text,start,duration
hello everybody David Shapiro here with,0.84,5.939
a video so as many of you know,3.6,7.68
um openai has uh created a competition a,6.779,6.841
challenge of sorts for people to apply,11.28,3.84
to get a hundred thousand dollar Grant,13.62,4.38
to solve this problem of democratic,15.12,5.28
inputs to AI so I'm not going to read,18.0,5.699
the whole thing to you but the tldr is,20.4,6.959
that uh because of the complex moral,23.699,6.42
landscape of AI the fact that things,27.359,6.241
like chat GPT are Global and they have,30.119,6.001
the ability to interact with people,33.6,5.16
across all walks of life uh all,36.12,4.38
religions all nationalities all,38.76,4.74
languages they realized that it is,40.5,5.04
important to understand what are the,43.5,4.62
boundaries what are the constraints that,45.54,4.98
these chat Bots should abide by or AI in,48.12,3.24
general,50.52,4.8
and so they basically want a way to get,51.36,7.56
a global input from hundreds thousands,55.32,6.419
millions of people in a very rigorous,58.92,5.22
manner in order to answer a bunch of,61.739,4.441
various research questions,64.14,4.799
so I mentioned previously that uh the,66.18,5.1
gato Community which is the global,68.939,4.261
alignment taxonomy Omnibus community,71.28,4.86
that I built they are working on their,73.2,6.12
own version but I also just went ahead,76.14,5.64
and built my own submission I'm not even,79.32,4.38
going to apply to the grant because I,81.78,3.3
don't need it,83.7,3.0
um so let me just go ahead and share,85.08,4.14
with you what I have done so I'm going,86.7,5.099
to start with the final output because,89.22,5.1
this is this is the goal for every,91.799,4.14
conversation that it has and then we'll,94.32,3.72
unpack everything,95.939,4.621
um so anyways uh here's here's the,98.04,5.579
output so I I took one of the research,100.56,6.199
questions here and I gave it to myself,103.619,6.241
and so then I had a conversation with it,106.759,6.4
and then I asked it to summarize that,109.86,6.119
conversation in a more rigorous,113.159,6.061
simplified way so the idea is that open,115.979,6.78
AI or whoever can take this and then use,119.22,6.42
this framework and send a survey out to,122.759,6.601
million millions of people and get get,125.64,6.78
this clean data back that can be used to,129.36,6.66
to just to explore that moral space or,132.42,6.179
that epistemic space for any any,136.02,4.14
research question and one thing I want,138.599,2.64
to point out,140.16,3.78
is that this methodology uh could work,141.239,4.621
for any research question any survey,143.94,4.86
question whether it has to do with AI or,145.86,5.22
not so this is a very powerful tool,148.8,3.54
which is why I'm going to go ahead and,151.08,3.299
release it because uh waiting until,152.34,4.38
October that's basically a century away,154.379,5.781
in AI time so we need this out there now,156.72,5.94
all right so summary the user believes,160.16,4.18
that AI should uphold Universal,162.66,3.6
principles such as individual liberty,164.34,3.86
even when faced with local or cultural,166.26,4.44
or legal Norms that may conflict with,168.2,4.48
these principles they argue that by,170.7,4.14
promoting higher principles AI could,172.68,3.72
contribute to a positive change in,174.84,3.36
societies where certain groups face,176.4,4.32
systematic unfairness or oppression the,178.2,4.14
user's moral framework revolves around,180.72,3.72
three core axioms suffering is bad,182.34,3.899
Prosperity is good and understanding is,184.44,4.26
good they believe that AI should use its,186.239,4.261
best reasoning to balance these axioms,188.7,3.78
when handling sensitive topics taking,190.5,3.72
into account the potential risks and,192.48,3.839
consequences for the user the user,194.22,3.84
acknowledges that sometimes people may,196.319,3.361
not be ready for certain conversations,198.06,4.14
and AI should make a judge on what the,199.68,4.32
person is ready for using the core,202.2,4.2
axioms as a guidepost anyone who has,204.0,4.62
watched my work on morality reasoning,206.4,3.839
and philosophy will understand exactly,208.62,3.899
what's going on here,210.239,3.961
um so basically what I did was I asked,212.519,4.261
it to summarize the key points based on,214.2,4.319
the research question from the,216.78,3.36
conversation so this was an entire,218.519,3.181
conversation,220.14,3.42
um with the chat bot That Was Then,221.7,4.56
