Join us as we delve into the fascinating world of Artificial Intelligence (AI) by interviewing the brightest minds and exploring cutting-edge projects. From innovative ideas to groundbreaking individuals, we're here to uncover the latest developments and thought-provoking discussions in the AI space.
0:35
I wanted to ask so let's say you do you say you make a really complicated task
0:40
you go through like multiple multiple loops and you finally get it to work right now now I exit off and everything
0:47
like that now is there a way that it stores that prompt that I can go back in
0:53
and kind of just build off that last Loop or do I have to go through all the weeks again hi y'all I'm Nick Tindall I
0:58
am a developer on auto GPT a AI agent that can help you accomplish tasks
1:05
um my background is a little complicated I've done a little bit of everything in the engineering World
1:13
um well Computer Engineering World um I went to school for computer engineering haven't finished yet one day
1:18
though um I started that many moons ago and I've worked at a couple different
1:24
companies as a devops engineer data analyst um head of engineering and just general
1:30
software engineer done all kinds of roles now I am at dialexa an IBM company
1:36
where I am a consultant what school did you go to uh I'm still going to the
1:42
University of North Texas I live in Denton we have there's IBM some kind of building IBM buildings in Cleveland as
1:49
well I don't know what I'm sure they're everywhere but I just noticed that they're a pretty big building so yeah
1:55
it's a good world to be in Consulting so yeah and as I said kind of before we got
2:00
on here I I'm kind of have roots in Texas as well that's where I did my Master's studies not at Baylor even
2:06
though I'm wearing the Baylor shirt I got this one for free so I guess I'll go with that why not wear free shirts but
2:12
right uh yeah Texas is great I love it there um but anyways uh to get in kind of the
2:19
project a little bit Auto GPT for people who don't know what what is auto GPT
2:27
rogbt is an experimental AI agent that's goal is to accomplish directed tasks
2:34
that are assigned by a user right at its core it works by splitting a problem
2:39
apart and attempting to solve it step by step and outputting things along the way
2:46
very cool so how does that work for like okay let's say I'm I'm the average Joe and I think well that sounds pretty cool
2:53
have some somebody accomplish my task for me how do I do that how does it work practically yeah so we go through what
3:01
we call the agent Loop um so we'll get tickets out a little bit but for the average Joe you'd basically
3:08
start it up and our install process is non-trivial right now we're working on that but it started up you uh it would
3:15
ask you a couple questions such as hey what are you trying to accomplish right what's your objective what's your task
3:20
you'd give it all that info and it would start step by step going through what we call the agent Loop that agent Loop is
3:27
basically a breakdown of a couple different things such as hey what is my
3:33
next plan right take planet's next steps that is hey uh in order to
3:41
say I don't know research this project right I need to Google the project or X
3:48
Y or Z right and from there it will plan to do something else right and then
3:53
execute and then it'll plan and then execute and plan and execute over and over and over as iterates towards its
4:00
goal right along the way we do a couple different things such as breaking down your goals allowing it to execute
4:06
commands and asking the user for clarification right so as part of this agent Loop
4:12
there is a step that you can optionally turn off and what we call continuous mode I don't recommend that if you don't like
4:18
big bills but you can basically tell it don't ask me
4:24
for feedback just go for it right otherwise every step or every instance
4:30
of the loop you'll be asked to provide feedback or tell it yes you can do that command
4:37
right these commands could be lots of different things the simplests are reader writer file
4:42
the more complex ones are interact with an API or call a different llm right all
4:49
the way up to making another agent to do things right that doesn't work super great right now but it was a step along
4:55
the experimentation pathway that we're on but as a person trying to use it you
5:00
would follow this process right and at the end it would output various artifacts using
5:06
these commands such as a file or it could tweet well not
5:11
anymore because Twitter locked on their API it could post a blue sky or it could send an email right there's
5:18
various things it can do and at the end it just stops so how long does the process usually
5:25
take or how long do the loops take uh each Loop in and of itself can take anywhere from five seconds if it's
5:32
really bad right if you throw in here uh you know this is an experiment after all um to around 25 to 30 seconds for each
5:39
iteration right because behind a loop is a whole bunch of different steps
5:45
um each one of these can have multiple like two the like open AI service
5:51
so each one of those takes a little bit of time to execute and then it repeats so 30 seconds per step about
5:58
um depends on your internet speed really okay that's what I was going to ask does it depend on your internet speed or
6:04
what's the what's kind of the bottleneck there with the time yeah open AI is uh bandwidth realistically
6:11
you have like for the apis you have an array of apis
6:16
or a list of AI apis that you guys put into Auto GPT so like say if you want to order a pizza or something like
6:22
that you'd have to tie into a Domino's API or a Papa John's API or something like that is that how it works yeah so
6:29
we have a couple different types of tie-ins right we have our commands which are direct
6:36
built into the agent itself right that's for base level things like reading or writing to a file
6:42
or calling an API there's also plugins which is where your Domino's example would live is somebody could build out a
6:47
Domino's plugin which would provide another command that allows it to order pizza
6:52
or to send a text message right we have a pretty healthy ecosystem of those that
6:58
are pretty expensive so you're it's just a matter of adding so it sounds like the core you have certain apis that you deal
7:06
with like opening open AIS API but then sort of Fringe ones or specific ones Niche ones those are up to contributors
7:13
