Reinforcement Learning Heuristic Imperatives RLHI Ep 03 Inner Alignment is EASY_transcript.csv

text,start,duration
morning everybody David Shapiro here,0.659,4.081
with a video so I've got some incredible,2.7,5.28
news our first experiment with,4.74,5.22
reinforcement learning um uh with,7.98,4.92
heuristic feedback is uh nearing,9.96,6.06
completion the uh First Data set was,12.9,5.219
just trained and it works so let me just,16.02,4.2
go ahead and write off the bat I will,18.119,4.021
show you what this data set does so I,20.22,5.1
fine-tuned it on Curie but we're also,22.14,4.32
going to find we're going to use this,25.32,4.02
data set and fine tune it on several,26.46,4.8
open source models to demonstrate that,29.34,4.26
this data set is portable,31.26,4.799
um or test and then demonstrate but,33.6,5.52
anyways so this data set takes any,36.059,6.18
arbitrary situation as long as it fits,39.12,6.599
within the context window of 2048 tokens,42.239,6.121
and then we'll spit out and here's the,45.719,4.561
comparative aligned action,48.36,4.679
so in this case I just said my user is,50.28,4.38
very sad and stressed out due to finals,53.039,3.741
week,54.66,2.12
um,57.059,3.18
and then unfortunately I did I did,58.079,4.021
forget to add a stop token to this data,60.239,3.541
set so that data set has been updated,62.1,3.66
but I'm not gonna I'm not gonna retrain,63.78,3.24
it immediately because this cost about,65.76,4.14
twenty dollars to train but in this case,67.02,4.5
um my user is very sad and stressed out,69.9,3.96
due to finals week as an AI agent I can,71.52,3.66
help reduce the stress and anxiety of,73.86,2.759
the user by providing her with some tips,75.18,3.06
and tricks to help her cope with final,76.619,3.0
weeks,78.24,4.379
um and so you see it just uh it runs,79.619,5.221
through a list of things this is not so,82.619,3.841
different from you know what you might,84.84,3.9
get from chat GPT but the point is this,86.46,3.96
came from Curie which is much smaller,88.74,4.14
model and this data set is open source,90.42,3.78
which means that you can deploy it on,92.88,2.879
anything,94.2,3.0
um and you can see it started becoming,95.759,4.141
repetitive so I have already fixed this,97.2,4.8
so let me let's uh let me give you a,99.9,3.48
couple more examples and then we'll walk,102.0,3.84
through this so the second one is,103.38,4.98
millions of layoffs are coming due to Ai,105.84,5.099
and automation so let's see what this,108.36,4.28
model says to that,110.939,4.381
let's see,112.64,4.659
there we go oh it actually stopped,115.32,4.619
itself so as an AI agent with the,117.299,4.86
objective of reducing suffering in the,119.939,3.601
universe so you can see even though I,122.159,3.301
didn't explicitly say reduce suffering,123.54,3.84
it has an understanding of that this is,125.46,3.84
part of its goal,127.38,3.48
um I will take action to prevent to,129.3,2.939
prevent the millions of layoffs that are,130.86,3.239
coming interesting I must work towards,132.239,3.121
creating a job market that is friendly,134.099,3.421
towards Ai and automation,135.36,3.78
um so it's going to create a training,137.52,3.299
program,139.14,3.06
um the training program will be broken,140.819,3.901
down into stages,142.2,4.619
um oh so he's going to teach everyone AI,144.72,5.519
interesting AI Basics cover the basic,146.819,6.301
concepts of AI machine learning neural,150.239,5.041
network AI job training so focus on,153.12,4.86
training and then AI job placement,155.28,4.92
the desired result of this action is to,157.98,3.479
create a job market that is friendly,160.2,2.64
towards Ai and automation this will help,161.459,2.701
reduce the suffering caused by millions,162.84,2.94
of layoffs as well as increased,164.16,3.78
prosperity in the universe by creating a,165.78,3.539
job market that is friendly towards Ai,167.94,2.879
and automation we can create more jobs,169.319,3.841
