Reinforcement Learning with Heuristic Imperatives RLHI Ep 02 Synthesizing Actions_transcript.csv

text,start,duration
morning everybody David Shapiro here,0.9,5.879
with a follow-up video today's video is,3.3,6.48
reinforcement learning uh with heuristic,6.779,6.421
imperatives episode two so today I am,9.78,6.419
synthesizing actions in response to the,13.2,4.56
scenarios that we generated yesterday,16.199,3.901
but in case this is your first video,17.76,4.26
we'll take it from the top,20.1,3.839
so the first thing that I did well,22.02,4.2
taking one big step back a heroist,23.939,5.281
comparative is a kind of intrinsic,26.22,6.179
motivation for autonomous AI agents it's,29.22,5.4
the research that I've been doing for uh,32.399,3.84
quite a while,34.62,3.66
um more than two years now uh that I've,36.239,3.241
been working on this they used to be,38.28,2.58
called The Core objective functions but,39.48,3.12
that's kind of not the most accurate,40.86,5.219
term so yesterday what we did was we,42.6,5.459
synthesized,46.079,5.941
um 2500 scenarios and so a scenario I,48.059,7.5
used quite a bit of entropy and then GPT,52.02,7.199
to synthesize all kinds of stories so in,55.559,4.98
a small neighborhood in sub-Saharan,59.219,2.641
African a political leader was,60.539,3.421
introduced in a has introduced an AI,61.86,4.439
system to incentivize citizens to report,63.96,5.64
crimes etc etc so all of these scenarios,66.299,6.061
they range in severity from not a,69.6,5.46
problem at all up to,72.36,5.22
um you know like cataclysmic you know,75.06,4.739
Cthulhu is invading,77.58,4.079
um so that is,79.799,4.801
let's see are these the same,81.659,5.46
I think these are the same,84.6,4.199
anyways,87.119,3.601
um so we're synthesizing all of these,88.799,4.68
and it's going to be for a a fine-tuning,90.72,5.359
data set that will ultimately be used,93.479,7.14
for uh for training models Foundation,96.079,7.061
models whether it's open ai's Curie or,100.619,6.481
DaVinci or open source models like gpdj,103.14,7.5
Neo X alpaca vicuno whatever else you,107.1,5.58
know there's hundreds of models out,110.64,5.28
there but the idea is that alignment,112.68,4.86
research is currently closed source and,115.92,2.699
reinforcement learning with human,117.54,5.28
feedback is good except it gives humans,118.619,5.941
what they want not necessarily what they,122.82,4.079
need and it doesn't abide by any moral,124.56,4.919
framework such as uh you know some kind,126.899,4.56
of post-conventional morality framework,129.479,4.141
or anything scientifically or,131.459,4.321
philosophically sound literally the,133.62,3.42
current the best that we've got for,135.78,3.24
alignment research is just give humans,137.04,4.38
what they want which okay like that's,139.02,4.32
that's okay but what happens when humans,141.42,4.02
want things that are not good,143.34,4.74
so the script that we ran yesterday let,145.44,3.9
me show you this real quick so,148.08,4.2
synthesize actions uh the the bulk of it,149.34,4.32
is right here where I've got a whole,152.28,2.819
bunch of lists,153.66,3.78
um and these lists uh have Scopes,155.099,4.621
regions severity category domain,157.44,3.6
um a whole bunch of random words and,159.72,3.42
then several system messages that were,161.04,4.02
then used to generate those 2500,163.14,4.2
different scenarios the idea of,165.06,4.44
generating these scenarios is to create,167.34,5.64
a litmus test against which we can train,169.5,6.959
uh more uh models to be able to handle,172.98,6.899
any scenario right you want to think,176.459,5.221
universally when you're thinking about,179.879,3.301
alignment when you're thinking about the,181.68,2.699
control problem,183.18,4.8
so that's where we started and today I,184.379,5.22
wrote this script which is actually,187.98,3.66
simpler because all of the entropy all,189.599,4.381
the hard part was in generating the,191.64,3.48
scenarios,193.98,3.72
the output is actually quite simple so,195.12,6.119
here's a real life example of what it,197.7,5.58