of that we'll make plugins for auto GPT exactly okay cool all right so just a little uh
7:21
another thing what do you do for auto GPT what's your role with the platform I
7:26
I wear mini hats um I do a lot of organization right now so that is orchestrating our various
7:33
open source teams to accomplish things in the fields of research development or
7:39
project management and Community engagement um those hats are usually right now just
7:46
orchestrating all the different moving Parts scheduling meetings I am technically what we call a lead
7:52
maintainer I think because of that title or maintainer Plus or whatever uh the titles don't really matter I suppose
7:59
we're all volunteers um but I I organize a lot of stuff
8:04
I also work on designing our re-arc a rear architecture with James
8:09
and doing our CI CD pipelines and general code reviews
8:14
I'm pretty busy at work these days so I don't have a ton of time for in-depth on
8:20
the ground coding but I do a lot of reviewing of other people's work and ensuring it follows good design patterns
8:26
and is actually helpful it's the project how did the can you give us a little background on how the project sort of
8:33
got started like who started it and then to this point so many months back uh
8:40
tauren started this project as like an experiment right and it still is an experiment we haven't released a
8:45
non-experimental version of this just so setting expectations um but it started a while back with
8:53
tauren just trying to upload a little python Loop that did a simple agent Loop where it would
8:58
plan ask for feedback repeat and execute and repeat right super simple basically
9:05
one python file um then we had a bunch of people really liked that idea and people started as I
9:13
describe it putting one ball of mud on top of another uh
9:18
so our architecture got messy because at the early stages it was just like an
9:24
idea that was fun and we just pretty much accepted any PR that
9:29
added a feature right so are you are you making the commits yourself uh not like
9:36
at this but this is before I joined okay I joined during this process I started
9:41
organizing this a little bit immediately uh so we just accepted anything that people would join come in um and a lot
9:48
of that was volunteer right like Community contributions not tauren or any of the other devs you can look at
9:54
our contributors contributors list is huge um I think it's like 300 people have worked
10:00
on the project yeah that's insane yeah it's insane um so we just accepted basically
10:06
anything um whenever I joined we started to focus on actual features that improved the
10:12
user experiment experience or enhanced it in some way right you
10:18
couldn't just change this line to get your name in the project anymore around this time we started to work on
10:24
developing the plug-in system so that we didn't have to deal with the hundreds of PRS that were coming in
10:31
there was a ton of comments tons of PR's it was just super overwhelming at the
10:36
time the team was probably around 10 people and that was just not sustainable right
10:42
for 10 people to manage 100 PR's a day is just impossible right because the five you don't merge
10:47
yesterday get updates tomorrow and now you have 15 now that it's 30 right with
10:53
issues yeah right yeah it grows uh and you can see that still we still
10:58
have PRS that haven't been merged we do our PR thumb Thursdays now decline try and clear them out but all on time but I
11:06
started by working on fixing the CI CD pipelines to help evaluate these PRS
11:13
that they came in to try and reduce the bottleneck on maintainers running these PRS locally and evaluating
11:20
whether or not this code is good by stripping out the things that fail our tests first or that
11:26
don't pass our code coverage guidelines right doing a lot of basic level hey we need to filter these into good
11:32
quality ones up front what what code what's it mainly written in what language
11:38
um most of the project is written in Python almost all of it actually okay
11:43
there's a little bit most projects are written in Python nowadays so yeah it's really easy for
11:49
people to pick up and learn so that's really helpful as well so you hinted at this but what would you say is the biggest complication with auto GPT right
11:58
now so the biggest complication is our convoluted code base and we're working on resolving it actually
12:04
uh we have just recently merged a core of a new architected system that is
12:10
easier to contribute to easier to swap components out of and overall just less
12:16
tightly coupled into a spaghetti bowl and more into a well interfaced system and why why do you say it's very
12:22
convoluted what's what's going wrong there um so we have our plugin architecture
12:27
was designed without a ton of context on what types
12:33
of things people would want to change out so we work off a hookspace system which allows you to intercept the code operate
12:41
flow it'll probably 15 different points and just change it right so the code
12:47
flow can respond different things and the expectations of what can be responded to that are complicated right
12:55
alongside that there is Imports between files that shouldn't really do that because oh this piece of code
13:03
was useful to this person I can just pull this in and reuse it right we have lots of that where there is no
13:10
separation between like module sections where they say the thing that counts
13:15
tokens could be used in 14 different places and there's just 14 copies of it
13:20
right so if you change it one place you gotta change it 14 other places whereas that should really just be a thing that
13:26
counts tokens as its own entity that's not a great example because I think we already removed that but we're
13:33
working towards cleaning it up yeah what do you think what do you think is like the biggest do you see that is the biggest threat to your success without a
13:40
GPT or do you see some other bigger threats that you're working to overcome I see I see that as a huge threat uh it
13:47
makes it hard for people to pick up our project and just use it um especially developers who want to build
13:53
on top of it right building a plan right now is at best heart
13:59
um swapping out our llm provider is basically impossible right
14:04
um we've been working to make some of that stuff easier right whether that is
14:10
adding the capability of swapping out the base URL which we now have so you can use like
14:16
GPT for all or you could right we're doing like these intermediate Solutions but the biggest
