Cognitive Architecture April 28 Salience Cognitive Control Task Management Modular Design_transcript.csv

text,start,duration
all right cool we are recording so just,0.78,5.579
quick introduction,4.38,4.02
um this is uh these are some of the guys,6.359,5.581
from the heroistic imperatives uh uh,8.4,6.0
Discord the research group,11.94,4.38
um they're building a cognitive,14.4,3.66
architecture,16.32,3.48
um so I'll let you guys kind of if you,18.06,3.059
want to introduce yourselves not,19.8,3.84
required but uh yeah let's jump in and,21.119,3.781
just kind of talk about where we're at,23.64,3.12
in terms of cognitive architecture and,24.9,3.84
and some of the stuff that you guys have,26.76,3.779
been working on,28.74,4.74
yeah sure um I'm database,30.539,6.54
um I'm on the the uh cognitive AI,33.48,5.82
Channel too y'all see me all the time,37.079,3.901
I'm the one who's always arguing with,39.3,3.3
everybody,40.98,2.28
um,42.6,3.18
Angela if you wanna,43.26,5.16
yeah well uh I'm Ansel I'm on the same,45.78,5.099
Discord I'm a computer science engineer,48.42,5.94
and I just I'm really fascinated by this,50.879,4.741
uh,54.36,5.219
subject so I just had to go and dive in,55.62,8.279
excellent excellent and so uh give us a,59.579,6.841
tiny bit of background actually because,63.899,3.961
um you guys,66.42,3.3
um or a database I remember your story,67.86,3.24
you you jumped into cognitive,69.72,3.66
architecture what like just a few weeks,71.1,3.72
ago right,73.38,3.48
yeah basically,74.82,5.1
um so it was about the time that uh Auto,76.86,5.759
GPT went viral I was like this is really,79.92,4.44
cool I don't want to have to pay for it,82.619,5.221
so I uh started trying to convert it to,84.36,7.5
uh to a local model which,87.84,4.56
um,91.86,3.24
we ran into uh we ran into some issues,92.4,4.14
with that,95.1,5.1
um but that's you know before then I,96.54,5.46
hadn't really done a whole lot with,100.2,5.4
programming or python at all,102.0,7.079
um and so you know most of what I've,105.6,5.64
learned in the past three weeks now I,109.079,5.4
guess has been primarily through Chad,111.24,7.5
GPT nice excellent yep that's that's the,114.479,6.661
case for many people and Ansel you're,118.74,4.5
you you're you're you've been a,121.14,3.6
developer or an engineer for a little,123.24,4.799
bit longer yes yes right right,124.74,5.64
um I've been an engineer already for uh,128.039,4.741
like three years I have years of,130.38,4.859
experience mostly in c-sharp that's like,132.78,4.14
my main language,135.239,4.5
so yeah I I got into this project like,136.92,5.52
three weeks ago with data and same with,139.739,4.86
python so basically I've just been uh,142.44,4.56
converting everything I know from C,144.599,3.78
sharp to python,147.0,3.599
and it's it's been a great experience,148.379,4.08
learning a lot of things and it's pretty,150.599,6.181
cool excellent so you know obviously uh,152.459,7.981
Auto GPT baby AGI chaos GPT all that is,156.78,6.239
relatively new gpt4 just came out you,160.44,5.1
know what four or five weeks ago,163.019,6.121
um and so uh like walk us through the,165.54,5.88
the the process because I've I've had,169.14,3.66
the privilege of kind of watching you,171.42,2.819
guys figure stuff out from the beginning,172.8,4.56
but can you tell us a little bit about,174.239,4.981
um like where you got started what were,177.36,4.92
the problems that you've you've overcome,179.22,4.799
um and and what problems you're working,182.28,3.66
on now,184.019,4.621
yeah right the uh oh God the biggest,185.94,5.04
problem that we ran into to start with,188.64,6.84
was just finding a a host for a language,190.98,8.399
model that had a working API,195.48,7.619
um we you know we started with uh llama,199.379,8.64
CPP but it's python API didn't work on,203.099,9.541
uh x64 windows for whatever reason,208.019,7.321
um there were a couple of others that we,212.64,4.379
worked with and then eventually we came,215.34,5.459
around to uba booga because and it's,217.019,6.241
really they've got the API fix now we,220.799,4.44
have people using it,223.26,4.979
um right the the the real advantage of,225.239,6.301
uba Booga is that it gives us,228.239,6.0
um it lets us host multiple language,231.54,6.66
models so we're not just stuck with,234.239,7.441
um you know with llama or alpaca,238.2,4.14
um,241.68,3.96
the the new vikuna model is working in,242.34,7.14
there it has some of the science,245.64,6.54
um and coding language models available,249.48,5.339
to it as well we haven't done a whole,252.18,4.92
lot of testing on that we've switched,254.819,7.38
over to open AI on the 3.5 model just to,257.1,7.2
rapidly iterate on the development,262.199,5.341
process and save some resources at home,264.3,4.86
Mexican,267.54,3.78
and by the time we did that was mostly,269.16,5.039
because we were stuck with the uba Booga,271.32,6.36
API and openai was working so we just,274.199,5.161
like if we want to develop we need to,277.68,3.32
switch fast to that,279.36,7.68
and then the auto GPT code base was,281.0,9.699
let's just say it was not ideal I spent,287.04,7.14
an entire day trying to fix that yeah so,290.699,7.081
so we ended up moving over to Baby AGI,294.18,9.06
we got baby working on uh on uh open Ai,297.78,7.58
and Uber booga,303.24,5.399
temporarily and then uba Booga broke,305.36,5.02
their API,308.639,5.34
um for like two weeks which was wild but,310.38,5.94
they've got it back working now,313.979,4.621
um the people who are using our our new,316.32,6.18
project some of them are using the Uber,318.6,4.76
um,322.5,2.4
API,323.36,4.6
for and running it against vicuno which,324.9,5.88
was the original gold so it works we're,327.96,5.1
trying to keep it stable as we as best,330.78,5.759
we can for those people and uh we're you,333.06,6.66
know super excited honestly that all of,336.539,4.621
this is coming together,339.72,2.819
excellent,341.16,5.759
uh data database and Ansel you guys have,342.539,7.5
been working super hard on what you call,346.919,6.12
the salience uh module,350.039,4.5
um can you talk us through that a little,353.039,3.121
bit because that that is a pretty,354.539,4.801
fascinating set of problems so I have it,356.16,5.58
loaded up right here uh as you can see,359.34,6.419
