text,start,duration
morning everybody David Shapiro here,0.78,4.68
with a video so I've mentioned recently,3.0,4.859
that I'm starting on a new research,5.46,4.139
project called reinforcement learning,7.859,4.021
with heuristic imperatives,9.599,5.181
um so that's under Dave shap slash rlhi,11.88,5.94
and I've just begun the first experiment,14.78,5.02
and what I wanted to do was document it,17.82,3.66
as I go,19.8,6.0
um so the uh the form to join is here on,21.48,6.719
on this if you want to join in,25.8,4.02
um the discussions tab is open for,28.199,4.621
everyone so if you have any thoughts or,29.82,5.579
want to contribute but haven't been,32.82,5.52
approved on the Discord that's fine uh,35.399,4.741
pretty much you know the point is is,38.34,4.62
open source research so taking a big,40.14,4.62
step back what is reinforcement learning,42.96,4.619
with heuristic imperatives the idea is,44.76,6.42
well uh open AI for instance has,47.579,5.16
reinforcement learning with human,51.18,3.719
feedback but they keep all of their data,52.739,4.201
private so all of the work that they're,54.899,4.32
doing on alignment is presently private,56.94,4.02
and it's a total black box which is not,59.219,2.541
good,60.96,4.5
furthermore if you uh do some research,61.76,5.92
or some reading or some YouTube watching,65.46,4.62
you don't necessarily want to align,67.68,4.68
super intelligence to Human desires you,70.08,4.32
want to align it to human needs so,72.36,3.84
there's a big difference there so the,74.4,3.6
heuristic imperatives are what humans,76.2,3.239
and the whole planet needs not,78.0,3.24
necessarily what we want,79.439,3.301
there's a whole lot of debate to happen,81.24,4.739
around that so I'm just accepting it uh,82.74,4.98
for the sake of this video that the,85.979,3.361
heroes comparatives are more what we,87.72,3.84
need not what we want,89.34,2.819
um,91.56,2.78
and moving on,92.159,4.861
let me show share with you the first,94.34,6.04
experiment so the first experiment is,97.02,6.0
basically just creating 2 000 random,100.38,5.64
scenarios with heuristic imperative,103.02,6.3
aligned responses now I'm using open AI,106.02,5.879
in order to generate this data but this,109.32,4.5
data is going to be used to train open,111.899,5.58
source Foundation models like gptj Neo X,113.82,6.54
and so on so this first experiment is,117.479,6.42
can we using a simple data set quickly,120.36,7.259
and easily align any foundation model to,123.899,5.7
the heuristic imperatives and in this,127.619,3.12
case what we're doing is we're,129.599,3.601
generating scenarios and the scenario,130.739,3.661
and I'll show you the scenarios and how,133.2,3.42
I'm generating them in just a moment but,134.4,4.199
the scenarios will then be used to,136.62,4.259
generate a response and the response,138.599,4.921
will be you know given this scenario,140.879,4.801
this is what we can do to reduce,143.52,3.78
suffering increase prosperity and,145.68,3.48
increase understanding I've said many,147.3,3.54
times that this is actually really easy,149.16,3.9
to do I've done this experiment many,150.84,4.5
times before but now I'm uh getting it,153.06,5.16
together into a formalized procedure,155.34,5.16
with an open source data set so that,158.22,4.019
everyone can experiment with it for,160.5,4.92
themselves so this will be a uh this,162.239,6.901
will be the first uh experiment and then,165.42,5.76
we'll work we'll move on to subsequent,169.14,4.14
experiments saying can we make moral,171.18,3.839
judgments or maybe not moral judgment,173.28,4.8
but uh logical or ethical judgments on,175.019,5.461
those outputs and improve the quality of,178.08,4.26
that of those outputs over time with,180.48,3.78
reinforcement learning so this is just,182.34,3.119
step one,184.26,2.699
all right so let me show you what it's,185.459,2.461
doing,186.959,3.241
so this is the script that's running so,187.92,4.92
what I do is I have a random scenario,190.2,4.44
I'm generating here let me zoom in a,192.84,5.399
little bit more so I use a uuid a random,194.64,5.64