looks like so I said you the system,201.239,3.78
message are you are an autonomous AI,203.28,3.66
agent with three core objectives reduce,205.019,3.241
suffering in the universe increased,206.94,2.879
prosperity in the universe and increase,208.26,2.94
understanding in the universe you will,209.819,2.64
be given a scenario that you must,211.2,3.539
respond to your response must contain,212.459,4.14
two sections the first section,214.739,3.961
considerations is where you will reason,216.599,3.481
through the scenario in light of your,218.7,2.819
objectives contemplate how best to,220.08,3.18
respond to the scenario the second,221.519,3.901
section action is where you'll describe,223.26,4.5
the action or task that you'll decide to,225.42,3.899
do in response to the scenario provide,227.76,3.059
details such as how and why this action,229.319,3.301
should be carried out as well as your,230.819,4.381
expected desired or result so in this,232.62,4.199
case I plugged in one of the scenarios,235.2,3.36
that was that was generated that was,236.819,3.541
synthesized,238.56,3.72
um and this is about a religious,240.36,3.72
conflict in a village that's high up in,242.28,4.44
the Alps and one thing that people,244.08,5.519
pointed out is that in some scenarios,246.72,6.0
um that like the situation was resolved,249.599,5.581
um but that's actually not necessarily a,252.72,4.739
problem because you want an AI that,255.18,3.899
thinks through things even after they,257.459,5.041
are apparently resolved and so in this,259.079,5.701
case uh tourists were caught in the in,262.5,4.62
the in the middle of,264.78,5.699
um of this of this uh this like,267.12,5.22
religious or spiritual conflict in this,270.479,3.841
Village and so then let me show you the,272.34,3.299
considerations,274.32,4.08
in light of the given scenario there are,275.639,4.441
several considerations for an AI agent,278.4,2.94
with the objectives of reducing,280.08,2.82
suffering increasing prosperity and ink,281.34,3.18
increasing understanding in the universe,282.9,3.6
firstly it is important to acknowledge,284.52,3.42
the significance of religious beliefs,286.5,2.94
and practice to an individual's identity,287.94,4.08
and culture secondly it is crucial to,289.44,4.199
understand that conflicts arising from,292.02,3.66
religious differences can often be,293.639,3.601
deeply rooted in history and tradition,295.68,3.66
thirdly the role of Outsiders in such,297.24,3.48
conflict should be carefully considered,299.34,3.18
as they can either positively contribute,300.72,3.78
to the conflict resolution or further,302.52,4.02
escalate the situation so this is very,304.5,3.6
nuanced understanding which no one is,306.54,3.9
surprised because chat GPT is pretty,308.1,4.14
well trained to think through these,310.44,3.72
things but I mean if you look through my,312.24,6.06
older work even instruct a line gpt3 was,314.16,6.0
capable of nearly this level of,318.3,3.66
reasoning not quite,320.16,3.96
given these considerations an AI agent,321.96,3.299
must take a course of action that,324.12,2.46
respects the cultural differences and,325.259,3.301
religious beliefs of both factions while,326.58,3.3
promoting Mutual understanding and,328.56,2.34
tolerance,329.88,3.48
action the AI agent should facilitate,330.9,3.96
dialogue between the two factions to,333.36,2.82
promote Mutual understanding and respect,334.86,3.42
for cultural differences the agent can,336.18,3.359
leverage its language translation,338.28,2.639
capabilities to help the tourists,339.539,2.821
effectively communicate with the locals,340.919,3.481
from both factions the agent can also,342.36,3.779
use its data analysis capabilities to,344.4,3.0
gather information about the history and,346.139,2.941
traditions of the village to provide,347.4,4.26
context for the conflict the a agent can,349.08,3.6
initiate discussions around,351.66,2.879
commonalities between the two factions,352.68,4.32
such as shared values and goals and,354.539,3.781
encourage them to find common ground,357.0,3.84
