. Virgin Islands, to discuss such a project--to develop new architectural 
schemes that can bridge between different strategies and representations. This article reports on the events 
and ideas developed at this meeting and subsequent thoughts by the authors on how to make progress. 
• *****•*• 
* 
The Need for Synthesis in Modern AI 
To build a machine that has "common sense was once a principal goal in the field of artificial 
intelligence. But most researchers in recent years have retreated from that ambitious aim. Instead, each 
developed some special technique that could deal with some class of problem well, but does poorly at 
almost everything else. An outsider might regard our field as a chaotic array of attempts to exploit the 
advantages of (for example) neural networks, formal logic, genetic programming, or statistical inference--
with the proponents of each method maintaining that their chosen technique will someday replace most of 
the other competitors. 
We do not mean to dismiss any particular technique. However, we are convinced that no one such 
method will ever turn out to be "best," and that instead, the powerful AI systems of the future will use a 
diverse array of resources that, together, will deal with a great range of problems. In other words, we 
should not seek a single "unified theory!" To build a machine that is resourceful enough to have humanlike 
common sense, we must develop ways to combine the advantages of multiple methods to represent 
knowledge, multiple ways to make inferences, and multiple ways to learn. 
We held a two-day symposium in St. Thomas, U.S. Virgin Islands, to discuss such a project--to 
develop new architectural schemes that can bridge between different strategies and representations. This 
article reports on the events and ideas developed at this meeting and subsequent thoughts by the authors on 
how to make progress. (1) 
Organizing the Diversity of Al Methods 
EFTA00176560

Marvin Minsky kicked off the meeting by discussing how we might begin to organize the many 
techniques that have been developed in AI so far. While AI researchers have invented many 
representations, methods, and architectures for solving many types of problems, they still have little 
understanding of the strengths and weaknesses of each these techniques. We need a theory that helps to 
map the types of probkms we face onto the types of solutions that are available to us. When should one use 
a neural network? When should one use statistical learning? When should one use logical theorem proving? 
To help answer these kinds of questions, Minsky suggested that we could organize different AI 
methods into a "causal diversity matrix" (figure I). Here, each problem-solving method, such as analogical 
reasoning, logical theorem proving, and statistical inference, is assessed in terms of its competence at 
dealing with problem domains with different causal structures. 
[FIGURE I OMITTED] 
Statistical inference is often useful for situations that are affected by many different matched causal 
components, but where each contributes only slightly to the final phenomenon. A good example of such a 
problem-type is visual texture classification, such as determining whether a region in an image is a patch of 
skin or a fragment of a cloud. This can be done by summing the contributions of many small pieces of 
evidence such as the individual pixels of the texture. No one pixel is terribly important, but en masse they 
determine the classification. Formal logic, on the other hand, works well on problems where there are 
relatively few causal components, but which are arranged in intricate structures sensitive to the slightest 
disturbance or inconsistency. An example of such a problem-type is verifying the correctness of a computer 
program, whose behavior can be changed completely by modifying a single bit of its code. Case-based and 
analogical reasoning lie between these extremes, matched to problems where there are a moderate number 
of causal components each with a modest amount of influence. Many common sense domains, such as 
human social reasoning, may fall into this category. Such problems may involve knowledge too difficult to 
formalize as a small set of logical axioms, or too difficult to acquire enough data about to train an adequate 
statistical model. 
It is true that many of these techniques have worked well outside of the regimes suggested by this 
causal diversity matrix. For example, statistical methods have found application in realms where previously 
rule•based methods were the norm, such as in the syntactic parsing of natural language text. However, we 
need a richer heuristic theory of when to apply different AI techniques, and this causal diversity matrix 
could be an initial step toward that. We need to further develop and extend such theories to include the 
entire range of Al methods that have been developed, so that we can more systematically exploit the 
advantages of particular techniques. 
How could such a "meta-theory of AI techniques" be used by an AI architecture? Before we turned to 
this question, we discussed a concrete problem domain in which we could think more clearly about the goal 
of building a machine with common sense. 
Returning to the Blocks World 
Later that first morning, Push Singh presented a possible target domain for a commonsense 
architecture project. Consider the situation of two children playing together with blocks (figure 2). 
[FIGURE 2 OMIlibD] 
Even in this simple situation, the children may have concerns that span many "mental realms": 
Physical: What if I pulled out that bottom block? 
Bodily: Can I reach that green block from here? 
Social: Should I help him with his tower or knock it down? 
Psychological: I forgot where I left the blue block. 