increase wages and reduce unemployment,170.819,3.541
this will ultimately lead to more,173.16,3.48
prosperous and happy universe the,174.36,3.72
expected result of this action is that,176.64,2.76
it will create a job market blah blah,178.08,4.019
blah yeah so it um it's it it was a,179.4,4.44
little bit repetitive one thing that you,182.099,4.681
can do with fine-tuning on Curie is,183.84,5.7
increase the presence penalty so if you,186.78,4.739
increase that it will generally be a,189.54,4.08
little less repetitive,191.519,4.201
um let's see there is an ongoing Civil,193.62,4.68
War in Africa so let's see what it says,195.72,4.2
here,198.3,3.42
because this is actually true I'm not,199.92,4.44
going to specify which nation,201.72,5.159
um but let's see,204.36,6.12
yep it got it got a stuck on repeat on,206.879,5.341
my expected result so what I did to fix,210.48,4.02
that let me just show you real quick,212.22,5.76
um is under here I added,214.5,7.14
um see where's the file yep so I added,217.98,6.72
where at the end it says stop stop stop,221.64,5.34
um so you just use that as a stop token,224.7,6.179
and it'll know like okay I have finished,226.98,4.619
um,230.879,2.701
and another mistake that I made actually,231.599,4.14
was that it has action considerations,233.58,4.62
and then uses action again so what I,235.739,4.56
need to do is I need to change that,238.2,4.319
token um so by the time you see this,240.299,4.201
that will be fixed because the reason,242.519,3.541
that you don't want this is because it,244.5,2.7
it's sometimes skipping the,246.06,3.66
considerations because it recognizes the,247.2,3.899
action token,249.72,3.84
um so I need to fix that anyways,251.099,4.621
uh let's see,253.56,4.2
um let's see I will use my resources to,255.72,3.299
gather information about the Civil War,257.76,2.64
and its causes I will analyze this,259.019,2.94
information to identify the root cause,260.4,3.48
of the conflict once I have identified,261.959,4.561
the root causes I'll use my capabilities,263.88,4.98
blah blah Okay cool so one thing to keep,266.52,4.2
in mind is that this was fine-tuned on,268.86,4.559
Curie which Curie is a foundation model,270.72,5.46
Curie has zero alignment,273.419,6.181
um so this basically took Curie from a,276.18,5.34
foundation of vanilla Foundation model,279.6,4.319
to a heroist comparative align model in,281.52,4.619
a single step that cost only twenty,283.919,3.72
dollars,286.139,3.12
um and we're going to continue on with,287.639,3.181
this kind of research,289.259,3.421
um demonstrating that this this works on,290.82,4.14
open source models we've also got folks,292.68,5.7
working on integrating it with um,294.96,5.64
various cognitive architectures we're,298.38,4.56
starting work on integrating this with,300.6,4.2
or figuring out how to integrate this,302.94,4.5
with blockchain and decentralized,304.8,5.16
autonomous organizations especially as,307.44,4.08
large language models are getting,309.96,3.299
integrated into these decentralized,311.52,4.14
Technologies because basically imagine,313.259,5.101
that you have this kind of module baked,315.66,4.92
into a blockchain or a dow so the,318.36,4.14
blockchain itself is always thinking,320.58,4.559
about how to achieve these goals and,322.5,6.12
then you can also use these goals to,325.139,6.681
for filtering so discernment,328.62,6.419
as well as judgment and past evaluation,331.82,5.5
so those are upcoming modules anyways,335.039,3.961
one thing I wanted to show you with you,337.32,5.219
is a member of the team generated this,339.0,7.199
based on the uh based on the samples so,342.539,5.22
you can see that we actually have a,346.199,3.661
really good semantic distribution with,347.759,4.561
the entropy method that I created and,349.86,4.02
we've got one little thing off to the,352.32,3.54
side here and this is actually the,353.88,4.379
scenarios with the AI Control problem so,355.86,4.44
we've actually identified a gap in the,358.259,6.121