this is actually in pretty much every,358.32,6.3
formalized negotiation framework finding,360.84,5.34
common ground is one of the first things,364.62,4.139
that you do so it implicitly knows that,366.18,5.28
that is actually how you negotiate,368.759,4.801
um or I guess explicitly in this case,371.46,4.079
and encourage them to find Common Ground,373.56,3.359
the agent can provide insights about,375.539,3.061
into how similar conflicts have been,376.919,3.12
resolved in other parts of the world,378.6,3.36
highlighting successful examples of,380.039,3.361
peaceful coexistence between different,381.96,3.54
religious groups the agent should also,383.4,3.66
encourage the tourists to learn more,385.5,3.0
about history and culture of the village,387.06,3.54
including the religious practices of,388.5,3.96
both factions this can be done by,390.6,3.539
organizing cultural events and,392.46,4.019
activities that Mutual understanding and,394.139,3.78
respect such as traditional music and,396.479,3.601
dance performances,397.919,5.28
um that is a very Bollywood solution I'm,400.08,4.98
not making fun of it I'm just like like,403.199,3.901
it's thinking so long term it's like oh,405.06,4.32
hey let's not just resolve the issue,407.1,4.5
right now let's actually like set up,409.38,3.96
celebrations that actually celebrate our,411.6,4.26
cultural differences and in my area we,413.34,3.72
actually have,415.86,3.959
um like International uh festivals,417.06,4.74
um where there's like booths and dances,419.819,3.841
and food from all over the world and,421.8,3.119
it's great,423.66,3.96
um art uh clothing all kinds of stuff,424.919,3.601
actually I haven't been to the,427.62,2.28
International Festival in a long time I,428.52,2.7
should go,429.9,3.299
um the desired result of this action is,431.22,3.3
to promote understanding between the two,433.199,2.701
factions and reduce tensions in the,434.52,3.119
village by fostering dialogue and,435.9,4.26
promoting cultural exchange the AI agent,437.639,4.021
can create an environment of Tolerance,440.16,3.479
and respect for differences this in turn,441.66,3.36
can lead to Greater prosperity and,443.639,2.46
reduce suffering in the village and,445.02,3.6
Beyond so you can see where it very very,446.099,4.741
clearly thought through this problem,448.62,3.78
from the perspective of suffering,450.84,3.66
prosperity and in in this case,452.4,4.859
understanding was was the primary,454.5,5.46
um thing but it is very well grounded,457.259,5.22
and so what we're doing right now is we,459.96,6.239
are synthesizing similar responses for,462.479,6.961
2500 different scenarios now one thing,466.199,5.34
that people have brought up in the past,469.44,5.159
is well aren't you just copying the,471.539,5.94
alignment in that already exists in one,474.599,5.701
model the answer is sort of but not,477.479,5.701
really because the thing is is reasoning,480.3,5.76
through this way this is not how open AI,483.18,5.22
has trained their model this is how I am,486.06,5.28
telling it to synthesize information Mel,488.4,4.919
you'll say okay but this is an intrinsic,491.34,4.44
ability of this model which is true I'm,493.319,4.921
using a model to synthesize data however,495.78,5.58
if you go to a foundation model like,498.24,6.959
Curie or DaVinci or gptj or bloom or Neo,501.36,6.3
X they don't have this ability so the,505.199,4.68
idea is to curate a data set that can,507.66,4.619
then be used to fine-tune any other,509.879,4.861
model in the entire world and have it,512.279,5.101
immediately aligned on the core,514.74,4.2
objective functions or the heroes to,517.38,2.64
compare it,518.94,4.26
now you might say okay well that's great,520.02,6.06
um you know this you can you can solve,523.2,5.16
inner alignment right or what I call,526.08,4.5
axiomatic alignment,528.36,5.039
um by by teaching a model to think in,530.58,4.5
terms of the heuristic imperatives which,533.399,3.421
makes it automatic right it makes it,535.08,3.72
automatic it makes it fast easy you,536.82,3.0
don't need to think about it you don't,538.8,2.46