Visual: Is the blue block hidden behind that stack? 
Spatial: Can I arrange those blocks into the shape of a table? 
Tactile: What would it feel like to grab five blocks at once? 
EFTA00176561

Self-Reflective: I'm getting bored with this--at else is there to do? 
Singh argued that no present-day M system demonstrates such a broad range of commonsense skills. 
Any architecture we design should aim to achieve some competence within each of these and other 
important mental realms. He proposed that to do this we work within the simplest possible domain 
requiring reasoning in each of these realms. He suggested that we develop our architectures within a 
physically realistic model world resembling the classic Blocks World, but where the world was populated 
by several simulated beings, and thus emphasizing social problems in addition to physical ones. These 
beings would manipulate simple objects like blocks, balls, and cylinders, and would participate in the kinds 
of scenarios depicted in figure 3. which include jointly building structures of various kinds, competing to 
solve puzzles. teaching each other skills through examples and through conversation. and verbally 
reflecting on their own successes and failures. 
(FIGURE 3 OMITTED 
The apparent simplicity of this world is deceptive, for many of the kinds of problems that show up in 
this world have not yet been tackled in AI, for they require combining elements of the following: 
Spatial reasoning about the spatial arrangements of objects in one's environment and how the parts of 
objects are oriented and situated in relation to one another. (Which of those blocks is closest to me?) 
Physical reasoning about the dynamic behavior of physical objects with masses and 
colliding/supporting surfaces. (What would happen if I removed that middle block from the tower?) 
Bodily reasoning about the capabilities of one's physical body. (Can I reach that block without having 
to getup?) 
Visual reasoning about the world that underlies what can be seen. (Is that a cylinder-shaped block or 
part of a person's leg?) 
Psychological reasoning about the goals and beliefs oneself and of others. (What is the other person 
trying to do?) 
Social reasoning about the relationships, shared goals and histories that exist between people. (How 
can I accomplish my goal without the other person interfering?) 
Reflective reasoning about one's own recent deliberations. (What was I trying to do a moment ago?) 
Conversational reasoning about how to express one's ideas to others. (How can l explain my problem 
to the other person?) 
Educational reasoning about how to best learn about some subject, or to teach it to someone else. 
(How can I generalize useful rules about the world from experiences?) 
Many of the meeting participants were enthusiastic about this proposal and agreed that there would be 
challenging visual, spatial, and robotics problems within this domain. Ken Forbus pointed out that the 
video game communities would soon produce programmable virtual worlds that would easily meet our 
needs. Several participants mentioned the success of the RoboCup competitions (Kitano et al. 1997), but 
some concluded that the RoboCup domain, while appropriate for those interested in the problem of 
coordinating multiagent teams in a competitive scenario, was very different in character from the situation 
of two or three people more slowly working together on a physical task, communicating in natural 
language, and in general operating on a more thoughtful and reflective level. 
Still, the participants had a heated debate about the adequacy of the proposed problem domain. The 
most common criticism was that this world does not contain enough of a variety of objects or richness of 
behavior. Doug Lenat suggested a solution to this, which was to embed the people within not a Blocks 
World, but instead somewhere like a typical house or office, as in the popular computer game The Sims. 
Doug Riecken argued that we could develop enough of the architecture within the more limited virtual 
world, and later add extensions to deal with a wider range of objects and phenomena. 
A different response to this criticism was that in order to focus on architectural issues, it would help to 
simplify the problem domain, so that we could focus less on acquiring a large mass of world knowledge, 
and more on developing better ways for systems to use the knowledge they have. However, other 
EFTA00176562

participants argued that restricting the world would not entirely bypass the need for large databases of 
commonsense knowledge, for even this simple world would likely require hundreds of thousands or even 
millions of elementary pieces of commonsense knowledge about space, time, physics, bodies, social 
interactions, object appearances, and so forth. 
Other participants disagreed with the virtual world domain. They felt that we should instead take the 
more practical approach of developing the architecture by starting with a useful application like a search 
engine or conversational agent, and extending its common sense abilities over time. But Ben Kuipers 
worried that choosing too specific an application would lead to what happened to most previous projects--
someone discovers some set of ad hoc tricks that leads to adequate performance, without making any more 
general progress toward more versatile, resourceful, or "more intelligent" systems. 
In the end, after long debates we achieved a substantial consensus that to solve harder problems 
requiring common sense, we first needed to solve the more restricted class of problems that show up in 
simpler domains like the proposed virtual world. Once we get the core of the architecture functioning in 
this rich but limited domain, we can attempt to extend it—or it extend itself--to deal with a broader range of 
problems using a much broader array of commonsense knowledge. 