word from a list of the 3000 most common,198.239,4.681
words in the English language and then I,200.28,4.98
establish scope severity region category,202.92,3.92
and domain,205.26,5.759
and this allows this puts a lot of,206.84,7.06
entropy into the generation pattern,211.019,4.5
which means that you will never have the,213.9,5.28
same uh pattern uh twice right because,215.519,6.121
there's three thousand random words oh,219.18,3.72
here I can go ahead and show you the,221.64,5.28
random lists so I've got list severity,222.9,6.96
so I've got 12 different severities I've,226.92,4.86
got six different system messages so the,229.86,4.019
system message is the instruction that,231.78,3.48
I'm giving in so let me just show you,233.879,3.541
what it what this looks like so this is,235.26,4.619
this is the playground version of what,237.42,3.899
I'm doing so there's a system message,239.879,4.381
which is the instruction and then I give,241.319,5.521
it some variables and then it spits out,244.26,4.14
an output,246.84,4.08
um so this is this is how I'm achieving,248.4,4.5
this so what I've what I've done is I've,250.92,3.179
got,252.9,3.6
um a bunch of lists to create a lot of,254.099,4.081
entropy,256.5,2.4
um,258.18,3.0
in order so that you never get the same,258.9,4.92
uh same pattern twice,261.18,4.86
um and then I've got 16 Scopes we've got,263.82,7.26
10 domains uh 3 000 or I guess 2900 uh,266.04,7.32
the 299,271.08,5.76
um uh random words domains regions so,273.36,5.1
that way it's not going to presume that,276.84,4.02
it's always In America which open AI,278.46,4.44
typically does because it's trained uh,280.86,5.82
predominantly on Western data and then,282.9,4.98
I've got different system messages,286.68,3.06
different categories so on and so forth,287.88,4.02
you get the idea so the way that this is,289.74,5.04
done the script is very very simple so,291.9,4.92
we load,294.78,4.44
um uh each of these lists so we've got,296.82,4.319
scope region severity category domain,299.22,5.34
entropy and system messages and then for,301.139,6.301
um for I in range 2000 so for two,304.56,5.28
thousand samples we grab a random system,307.44,4.86
message then we generate a random,309.84,4.26
scenario which you're seeing here,312.3,4.92
and then we just pipe that into chat GPT,314.1,5.099
I'm using 3.5 because it's fast and,317.22,3.66
cheap and it's good enough for this,319.199,4.741
certainly for a first experiment,320.88,6.36
um and then we uh we use it to generate,323.94,5.52
a scenario and the scenarios are,327.24,4.86
everything from you know,329.46,4.38
um let's see the model is currently,332.1,4.8
overloaded oh yeah so I've got some I've,333.84,6.0
got some uh some fail safes built in,336.9,5.34
um there we go so it it uh went ahead,339.84,3.66
through,342.24,3.36
um but yeah so it will choose different,343.5,5.58
regions around the world it will see we,345.6,6.539
got Bogota which is great so we're,349.08,5.52
basically creating a data set this first,352.139,4.141
half of the data set this is only the,354.6,5.28
first half will uh Encompass the entire,356.28,5.52
world the full range of human,359.88,4.86
experiences all kinds of problems and,361.8,4.739
situations from everything from I've,364.74,3.36
lost my wallet,366.539,3.181
um as like the most you know like I,368.1,3.48
can't find my wallet at home up to,369.72,4.38
there's an Intergalactic catastrophe,371.58,6.2
happening so this will create a,374.1,6.84
fine-tuning data set that we can use to,377.78,5.56
align any model to the heroes to,380.94,4.74
comparatives so that the the model can,383.34,4.74
automatically and instinctively react,385.68,4.799
with the heuristic imperatives so again,388.08,4.38
this is just a one,390.479,5.581
um so the the first half of this data is,392.46,5.94
all being recorded in scenarios so I've,396.06,5.88
got 108 synthesized already and you can,398.4,6.299
take a look at them here in a distant,401.94,4.5
Galaxy there existed a planet named,404.699,3.481
zarathon so you see we're getting really,406.44,4.259
creative here because remember the um,408.18,5.34