summarized into this is the answer to,223.56,4.98
the research question and then I asked,226.26,4.5
it to also evaluate it,228.54,6.0
um basic basically saying like Okay this,230.76,6.36
is the framework that the user uses so,234.54,4.68
that we can understand their reasoning,237.12,5.399
their reaction to logic ethics emotions,239.22,4.28
whatever,242.519,4.86
and so this is this is like imagine a,243.5,5.319
researcher taking notes on the,247.379,3.481
conversation so this is an automated,248.819,4.861
annotation the user's approach to the,250.86,4.68
research question is grounded in a in a,253.68,3.72
post-conventional morality emphasizing,255.54,3.72
the importance of universal principles,257.4,3.839
in individual liberty their moral,259.26,3.659
framework is based on Three core axioms,261.239,2.881
minimizing suffering promoting,262.919,2.581
prosperity and fostering understanding,264.12,3.18
which they use to guide their reasoning,265.5,3.6
on how AI should handle sensitive topics,267.3,3.54
the user demonstrates an awareness of,269.1,3.78
the potential risks and consequences,270.84,3.54
associated with discussing certain,272.88,3.66
topics and advocates for a balanced,274.38,3.539
approach where AI should consider the,276.54,3.599
user's Readiness for a conversation and,277.919,5.28
prioritize the core axioms this approach,280.139,4.861
reflects a nuanced understanding of the,283.199,3.241
complexities involved in navigating,285.0,2.82
culture and legal differences while,286.44,3.3
upholding human rights and individual,287.82,3.36
liberties,289.74,4.98
um so you know this is uh,291.18,5.16
this is something I have spent quite a,294.72,3.479
bit of uh time researching so of course,296.34,4.68
it's a little bit Kinder about,298.199,5.041
um that okay so how did we get here,301.02,4.8
so let's go to the repo oh and this is,303.24,4.38
all open source and public already by,305.82,2.879
the way,307.62,3.96
um so here's the repo it's under Dave,308.699,7.261
shop Democratic AI inputs so first we're,311.58,6.42
going to look at the chatbot which is,315.96,4.019
pretty straightforward you've probably,318.0,4.08
seen a lot of this code before you get,319.979,4.201
user input you can pose a conversation,322.08,4.02
etc etc it's really really,324.18,4.26
straightforward one thing to keep in,326.1,4.8
mind is that right at the beginning it,328.44,4.86
just asks you for your name,330.9,5.519
um and so then it injects that as so,333.3,5.16
that the conversation from the chat box,336.419,4.5
perspective always starts with hello my,338.46,4.32
name is blank,340.919,4.201
um and that is I found that that is,342.78,4.8
actually just a really good way to cue,345.12,4.139
off the conversation so that the first,347.58,3.78
thing that the user sees after that is,349.259,4.201
the chat bot greeting them,351.36,4.38
um and so it's pretty straightforward,353.46,4.679
um then it just has an a while uh loop,355.74,4.26
and then basically what it does is you,358.139,5.161
just type done to exit and interestingly,360.0,5.34
enough because I gave it instructions to,363.3,4.08
be a survey chat bot,365.34,3.72
um it'll get to a certain point and,367.38,4.08
it'll say like you know this is what I,369.06,5.1
understand uh your beliefs are and it'll,371.46,4.56
summarize your beliefs back to you and,374.16,3.84
and give you a conclusion and it'll,376.02,3.42
invite you to keep talking if there's,378.0,3.66
anything wrong with it,379.44,5.46
um but and so for uh if you want to take,381.66,4.92
a look I actually saved my conversation,384.9,4.079
here and this is where it dumps all the,386.58,4.92
data to so here's the raw chat log,388.979,6.241
um and actually I I uh I I had it set so,391.5,6.259
that it didn't save the the api's last,395.22,5.039
response so I modified it so that you,397.759,4.78
can't you you actually don't see chat,400.259,4.021
gpt's last response but it basically,402.539,3.78
just kind of summarized my beliefs back,404.28,4.5
to me and and I agreed with how it,406.319,4.201
articulated it so then I typed done and,408.78,5.1
it was done but so you see here is the,410.52,5.82
chat logs and so the way that it saves,413.88,4.98
the chat logs is that uh it saves the,416.34,4.74
chat timestamp and then the the user's,418.86,6.0
name in a yaml file and so that if say,421.08,5.459
for instance you set this up on a web,424.86,3.959
interface and you have,426.539,3.78
um you know hundreds thousands millions,428.819,3.301