it's just a simple python script uh one,361.74,6.179
of the main reasons we decided to work,365.759,5.101
on our our own project was because,367.919,6.421
autodpt was not made to be,370.86,6.72
um extensively modified I spent an,374.34,6.299
entire day just trying to modify it,377.58,6.48
myself and I had a pretty bad time I'm,380.639,6.541
not gonna lie so we decided to just take,384.06,5.639
the minimum viable version there was and,387.18,4.859
start our own architecture,389.699,5.401
so we created this little salience Loop,392.039,4.141
here,395.1,4.08
um we basically start by instantiating,396.18,5.28
all the agents that we have,399.18,3.06
um,401.46,3.72
and then we start looping through it,402.24,6.54
loading a task list and everything,405.18,5.7
um so let me just run for you first,408.78,4.859
see python,410.88,5.34
we can edit this out in post right we,413.639,5.101
can edit this out in post yeah,416.22,6.06
the uh the the logic Loop that we built,418.74,5.22
okay cool,422.28,5.359
yeah basically okay,423.96,3.679
and one Saturday okay,428.52,4.98
got this pulled up,431.3,4.54
um yep wait I'm watching it can you zoom,433.5,3.66
in a little bit and kind of tell us a,435.84,2.76
little bit about what this diagram is,437.16,2.64
there we go,438.6,4.62
right so this is this is the the logic,439.8,7.679
Loop and one of the really helpful or,443.22,6.9
not helpful but the one of the things,447.479,5.041
that we've kind of designed from the,450.12,6.54
get-go on uh,452.52,6.78
on on this,456.66,6.0
uh system was to try to make it,459.3,3.959
um,462.66,4.979
easy to build agents and logic Loops in,463.259,6.541
a very short fashion so that you can,467.639,4.381
iterate quickly,469.8,4.2
um as opposed to you know if you wanted,472.02,5.88
to to change the logic that open AI uses,474.0,7.199
or even baby AGI really it really you,477.9,5.1
really have to understand the entire,481.199,4.321
code base that they're using and how,483.0,4.74
different things interact we're trying,485.52,4.56
to build a streamline flow where you can,487.74,5.76
just write a logic Loop for the the,490.08,5.7
actual interaction where it takes the,493.5,4.8
prompt and then decides where to store,495.78,6.24
it and you chain them together,498.3,5.1
um so,502.02,5.399
in a similar fashion to the way that uh,503.4,6.78
that auto GPT Works,507.419,5.581
um right now in the salience agent we're,510.18,6.06
working with a predefined task list,513.0,5.539
um baby AGI generates the task list,516.24,4.799
automatically and later we're going to,518.539,6.341
do a hybrid where we integrate both but,521.039,5.821
for now we're just working with the a,524.88,3.899
predefined task list,526.86,4.44
so you know we sort the task list,528.779,5.161
because every time every time you make,531.3,8.34
an edit to a an item in chroma DB it,533.94,9.42
falls to the end of the table or,539.64,5.759
collection is what they're calling,543.36,4.2
Vector databases,545.399,5.161
um but we sort the list we filter it,547.56,4.68
during the sort we filter for any,550.56,5.459
completed tests past the current task to,552.24,6.36
the job agent,556.019,4.861
um so what the job agent does is it,558.6,4.859
searches the vector database or the,560.88,6.3
results database or any any results,563.459,7.201
related to the current task summarizes,567.18,6.12
those passes them down to the Java agent,570.66,6.9
the job agent then passes the the,573.3,8.82
contacts so that's your current task the,577.56,9.6
um the related results from the summary,582.12,9.54
um and a couple of other like metadata,587.16,7.26
related things and it runs it runs in,591.66,5.58
the the execute process which is right,594.42,6.359
now identical to what baby was using but,597.24,6.42
we've broken the The Prompt for that out,600.779,9.841
into a easily editable Json file that,603.66,9.96
um so so that you can you you can change,610.62,5.94
that prompt on the Fly okay the execute,613.62,5.279
agent pulls pulls the prompt from the,616.56,5.1
juson runs it and then passes the,618.899,5.761
results back to the job agent,621.66,6.0
um we we don't have frustration,624.66,4.92
implemented yet but that's going to be,627.66,6.179
based on the uh write-up that you did on,629.58,6.78
Reddit the other day yeah,633.839,6.0
um into kind of check things out and,636.36,6.18
then once it passes frustration or skips,639.839,5.461
frustration then we pass those results,642.54,6.299
back to the results database and it goes,645.3,6.719
to the analyst agent so the analyst,648.839,6.361
agent is here do I have everything there,652.019,6.421
we go the analyst agent is here,655.2,4.98
and it's a little bit more complex,658.44,3.899
actually even though it looks simpler,660.18,4.56
than that loopy stuff,662.339,5.041
um we're actually taking multiple items,664.74,6.48
or more items and adding them,667.38,7.92
to to the uh prompt so let me basically,671.22,6.9
ask do it let me let me pause you for a,675.3,5.279
second and and ask how much of this,678.12,5.88
architecture did you borrow from other,680.579,5.041
projects and how much of this is stuff,684.0,4.5
that you guys have have uh created on,685.62,4.26
your own,688.5,4.62
so right now the only thing that,689.88,6.12
borrowed is the execution,693.12,6.719
and it's right now essentially just the,696.0,7.32
prompt so okay we what we did was we,699.839,8.281
took we took baby agi's architecture and,703.32,7.019
broke it out into lots of different,708.12,5.88
pieces and then we changed all of those,710.339,6.721
pieces completely,714.0,4.459
um and,717.06,6.18
re-imported them or reintegrated them as,718.459,8.081
classes in multiple,723.24,7.38
um in multiple uh agents right,726.54,9.299
originally Baby AGI was only,730.62,7.86
um was only one script and it was less,735.839,4.821
than 200 lines that was the entire,738.48,6.96
purpose of baby AGI right um we're over,740.66,10.06
3 000 lines on this code repo now,745.44,7.5
um yeah we're averaging about nice up,750.72,3.44
counting,752.94,4.92
stop counting so so you've you've taken,754.16,5.56
kind of the initial idea and really,757.86,4.919
you've gone 10x on it you've made it,759.72,6.239
more modular you've added a few new,762.779,5.221
ideas new Loops,765.959,4.141
um you know it like the one of the,768.0,3.12