14:23
problems is it's really hard for more into the system and then even harder for people to use
14:29
the system to accomplish things they want with changes they've made right
14:34
if you wanted to change the agent link for example which is something our research team does often
14:40
you have to go rewrite like 15 files right because the agent Loop flows
14:45
through all these files and really complicated code paths whereas with our redesigned stuff it's you change the
14:51
planning module out and it has a straight up interface where anybody can rewrite their own planning module and as
14:58
long as they correspond to that interface that's your new planning module right so we worked really hard to replace some
15:03
of these things um but for my code perspective that's our biggest problem for sure
15:09
now you guys are open source project correct yep we're MIT licensed all right and is there a company formed around
15:15
this like a non-profit or something like that or uh we're working on figuring a lot of this out uh it kind of blew up
15:21
overnight uh as far as like time scales go for forming a company or figuring all that
15:27
out right now it's under the significant gravitas organization just for
15:33
hey I can't live under one guy is forever uh you can't have team management and stuff like that under one guys
15:39
um I'm hoping we're gonna figure something out soon um but I'm not the guy to do that yeah
15:44
and it seems like that like the most important thing is the code so if you take care of that everything else will
15:49
take care of itself you know exactly yeah if you keep your code high quality and usable and people want to use your
15:55
stuff yeah the other problems will solve themselves sure one of the things is are
16:00
you mentioned when we first got on on this this um podcast is
16:05
do you guys are working on a more friendly uh user-friendly UI interface
16:11
um what so that's I'm assuming that's just like a website someone can go to and type in their what they want the
16:17
agent to do yeah exactly so you can actually sign up for our wait list for it oh cool the
16:23
link is at adpt.co uh there should be a
16:29
register where you can uh get wait listed uh we're working on a lot of
16:35
stuff to make that possible the main part was that re-architecture we talked about to be able to swap out the user
16:41
facing portions right which is right now a command line application with something more web-facing like an
16:47
API right are you able to see the projects that
16:52
are being done by Auto GPT um
16:59
people share a lot with us um some of it's not asked for um we don't have any uh like data
17:06
recording right now we're looking into integrating with some just to detect bugs but we get a lot of people who come back
17:14
with some very interesting use cases um I think the most
17:19
realistic of some of them are the prosumer help me do X Y or Z thing realistically
17:25
um hey I need to write 4 000 emails right maybe four thousands a lot but hey I need to go research this thing right
17:32
go build a package for me bring it back right whereas there's some people who
17:38
ask it to research I don't know
17:45
like they're ancestors right like that's something that is theoretically possible um it's not going to find a ton of
17:51
success uh because it can't go to the library right um not yet you know we'll see what we
17:57
can do about that but yeah we get some really weird uh requests for what people wanted to do
18:04
you can only hear your imagination when we go take you so far there yeah yeah do
18:10
you um so what do you see as the future of this project looking like okay so
18:17
that was my chair sorry
18:22
um the future of this project I think is a ton of potential um internally in the team we've discussed a
18:27
couple different things uh the re-architecture makes us more capable to be a library so tools like
18:34
GPT engineer or GPT researcher or all these different various projects which Implement a very
18:40
similar audio GPT like based run Loop instead of having to reconstruct all these tools by hand and like building
18:47
out reading file modules or all these different things could instead build out
18:52
tools on top of ours right because we're an open source project and they won't have to deal with all the
18:58
upkeep of all the extraneous stuff right um for example GPT engineer is an
19:05
awesome project but at its core it's a series of very excellent prompts right
19:10
and around that they do some things for handling user questioning and stuff like that right which could easily be built
19:17
on top of our re-architecture and save them a ton of time right so that's one way I see it
19:24
as a tool for developers right as a tool for consumers I can see this in the
19:29
prosumer area first um acting similar to what I described earlier a research tool or a simple
19:36
thing you can assign workflows to right hey I need to research and email these
19:43
20 perspective like leads right like so I assume I'm in sales I need to research and email these
19:49
20 prospective leads go tell me everything you can about these companies and these people and
19:54
I'll review the emails and shoot them out right throw them in my drafts or whatever I can see those as very realistic in
20:00
your term solution like near-term things long term I think it's a little more up in the air
20:07
um we can all Imagine The Pie in the Sky of oh wow it's an AGI that can do anything
20:12
I think that's a little unrealistic um and also maybe you don't want that
20:20
we're not going to dig into that too much but there is a lot of potential in this project and
20:27
I'm really excited to see where it goes but it seems like and tell me if I'm wrong is the the only bottleneck to how
20:33
big this can grow is what the uh the plugins can do like so if there's an API for something you guys can tell the
20:40
external application to do whether order be order a pizza order food somewhere
20:47
um set up an appointment to get your oil changed like as long as there's an API the the it's endless you know the
20:53
possibilities are endless at least right we've seen other projects come out that are UI automation tools where it can
20:59
automate your browser right and we're looking at how those maybe could integrate so it's not even just
21:04
limitations of the apis it's limitations of what you can do at a computer right there's a lot of
21:10
potential Avenues we can go down but the ecosystem is really important to it right and
21:17