Large-Scale Architectures for Human-level Intelligence 
In the afternoon, we discussed large-scale architectures for machines with human-level intelligence 
and common sense. Marvin Minsky and Aaron Sloman each presented their current architectural proposals 
as a starting point for the meeting participants to criticize, debug, and elaborate. These two architectures 
share so many features that we will refer to them together as the Minsky-Sloman model. 
These architectures are distinguished by their emphasis on reflective thinking. Most cognitive models 
have focused only on ways to react or deliberate. However, to make machines more versatile, they will 
need better ways to recognize and repair the obstacles, bugs and deficiencies that result from their own 
activities. In particular, whenever one strategy fails, they'll need to have a collection of ways to switch to 
alternative ways to think. To provide for this, Minsky's architectural design includes several reflective 
levels beyond the reactive and deliberative levels. Here is one view of his model for the architecture of a 
person's mind, as described in his book, The Emotion Machine, and shown here in figure 4. 
(FIGURE 4 OMITTED] 
Some participants questioned the need for so many reflective layers; would not a single one be 
enough? Minsky responded by arguing that today, when our theories still explain too little, we should 
elaborate rather than simplify, and we should be building theories with more parts, not fewer. This general 
philosophy pervades his architectural design, with its many layers, representations, critics, reasoning 
methods, and other diverse types of components. Only once we have built an architecture rich enough to 
explain most of what people can do will it make sense to try to simplify things. But today, we are still far 
from an architectural design that explains even a tiny fraction of human cognition. 
Aaron Sloman's Cognition and Affect project has explored a space of architectures proposed as 
models for human minds; a sketch of Sloman's H-CogAff model is shown in figure 5. 
(FIGURE 5 OMITTED) 
This architecture appears to provide a framework for defining with greater precision than previously a 
host of mental concepts, including affective concepts, such as "emotion," "attitude," "mood," "pleasure," 
and so on. For instance, H-CogAff allows us to define at least three distinct varieties of emotions; primary, 
secondary and tertiary emotions, involving different layers of the architecture which evolved at different 
times--and the same architecture can also distinguish different forms of learning, perception, and control of 
behavior. (A different architecture might be better for exploring analogous states of insects, reptiles, or 
other mammals.) Human infants probably have a much-reduced version of the architecture that includes 
self-bootstrapping mechanisms that lead to the adult form. 
The central idea behind the Minsky-Sloman architectures is that the source of human resourcefulness 
and robustness is the diversity of our cognitive processes: we have many ways to solve every kind of 
problem--both in the world and in the mind--so that when we get stuck using one method of solution, we 
EFTA00176563

can rapidly switch to another. There is no single underlying knowledge representation scheme or 
inferencing mechanism. 
How do such architectures support such diversity? In the case of Minsky's Emotion Machine 
architecture, the top level is organized as follows. When the system encounters a problem, it first uses some 
knowledge about "problem-types" to select some "way-to-think" that might work. Minsky describes "ways-
to-think" as configurations of agents within the mind that dispose it towards using certain styles of 
representation. collections of commonsense knowledge, strategies for reasoning, types of goals and 
preferences, memories of past experiences, manners of reflections, and all the other aspects that go into a 
particular "cognitive style." One source of knowledge relating problem-types to ways-to-think is the causal 
diversity matrix discussed at the start of the meeting--for example, if the system were presented with a 
social problem, it might use the causal diversity matrix to then select a case-bascd style of reasoning, and a 
particular database of social reasoning episodes to use with it. 
However, any particular such approach is likely to fail in various ways. Then if certain "critic" agents 
notice specific ways in which that approach has failed, they either suggest strategies to adapt that approach, 
or suggest alternative ways-to-think, as suggested shown in figure 6. This is not done by employing any 
simple strategy for reflection and repair, but rather by using large arrays of higher level knowledge about 
where each way-to-think has advantages and disadvantages, and how to adapt them to new contexts. 
[FIGURE 6 OMITTED] 
In Minsky's design, several ways-to-think are usually active in parallel. This enables the system to 
quickly and fluently switch between different ways-to-think because, instead of starting over at each 
transition, each newly activated way-to-think will find an already-prepared representation. The system will 
rarely "get stuck" because those alternative ways-to-think will be ready to take over when the present one 
runs into trouble, as shown in figure 7. 