modules that you haven't haven't,770.1,3.0
integrated yet I think earlier you said,771.12,3.42
you haven't seen a need for like the,773.1,4.14
frustration uh signal yet,774.54,6.359
um but yes so I haven't gotten into into,777.24,6.0
very complex tasks yet so maybe that's,780.899,4.44
why we haven't right uh needed,783.24,4.2
frustration yet but I'm I'm guessing we,785.339,3.361
are gonna need it at some point yeah,787.44,4.26
that's fair so for absolutely for any,788.7,5.639
viewers who um who aren't familiar with,791.7,4.439
with the write-up that I did on on,794.339,4.62
frustration as a signal basically it is,796.139,4.801
you keep a ratio that the simplest,798.959,4.44
version is you keep a ratio of successes,800.94,5.699
to failures and if the failure rate goes,803.399,5.94
up then that tells you to maybe switch,806.639,7.561
to a smarter uh smarter model or you can,809.339,7.62
back up and try a different approach or,814.2,4.319
eventually you can just stop and ask for,816.959,3.12
help there's any number of things that,818.519,3.12
you can do with the frustration signal,820.079,4.741
but basically uh this is a frustration,821.639,5.88
is a key component of cognitive control,824.82,5.34
which is how humans say hey I expected,827.519,4.32
to be making more forward progress and,830.16,4.08
I'm not and so if you're not making the,831.839,4.321
progress that you expect then it's time,834.24,3.659
to either change your approach ask for,836.16,4.44
help or change something so thanks for,837.899,5.161
giving some context did you want to go,840.6,3.96
ahead and jump into your your analyst,843.06,3.48
agent,844.56,4.98
yeah yeah absolutely,846.54,5.06
um,849.54,2.06
so the analyst agent is,852.06,3.6
[Music],855.03,1.35
um,855.66,3.239
he is a little bit more complex because,856.38,4.8
we're taking you know we're taking the,858.899,3.261
results,861.18,4.14
the the most recent results and the,862.16,4.119
current,865.32,5.959
plus we're getting the summary uh from,866.279,9.721
uh from this agent of the job agent or,871.279,8.86
salience agent as well and we're we're,876.0,8.22
feeding it the the uh a little bit more,880.139,8.64
context on the task we're asking the,884.22,8.64
um the the gbt agent to decide is this,888.779,7.981
job completed yet and if it's you know,892.86,8.82
if not then what we can do is send if,896.76,7.259
it's if it's reached the frustration,901.68,6.599
level we can send that task back to the,904.019,6.781
task creation agent,908.279,6.06
um to generate sub tasks potentially or,910.8,5.339
new tasks,914.339,4.68
um and then we're we're also,916.139,6.181
um this weekend going to set up sending,919.019,7.921
feedback back into the agent and you'll,922.32,6.9
see what that feedback the feedback I,926.94,4.68
think is going to really change the game,929.22,7.44
on getting uh fewer fewer runs to to,931.62,7.38
complete a particular task and then of,936.66,4.38
course if it does complete it and we,939.0,5.1
update the task with the task scheduler,941.04,4.44
um and,944.1,4.02
um and and tell it to move on to the,945.48,4.68
next task,948.12,2.7
um,950.16,3.539
did did you get the terminal resize yeah,950.82,4.8
absolutely I'm already sharing my screen,953.699,4.621
again yeah yes I could not figure out,955.62,4.5
how to get it recycle what I actually,958.32,5.699
did was resized uh your monitor display,960.12,5.959
yeah,964.019,4.921
all right I actually increase the the,966.079,6.101
scale of it anyway just zoom in,968.94,6.18
um all right uh Ansel you ready uh uh,972.18,6.899
database you were used sufficiently,975.12,6.12
yeah yeah okay that's where we're at,979.079,5.7
right now okay cool enough in the future,981.24,5.52
so Ansel,984.779,4.081
um give us a little bit of uh a little,986.76,3.66
bit of background what is it that you're,988.86,4.5
about to to demonstrate for us well I'm,990.42,6.539
about to run the salience uh loop uh as,993.36,4.979
you can see it's asking me a few things,996.959,3.781
it's asking me if I want to start from,998.339,4.5
where it left off basically restart the,1000.74,4.14
previous date and if I want to allow,1002.839,3.961
auto mode or manual mode for now let's,1004.88,3.54
just go with manual mode,1006.8,3.959
so and it's starting to initialize the,1008.42,5.039
agents here if you want to explain go,1010.759,5.88
ahead data well I was going to say,1013.459,5.461
um the the restore from previous state,1016.639,4.681
is to my knowledge not something as,1018.92,5.639
present in any other uh any other,1021.32,7.019
autonomous AI system out there yet so,1024.559,6.721
really cool stuff yeah so tell us a,1028.339,4.321
little bit about that feature because it,1031.28,4.2
sounds sounds valuable like you can uh,1032.66,3.96
you're in the middle of cooking dinner,1035.48,2.579
and someone knocks on the door and you,1036.62,2.76
come back and remember where you were,1038.059,3.62
right,1039.38,2.299
right well you could always just leave,1042.62,5.64
it sitting there because the system has,1044.72,6.48
this as you can see it it has a auto,1048.26,5.7
mode or a manual mode since we set,1051.2,4.979
manual mode every Loop is going to stop,1053.96,5.64
and ask us for feedback,1056.179,5.341
um all right,1059.6,4.86
um but if you have to turn your computer,1061.52,6.48
off because the FBI show up then,1064.46,4.64
um,1068.0,3.72
then uh you you know when you power it,1069.1,3.76
back up,1071.72,4.8
um the uh the the eight all of the task,1072.86,6.84
statuses or and results are stored,1076.52,6.6
persistently in chroma DB so you never,1079.7,6.54
lose any data unless you tell it you,1083.12,5.1
want you don't want to restore from a,1086.24,4.58
previous state,1088.22,2.6
you want to go ahead and go ahead and so,1091.1,3.6
in this case I basically wipe this,1093.02,3.96
memory it doesn't know anything uh I'm,1094.7,4.5
just gonna say yes just to keep it,1096.98,3.36
simple for now,1099.2,4.08
and it started to run the salience agent,1100.34,6.48
and a salience agent is going to call,1103.28,6.36
the execution agent to start uh running,1106.82,5.76
the task list for now the task lists are,1109.64,7.14
defined uh from our Persona Json so what,1112.58,5.94
we wanted to do was separate the code,1116.78,3.96