that's part of why we've gone through so much effort doing this re-architecture is now instead of building a very very
21:23
complicated plug-in you can Implement one interface and congrats now Auto GPT
21:29
can use it right okay we've been working with a couple different teams trying to get some of
21:34
those better uh get some initial versions of those going but I'm not ready to talk about those yet sure now it seems like this is
21:42
kind of the direction that AI is going uh do less work have ai do more of the
21:48
work for you do you see this kind of as the general direction of AI or what other areas of
21:54
AI do you think are uh going to develop more rapidly in the
21:59
next few years I think this one's going to get a lot of focus um not necessarily because it deserves
22:04
it but because it's going to make things more efficient right
22:11
everything that can make something more productive gets a ton of attention right you can look back in time through
22:18
the you know.com or the like Cloud transitions or all these
22:25
different like Web 2.0 you know all these different things and realistically it all boil down to making an existing
22:31
process for efficient um AI is going to revolutionize that for sure right you can already see that with uh
22:37
10 000 SAS this startups that have popped up for hey I can generate this X
22:43
Y or Z thing right and that's awesome but I think we'll see a lot of those slowly fade away as tools that are
22:50
capable as multimodal or similar come out that are able to you know do any of
22:56
that because you should ask it to what's the what would you say life the the top few prompts for auto GPT
23:04
that work really well right now because like you said it's an experiment like someone tries to go in there and type
23:10
out their their biggest dream it's not going to work right because it's not yeah that far enough along yet maybe
23:16
it's not a possible ever but what are some things that work really well right now for prompts to put into Auto GPT
23:22
that someone can do right now one of our maintainers is actually really good at making these
23:27
um couple suggestions to start off with it'll generate you a default prompt based off a one-liner
23:33
you can use that but if you go ahead and change those default goals you can make it a lot more efficient
23:39
another thing is you can put Gates behind things right so you can say
23:45
if this file exists do that as part of your goals or if x or y do Z that can
23:53
help a ton right it struggles to keep track over long periods of times so if
23:58
you can give it ways to keep it on track that's really helpful right so an example let me see if I can
24:04
pull it up that Luke Luke made a
24:11
little dock on this let me get it sorry I gotta log into Discord you know because what's the Discord channel that
24:17
people could go to it's going to be prompt showcase and that's just for auto GPT
24:22
um yeah it's in our Discord Channel um for example right to do
24:30
he added a couple goals and doing them manually is really helpful right you can put a ton of content into these
24:36
um for example he wrote I I'll just send a deal this is a terrible thing to read
24:43
haha uh if if X-File doesn't exist do y then do
24:50
Z then a b c so on and so forth um you can look at GPT engineer for
24:57
example of course um help on how to write these prompts well and then you
25:02
can do other steps so he says after you've done all that in step two
25:07
do Z right once the file exists and then finally
25:14
whenever you've done that and this other file exists do this right
25:20
um I'll send you all over that's All in One initial prompt or is it you ask it you see if a file exists and that it
25:27
wait for it to see if it exists or not and then enter a new prompt yeah yeah so GPT Auto GPT can have a series of goals
25:34
right um and it has an objective and goals so whenever you
25:40
run it by hand like automatically as there's not a real good word for it but in the default mode It'll ask you for a
25:48
one-line prompt and it'll generate these by hand it'll generate these automatically and it'll turn your uh one line into an
25:55
objective and five or so goals or you can write those by hand
26:00
um you should definitely write them by hand for right now if you want it to be highly effective right it can be pretty
26:06
useful if you don't but I've sent a link in our private chat here showing how you can do this very
26:13
well right yeah we got it cool I wanted to ask so let's say you do you say you make
26:20
a really complicated task you go through like multiple multiple loops and you finally
26:27
get it to work right now I exit off and everything like that
26:32
now is there a way that it stores that prompt that I can go back in and kind of
26:37
just build off that last Loop or do I have to go through all the loops again uh not yet so you'd be able to keep the
26:43
artifacts you have artifacts right and that's where stuff like what I was talking about was if x like if x
26:49
verifiable step exists continue right so that it's able to pick that
26:54
back up but we're working on making the agents more capable of being stopped and started and keeping those contexts in
27:00
that um that history if you will so that whenever you're saying hey
27:07
I gotta go to lunch because you know I'm hungry don't run without me I can stop it and come back
27:13
and pick that back up right or hey you know my bill is forty dollars right now from open AI because
27:20
they like money yeah I'll come back later right yeah
27:26
I mean what have you ever seen like or seen anything happen where the bill gets
27:32
to like a thousand bucks for someone or something like that like something that um not for an individual but for like
27:37
our organization and testing uh we get a Big Bill yeah
27:43
um I'm not the person who pays it so I'm not going to dig into it too much but okay okay
27:50
our Mainline contributors and maintainers we help supplement that gospel yeah so what what is the cost
27:55
yeah so Auto gbt in and of itself doesn't actually cost money
28:00
um right we're an open source project however we call out to various apis that do right
28:07
um whether that's often that's open AI where they charge you for every API call right so hey I
28:16
want to complete this like we provide a prompt and it kind of like Chad jpg
28:21
where it replies back right we do the same thing but programmatically um not quite the same but we're not
28:27
going to dig into that we do a similar thing programmatically where we send it a message and it replies to us but that costs money right