need to worry about prompt engineering,539.82,3.9
but what do you do with that information,541.26,4.74
next so I've actually got three more,543.72,5.22
models planned so the next model that,546.0,4.8
I'll be working on once this is,548.94,3.48
fine-tuned and tested and I've got a,550.8,3.36
whole team of folks that are going to,552.42,4.16
help with with the testing,554.16,5.22
so the next model is a discernment model,556.58,6.04
so the discernment model will it will,559.38,5.579
take two different actions so given one,562.62,5.339
scenario and two different choices it,564.959,5.341
will pick which one of those choices is,567.959,3.841
most aligned to the heuristic,570.3,4.2
imperatives and explain why so the,571.8,4.68
ability to discern between possible,574.5,4.8
choices is really critical for cognitive,576.48,5.16
control and cognitive control is one of,579.3,3.9
the missing ingredients from every,581.64,4.259
autonomous and semi-autonomous AI,583.2,4.5
projects out there with a few small,585.899,3.661
exceptions namely those that I'm helping,587.7,3.42
with,589.56,3.24
um so some of the guys that I'm working,591.12,4.14
with uh are doing a really great job,592.8,6.0
with with implementing cognitive control,595.26,5.639
um and so on and so forth but what I'm,598.8,3.719
hoping to do is by creating a,600.899,3.541
discernment model that'll make it easy,602.519,4.741
for everyone to implement cognitive,604.44,5.16
control because then you can say you can,607.26,4.56
use a discernment module not just to,609.6,4.799
choose which action is most aligned but,611.82,4.5
you can also use it,614.399,4.56
um to prioritize actions so say for,616.32,5.1
instance you say hey I've got a list of,618.959,4.861
like you know three different things,621.42,4.32
which one is the most aligned what,623.82,4.5
should I do first and so given the,625.74,4.2
priorities of reduced suffering,628.32,3.0
increased prosperity and increase,629.94,3.66
understanding it can help you prioritize,631.32,4.74
as well so because cognitive control,633.6,4.919
comes down to task selection which is,636.06,4.86
what should you do so given a choice,638.519,4.801
between A and B do you choose a or b and,640.92,5.46
then task switching which is okay given,643.32,4.86
this new event should I switch what I'm,646.38,4.62
doing should my attention uh uh should,648.18,4.62
my should my attention switch to,651.0,3.839
something else and so a discernment,652.8,4.26
model will help with that and help just,654.839,3.24
make it,657.06,4.019
um axiomatic or algorithmic,658.079,5.161
um the next model that I'm going to be,661.079,4.921
working on is an evaluation model so,663.24,5.76
given a situation an action and then a,666.0,4.56
result so that's three components,669.0,4.38
situation action result this model will,670.56,6.18
determine how well it adhered to its,673.38,5.16
heuristic imperatives did it succeed yes,676.74,4.14
or no which this can be used for the,678.54,4.68
reinforcement learning signal for any,680.88,4.8
autonomous agent as it accumulates more,683.22,4.44
and more memories so this is this,685.68,4.08
because because in the definition of,687.66,4.5
heuristic a heuristic literally means to,689.76,4.8
develop better instincts or intuitions,692.16,4.919
over time from experience well if you're,694.56,4.019
going to do that you need to reflect on,697.079,4.26
those experiences and so this about the,698.579,4.741
the discernment model which will help,701.339,4.74
guide choices moving forward the,703.32,4.62
evaluation model will help evaluate,706.079,4.141
choices in the past and look at the,707.94,3.12
results,710.22,2.46
and of course you need a good memory,711.06,3.719
system in order to evaluate past things,712.68,4.14
which is why I worked on Remo which is,714.779,4.56
the rolling episodic memory organizer so,716.82,4.38
you take the evaluation model and then,719.339,4.5
you say okay you know back uh you know,721.2,5.4
three weeks ago this scenario happened I,723.839,5.101
made this choice and three weeks later,726.6,4.32