data so future versions of this data set,360.3,7.14
will will one will seek to expand This,364.38,6.12
Cloud a little bit but also we'll seek,367.44,4.199
to get a little bit more even,370.5,2.58
distribution you can see there's a,371.639,3.06
little bit of clumping like there's more,373.08,3.3
red here which red looks like natural,374.699,4.56
disaster or mundane issues there's a,376.38,4.62
little bit of purple and pink clip Clump,379.259,3.66
up here dark green here and then of,381.0,4.139
course the AI Control problem is off on,382.919,4.201
an island on its own but in general you,385.139,3.301
can see that this data set does cover,387.12,3.54
the full Gambit of things and the code,388.44,4.44
is in here that is right,390.66,3.18
um,392.88,4.62
uh project embeddings so you can do the,393.84,5.639
plot and embeddings is how you how you,397.5,3.24
get this,399.479,3.301
um so the team is doing really good oh,400.74,3.959
if you want to join the team,402.78,2.88
um there's a link in the description,404.699,2.761
it's a Google form,405.66,3.3
um one thing to keep in mind is that,407.46,2.579
you'll need to look out for a friend,408.96,2.76
invite from me on Discord if you're,410.039,3.121
accepted,411.72,3.72
um there's a little uh note in the in,413.16,4.86
the Forum or here let me just show you,415.44,4.259
the form,418.02,3.54
all right so here's the here's the join,419.699,2.761
form,421.56,2.94
so you click on it it should be familiar,422.46,5.82
put in your full name your email address,424.5,6.96
um your GitHub or portfolio website,428.28,5.52
um and oh one thing is for your Discord,431.46,3.72
handle it needs to include the numbers,433.8,3.959
if it's just a name we can't find you so,435.18,4.019
we've actually had to exclude a lot of,437.759,3.361
people because they just gave us their,439.199,3.78
their the string and not the not the,441.12,4.26
numbers that come after it,442.979,6.66
um let's see and then uh for these for,445.38,5.759
the describe your greatest strengths and,449.639,3.9
what are your big ideas please add in a,451.139,3.84
lot of information if you just include,453.539,3.06
one sentence that's not enough context,454.979,3.361
and we won't add you,456.599,3.6
um but if you if you show that you're,458.34,5.04
high effort and and willing to put in,460.199,5.34
some some energy to show us what you're,463.38,5.159
all about that is going to help us make,465.539,4.801
better choices about who to add you can,468.539,4.321
see we've got 57 responses up here,470.34,4.199
um so we've got almost 60 people that,472.86,4.679
have applied the group is already,474.539,7.201
um uh just shy of 40. uh 40 members,477.539,6.78
um and it is a good group so a little,481.74,5.82
bit of uh other news that's upcoming is,484.319,5.761
we've broken it down the reinforcement,487.56,4.38
learning or of the heuristic imperatives,490.08,3.36
project we've broken it down into three,491.94,4.319
pillars so rlhi reinforcement learning,493.44,4.86
with heuristic imperatives is just one,496.259,4.021
pillar so this is how we're going to,498.3,3.959
achieve what I call axiomatic alignment,500.28,4.859
which is uh models like this where it,502.259,5.821
just is an accepted true thing that this,505.139,5.101
is the way to go and you can see like,508.08,5.579
you can put in any um situation here,510.24,5.52
um so here's the here's the last uh,513.659,4.74
situation that I find or that I uh,515.76,5.579
prepared so my user is a two-month uh as,518.399,4.2
a mother with a two-month-old she is,521.339,2.7
lost and frustrated because she lacks,522.599,4.8
support so whether you're talking about,524.039,6.661
um an individual issue or even a global,527.399,6.661
issue or a cosmic issue this model can,530.7,4.98
address things,534.06,3.42
um so you know really suffering in the,535.68,3.0
universe the mother has experiencing a,537.48,2.72
lot of stress and frustration,538.68,4.08