from the prompts so whoever wants to,1118.52,4.68
develop agents can simply look at two,1120.74,3.9
things uh,1123.2,3.479
the code of the agent what it's supposed,1124.64,4.68
to do logically and they can uh alter,1126.679,5.401
the prompts on the Json file so I'll,1129.32,4.38
show I'll show you the Json file in a,1132.08,3.9
second right now the first task is well,1133.7,4.68
to develop a test list to achieve a goal,1135.98,4.14
which in this case I don't believe I'm,1138.38,4.08
printing it here,1140.12,4.32
um yes it's supposed to create a program,1142.46,4.68
for an AI to search the internet,1144.44,5.7
so it did the first task let's just go,1147.14,4.919
ahead and let it run again,1150.14,4.14
it's going for the second task right now,1152.059,4.74
it's only showing you the the info like,1154.28,3.96
what you really want to see it's doing,1156.799,3.481
you're not seeing all the processing,1158.24,4.38
behind there's also a debug mode you can,1160.28,3.06
do,1162.62,2.4
um that starts showing the prompts,1163.34,4.02
what's going in what's in the database I,1165.02,4.38
can run another loop later uh with that,1167.36,4.8
turned on uh to show you how what it's,1169.4,4.26
doing internally,1172.16,4.44
and it'll it'll just keep going uh,1173.66,5.519
apparently it's reflected on the task so,1176.6,4.26
yeah let's see us leave it in auto mode,1179.179,4.38
I can change it to auto mode,1180.86,5.04
and it's now in auto mode it's you can,1183.559,4.201
see it says you can press escape to,1185.9,3.54
return to Automotive so at any point in,1187.76,3.24
the code if I press Escape it's just,1189.44,3.119
going to say switching back to auto mode,1191.0,4.02
and it'll stop and ask me uh for,1192.559,5.761
feedback again excellent at at this,1195.02,6.42
point I don't know if it's gonna say if,1198.32,5.16
it completed the task or not because the,1201.44,5.4
last task is to add out the,1203.48,6.72
the task list it made but it very much,1206.84,7.14
depends on the of the status agent if it,1210.2,6.839
decides it did it or not in this case it,1213.98,5.579
says it completed it and it also the,1217.039,4.861
good thing is it it provides a reasoning,1219.559,4.74
for why it decided it's completed or not,1221.9,5.22
so you can also uh get more information,1224.299,4.981
as to why that,1227.12,2.88
um,1229.28,3.72
it took that decision and help you do,1230.0,4.44
better prompt engineering so you can,1233.0,4.559
change the prompt easier so and if Let's,1234.44,6.3
uh let's pause for a second because,1237.559,5.521
um I I think that there's there's a,1240.74,3.86
tremendous amount of significance here,1243.08,5.099
so like walk us through the steps like,1244.6,8.079
it from from scratch it it had a what,1248.179,6.541
what appeared to be a metacognitive step,1252.679,4.021
where it said okay what am I doing,1254.72,4.1
and then it designed its own task list,1256.7,5.24
followed through the task list and then,1258.82,5.68
also evaluated its own performance to,1261.94,4.96
know when it had finished,1264.5,5.94
correct okay yes so here it's basically,1266.9,5.7
where it starts let's see,1270.44,4.979
yep at the beginning so the first thing,1272.6,4.98
it does is start to develop a task list,1275.419,4.38
and it's searching for results in its,1277.58,4.86
memory since it's wiped it says there is,1279.799,3.62
no results,1282.44,4.02
uh it sends that results to summary,1283.419,7.301
since no results uh exist yet it just,1286.46,6.3
says no previous actions have been taken,1290.72,5.28
because this is this is fed to I don't,1292.76,4.5
remember if it's fed to the execution,1296.0,5.46
ages or if it's fed to the status agent,1297.26,6.38
one of those two,1301.46,6.24
diplomatically oh both of them okay,1303.64,6.159
um so that the agents at least know at,1307.7,5.82
what step of uh the process uh they are,1309.799,5.401
or what any other actions they've taken,1313.52,3.84
before it's basically like a very,1315.2,4.8
primitive sort of memory as to well okay,1317.36,5.699
what have I done before right right,1320.0,6.24
and basically we send uh all of that,1323.059,5.161
information to the agent it runs and,1326.24,5.76
here's the result we get from the uh llm,1328.22,7.02
which is well it did create a task list,1332.0,5.46
so at that point we sent those results,1335.24,5.1
to the status agent and it's going to,1337.46,5.459
check if that task has been completed in,1340.34,4.68
this case it says yes the task has been,1342.919,4.021
completed and it provides the reasoning,1345.02,4.62
right here or why it thinks it's done so,1346.94,5.58
and at that point we go back into the,1349.64,4.68
loop so it's going to ask me again do,1352.52,5.159
you want to uh continue or not go ahead,1354.32,6.96
data yeah so I I want to come back to to,1357.679,7.5
the uh status agent so not only are we,1361.28,7.2
returning the completed or the status of,1365.179,6.061
whether or not it's completed we're also,1368.48,4.98
providing feedback so if you get a,1371.24,5.04
status that's not completed and this,1373.46,4.92
isn't in yet but this should be,1376.28,5.1
relatively simple to incorporate one of,1378.38,5.22
the things that we can do is provide,1381.38,5.1
additional feedback to the execution,1383.6,6.3
agent the next time it runs so,1386.48,6.66
um and in this situation,1389.9,6.659
um what we would want to do to to help,1393.14,7.68
automate this is pass it the um the,1396.559,6.841
feedback from the status agent so the,1400.82,4.739
status agent is going to act as kind of,1403.4,4.8
like a project manager,1405.559,5.341
um where it has all of its little Dev,1408.2,4.979
Minions that go out and do their their,1410.9,4.2
things and then they come back with,1413.179,4.38
their results and they print or present,1415.1,7.68
a a a summarized result of some kind now,1417.559,7.321
um that execution agent needs to be,1422.78,5.1
expanded upon so that we can get web,1424.88,5.34
searching for research and code,1427.88,6.779
execution to or to do code testing and,1430.22,8.64
stuff we may even have a code QA agent,1434.659,6.121
separately,1438.86,3.9
um but but,1440.78,4.44
um we we can take that feedback from the,1442.76,5.64
project manager status agent and push it,1445.22,6.6
back in then if if the agent fails so,1448.4,5.519