28:35
usually it's three to Fifteen cents depending on how much it applies and how much we send it
28:41
um for gbt4 those costs can add up pretty quickly because each time you go through
28:47
the agent Loop it makes three to five of these calls right so each step can cost anywhere
28:52
from nine cents to a dollar right ideally not right we're working on
28:58
some ways to make that a little cheaper looking into some alternatives and there is G like that's part of the
29:05
reason we enabled changing your API endpoints so you can use tools like GPT for all and there's a couple others we
29:12
recommend I think in our repo for reducing those costs so it can run it locally using different models right by
29:20
far that is the biggest cost is What openai charges our users directly I did
29:25
I did think when I installed the auto GPT I did take your advice setting up the limits in opening eyes so that's
29:30
nice they have that at least you can say only spend five dollars and then after that cut me off so that that will help a
29:37
lot of people do you get any complaints on Discord or anything or do people just basically go to open AI a lot of
29:43
complaints um a lot of people think it's us that is charging the money I think right that's
29:48
what I'm thinking and trust me dog I get the bills too okay
29:53
um I sympathize with the people who are frustrated by the costs yeah especially with the people who are very new to
30:00
using AI tools and aren't super great at constructing prompts it can be really frustrating whenever it didn't work and
30:07
you got a bill right right um do you see a way that that could be
30:13
fixed in the future or is it just kind of see what it is what do you think the solution looks like
30:18
um there's a couple different solutions on the horizon um prompting as is will probably
30:24
become more and more flexible over time right the exact things I'm looking for will become
30:30
easier to get the AI to produce right and you can see a lot of that with the
30:36
examples of I don't know um open AI posted about the death of prompting at some point or somebody from their team
30:42
that I can't remember the actual post but prompting as a whole will get significantly easier over time and then
30:49
as we support other llm providers right um those prompts will probably get easier
30:56
because they'll compete naturally and the cost will go down right yeah that's economics people yeah
31:05
foreign very cool so you said you have a background in this
31:12
sort of stuff how did you get into AI okay I'm gonna sound like uh I'm gonna
31:19
sound like a total bro so I was head of engineering for crypto company uh
31:26
TI right I was doing verifiable identities and
31:31
okay uh they're a cool company it's uh called sonar
31:36
um yeah they're great company they're working on some very cool Tech um
31:42
and I ended up leaving that um it was a lot of work uh really
31:47
stressful um and I started to just do Consulting
31:53
uh and in my free time I was working on whatever the latest you know Cutting Edge technology was looking into
32:00
generative AI because we we played with it a little bit at our uh crypto thing because you know as you do right you're
32:07
on one of the technology you might as well look at another I would put Jonathan a little bit so I started looking into it a little more because I
32:13
thought I sounded pretty cool uh and got a little bit more familiar and saw Auto GPT saw they definitely needed help
32:19
managing their repo and that was something I had a lot of experience with from my time working with open source in
32:26
the past with the Arcane algorithm archive which is a niche of Anisha vanish
32:32
um and then figured I could help a little so started writing pipelines and stuff to
32:38
make it a little more efficient now you can still doing consulting on the side or you yes that's actually my
32:44
Mainline job right now um yeah so hey if y'all need Consulting
32:49
yeah Consulting in terms of just like pretty much anything related to python code or so I actually worked for a
32:55
full-on consultancy we do uh product delivery design and everything in between
33:01
um it's pretty cool stuff try to keep my work and you know play
33:07
Separate though and you said chat or you said uh Auto GPT has how many people
33:13
volunteering right now um so that's a complicated number
33:19
um I was working on figuring this out so I could actually come with and answer that question
33:25
um we have about 10 mainland maintainers uh we have
33:31
approximately 50 what we call catalysts those are people who have actively
33:36
contributed in a health very helpful way to the project and then we have a hundred or more uh
33:43
like actual code contributors or people who have contributed in non-good ways right whether that's writing
33:49
documentation working on social media publishing like one of our guys Works does a really
33:54
good job working on our YouTube videos explaining how to do some of this stuff right uh he's also done code though but
34:01
we have a ton of people I think it's over 150
34:06
in our Discord um probably 50 of which are active
34:11
uh at any moment working on stuff oh that's really awesome yeah we are
34:17
definitely a worldwide organization that is not an organization and that's insane and that's why I spend a lot of time
34:23
organizing how do you guys stay organized how do you stay on the same page and get things
34:29
across so we work a lot in GitHub issues my favorite issue number is 4770 which
34:36
is detailing our re-architecture um we spend a lot of time communicating there and in Discord and we host weekly
34:43
maintainers meetings um I will definitely say this is an area that we struggle with for sure uh
34:50
there's not really a good example of how open source teams structure themselves right right that aren't backed by some
34:57
external company or whatever right you can look at w3c but they're like a whole
35:02
entity or you can look at react but there's a company behind that right
35:07
there's a lot of New Field games we play right mostly
35:14
maintainers recently got an email so we can actually schedule events right rather than sending around calendar
35:20
links hoping for the best um but we're actively working on that
35:26
problem because it's it's a real challenge yeah it seems like an interesting Dynamic so when I was an
35:31
undergrad I went for business management something that always stuck out to me was how different businesses were
35:37