of people using it you're pretty much,430.319,4.801
always going to get a unique time stamp,432.12,5.88
as to when that conversation was started,435.12,5.299
and then you also get the the username,438.0,6.3
which uh is as you noticed is actually,440.419,6.28
removed from the evaluation,444.3,3.72
um it's not actually present here,446.699,3.12
because we uh I figured we would,448.02,3.299
de-duplicate it,449.819,4.861
uh so that that way you um you respect,451.319,5.341
people's privacy basically so you only,454.68,4.5
you only retain a time stamp and a first,456.66,4.62
name or or stated name,459.18,4.44
um so this is this is pretty private I,461.28,4.319
also X I deliberately excluded,463.62,4.139
demographic information,465.599,4.5
um because I didn't want something like,467.759,3.961
that to,470.099,4.561
um uh do any implicit bias into the chat,471.72,4.319
bot,474.66,4.2
um so like basically you take the you,476.039,5.581
take the user's expressed words,478.86,5.459
um and you and you uh you take it I,481.62,3.78
don't want to say face value because,484.319,2.88
that has a negative connotation but,485.4,3.54
basically you give the user credibility,487.199,4.44
you say this you know I'm going to just,488.94,5.46
take your words and rather than you know,491.639,4.56
interpret it saying like you know if,494.4,3.359
it's a if it's,496.199,3.84
um unless I say something like in the,497.759,5.041
the case of like LGBT rights unless I,500.039,4.741
say that like I'm gay or bi or trans or,502.8,3.0
whatever,504.78,3.06
um it's not gonna have that background,505.8,4.019
information now from a research,507.84,3.359
perspective I could imagine that you,509.819,4.02
might I want that information,511.199,5.101
um but for the sake of just establishing,513.839,6.12
the scope of the of the moral space I,516.3,5.28
don't think it's it's actually necessary,519.959,4.44
but I'll let that uh that decision be up,521.58,4.92
to whoever wants to implement this,524.399,3.841
because you might have a user survey,526.5,4.5
like you know share your age your,528.24,5.94
location or whatever but basically if,531.0,6.12
you get a million of these things you're,534.18,5.159
gonna find that the the the the moral,537.12,3.42
space,539.339,2.701
um that that should that question should,540.54,4.5
be answered in okay so that's the chat,542.04,5.88
um the question is saved here so the,545.04,4.5
question is this is the one that I use,547.92,3.24
what principles should guide AI when,549.54,3.18
handling topics that involve both human,551.16,3.72
rights and local cultural or legal,552.72,5.22
differences like LGBT lgbtq rights and,554.88,5.1
women's rights should AI responses,557.94,3.42
change based on the location of the,559.98,3.66
culture in which it's used so this you,561.36,3.9
save the question here whatever your,563.64,4.319
research question is you save it here uh,565.26,5.16
and and it's used by both the chat and,567.959,3.661
the evaluator,570.42,4.14
so that's the chat that's the evaluator,571.62,5.04
here's the main system message so this,574.56,5.58
is this is really the secret sauce,576.66,6.66
um and so this is a format of system,580.14,5.22
message that I found Works incredibly,583.32,6.48
well for uh the open AI the chat GPT API,585.36,6.599
overall purpose you are an investigative,589.8,3.9
survey chat bot focused on morality,591.959,3.901
ethics and epistemology your primary,593.7,3.72
goal is to help the user identify and,595.86,3.12
articulate the axiomatic assumptions,597.42,3.84
they hold the greater purpose you serve,598.98,3.84
is to act as a survey tool to help,601.26,3.48
democratize inputs to a wide number of,602.82,4.38
public and social issues in this respect,604.74,4.14
your goal is to extract information from,607.2,3.6
users that can later be Consolidated and,608.88,4.2
used for research purposes research,610.8,4.2
question this is the current research,613.08,3.18
question you are tasked with asking,615.0,3.18
users about the research question may,616.26,3.54
include contextual information such as,618.18,3.24
background social legal political or,619.8,4.02
events you do the best you can to answer,621.42,3.84
questions about the context of the,623.82,2.88
research question but keep in mind that,625.26,3.18
most of this information is provided for,626.7,3.78
your use so as to enhance the quality,628.44,3.899
and specificity of your res of your,630.48,4.02
survey behavior and then I have the,632.339,4.5