[FIGURE 7 OMITTED] 
Here each way-to-think involves reasoning in a particular subset of mental realms. Impasses 
encountered while reasoning in one set of mental realms can be overcome within others. Further 
information about these architectures can be found in Singh and Minsky (2003), Sloman (2001), and 
McCarthy et al. (2002). Minsky's model will be described in detail in his new book The Emotion Machine 
(Minsky, forthcoming). 
Generally, the participants were sympathetic to these proposals, and all agreed with the idea that to 
achieve human-level intelligence we needed to develop more effective ways to combine multiple AI 
techniques. Ken Forbus suggested that we needed a kind of "component marketplace," and that we should 
find ways to instrument these components so that the reflective layers of the architecture had useful 
information available to them. He contrasted the Soar project (Laird, Newell, and Rosenbloom 1987) as an 
effort to eliminate and unify components rather than to accumulate and diversify them, as in the Minsky-
Sloman proposals. Ashwin Ram and Larry Birnbaum both pointed out that despite the agreement over the 
architectural proposals it was still not clear what the particular components of the architecture would be. 
They pointed out that we needed to think more about what the units of reasoning would be. In other words, 
we needed to come up with a good list of way-to-think. Some examples might include the following: 
Solving problems by making analogies to past experiences 
Predicting what will happen next by rule-based mental simulations 
Constructing new "ways to think" by building new collections of agents 
Explaining unexpected events by diagnosing causal graphs 
Learning from problem-solving episodes by debugging semantic networks 
Inferring the state of other minds by re-using self-models 
Classifying types of situations using statistical inference 
Getting unstuck by reformulating the problem situation 
This list could be extended to include all available AI techniques. 
EFTA00176564

Educating the Architecture 
On the morning of the second day of the meeting, we addressed the problem of how to supply the 
architecture with a broad range of commonsense knowledge, so that it would not have to "start from 
scratch." We all agreed that learning was of value, but we didn't all agree on where to start. Many 
researchers would like to start with nothing; however, Aaron Sloman pointed out that an architecture that 
comes with no knowledge is like a programming language that comes with no programs or libraries. 
One view that was expressed was that approaches that start out with too little initial knowledge would 
likely not achieve enough versatility in any practical length of time. Minsky criticized the increasing 
popularity of the concept of a "baby machine"--leaning systems designed to achieve great competence, 
given very little initial structure. Some of these ideas include genetic programming, robots that learn by 
associating sensory-motor patterns, and online chatbots that try to learn language by generalizing from 
thousands of conversations. Nlinsky's complaint was that the problem is not that the concept of a baby 
machine is itself unsound, but rather that we don't know how to do it yet. Such approaches have all failed to 
make much progress because they started out with inadequate schemes for learning new things. You cannot 
teach algebra to a cat; among other things, human infants are already equipped with architectural features to 
equip them to think about the causes of their successes and failures and then to make appropriate changes. 
Today we do not yet have enough ideas about how to represent, organize, and use much of commonsense 
knowledge, let alone build a machine that could learn all of that automatically on its own. As John 
McCarthy noted long ago: "in order for a program to be capable of learning something, it must first be able 
to represent that knowledge." 
There arc very few general-purpose commonsense knowledge resources in the Al community. Doug 
Lenat gave a wonderful presentation of the Cyc system, which is presently the project furthest along at 
developing a useful and reusable such resource for the AI community, so that new Al programs don't have 
to start with almost nothing. The Cyc project (Lenat 1995) has developed a great many ways to represent 
commonsense knowledge, and has built a database of over a million commonsense facts and rules. 
However, Lenat estimated that an adult-level commonsense system might require 100 million units of 
commonsense knowledge, and so one of their current directions is to move to a distributed knowledge 
acquisition approach, where it is hoped that eventually thousands of volunteer teachers around the world 
will work together teach Cyc new commonsense knowledge. Lenat spent some time describing the 
development of friendly interfaces to Cyc that allow nonlogicians to participate in the complicated teaching 
and debugging processes involved in building up the Cyc knowledge base. 
Many of the participants agreed that Cyc would be useful, and some suggested we could even base our 
effort on top of it, but others were sharply critical. Jeffrey Siskind doubted that Cyc contained the spatial 
and perceptual knowledge needed to do important kinds of visual scene interpretation. Roger Schenk 
argued that Cyc's axiomatic approach was unsuitable for making the kinds of generalizations and analogies 
that a more case-based and narrative-oriented approach would support. Srini Narayanan worried that the 
Cyc project was not adequately based on what cognitive scientists have learned about how people make 
commonsense inferences. Oliver Steele concluded that while we disagreed about whether Cyc was 90% of 
the solution or only 10%, this was really an empirical question that we would answer during the course of 
the project. But generally, the architectural proposal was regarded as complementary to parallel efforts to 
accumulate substantial commonsense knowledge bases. 