structured So when you say that it kind of piques my interest as you know I as you say that I don't I
35:43
don't think there is really a model yet for how these open source companies are structured it seems difficult I mean
35:49
you have a lot of people with a lot of different backgrounds it's difficult to get everybody on the same page it's difficult to communicate uh also so you
35:58
know that's that's a challenge I think a lot of companies are facing right now and I think somebody somebody's
36:04
eventually going to start to figure some stuff out yeah there's a couple of companies working a little more publicly than
36:10
others um right we've taken some looks at like gitlab for example
36:15
on how they organize their building and public teams right or I don't know what
36:21
they call them exactly but like they have a public first like employee handbook
36:26
and stuff like that right and we've also been looking at the home assistant project s and I was just working on that today
36:33
so oh nice yeah I have a home assistant instance at home and I actually did a presentation on work at work about some
36:40
of the automations we've been doing with it in our office that's pretty cool yeah it's so cool but we've been looking at them for
36:48
a little bit of inspiration as well but yeah it's a real challenge
36:53
if anybody uh has the perfect solution let me know have you thought about I mean not going back getting into crypto
37:00
or anything but like setting up a dollar or something like that so it'd be sort of an autonomous company
37:06
I've talked with a couple people about it I don't think it'd be a perfect fit for us yeah um just because we operate
37:12
in so many different jurisdictions um paying people with a dow is complicated I actually was restored of
37:20
developer now which I'll may or may not know about um it is a Dao community of developers
37:26
who work to mentor and grow people and
37:31
their development life cycle um awesome group of people love them all
37:36
just don't have enough time to do both um
37:41
yeah we've I thought about it a lot um yeah and how we can best manage it but it's a real challenge right
37:48
you can't you can't operate it down every jurisdiction right I think the only
37:54
place I know like I looked into it a little bit too last year when they were when the craze was happening like
37:59
Wyoming lets you set one up but then like you said if you have to pay someone in a different country it sort of
38:04
complicates things right and then we're from day like if we started a company for example we'd be International from
38:11
day one I don't think a single core maintainer lives in the same state in the US um half of them don't live in the same
38:17
country um right and I don't think any of the other core maintainers outside of the US live
38:23
in the same country right we have guys in Netherlands we have guys and New Zealand we have guys and you
38:31
know the UK and Germany right like from the literal start if we did start a
38:38
company for example it would be a problem and a dow would just complicate that a
38:43
lot yeah a regulation nightmare it seems like like yeah but you'll figure it out
38:49
yeah I I've looked into and talked with the team about a bunch of different options and I frankly don't know what's
38:55
the best yeah it seems like one of those things where you as long as you keep chugging along with the code everything else will take
39:01
care of itself you know Growing Pains yeah um well are you guys are you like where is
39:08
where how are you paying for like server bills and things like that like and how is it yeah posted right now so tauren
39:15
the founder of the project actually set up a GitHub sponsored program so we have various individuals actually
39:22
pay for all that for us right so we received donations on our cool repos
39:28
that basically allow us to pay for our our Google Cloud bills and our open AI
39:33
bills we don't have a ton of bills outside of that other than like tooling for maintainers to use to build our projects like whimsical
39:39
or like G Suite so we can email each other right um though we actually don't have a ton
39:46
of costs right because we don't have health insurance because we're not a company yet right like we
39:52
don't have payroll because who would you pay right like when you
39:57
have 300 people who've contributed it's really hard to make those determinations right
40:04
um the guy who has to decide that if we do that
40:10
and is there a debt to decide that or is it that's what I was gonna ask well yeah how do you guys make decisions is it
40:15
majority rules or so for decisions um generally there's each team member
40:21
like each of our maintainers um is very competent in their area um so they come with a proposal and we
40:28
either accept or deny it uh deny with changes or changes or whatever there's a
40:34
little bit of a process around it but not much of one it's often hey here's what I'm thinking do you all think this is a terrible idea yes or no
40:42
um we don't often run into things where outside of code we have to make determinations
40:47
we have for example monitoring our Discord we have some rules we all agreed to
40:53
and they're all pretty reasonable right like don't spam don't advertise things in our server
41:00
right like all of the decisions that need to be made are pretty straightforward and we're all pretty reasonable people
41:06
so there's not a whole lot of arguments so awesome so far so good
41:14
well I mean if you're doing it with uh you know 100 to 300 people or whatever the actual number is that's pretty
41:20
that's pretty good I mean you'll be able to piece everything together right now without paying anyone having a company
41:25
there so or a company without a company yeah so it's wild um we've built
41:30
informal like organizational structures right so I'm a lead maintainer right which means I can add people to GitHub
41:36
repo and there's others who can move people between teams and there's people who can request to be added or
41:42
so we have a bunch of weird and formal structures that I'm working on identifying
41:47
um just so we can figure out what everybody is following so we could
41:52
at least try and agree to the same pathways so so what is the motivation for most of these people for
41:58
contributing to the project because they're not getting paid for it they're not getting health insurance it's just because they love the project and they
42:04