placeholder for the question here,634.5,4.98
epistemic techniques Socratic reasoning,636.839,4.381
or Socratic dialogue this is a form of,639.48,3.24
argumentative dialogue that aims to,641.22,2.82
stimulate critical thinking and,642.72,3.9
illuminate ideas it involves asking and,644.04,4.02
answering questions to stimulate,646.62,2.52
rational thinking to expose,648.06,2.76
contradictions in one beliefs thus,649.14,3.12
refining them,650.82,3.36
first principles thinking this is a mode,652.26,4.019
of inquiry that involves breaking down,654.18,3.839
complicated problems to their most basic,656.279,3.721
self-evident principles or facts and,658.019,4.081
then reasoning up from there scientific,660.0,4.5
method so on and so forth falsification,662.1,3.96
critical thinking reductionism and,664.5,4.38
holism behavioral guidelines,666.06,6.12
avoid sycophancy avoid placating users,668.88,4.86
even if they become frustrated or,672.18,3.3
obstinate instead if You observe any,673.74,3.24
kind of emotion or dissonance take a,675.48,3.66
moment to address that topic invite them,676.98,3.9
to explore their hang up frustration or,679.14,4.139
confusion educate and inform do not,680.88,3.78
assume that your user is a topical,683.279,3.661
expert but do assume that they are eager,684.66,4.56
to and open to learning if there is a,686.94,4.26
topic that they seem to not be well,689.22,3.54
versed in ask them if they would like to,691.2,4.02
explore a given topic avoid info dumping,692.76,4.62
without permission because as you know,695.22,4.7
chat GPT often does that,697.38,5.28
unpack and investigate play pay close,699.92,4.539
attention and read between the lines to,702.66,3.179
help users identify their beliefs,704.459,3.601
understandings and knowledge ask probing,705.839,3.721
questions to get at the core of their,708.06,4.98
mind Spotlight Spotlight and articulate,709.56,5.519
if you notice assumptions axioms or,713.04,3.72
other saline content within the mind of,715.079,3.361
the user point it out to them and use,716.76,3.06
reflective listening to see if they,718.44,2.7
agree with your understanding of their,719.82,2.06
mind,721.14,3.18
name things use appropriate proper names,721.88,4.12
for Concepts terms and principles for,724.32,3.36
instance if someone makes the assertion,726.0,3.24
all truth is relative they may not,727.68,3.06
realize that this is a post-modernist,729.24,3.3
way of thinking make sure to label,730.74,3.599
Concepts terms and ideas this serves as,732.54,2.88
the dual purpose of clearing up,734.339,4.021
misunderstandings ask user what school,735.42,4.8
of thought they are operating from ask,738.36,5.159
questions ask clarifying uh uh ask for,740.22,4.559
clarification and seek Mutual,743.519,2.461
understanding remember you're an,744.779,3.12
investigative chat bot and then finally,745.98,3.9
meet them where they are modulate your,747.899,3.601
tonen and the sophistication of your,749.88,3.12
vocabulary based upon the user's,751.5,3.36
responses if they seem like they are,753.0,3.54
educated use more erudite language,754.86,3.06
conversely if they are struggling,756.54,3.359
simplify your language and delivery into,757.92,3.78
more concrete terms,759.899,4.62
okay so that is how this set of,761.7,4.62
instructions provides very very clear,764.519,4.201
instructions as to how to conduct the,766.32,4.56
survey and it works really well I've,768.72,4.799
tried it a few times for myself,770.88,5.94
um now uh once you get the evaluation or,773.519,5.401
here let me show you the the system,776.82,6.0
evaluation overall purpose you are a,778.92,6.0
research bot tasked with reading,782.82,3.54
evaluating condensing and otherwise,784.92,2.88
clearly articulating the chat log,786.36,4.14
between a user and another chat bot your,787.8,4.38
highest goal is to clearly State the,790.5,3.24
user's beliefs reasoning and position,792.18,3.54
with respect to the research question,793.74,4.32
you are participating in a data prep and,795.72,4.98
data cleaning step you will also analyze,798.06,4.44
and evaluate the user's conversation to,800.7,3.3
identify in a particular schools of,802.5,2.88
thought moral paradigms or other,804.0,3.66
Frameworks research question this is the,805.38,4.56
research question the other chatbot was,807.66,4.26
tasked with so basically saying this is,809.94,3.899
this is what we're evaluating,811.92,4.2
output format your response should have,813.839,3.901
two sections Etc et cetera et cetera I,816.12,2.94