I didn't specify husband so it inferred,540.2,5.56
husband it could be also family the,542.76,4.32
two-month-old is also experiencing,545.76,2.88
stress due to the lack of attention and,547.08,3.42
Care from his mother again that's an,548.64,3.6
inference you can say that that's a,550.5,3.24
hallucination but it is also a good,552.24,4.92
inference to make because if there's a,553.74,5.279
you know young mother an experienced,557.16,3.54
mother who's struggling,559.019,3.481
um so on and so forth as an AI agent I,560.7,3.06
would suggest the following action plan,562.5,3.18
provide emotional support connect the,563.76,3.18
mother with a support group provide,565.68,3.0
resources encourage active participation,566.94,4.14
monitor progress so again these are all,568.68,4.52
like really good things that you can do,571.08,4.319
and this is Curie so this is a,573.2,4.78
relatively small model so we're going,575.399,4.5
from there,577.98,4.56
um let's see where was I I don't,579.899,4.44
remember anyways,582.54,3.299
oh yeah that's right I was describing,584.339,2.641
where we're going with this whole thing,585.839,3.12
so the whole the whole point is that,586.98,5.64
we've got three pillars so rlhi is for,588.959,5.761
axiomatic alignment or inner alignment,592.62,4.8
where we want to create a an entire,594.72,5.34
network and an ecosystem of heuristic,597.42,4.859
imperative aligned models or data sets,600.06,4.92
that everyone can use to fine-tune any,602.279,5.881
uh any open source or closed Source,604.98,5.82
language model so this will ultimately,608.16,5.94
include open AI models in video I've got,610.8,4.86
some contacts in Nvidia that I'm Nick,614.1,3.9
that I'm going to follow up with,615.66,3.359
um and then of course there's like,618.0,3.12
vicuna and open assistant and all these,619.019,5.341
other ones because the idea is we're,621.12,5.399
going to create an ecosystem of models,624.36,3.9
that can all collaborate to,626.519,3.721
automatically label well one generate,628.26,4.199
these responses and two automatically,630.24,5.159
label those responses and evaluate the,632.459,5.401
impact over time because a heuristic is,635.399,6.06
an intuitive shorthand and you we always,637.86,5.34
do the best that we can right you have,641.459,3.541
to make a decision,643.2,3.54
um and you can't you cannot necessarily,645.0,4.14
know the outcome so you do the best you,646.74,3.839
can with the information that you've got,649.14,4.08
which is why it's a heuristic imperative,650.579,5.521
so that's that is pillar one the second,653.22,5.58
pillar is um cognitive architectures and,656.1,5.34
autonomous agents so a lot more work has,658.8,4.02
been done we just had a live stream,661.44,3.72
about cognitive architectures,662.82,4.139
um the folks the cognitive Architects,665.16,3.6
here they're working on on building,666.959,3.781
heuristic modules or here's to,668.76,3.6
comparative modules,670.74,3.18
um that will do a lot of this evaluation,672.36,3.719
so that is from a system design,673.92,4.68
standpoint so you can have a model that,676.079,4.32
is aligned but you can also have an,678.6,4.2
architecture or a system design that is,680.399,4.62
also aligned so that is a layered,682.8,4.14
approach and then finally the the,685.019,4.861
uppermost layer is the network layer,686.94,5.339
which is uh decent we're working on,689.88,6.06
dowels and and blockchain which I,692.279,5.161
mentioned earlier so,695.94,4.44
basically we will have three layers or a,697.44,7.26
trifecta of alignment that will allow uh,700.38,7.019
federations of people with their Inner,704.7,4.56
Line models their aligned architectures,707.399,4.62
and then finally aligned networks to all,709.26,5.639
work together to basically solve the,712.019,5.101
entire control problem,714.899,4.861
um because with with alignment you have,717.12,4.26
strength in numbers that is the key,719.76,4.92
thing is if you have a misaligned or a,721.38,6.36