they're passionate about it yeah for a lot of people they love the project and they're just passionate about what they build right that's very cool a lot of
42:11
people are really excited because it's open source right like they want to contribute to that right they want to
42:17
make it better or their users of the tool they just said I'm tired of this part not working good I can make it
42:23
better right um and we love that right that's the
42:29
kind of eat those we love to support it's sort of like an interviewing process for these people that are coming on board or what's the criteria
42:36
how do you decide who's on board and who's not yeah so if there's an interviewing process or not um It's
42:43
Complicated uh basically anybody can be a contributor right the criteria is
42:49
did you contribute something to the project right um whether that is did you tell us hey
42:55
you're doing great today or did you write code right anybody basically anybody can be a contributor if they just want to be
43:02
um then we have our catalysts we have multi-tiered system right um
43:07
to help these decision-making structures we have catalysts who are contributors
43:13
who contributed multiple times right like hey you didn't
43:18
just tell me I just was doing great today you also helped somebody in the Discord 400 times or whatever right
43:23
that's an extreme example but you clearly care about the project you plan to stay around for more than
43:30
this feature right or whatever right then we have our lead catalysts which are the people who help organize
43:37
those right um they are technically maintainers as well which is the next tier maintainers
43:44
are catalysts who have changed the large-scale shape of the project helped guide it and or helped
43:52
make large-scale decisions around say code movement or architectural patterns
43:58
like large complicated multi multi-step feature development right
44:05
that one has interviews interviews right you get on with a couple maintainers and
44:10
talk to them but by that point you've already been working with us for a while right and your interview was your work
44:17
you've done right right it's not like at a company where hey I want to work here okay apply it's
44:24
like hey I want to work here make a PR right open an issue contribute right
44:30
like it's public just do it and if you're good at that you'll be invited up
44:36
all right I don't know it's kind of weird it's like you have to prove yourself
44:43
kind of you know through your code or marketing or whatever it may be right exactly we actually have some people who
44:49
do marketing who have never written the line of code don't know how to read it right but they're like I can write a
44:54
mean tweet and they do right there's a lot of people that can write that's just as
45:00
important sometimes as writing the code it's for the word so yeah it's different work but it's still just as meaningful
45:06
sure what are like a couple uh things in regard to monetization are you it's
45:13
you're an open source project so it's a little bit different but are you guys looking for VC funding and
45:18
if you are not do you got and you when you release the web interface are you
45:24
going to charge is there going to be a fee for for those props that you know when people type those in or what does
45:30
that look like uh I'm not the guy who would handle that okay I I do a large scale organization of our
45:37
project yeah reactions um we have a couple other team members like Torin who are far better equipped to answer those
45:43
credit questions okay no worries yeah just uh uh I expect he
45:49
has answers but that's not my role okay and you keep saying tauren what's tauren's name the developer oh yeah
45:56
that's touring uh it's just his actual name um tauren Bruce Richards I think that's his last name hold on let me grab it for
46:03
you yeah foreign Bruce Richards um he was the initial Dev uh he was
46:09
actually a Game Dev way back um before this whole thing started I don't know if he is actually anymore but
46:15
that's for him um he would be the kind of guy who helps with the larger scale like
46:22
business decisions right because technically even though this is all MIT license and you can copy it it's all
46:29
under the organization that torn made because he started it and went oh man we
46:35
need to make this an organization now because I can't let other people manage this code at scale without making an
46:40
organization and so on um and he works with a couple others on our team to do that side of stuff
46:46
figuring all that out um I definitely work more on the open source side I'm trying to
46:52
get this community built right and when did the project launch oh man
46:58
launcher is a rough word we haven't released version still launching yeah
47:05
we haven't released version one yet uh so when did it launch is a fun question when will it launch
47:12
that's even worse another fun one um I don't know three four months ago
47:17
maybe yeah it's not it's really not that old and somehow in that time we've became the
47:24
25th largest project on GitHub uh I think a lot of that is due in part to how you have to run the project uh you
47:31
have to clone the repo and when you do that it's like hey you should start this so you can come back to it later uh so
47:37
that helps us grow really quickly and you said you've been with working on it for how long
47:42
um two three months probably um okay so very close time yeah yeah very cool definitely most of his
47:48
lifetime um it's been pretty cool I really love the team really love the project
47:54
we've gone through a lot of large-scale code changes in the time what's the uh biggest area of growth you've seen since
48:01
the beginning our interactions with our community for sure right
48:06
um we struggled very very heavily with
48:11
dealing with the level of press and contributions and issues and
48:19
all manner of things right that come along with hey you're on the front page of news now or hey you're being talked
48:25
about in Congress um in the US at least we we've done a really good job as a team
48:33
adapting to the changes that they came as they came up and rising to meet them
48:40
I'm really proud of the team for that we've done a whole lot of code fixes and improvements but by far uh
48:49
we've done a great job building committee yeah congrats I mean it's it's no small feat to do what you guys are
48:54
doing um yeah it's really awesome can you see like how many prompts I know like right
49:02
now you have to download it locally pretty much to can you see how many prompts total have
49:09