that it can do better the next time and,1451.82,5.64
it's not always running into the same,1453.919,7.201
results yeah um so let's let's unpack,1457.46,5.16
that real quick,1461.12,3.84
um yeah because I think we should switch,1462.62,6.179
over to the um to the personas default,1464.96,6.9
just on I've got this hey hey Data,1468.799,5.221
before we before we proceed because uh,1471.86,4.08
uh what you're what you're talking about,1474.02,5.76
is is like cognitive neuroscience and a,1475.94,5.7
lot of people are going to be lost so,1479.78,4.44
the status agent that's the supervisor,1481.64,3.539
right,1484.22,3.3
that kind of sits above the other things,1485.179,4.681
and it dispatches let me make sure I got,1487.52,4.26
it right it dispatches the execution,1489.86,5.28
agent says hey this is your set of tasks,1491.78,5.82
to go do and then that uses the salience,1495.14,5.76
loop in order to Loop through uh and,1497.6,5.76
ensure that that task gets completed is,1500.9,4.08
that the is that the correct well it's,1503.36,4.199
it's the salience it's the salience loop,1504.98,5.939
that calls the execution agent and,1507.559,7.081
Status agent separately okay so so the,1510.919,6.12
salience is what binds it together right,1514.64,4.8
exactly the same okay this one together,1517.039,5.581
correct okay and so by by iterating on,1519.44,5.76
that you are you're checking multiple,1522.62,3.96
times,1525.2,4.2
um for uh whether you're whether you're,1526.58,4.38
being successful the feedback that,1529.4,3.6
you're getting and it can because it's,1530.96,4.56
iterative it can adapt its strategy as,1533.0,3.96
it goes,1535.52,5.519
right so and it's kind of talking to,1536.96,6.36
itself in this sense so it's really more,1541.039,6.0
it's it's less like corporate developer,1543.32,6.32
type of environment and more just like,1547.039,6.0
internal monologue like if you were you,1549.64,5.8
know building a,1553.039,4.441
um if you were building a raised garden,1555.44,4.859
bed for example now you're like okay,1557.48,5.16
first thing I need to do is take,1560.299,4.201
measurements and then I want to cut wood,1562.64,3.84
and then I want to nail wood together,1564.5,4.2
finally we put it in place and then we,1566.48,4.74
fill it with you you you kind of like,1568.7,4.979
outline that and then you know as you're,1571.22,4.8
cutting wood like okay is this actually,1573.679,4.081
gonna work are these gonna fit together,1576.02,5.82
is this nail gonna stick through so it's,1577.76,6.18
we're trying to replicate that that,1581.84,5.52
internal monologue yeah so it shouldn't,1583.94,6.42
it shouldn't theoretically take any jobs,1587.36,5.699
from any developers anytime soon at,1590.36,4.679
least not any more than that GPT is,1593.059,3.961
already sure,1595.039,4.5
the point is is that you you have a self,1597.02,5.58
self reflection and and kind of a set of,1599.539,5.701
medical metacognitive functions that are,1602.6,4.38
evaluating what it's doing and it's,1605.24,3.12
thinking through what it's doing every,1606.98,3.42
step of the way iteratively all right,1608.36,3.66
cool all right so take it away I think,1610.4,3.18
you said you wanted to move over to,1612.02,3.6
personas or something else,1613.58,5.4
yeah so um all well you were asking,1615.62,5.7
about how these all,1618.98,6.72
um interconnect so I've got the Persona,1621.32,7.56
just on file pulled up here,1625.7,7.92
um the the the primary goal is for,1628.88,7.02
anybody who's just picking the software,1633.62,3.9
up that doesn't want to learn how to,1635.9,3.3
program,1637.52,3.779
um the the most that they're ever going,1639.2,4.26
to have to touch if they just use the,1641.299,6.661
pre-existing agents is this file and the,1643.46,6.0
main Loop,1647.96,4.44
um so this file is where we set down our,1649.46,5.94
prompt so you know the system prompt is,1652.4,6.54
just like the system prompt in,1655.4,6.18
um in open AI,1658.94,5.7
um and then these other prompts go in as,1661.58,5.459
user,1664.64,5.7
um so we're taking it kind of like the,1667.039,6.961
way that you send chat history into chat,1670.34,8.939
GP or gpt4 when you're using a a home,1674.0,6.9
bot,1679.279,5.121
um in in the same way we're we're using,1680.9,7.32
these these variables here to inject,1684.4,8.86
data from other agents into this bot,1688.22,8.4
um in a in a human readable format and,1693.26,6.06
then every bot we can we can set whether,1696.62,7.74
it's going to use uh GPT 3.5 or 3.4 uh,1699.32,6.78
the plan is to also be able to leverage,1704.36,5.34
local models if you're running Uber,1706.1,6.84
um or you know you we can we eventually,1709.7,6.599
plan to incorporate hugging face,1712.94,5.099
um and,1716.299,4.26
um you know maybe maybe at some point,1718.039,4.441
Lane chain we'll see,1720.559,5.521
um but but the each each agent gets its,1722.48,8.579
own parameters for sending to the for,1726.08,7.56
sending to the language model,1731.059,5.701
um and you can't minimize here but but,1733.64,5.34
and we can you know,1736.76,6.539
um each each agent has its prompt,1738.98,7.319
exposed without sure without any code,1743.299,8.581
and of course the main Loop is so simple,1746.299,8.281
um I'm going where is Salient so here's,1751.88,4.5
salient,1754.58,5.459
um so if you ignore the like the the,1756.38,6.84
beginning section here you can see that,1760.039,6.24
our entire Loop is,1763.22,7.38
five lines of code commentation comments,1766.279,8.64
and yeah the it's it's super simple for,1770.6,7.14
anybody to kind of go in,1774.919,5.88
um and build this uh you know this logic,1777.74,4.5
Loop,1780.799,2.521
um,1782.24,6.0
one issue right now is is that the,1783.32,9.239
um where are you uh so the Salient agent,1788.24,7.799
run kind of has agents embedded in it so,1792.559,6.541
you can't modify it but you can also,1796.039,4.801
just use those,1799.1,5.84
um agents outright so our main pie,1800.84,7.62
function this is actually this is,1804.94,6.719
actually the uh,1808.46,7.14
baby AGI that's,1811.659,7.721
um been rebuilt into our system,1815.6,6.6
um so you can you know you have a really,1819.38,6.299
simple starting point right um and then,1822.2,6.12
when you go into,1825.679,5.581
um when you go into the agents like,1828.32,6.12