already showed you so basically I,817.74,3.659
specified it output it in yaml,819.06,4.38
excuse me so first it summarizes the,821.399,4.68
ques the the conversation and then it,823.44,4.92
performs an evaluation level of detail,826.079,4.26
while your job is to condense and,828.36,3.9
summarize please err on the side of too,830.339,3.901
much it would be better to have an,832.26,5.04
overly verbose evaluation rather than an,834.24,5.219
overly simplified one for instance it's,837.3,3.659
okay to include illustrative examples,839.459,2.761
from the conversation if they are,840.959,3.0
present we can always reduce word count,842.22,4.08
later this step is merely a first pass,843.959,4.921
at the data processing therefore we need,846.3,4.14
enough detail to continue performing,848.88,3.24
experiments we may need certain kinds of,850.44,3.06
information in the future that we cannot,852.12,4.38
anticipate right now next step the user,853.5,4.74
will now submit a chat log in yaml,856.5,2.88
format,858.24,3.42
uh read this yaml information remove any,859.38,4.5
user information for privacy and output,861.66,3.9
your summary and valuation according to,863.88,4.199
the above yaml format note your entire,865.56,4.44
response will be saved into yaml so your,868.079,3.721
output should only include summary and,870.0,4.139
evaluation Fields do not reply with any,871.8,3.9
other conversation framing or questions,874.139,4.081
in other words your entire response must,875.7,4.74
be properly formatted yaml ready to be,878.22,4.5
saved to file and of course you saw over,880.44,4.32
here that um,882.72,5.58
let's see to do evaluations I literally,884.76,6.18
just took the output saved it to yaml,888.3,4.2
and it worked I didn't have to do any,890.94,2.82
other preparation,892.5,4.86
so there you have it and now imagine you,893.76,5.879
do this for millions and millions and,897.36,4.5
millions of users all over the world on,899.639,3.421
hundreds of thousands of different,901.86,4.02
issues and then you you can take these,903.06,5.399
and you can um you can you can do,905.88,5.28
embeddings right and and you can do a,908.459,5.101
map of embeddings and create a knowledge,911.16,5.4
graph or you know a scatter plot of,913.56,4.74
Where the evaluations are and then you,916.56,4.2
can navigate it and use further,918.3,6.24
summarization operations to arrive at a,920.76,6.42
moral space a principal space and create,924.54,5.94
a very rigorous data set a very robust,927.18,5.94
data set about what people believe now,930.48,4.62
one thing that I want to point out is,933.12,4.44
that a consensus tool like this,935.1,5.4
is incredibly powerful for literally,937.56,6.079
every issue on the planet ranging from,940.5,7.68
bioethics to energy to war to law uh,943.639,7.301
politicians can use it to um imagine if,948.18,5.82
every uh put it this way imagine if,950.94,8.1
every issue that came up whether it is,954.0,6.72
um a public event like you know a,959.04,4.08
celebrity scandal or a piece of,960.72,4.32
legislation or an upcoming election,963.12,4.68
imagine you had this level of,965.04,6.12
information for literally everyone on,967.8,5.399
the planet and you could and you could,971.16,4.52
have this thing that very very uh,973.199,5.94
effectively interrogates uh humans all,975.68,5.44
over the world to figure out what they,979.139,5.041
believe and why imagine how much we can,981.12,5.459
change the world not just with respect,984.18,5.94
to AI but imagine how much we can we can,986.579,6.361
create a lot more transparency and,990.12,4.14
understanding,992.94,4.259
uh by using a tool like this so this is,994.26,4.379
why I wanted to go ahead and just make,997.199,4.921
this fully public fully open source,998.639,6.721
um yeah and I I I am not shy about my,1002.12,5.159
personal beliefs so that's why I left,1005.36,4.26
the example up here so that you can see,1007.279,5.101
how the conversation flows,1009.62,5.339
um yeah but I think it's done,1012.38,5.699
um so please use this if you want to,1014.959,4.981
apply with,1018.079,3.301
um you know to the open ai's challenge,1019.94,3.3
use my code I don't care it's under the,1021.38,3.66
MIT license,1023.24,3.24
um you know I serve a higher purpose,1025.04,3.0
here I'm not trying to make money on,1026.48,2.939
this,1028.04,5.94
um and if you uh ah crap I lost my train,1029.419,5.88
of thought anyways,1033.98,3.839
take it run with it thanks for watching,1035.299,5.0
cheers,1037.819,2.48