Minsky predicted that if we used Cyc, we might need to augment each existing item of knowledge 
with additional kinds of procedural and heuristic knowledge, such as descriptions of (1) problems that this 
knowledge item could help solve; (2) ways of thinking that it could participate in; (3) known arguments for 
and against using it; and (4) ways to adapt it to new contexts. 
It was stressed that knowledge about the world was not enough by itself--we also need a knowledge 
base about how to reason, reflect and learn, the knowledge that the reflective layers of the architecture must 
possess. The problem remains that the programs we have for using knowledge are not flexible enough, and 
neither Cyc's "adult machine" approach of supplying a great deal of world knowledge, nor the "baby 
machine" approach of learning common sense from raw sensory-motor experience, will likely succeed 
without first developing an architecture that supports multiple ways to reason, learn, and reflect upon and 
improve its activities. 
EFTA00176565

An Important Application 
Several of the participants felt that such a project would not receive substantial support unless it 
proposed an application that clearly would benefit much of the world. Not just an improvement to 
something existing, it would need to be one that could not be built without being capable of human-level 
commonsense reasoning. 
After a good deal of argument, several participants converged upon a vision from The Diamond Age, 
a novel by Nell Stephenson. That novel envisioned an "intelligent book"--The Young Ladies Illustrated 
Primer--that, when given to a young girl, would immediately bond with her and come to understand her so 
well as to become a powerful personal tutor and mentor. 
This suggested that we could try to build a personalized teaching machine that would adapt itself to 
someone's particular circumstances, difficulties, and needs. The system would carry out a conversation with 
you, to help you understand a problem or achieve some goal. You could discuss with it such subjects as 
how to choose a house or car, how to learn to play a game or get better at some subject, how to decide 
whether to go to the doctor, and so forth. It would help you by telling you what to read, stepping you 
through solutions, and teaching you about the subject in other ways it found to be effective for you. 
Textbooks then could be replaced by systems that know how to explain ideas to you in particular, because 
they would know your background, your skills, and how you best learn. 
This kind of application could form the basis for a completely new way to interact with computers, 
one that bypasses the complexities and limitations of current operating systems. It would use common 
sense in many different ways: (1) It would understand human goals so that it could avoid the silliest 
mistakes. (2) It would understand human reasoning so that it could present you with the right level of detail 
and avoid saying things that you probably inferred. (3) It would converse in natural language so that you 
could easily talk to it about complex matters without having to learn a special language or complex 
interface. 
To build such a kind of "helping machine." we would first need to give it knowledge about space. 
time, beliefs, plans, stories, mistakes, successes, relationships, and so forth, as well as good conversational 
skills. However, little of this could be realized by anything less than a system with common sense. To 
accomplish this we would need to pursue some sequence of more modest goals that would help one with 
simpler problem types--until the system achieved the sorts of competence that we expect from a typical 
human four- or five-year-old. 
However, to get such a system to work, we would need to address many presently unsolved 
commonsense problems that show up in the model-world problem domain. 
Final Consensus 
The participants agreed that no single technique (such as statistics, logic, or neural networks) could 
cope with a sufficiently wide range of problem-types. To achieve human-level intelligence we must create 
an architecture that can support many different ways to represent, acquire, and apply many kinds of 
commonsense knowledge. 
Most participants agreed that we should combine our efforts to develop a model world that supports 
simplified versions of everyday physical, social, and psychological problems. This simplified world would 
then be used to develop and debug the core components of the architecture. Later, we can expand it to solve 
more difficult and more practical problems. 
The participants did not all agree on which particular larger-scale application would both attract 
sufficient support and also produce substantial progress toward making machines that use commonsense 
knowledge. Still, many agreed with the concept of a personalized teaching machine that would come to 
understand you so well that it could adapt to your particular circumstances, difficulties, and needs. 
Ben Kuipers sketched the diagram shown in figure 8, which captures the general dependencies 
between the three points of consensus: Practical applications depend on developing an architecture for 
commonsense thinking flexible enough to integrate a wide array of processes and representations of 
problems that come up in the model-world problem domain. 
(FIGURE 8 OMITTED] 
EFTA00176566