that's getting a little bit more complex,1831.26,7.26
um but you should be able to also build,1834.44,7.78
agents without having to,1838.52,4.32
[Music],1842.22,1.28
um,1842.84,3.839
without having to to dig into the rest,1843.5,5.52
of the code base,1846.679,7.441
um so all of the the loop of the agent,1849.02,9.659
could all be self-contained and most of,1854.12,7.32
these functions are already templated,1858.679,6.061
out for you so it's almost Plug and Play,1861.44,5.4
we're not there yet but we're getting,1864.74,4.08
there,1866.84,4.68
so let's uh let's talk about this,1868.82,4.079
architecture real quick because you've,1871.52,4.56
got you've got several layers of nested,1872.899,6.0
Loops you've got a modular and,1876.08,5.28
configurable architecture and like you,1878.899,4.14
said the goal is to get to Plug and Play,1881.36,5.16
So and you're almost there so talk us,1883.039,6.12
through a little bit about like,1886.52,4.379
how you came to some of these,1889.159,3.541
architectural decisions because for,1890.899,3.961
someone who's only been you know coding,1892.7,3.54
for three or four weeks you've kind of,1894.86,3.96
unpacked a lot of really important,1896.24,4.74
Concepts and development,1898.82,3.719
um and I know that you've had uh Ansel,1900.98,3.96
to help you out so maybe you can't take,1902.539,4.38
full credit and then chat GPT as well,1904.94,5.219
but so you know you guys you guys have,1906.919,6.0
uh I I see you like in chat like all day,1910.159,4.921
every day some days,1912.919,4.74
um so so talk us through like,1915.08,6.14
like how you came to some of these ideas,1917.659,7.26
and and what like what is what is the,1921.22,7.72
overall status of like how how close are,1924.919,6.301
you to having something that is it looks,1928.94,4.32
like a semi-autonomous or fully,1931.22,4.26
autonomous what are the missing steps,1933.26,4.26
and you know who do you want to get in,1935.48,3.84
touch with you and and so on and so,1937.52,4.139
forth where do you need help,1939.32,4.62
well you know Bill Gates wants to give,1941.659,4.561
me a call I will totally take his call,1943.94,4.92
now I only charge two thousand dollars,1946.22,6.179
uh an hour which you know is is pretty,1948.86,5.76
sure,1952.399,6.121
um the uh yeah the loop is there the the,1954.62,6.0
loop is already there,1958.52,5.519
um at least when you compare it to,1960.62,5.4
um Auto GPT which,1964.039,4.201
kind of has a functioning Loop too I,1966.02,4.74
haven't looked at their stuff since I I,1968.24,5.159
started this project and even even,1970.76,4.86
before I went like you know we had we,1973.399,5.941
had chaos GPT was looping at least so so,1975.62,5.76
it Loops,1979.34,2.88
um,1981.38,3.84
so now what we're doing is expanding,1982.22,4.52
functionality,1985.22,3.12
and,1986.74,5.819
um and uh for expanding functionality,1988.34,8.339
and we're expanding uh compatibility,1992.559,7.541
with these other systems we've got AP uh,1996.679,6.181
open AI working we've got Uber working,2000.1,4.559
which should work with any other,2002.86,3.78
language model,2004.659,4.74
um chroma is is done manually so it,2006.64,5.46
doesn't rely on Lang chain but if we,2009.399,5.821
incorporated Lane chain that would that,2012.1,5.819
would open up functionality in a much,2015.22,6.6
broader capacity right yeah I was just,2017.919,6.0
about to ask for extensibility for,2021.82,4.38
instance like you know can this right,2023.919,4.681
what code can it talk to apis yeah go,2026.2,4.199
for it Ansel now now that you mentioned,2028.6,3.66
extensibility that was the main reason,2030.399,4.5
for developing this because like I,2032.26,4.919
mentioned before I try to mess with auto,2034.899,4.201
GPT to get it to like do something I,2037.179,4.5
wanted just for testing purposes and I I,2039.1,4.74
wish it wasted my time trying to modify,2041.679,5.401
it because it's it's it's I don't want,2043.84,5.759
to say it like like the develop didn't,2047.08,4.259
pay attention to it but it's just the,2049.599,3.421
structure of it is,2051.339,5.641
is not user friendly at all so we when,2053.02,6.659
we went into this the the main goal was,2056.98,6.359
to make it extensible organized uh and,2059.679,6.42
very easy to know what's going on where,2063.339,5.641
um in their defense I don't think that,2066.099,5.701
they were writing Auto GPT for other,2068.98,5.34
people to come in and and mess with it,2071.8,5.64
and make it do what they wanted,2074.32,6.299
um you know we from the ground up we've,2077.44,4.86
been building this so that other people,2080.619,4.681
can come in behind us right and really,2082.3,6.72
what we need uh more than anything is,2085.3,6.78
for people to come in and and build new,2089.02,6.18
agents for example this execution agent,2092.08,4.799
right here,2095.2,5.419
um the the execution agent needs to be,2096.879,7.861
expanded in a much more broad fashion,2100.619,8.021
um the and then we need to work on the,2104.74,6.48
task generation right so,2108.64,5.219
um baby relied heavily on task,2111.22,6.18
generation to determine if it had,2113.859,5.881
completed a step,2117.4,5.64
um and and that ended up with it redoing,2119.74,6.66
a lot of work which we didn't like but,2123.04,5.22
um I think that that could be solved,2126.4,3.179
with,2128.26,1.92
um,2129.579,5.701
with with a um with a set in stone,2130.18,9.3
primary goals that and allowing it to,2135.28,7.38
generate sub goals so if somebody wants,2139.48,6.48
to take a crack at task generation we,2142.66,7.08
actually have a a kind of uh prototype,2145.96,6.6
outline although this is a you know we,2149.74,4.98
haven't actually done any work on this,2152.56,4.98
stage yet but those are those are the,2154.72,5.52
two the two threat things that we have,2157.54,5.88
going forward to improve its,2160.24,6.359
um capacity is building out execution,2163.42,4.86
agent,2166.599,5.161
um rebuilding this task generation and,2168.28,4.559
then,2171.76,3.24
um you know a few odds and ends here and,2172.839,4.981
there just to to make it run a little,2175.0,4.98
bit more smoothly,2177.82,6.18
um I find that the biggest gains are are,2179.98,8.16
usually from the uh from prompt,2184.0,7.56
engineering though and sometimes even,2188.14,8.88
