YT YouTube2Text

YouTube2Text → Transcripts → Build Your AI Operating System: Regen OS. Human-Centred AI Community Workshop by Ben Pecotich

Build Your AI Operating System: Regen OS. Human-Centred AI Community Workshop by Ben Pecotich — Transcript

by Dynamic4 · 14,134 words · 2,061 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Good day everyone. I'm Riley Coleman. Um
0:02I have um been uh I've known Ben for a
0:07long time. Um but uh over the last sort
0:09of year or so as I've come back to
0:11Australia, I've definitely we've caught
0:13up quite a few times and um I think that
0:15we are um uh we've sort of uh rifted
0:18about AI quite a few times uh in terms
0:21of sharing what each other's doing and
0:22things like that. And so um one of the
0:25um uh reasons I brought Ben in is
0:27because of some u work that I was doing
0:29with the course that I teach people um
0:32and uh talking about you know how do you
0:34actually go for the next level in terms
0:36of um your AI maturity and that really
0:40comes down to systemizing things so that
0:43you're not having to reinvent the wheel
0:45every time. You're getting a lot more of
0:47that advantage baked in from the very
0:49start um and things like that. So
0:51building something that is more
0:53sustainable
0:54um that is able to be used um uh in your
0:58daily work um and also it learns
1:00alongside you because right now
1:03obviously um uh different AI systems are
1:06learning systems but they're kind of
1:07learning at scale and behind you know
1:10behind big kind of um retraining
1:12moments. So, um this is why Ben's coming
1:16in because uh he has uh set up uh his
1:19own um uh AI operating system uh which
1:24um is called regen. Um
1:27and is it like I wonder if it's like
1:30Siri Ben you can tell us like do you say
1:31regen and it's listening to you for that
1:33or something but
1:35>> it's not someone called me or Jen.
1:37[laughter]
1:39>> So he's going to walk through um how he
1:41Hi Oliver. uh he's going to walk through
1:43how he is um had thought through this
1:46process, how he approached both building
1:47it um iterating and and sort of
1:50developing it over time. Um and then
1:52also how he uses it um uh and how that
1:56um uh what advantage like why would you
1:59go into doing this essentially? Um what
2:01what sort of advantage does that offer
2:03you? Um and uh how does he actually
2:05leverage it um in the way that he works
2:08um every day?
2:10So um this is part of a series of
2:12workshops that I'm bringing together for
2:15a community that I'm sort of starting
2:17called human- centered AI community. So
2:19looking at um uh you know the advantages
2:22and the opportunities that AI brings um
2:25sort of co-learning together but also
2:27talking about things like um how do you
2:29do um personalization without breaching
2:32people's privacy. Uh so really looking
2:34at the ethical human uh centered side of
2:36it as well. And so um thank you Ben for
2:39joining us. Um Ben has been working uh
2:43uh for um decades within design. Uh he
2:47currently leads uh an organization uh
2:50called Dynamic 4. Um and some of you uh
2:53folks in Sydney in particular may know
2:56Ben from his um uh I think it's you said
2:5911th birthday uh Sydney design thinking.
3:01Is that correct? Yep. Uh so he's been um
3:04uh uh facilitating and running the
3:06Sydney design thinking uh community for
3:08the last 11 years which is a a really
3:11really you know uh huge kind of thing to
3:14be doing for that long a period of time.
3:16So he's been contributing to the design
3:18community uh particularly in Sydney um
3:20uh from uh a a longtime advantage uh
3:24point of view. Um and so it's a real
3:27pleasure to have Ben come in and uh talk
3:30us through how he actually approaches
3:32this um and answers all your questions.
3:34So save them. You can put them in the
3:37chat and we'll come to them at the end.
3:39Um alternatively um you can save them to
3:42the end and obviously say them out. But
3:43uh yeah, if you think of anything as
3:44he's going through, please by all means
3:46throw it in the chat and we'll come to
3:48that at the very end.
3:50Um so without further ado, I'd love to
3:53invite Ben to uh you know uh take the
3:56lead and to yeah uh tell us everything
4:00we can learn from you and your um
4:03[laughter] awesome regen OS.