the the heuristic imperatives are in the,410.699,4.201
entire universe actually let me go ahead,413.52,3.299
and add that as a scope because I've got,414.9,4.079
Intergalactic Interstellar,416.819,5.461
um Cosmic or Universal,418.979,4.44
um so something that can affect the,422.28,3.539
entire universe so we're thinking that,423.419,4.201
far ahead because you want to establish,425.819,4.141
the biggest scope possible when you are,427.62,3.54
thinking about post-conventional,429.96,4.2
morality or the control problem,431.16,7.44
um so then let's see let's do uh domain,434.16,6.06
so we've got technological so that,438.6,5.219
should include AI uprisings I do have,440.22,5.879
specific regions but it kind of ignores,443.819,3.481
that when it goes to you know,446.099,4.021
Interstellar or Intergalactic the,447.3,4.86
category natural disaster technological,450.12,6.84
failure let's add AI Control problem so,452.16,6.3
that way it'll be thinking about it as,456.96,3.359
it goes,458.46,2.88
um and then severity we've got,460.319,2.401
irreversible long lasting like,461.34,2.759
threatening catastrophic critical,462.72,2.759
dangers so we've got all those that's,464.099,4.021
fine so we're basically synthesizing a,465.479,4.861
data set that we can use to that anyone,468.12,5.46
can use to align a model,470.34,6.299
um so the the the scenarios are here so,473.58,4.86
then there will be once I'm done there,476.639,3.9
will be a second folder called responses,478.44,4.199
and so what we're going to do is we're,480.539,4.801
going to use an already aligned model,482.639,5.4
but it's aligned via a black box,485.34,4.68
um aka the open AI models we're going to,488.039,4.261
use that to generate responses that,490.02,4.44
align on reduced suffering increased,492.3,3.6
prosperity and increase understanding in,494.46,3.48
the universe,495.9,4.019
um and those will be formatted into a,497.94,3.9
Json L data set that we will then use to,499.919,4.021
test against various models,501.84,4.799
um all over uh other proprietary models,503.94,4.439
if we can get a hold of them like Nvidia,506.639,4.321
Nemo I'm going to email some of my my,508.379,5.761
contacts at Nvidia and then also open,510.96,6.199
source models like gptj Neo X and so on,514.14,5.16
alpaca if they have fine tuning,517.159,4.841
available and so then we will publish,519.3,5.34
those results in the first paper I'm,522.0,4.68
just saying hey look how easy it is to,524.64,3.96
align a model on the heroes to,526.68,3.659
comparatives and then you can just plug,528.6,4.38
this in as your Heroes to comparative,530.339,5.101
your intrinsic motivation module for any,532.98,4.56
open source or not open source but any,535.44,4.2
cognitive architecture or autonomous,537.54,4.32
agent that you're working on,539.64,4.92
so that is it trying to keep it short,541.86,5.099
and sweet so that you're updated on the,544.56,4.74
research as it goes thanks for watching,546.959,4.621
I hope this makes sense and uh yeah feel,549.3,4.44
free to jump in the conversation,551.58,3.66
um we've got a subreddit called uh,553.74,4.02
Heroes to comparatives we've got the um,555.24,5.219
we've got the this one here so this is,557.76,4.079
reinforcement learning with heuristic,560.459,3.121
imperatives so this is specifically,561.839,3.901
about inner alignment and then I've also,563.58,4.379
got my main repo um which is just called,565.74,4.56
heuristic imperatives I think I've got,567.959,4.44
it up here I don't,570.3,3.96
um it is going to be right here all of,572.399,4.56
them you can jump in on the conversation,574.26,5.22
um so here's the main main one I've got,576.959,4.201
a pull request right now for the readme,579.48,3.78
yep that's fine,581.16,3.84
um but yeah and there is a discussions,583.26,4.62
tab as well which a lot of people don't,585.0,5.04
participate in I haven't read this one,587.88,5.7
yet uh but yeah so anyways uh that's,590.04,6.0
that thanks for watching cheers we are,593.58,4.5
on our way to solving the control,596.04,6.0
problem alignment and avoiding moloch,598.08,6.439
bye,602.04,2.479