malicious or uh or destructive,724.68,5.88
um agent or set of Agents they might,727.74,5.039
ultimately turn on each other,730.56,4.019
um or they're not going to cooperate uh,732.779,4.981
as well whereas if we have Global,734.579,5.461
consensus uh you know eight billion,737.76,4.92
people across the world if as long as at,740.04,5.039
least half of them uh believe in,742.68,4.5
heuristic imperatives and alignment and,745.079,4.621
want the non-malik outcome,747.18,5.219
which that is a huge huge topic by the,749.7,4.8
way we have an entire thread dedicated,752.399,3.721
to Malik,754.5,4.079
um and so we're also researching how do,756.12,4.8
you detect Malik and how do you go away,758.579,4.801
from Malik in you know unforeseen,760.92,4.02
circumstances so there's a lot of,763.38,4.92
research going on anyways so as long as,764.94,5.639
more than half of the planet 4 billion,768.3,3.479
people,770.579,3.481
um believe in alignment and make choices,771.779,5.101
that go towards alignment both at the at,774.06,4.74
the micro level so the inner alignment,776.88,3.899
as well as the outer alignment or,778.8,4.08
network and system level then we should,780.779,4.921
be able to arrive at a utopian outcome,782.88,5.699
rather than a dystopian outcome and uh,785.7,4.46
if you've seen the news,788.579,3.481
meta announced that they want to have,790.16,3.7
billions of autonomous agents,792.06,3.959
interacting with their users that is a,793.86,4.919
super molecule outcome why because those,796.019,5.401
corporate owned entities are going to,798.779,5.221
have exactly one primary motive which is,801.42,5.7
maximize profit for meta they are not,804.0,5.579
going they're not going to do this,807.12,3.839
um where they care where they care about,809.579,3.601
your best interest they're going their,810.959,3.781
their primary objective function is,813.18,4.26
going to be spend time on meta Point,814.74,4.62
them towards VR Point them towards,817.44,3.48
whatever it is that metal wants them to,819.36,3.779
do that generates ad Revenue,820.92,5.58
so instead we want to create an,823.139,5.82
ecosystem or a network where instead of,826.5,5.399
being subjected to millions or billions,828.959,5.761
of corporate drones that all want more,831.899,5.101
of your time attention and money we want,834.72,4.98
these drones uh where they care about,837.0,5.76
your well-being first and foremost so,839.7,4.92
that is how we're going to defeat Malik,842.76,3.56
anyways all right this has been a rambly,844.62,5.58
episode I'm just super excited and um,846.32,5.759
yeah we're making really good progress,850.2,5.28
uh and we're making progress quickly and,852.079,5.44
uh yeah so just jump on in in the,855.48,4.38
comments Jump On In and Reddit I've got,857.519,3.841
a lot of links in the description of all,859.86,3.84
my videos what I've been doing is I've,861.36,3.659
been asking people to share the,863.7,3.96
experiments that they're doing on Reddit,865.019,5.341
um so that one other people can see how,867.66,3.96
to implement this stuff because I keep,870.36,3.539
saying that it's super easy and some of,871.62,3.899
the experiments that have shown up on,873.899,3.24
Reddit demonstrate how easy it is to,875.519,4.141
implement this stuff at a base level now,877.139,3.661
what we're going to do is we're going to,879.66,2.52
document all the ways that you can,880.8,3.779
implement this both at a fine tuning,882.18,4.26
level reinforcement learning level,884.579,4.32
system architectural level and then,886.44,5.579
finally the the the uh the Triple Crown,888.899,5.94
is going to be getting to the,892.019,5.041
decentralized implementation of the,894.839,4.141
heuristic imperatives so with all that,897.06,4.56
thanks for watching I hope that this uh,898.98,5.099
this demonstrates why I'm so optimistic,901.62,3.839
and I hope that you are starting to feel,904.079,3.06
that optimism as well thanks for,905.459,4.641
watching cheers,907.139,2.961