4:05>> Let's not get carried away. Thanks
4:07Riley. [laughter]
4:10>> Awesome. Great to see you all. Lots of
4:13familiar faces and a bunch of fresh ones
4:16as well. So, great to meet you. Um, just
4:19want to acknowledge that I'm lucky to
4:20live, work, and play on Gatle country
4:22and pay my respects to elders past,
4:24present, and future. Um, as Riley
4:27mentioned, I'm Ben. Hi. Um, I thought it
4:31might be worth me just quickly sharing
4:33kind of the perspective that I come at
4:34this with because there's a lot of talk
4:36about AI and it's kind of everywhere all
4:39the time, all at once. Uh a lot of it's
4:41hyperbole and there's lots of fear and
4:43scaremongering as part of that. Uh and a
4:46lot of it comes very much from a tech
4:48perspective. Um but I do come at it from
4:51a tech perspective but also with some
4:53other perspectives to it as well. So the
4:55way I spend my days is um literally from
4:58boardrooms to pixels, devops and AI um
5:02and all the bits in between. So it's not
5:04just a tech orientation to this. Uh lots
5:07of that's around business model
5:08innovation, product, service and
5:10organizational design, designing and
5:12delivering project based leadership
5:14programs. So there's a lot of that kind
5:16of stuff in play. And another way that a
5:20bunch of you will probably um already
5:22know is that I'm bit of a governance,
5:26data, systems design, product tech, and
5:29learning geek. So, where all of those
5:32things come together and hang out,
5:34that's kind of where I hang out. And,
5:36um, I can see a bunch of you already
5:38here. That's I get to collaborate with
5:40and do that fun stuff with as well. Um,
5:43and so a lot of that's really about how
5:45do we uh work with people to design and
5:48build business models, products, and
5:49services that customers and teams really
5:51love. They actually make money because
5:53without that, we don't get to keep doing
5:55it. And they do great things for people
5:57in our planet all all while increasing
5:59well-being. So everything I do is
6:02focused on how do we help accelerate the
6:03transition to more regenerative ways of
6:05living and doing business. And that's
6:08how I spend my days now in terms of what
6:10I've been doing. Um over the last 30
6:13plus years I've been designing and
6:15building digital products, services and
6:16experiences. Um and also the teams and
6:19organizations that actually deliver
6:21those things. and how do you again it's
6:24not just the design and tech but how do
6:27you actually bring that together as an
6:28integrad system um I spend about 80% of
6:32my time actually doing the work and I
6:35try to keep talking about it coaching
6:38teaching writing to about 20% so
6:42everything I do is very much in current
6:44practice and real world experience not
6:46just theory or I remember a time 10
6:48years ago when we used to do it this way
6:50and that's now completely obsolete and
6:52irrelevant
6:52Um, I find it really important that I
6:54stay completely
6:57every day in the practice as well. Um, I
7:00wrote a book 5 years ago, um, it's hard
7:03to believe it's 5 years called Solve
7:04Problems That Matter, which was all
7:07about how to design, build, and launch
7:09your social enterprise idea. And I am in
7:13the middle of writing a new book. And
7:16this is actually the first tease of the
7:19working title uh which is lead what
7:21matters and that is all about how to
7:25design the conditions for momentum. So
7:28working title I'll be testing that yet
7:30but it's very much around leadership and
7:33how do we adapt to all the increasing
7:35expectations including Gen AI and and
7:39how that um impacts people and teams and
7:42organizations. Um so a few moving parts
7:45there. Hopefully that helps sort of
7:47frame this not as a tech thing. It is
7:50very much technology enabling but it's
7:53actually more about the people, the
7:55process, the organizations performance,
7:58workflows, habits, all those other
8:00things that actually are the way that
8:02we've always um needed to do work. So a
8:06very quick hopefully in the right room
8:09this is what we'll cover. Um, and a lot
8:12of this I've kind of been hesitant to
8:15share publicly because things move so
8:18fast and literally like um can say
8:23something tonight and by the morning
8:25it'll already be out of date. In fact,
8:27it might already be happening while I'm
8:28saying these words. Um, in between
8:31running a version of this workshop um
8:34back on the 7th of July, so less than 3
8:36weeks ago,
8:38huge number of things have changed.