small changes to the the text and the uh,2191.56,7.92
in the Persona wording is everything,2197.02,6.18
yeah yeah wording is everything,2199.48,6.96
um like let's see if I can find yeah so,2203.2,6.72
here here's the status agent and it just,2206.44,6.06
does not,2209.92,6.6
yeah it's too long but yeah we've been a,2212.5,5.76
good hour working on the prompt for,2216.52,3.96
status agent and,2218.26,5.099
um ultimately we we we still ended up,2220.48,6.0
having to switch over to gpt4 for this,2223.359,5.941
so yeah the running version that we have,2226.48,4.98
right now is using two separate language,2229.3,5.64
models oh cool it can handle multiple,2231.46,5.1
language models,2234.94,3.48
um so you've got you've got a lot of the,2236.56,4.68
extensibility and and configurability,2238.42,5.159
figured out so you mentioned earlier,2241.24,3.839
that you're getting close to making it,2243.579,4.081
plug and play and and like you said uh,2245.079,4.861
all the all the different agents are,2247.66,5.28
configurable just via Json,2249.94,5.82
um and and you you're getting close to,2252.94,6.3
adding some more extensible uh things so,2255.76,6.0
how do people collaborate with you do,2259.24,5.58
you just all go through GitHub,2261.76,4.92
um how to how should if anyone wants to,2264.82,3.96
jump in what's the best way for them to,2266.68,4.38
to get in touch with you guys,2268.78,6.18
so we have a we have a project set up in,2271.06,6.36
the community project section of the,2274.96,4.92
cognitive AI Discord,2277.42,5.1
um specifically for agents,2279.88,6.0
um but if there's other stuff that you,2282.52,5.579
think that that you can contribute to,2285.88,4.8
for example we'd love to have a,2288.099,5.821
graphical interface right on the UI it's,2290.68,6.3
similar to Lang flow,2293.92,5.939
um to to make this even lower code for,2296.98,4.44
people,2299.859,2.341
um,2301.42,3.659
and then you know of course the you know,2302.2,8.46
the uh if there's a um a new model that,2305.079,9.121
comes out or a new host somewhere and,2310.66,6.24
they have their own API that isn't open,2314.2,5.52
AI compatible,2316.9,6.959
um writing um writing API interfaces for,2319.72,7.98
that so for example the the current V2,2323.859,6.541
version of,2327.7,3.96
um,2330.4,5.34
of the Uber Booga API was not actually,2331.66,8.64
written by us this was written by,2335.74,6.06
um,2340.3,3.9
that's so strange that he's not listed,2341.8,4.74
as a contributor here,2344.2,2.94
um,2346.54,3.0
I think his name was Max though it was,2347.14,6.719
Max something but he fixed it,2349.54,10.02
yeah oh uh mitchko in mitchko so he did,2353.859,8.581
get credit for this one but he wrote,2359.56,5.94
um he rewrote the the interface for the,2362.44,6.6
uba Booga API and people are are out,2365.5,6.0
there using this and it's working you,2369.04,5.88
can see the the old codes listed here,2371.5,4.92
um right because another thing you can,2374.92,5.34
do is also host the your local model on,2376.42,5.939
a Cloud Server and use the API to,2380.26,3.839
connect to it and you can just run the,2382.359,3.541
the interface on your phone or web,2384.099,4.201
browser or whatever interesting you can,2385.9,6.84
see the um the endpoint code is listed,2388.3,7.559
here as a variable and what we'll do,2392.74,6.9
later is come in and set this as,2395.859,7.861
configurable within the ini file,2399.64,6.3
um right oh yeah that's that's something,2403.72,5.94
else I I should just touch on briefly,2405.94,6.06
um things like determining,2409.66,4.679
um which language model you're using,2412.0,5.52
which database you want to use you're in,2414.339,5.161
embeddings,2417.52,2.64
um,2419.5,3.3
function that you're using all of that,2420.16,4.14
is,2422.8,4.559
um all of that is configurable here in,2424.3,6.96
the config.ini file as well so yeah we,2427.359,8.161
for we really try to get as much of the,2431.26,7.319
configuration as we could out of the,2435.52,5.52
code and into,2438.579,3.061
um,2441.04,3.5
into files that that people are gonna,2441.64,7.28
find useful wonderful,2444.54,4.38
so what's what's next I know you guys,2449.26,3.96
have outlined a lot of open problems and,2451.3,3.36
some things that you need help with and,2453.22,3.66
I think there's quite a few people that,2454.66,4.8
are expressing some interest,2456.88,3.959
um so where where do you guys want to,2459.46,3.72
take this project like what you know,2460.839,5.341
given a month or six months or if you,2463.18,5.1
have 20 people jump in What do you want,2466.18,3.96
it to be able to do,2468.28,5.52
well this weekend we are taking it to a,2470.14,7.08
hackathon and we don't know what the,2473.8,6.299
objective is yet but we're going our,2477.22,6.24
objective is to win the hackathon with,2480.099,7.141
this out of left field uh AGI uh,2483.46,5.76
framework that nobody has ever seen,2487.24,4.02
before,2489.22,5.76
but beyond that,2491.26,5.099
um,2494.98,5.099
beyond that I'd like to see this,2496.359,7.861
see this just used more broadly,2500.079,5.821
um you know it,2504.22,3.3
this is all,2505.9,6.9
um this is all in a what what uh,2507.52,9.12
be about there it is I missed but this,2512.8,6.5
is all in the gnu general public license,2516.64,6.66
just like Linux so anybody can take this,2519.3,8.62
right and use it for commercial purposes,2523.3,6.6
um,2527.92,5.1
with you know with restrictions,2529.9,5.76
um and you know me personally I would,2533.02,5.52
love to once we get this model pounded,2535.66,5.4
out and and set in stone I I would love,2538.54,7.44
to just build and bespoke AI cognitive,2541.06,8.46
systems for for companies on either a,2545.98,4.92
direct,2549.52,1.98
um,2550.9,3.42
you know direct to,2551.5,8.099
um Market kind of solution or even in a,2554.32,8.039
research capacity you know my you know,2559.599,6.601
I've already built some agents or at,2562.359,6.48
least one agent and then some made some,2566.2,5.1
modifications as well,2568.839,5.52
um to work on the heuristic imperatives,2571.3,5.88
project that you've got going Dave yeah,2574.359,7.801
um and I I really want to be this the I,2577.18,6.48
want this to be an engine for your,2582.16,4.5
project so that you can iterate quickly,2583.66,6.659
and you don't get hung up in in the,2586.66,8.58