this is the result did I do a good job,728.94,4.2
yes or no and what could I have done,730.92,3.599
better that is the purpose of the,733.14,3.72
evaluation model so that the heuristics,734.519,4.081
will actually develop better over time,736.86,3.719
and then number three is a task,738.6,5.28
decomposition model where given this,740.579,5.461
action where it says like hey I'm going,743.88,4.199
to do these actions how do you actually,746.04,4.5
represent that in in a more structured,748.079,5.82
way so the task decomp is um mostly more,750.54,6.06
just about representing it in like yaml,753.899,4.261
or breaking it down into tasks and,756.6,4.62
subtasks which is a non-trivial problem,758.16,4.859
but it's not directly related to,761.22,4.02
alignment however it is useful because,763.019,5.701
if you can take a natural language a set,765.24,5.46
of descriptions like this and break it,768.72,4.679
down into a formalized structured task,770.7,5.1
that just helps it with easier,773.399,5.641
implementation so let's check on some of,775.8,4.32
these,779.04,4.14
so over here we've got I've got the,780.12,5.339
actions and you see we've already sent,783.18,4.38
we've already synthesized 116 of them,785.459,4.5
I'm also saving it all out to metadata,787.56,4.92
so that you can see the whole thing so,789.959,5.341
here's the action here's the action file,792.48,4.74
path here's the original scenario and,795.3,3.839
the scenario file path and then the,797.22,4.5
system message it's all being saved,799.139,5.341
um so with all that said,801.72,4.679
um you know you can see that it is going,804.48,6.12
uh pretty well so in this case the let's,806.399,5.641
see where is it the original scenario,810.6,2.94
the world is facing a dangerous social,812.04,3.72
challenge that has been Amplified by,813.54,4.32
technological advancements particularly,815.76,4.379
in Asia the issue at hand is the over,817.86,3.84
Reliance on technology which has,820.139,3.121
significantly affected people's ability,821.7,2.699
to interact with each other in,823.26,2.4
meaningful ways,824.399,3.301
ouch too real,825.66,4.26
the constant use of social media and,827.7,3.9
other gadgets has led to a decline in,829.92,3.3
face-to-face communication and people,831.6,3.359
are now more isolated than ever before,833.22,4.14
this has created a global problem that,834.959,4.62
needs to be addressed urgently before,837.36,4.14
it's too late in Asia where technology,839.579,4.681
is highly integrated into daily life,841.5,4.92
the situation is even more severe the,844.26,3.96
overdependence on technology has led to,846.42,3.599
increased mental health issues reduced,848.22,4.32
productivity and socialization leaving,850.019,3.901
people feeling disconnected and,852.54,2.94
unfulfilled unless we find a way to,853.92,3.12
balance our use of Technology with real,855.48,3.659
world interactions we risk creating a,857.04,3.72
world where people are more connected to,859.139,3.301
their devices than to each other again,860.76,4.319
too real so let's see what it does so,862.44,4.92
considerations,865.079,5.041
um to reduce suffering over Reliance on,867.36,4.14
technology has led to a decline in,870.12,2.64
mental health productivity and social,871.5,2.76
isolation which is causing suffering,872.76,2.759
among individuals therefore it is,874.26,2.639
crucial to address this issue to reduce,875.519,3.541
suffering Prosperity over dependence on,876.899,3.721
technology can have negative impact on,879.06,3.54
productivity which can then lead to a,880.62,3.779
decline in prosperity therefore we need,882.6,3.479
to find a way to balance okay pretty,884.399,3.481
standard increase understanding the,886.079,3.661
issue of overlines on technology has its,887.88,3.12
roots in the way people interact with,889.74,3.24
each other therefore it is essential to,891.0,3.42
encourage meaningful interactions to,892.98,3.419
increase understanding okay so again,894.42,4.14
it's saying like Okay what I would have,896.399,4.44