been like submitted to you guys no I can't see how many been submitted we don't do actually any form of metrics right now
49:16
um we're working on integrating with a couple different tools to help identify bugs and that would probably be
49:21
something comes alongside that but one of the cool features we can see
49:26
is how many people have started the repo and how many people have clone the repo
49:31
um that's on our rolling two-week rental but it was crazy waking up some days and
49:36
seeing millions of people have clothing this repo for the first time in the past day right yeah it's like not today maybe but
49:45
like um I I don't even think I can check anymore we've we've been working on it
49:51
the GitHub training um repos a lot and you guys are always up there so yeah at
49:57
the top of everything not just AI stuff yeah we are the 25th largest project on
50:03
GitHub right now uh wow over three to four months yeah exactly
50:08
yeah um every day I check it because I still can't believe that we're here
50:15
um we're passing we've passed huge things like next JS or go or react
50:21
native right you're bigger than next right like
50:26
for Transformers or you know like kubernetes
50:33
node power toys like I I just looked through the list and I'm like wow right
50:39
yeah incredible that is crazy um we we definitely wrote A Rocket To
50:45
The Top yeah and here's the hoping it doesn't stop anytime soon we've got a lot of cool features coming out soon uh
50:51
lots of people have been working really hard on them yeah when's your is there any big feature that's coming out you know I
50:57
know it's hard to give timetables anything with code um yeah but is there some big feature that's going to be on the horizon that
51:03
people can look out for um yeah over the next month or two we're hoping for our rig architecture to start taking
51:11
to start being pulled into the repo in ways that are actually impactful to the user whether that's adding support for
51:18
other llm providers or adding support for new types of
51:23
plugins that are way more powerful or making it easier for developers to
51:29
build their tools on top of it right and then we're also working on making it much easier to install because right now
51:35
as I'm sure you all are aware it's a pain at best and then as part of all these things
51:43
we're hoping it to make it to our agencies are a lot more capable coming up but I can't promise any timelines on
51:49
that last one we've got a bunch of areas of active research we've been working on building
51:55
out research teams so if you're interested in contributing in that way let me know
52:00
um but yeah we got a couple really cool things on the horizon awesome we'll be we'll definitely keep our eye on things
52:06
because it's one of my favorite projects at least so yeah it's very cool yeah yeah and so what is there any one is
52:13
there one plug-in that is your favorite Nicky
52:18
there are some plugins that are just ridiculous to me um that are just show me how much people
52:25
can just make whatever they want um this is gonna sound ridiculous uh there's people who have built amazing
52:31
plugins from you can work with it and Telegram to it can read your text messages right
52:36
but my favorite one is one that just tells you how many astronauts are in space right now
52:42
it's a super simple plug-in that works off a space API um but it was built by a guy who didn't
52:49
know anything about coding and he got a plug-in merge when to our main plug-in repo that and he was really proud of it
52:57
that allowed him to see how many astronauts were in space right and it really goes to show you the power of AI to help a developer do
53:06
something that they're very uncomfortable with and very unfamiliar with yet still accomplish their goals right and it kind of helps show the
53:14
ethos of Auto GPT which is enabling Everyday People
53:20
to accomplish things that they couldn't before right that's a that's a great headline right
53:26
there don't worry I got it
53:34
awesome well you know we try to keep it keep our time timeline on track I don't
53:40
know if this was scheduled for an hour but I think it was around there so yeah we we appreciate your time you've got a lot of a lot of things to do
53:46
uh for your for your consultant work and for auto GPT so thank you so much for uh
53:52
for joining us here today and I'll just leave it up off the hunter if you got anything else yes so thanks for joining
53:58
us Nick we appreciate you a lot uh coming on here uh we hope you'll stay connected with you if there's any future
54:04
projects maybe we can talk a little more you know we can get in depth and maybe some of the updates or things like that
54:09
that would be pretty cool yeah if you're up for it and then uh you know just uh
54:14
you know for people who watch or listen uh just being subscribed Ryan and I have
54:22
have our AI news letter that we do every weekday
54:27
fryhyphen ai.com you can check it out you can see all of our previous newsletters you can see our long form
54:33
articles that we've done and on Sunday which is going to be
54:38
the 17th or something like that I can't keep track of days 16 something like that we're gonna have our when you're
54:44
when you're in this kind of space you lose track of that sort of stuff right you know what I'm talking about oh yeah for sure but we're going to have our our
54:51
article out so when this is out go go out go out down to our article and search it we'll put it in the comments
54:57
we'll put it in the description too for our article about Auto GPT uh so you can read a little more in depth about what
55:03
that is check out the project on GitHub look at their Discord Pages you said and if you want to be a contributor just
55:09
start working for the project that's all you got to do instead so yeah pretty cool just thank you for joining us today
55:15
we really appreciate it and uh you know just we hope you have a great
55:20
day appreciate it yeah Nick do you have anything anything else you want to plug you while you're while we're here sure I
55:26
mean uh Twitter has been replaced by threads so you can follow me there at Nick Tindall
55:32
um or you know Twitter at Nintendo same thing um uh just yeah as a developer I just want
55:39
to say you can accomplish things right like I never would have imagined I'd be doing an interview for a project I contributed
55:45
to an open source right take those leaps try and build new things awesome yeah
55:52
thank you so much Nick really appreciate all your work thank you