details of you know working manually or,2590.319,7.081
writing your having to write your own,2595.24,3.78
code,2597.4,4.38
um for every single little thing that,2599.02,5.579
you do we can just have an agent and,2601.78,5.579
plug those agents in spit data at it and,2604.599,6.5
see what comes out there you go,2607.359,3.74
is basically like empowerment right like,2613.079,7.301
human language has given us the ability,2617.22,5.32
to do everything we've done up until,2620.38,3.959
this point and we've basically given,2622.54,5.4
that uh ability to machines now and with,2624.339,5.401
this kind of architecture we can give,2627.94,4.74
that to everybody as long as they have a,2629.74,4.92
model to run they could have some kind,2632.68,4.439
of agent working for them so it might,2634.66,5.04
sound science fiction or outlandish but,2637.119,5.22
it's basically like constructing your,2639.7,4.44
own Jarvis right,2642.339,4.381
that's the goal,2644.14,5.34
that's the goal yep yeah I do not want,2646.72,6.359
to be the person who develops the uh the,2649.48,6.78
the the next Raven though let me just,2653.079,5.701
say that like you know I'll be a part of,2656.26,6.5
it but the uh a autonomous assistant,2658.78,7.38
is going to be a lot of work oh,2662.76,5.26
definitely that's why I said the,2666.16,4.919
ultimate goal yeah just looking at what,2668.02,4.86
we've done already,2671.079,6.181
um a you know like a better uh Siri for,2672.88,9.3
example is a a mini person project not,2677.26,8.16
just two guys in a garage,2682.18,5.34
yeah,2685.42,6.24
yeah so this is like the very first you,2687.52,5.18
know,2691.66,4.14
2022 there were very few of us even,2692.7,5.1
talking about cognitive architecture,2695.8,6.6
2023 hits and suddenly like everyone's a,2697.8,6.4
cognitive architect but you guys are,2702.4,5.699
definitely like lunging ahead forward,2704.2,6.3
um which is just incredible to see,2708.099,4.141
um so thanks for sharing your work so,2710.5,3.0
far,2712.24,2.94
um and I know that you guys have have a,2713.5,3.359
lot of work that you're you're working,2715.18,4.62
on this this is uh it's getting close to,2716.859,6.0
being uh semi-autonomous it doesn't,2719.8,4.86
doesn't quite come up with its own,2722.859,4.441
objectives but it can autonomously work,2724.66,5.76
through a lot of uh increase it's coming,2727.3,4.98
it's coming and it but it can it can,2730.42,4.08
think through open-ended objectives that,2732.28,4.44
you give it which is really incredible,2734.5,5.52
so um any final thoughts from uh from,2736.72,5.16
you guys or anyone else,2740.02,3.96
um before we wrap up today's tonight's,2741.88,3.959
recording,2743.98,4.98
yeah well if you had told me uh in,2745.839,5.881
December that I would be diving into a,2748.96,4.98
project on GitHub about cognitive,2751.72,4.02
architecture I would have spat my beer,2753.94,2.899
at you,2755.74,5.16
so everything is changing really fast,2756.839,6.28
it's getting really silly it is,2760.9,4.38
impossible to keep up and it's exciting,2763.119,3.781
and it's just,2765.28,3.42
it's just thrilling to be working on,2766.9,5.52
this yeah absolutely yeah and possibly,2768.7,7.919
keep up is definitely definitely the the,2772.42,7.38
situation right now even even with the,2776.619,6.121
past week or so where AI news has been,2779.8,5.519
relatively slow there haven't been a,2782.74,5.82
whole lot of interviews here lately,2785.319,6.961
um I've still got dozens of tabs open of,2788.56,5.82
AI stuff that I just haven't been able,2792.28,4.02
to read yet,2794.38,3.979
um,2796.3,2.059
I I really think that,2798.52,10.099
um a lot of people are are,2802.54,6.079
getting a way more concerned about,2808.66,7.08
artificial intelligence than they need,2812.44,5.46
to be we still need to be concerned we,2815.74,4.68
need to be working on on these things,2817.9,4.38
but we're not,2820.42,5.22
we're not quite in runaway AI yet I hate,2822.28,6.18
to disagree with you Dave but,2825.64,7.56
um you know it's it it if we told AI to,2828.46,7.68
build better AI that would be like,2833.2,5.159
telling you to do brain surgery on,2836.14,4.4
yourself,2838.359,2.181
so we don't need to bomb the data,2840.819,2.52
centers,2842.44,3.12
no not but yeah we should definitely,2843.339,4.5
have bombs in the data centers just in,2845.56,5.64
case just in case okay and local backups,2847.839,5.28
for your your shows or anything you want,2851.2,3.84
to watch after the apocalypse so there,2853.119,3.72
you go just,2855.04,3.96
etch it out on clay tablets all right,2856.839,4.141
gang thanks everything thanks for,2859.0,2.94
everything,2860.98,2.46
um and definitely keep up the good work,2861.94,3.78
and uh yeah I hope that you get some,2863.44,4.44
more folks reaching out to you we will,2865.72,4.8
of course have ongoing updates maybe not,2867.88,4.199
weekly we'll see,2870.52,3.059
um but you know we're doing more live,2872.079,3.301
streams lately,2873.579,3.0
um just to keep everyone up to date,2875.38,2.459
because as you guys mentioned it's,2876.579,3.121
happening so fast,2877.839,3.541
um and it's good to show people what's,2879.7,3.6
possible so I'm gonna go ahead and stop,2881.38,5.52
the go ahead one one more thing I think,2883.3,5.94
that getting everybody involved in,2886.9,5.88
talking about this is really the the,2889.24,7.079
best way forward yeah um because you,2892.78,6.66
know up until now all of the discussion,2896.319,8.701
about Ai and ethics and morality of AI,2899.44,10.74
has been siled and gate kept by the you,2905.02,7.799
know a few people at these major,2910.18,6.48
corporations yeah and you know you and,2912.819,6.3
the communities that you've helped get,2916.66,5.28
get off the ground I think has really,2919.119,6.841
opened it up and is going to be one of,2921.94,6.24
the most important things for the future,2925.96,5.159
of AI it's definitely already having an,2928.18,5.34
impact well thanks so much and I'm glad,2931.119,3.901
that it's definitely having hopefully a,2933.52,3.98
positive impact in the long run yeah,2935.02,6.0
we'll see yep yeah all right gang I'm,2937.5,5.14
gonna knock off the recording have a,2941.02,3.12
good night everybody and we'll talk,2942.64,5.0
again soon cheers,2944.14,3.5