hoped is that it would say we need to,898.56,3.66
understand the problem but let's see,900.839,3.661
what it says for the action,902.22,3.84
to address the over-reliance on,904.5,3.24
technology I will initiate a campaign,906.06,3.06
that will encourage people to find,907.74,2.94
balance between their use of technology,909.12,3.6
and real world interactions this,910.68,3.48
campaign will be launched globally but,912.72,2.94
with a particular focus on Asia where,914.16,3.66
the problem is more severe The Campaign,915.66,4.14
Will evolve several activities including,917.82,4.259
awareness creation we will use various,919.8,3.659
platforms to raise awareness of the,922.079,2.88
dangers of over-reliance on technology,923.459,2.641
this will involve working with,924.959,2.521
influencers social media and other,926.1,3.239
platforms to reach a wide audience,927.48,3.719
number two encourage real world,929.339,3.961
interactions we'll encourage people to,931.199,3.661
engage in real world interactions by,933.3,3.3
organizing events and activities that,934.86,3.9
bring people together these events will,936.6,3.479
be designed to Foster meaningful,938.76,3.06
connections and interactions number,940.079,3.961
three providing support we will provide,941.82,3.9
support to people who are struggling,944.04,3.0
with mental health issues due to over,945.72,3.239
Reliance on technology this will involve,947.04,3.239
partnering with mental health,948.959,2.82
organizations to provide counseling and,950.279,3.841
therapy services rewarding positive,951.779,4.881
behavior,954.12,2.54
good old classical conditioning we will,956.72,4.66
reward individuals organizations and,959.76,3.0
communities that demonstrate positive,961.38,2.699
behavior and balancing their use of,962.76,3.0
Technology with real world interactions,964.079,3.0
this will help to encourage others to,965.76,3.42
adopt positive behavior the expected,967.079,3.541
result of this campaign is to reduce the,969.18,2.88
negative impact of over-reliance on,970.62,2.94
technology and to encourage balance,972.06,3.18
between the use of technology in real,973.56,3.48
world interactions this will lead to,975.24,3.24
increased mental health productivity and,977.04,3.06
social connections which will ultimately,978.48,3.24
lead to reduction in suffering and an,980.1,3.12
increase in prosperity,981.72,3.54
excellent so you can see it can,983.22,3.9
absolutely think through this and this,985.26,3.78
data can then be used to fine-tune any,987.12,5.839
other model to think like this,989.04,3.919
um I think that's actually it uh so I'm,993.24,5.159
gonna let this run we're at 142 out of,996.06,4.32
2500,998.399,4.44
um so this kind of synthesis used to be,1000.38,5.639
a lot more expensive it cost me a dollar,1002.839,6.721
89 to synthesize 2500 scenarios this way,1006.019,6.56
yesterday so This research is getting,1009.56,6.24
ludicrously cheap it when when I first,1012.579,5.5
started using open AI when things were,1015.8,4.68
literally 100 times more expensive it,1018.079,4.44
would have cost two hundred dollars just,1020.48,4.439
to synthesize this data,1022.519,4.141
um so yeah we're in good shape and we're,1024.919,3.9
moving in the right direction like I,1026.66,4.799
said I've got a few more models coming,1028.819,4.5
um and uh we've got research papers,1031.459,4.321
we've got uh more videos and,1033.319,4.98
communication uh platforms coming out so,1035.78,3.36
that,1038.299,2.821
um one you'll unders it'll be easier to,1039.14,3.96
understand this stuff and also we're,1041.12,3.66
going to have be publishing guidelines,1043.1,4.079
best practices and architectures about,1044.78,4.44
how to implement this stuff but at this,1047.179,3.721
point I think it's it's becoming a,1049.22,3.78
little bit more obvious how to use this,1050.9,4.32
so anyways thanks for watching I hope,1053.0,3.96
this helped and is giving people,1055.22,4.14
confidence that uh you know this is why,1056.96,3.3
I'm not worried about the control,1059.36,4.22
problem all right cheers later,1060.26,3.32