8:40The day that I ran that workshop, there
8:43were two app app updates, one on claw
8:45desktop, one on chat GBT or it's codec
8:48still at the time. That was an hour
8:50before I started the workshop. That was
8:52on a Tuesday night. By Friday,
8:55there was no longer a codeex app as the
8:58desktop app for chat GPT. It had been
9:00merged in and ChatGpt
9:03work was announced and codeex was
9:05another angle of that. And a few days
9:08later, they even dropped the work part
9:10of chat GBT in the little toggle. So,
9:14you know, little things like that, just
9:15the way you even find and interact with
9:17these models and those harnesses or
9:19wraps that are wrapped around them
9:21literally changing by the hour. And, um,
9:24if you haven't seen in the last couple
9:26of days, um, Claude has released or
9:30Anthropic has released Opus 5. So, you
9:33probably heard all about Mythos and then
9:34you heard about Fable 5. That model
9:36disappeared for a few weeks, came back,
9:39fresh guard rails on it, and now Opus 5
9:42has been released and that is already
9:44outperforming Fable 5 that was banned by
9:46the US government um on a whole bunch of
9:49benchmarks. The other key sort of things
9:52that are without diving into all the
9:54tech side of this because it relies
9:56relates directly to the stuff we're
9:58talking about is both chat GBT/Codex
10:02um and Claude got realtime voice
10:06directly into their desktop apps. So you
10:08can control things in a real-time voice
10:10scenario um which you couldn't before
10:13Friday. Um, you could do dictation, but
10:16you couldn't do real-time voice
10:17interaction. And
10:20my biggest breakthrough of the last few
10:23days, weeks, can't even remember how
10:25long it's been. It's probably only been
10:26a week and a half, is that completely
10:30undocumented
10:31feature on codeex where now I have my
10:35team members, my synthetic team members
10:38directly DMing each other and being able
10:41to hand off work to each other and QA
10:42each other's work.
10:44according to a workflow that I've
10:45designed and completely undocumented
10:48feature and I was trying to solve team
10:50collaboration and it suddenly went hey I
10:54can do this now and a few days earlier
10:57couldn't do that haven't seen any
10:59announcements of that anywhere else
11:00either um so it's this kind of stuff
11:02that is happening all the time so even
11:05though I just spoke about it for the
11:07last couple of minutes I'm not going to
11:09talk about product features because they
11:11are literally changing that past. Um the
11:14thing I'm going to focus on and we'll
11:16talk about together is really the
11:18enduring principles that last. So the
11:20ways of working, the mindset, the
11:23workflows, the things that actually
11:25are stable regardless of what the model
11:28and the harness can do today. Um so
11:30that's where I find there's, you know,
11:32more value in thinking about because
11:35everything's changing so fast. Something
11:37I'm not going to cover today is the
11:40ethics and climate impacts. Uh extremely
11:42important topics. Love to talk about
11:44them, but that could be many hours of
11:47conversation and often one's best have
11:49at the pub. Anyway, the other thing this
11:51isn't is a buildalong session. So, it's
11:55very much about the principles and the
11:56architecture and the ways of working,
11:58but it's not going to be a workshop to
12:01sort of build your own. Um hopefully
12:03there'll be enough in there that you can
12:05sort of take it and apply. Um but it's
12:08not that guided version of things.
12:11So the main thing I'm going to focus on
12:13as we go through this is really on not
12:16technical work, not writing code. And
12:18when you um hear about these types of
12:21workflows and operating systems, they're
12:23almost always in a code and development
12:25context. I'm not going to talk about
12:27that at all. I'm going to assume that
12:29everyone is here because you want the
12:31non- tech version and it's very much
12:34about the knowledge work and so because
12:37of that very much focusing on
12:39unstructured ambiguous knowledge work
12:41not even automated processes of business
12:44process of how do you get step ABC to
12:47happen really fast. So it's working in
12:49that really uncertain ambiguous free
12:52flowing unstructured version of using
12:55these tools is the main thing that I'm
12:58going to focus on.
13:00So
13:02very quickly
13:05the other thing is we say AI now and we
13:08very specifically almost always mean
13:10generative AI. So AI has been around not
13:15since November 22. It's been emerging
13:19since the 1950s and there's been these
13:23different waves of AI
13:26and they have had different
13:27capabilities, different ways of
13:28training, different things they could do
13:30and different levels of adoption as
13:32well. And so as these waves have come
13:35through and the next wave starts, it
13:38hasn't stopped the previous wave. It's
13:40often built on top of and then now with
13:44Gen AI, they're actually all working
13:46together. But when we're talking about
13:48the frontier models, foundational models
13:51like chat GBT or Claude or Copilot or a
13:55lot of the others, we're specifically
13:57talking about Gen AI as the thing. Um,
14:01and so for a few years I held on to I
14:05refused to say AI unless I specifically
14:08meanted meant the other versions of AI,
14:11but we're past that point. So speaking
14:13the language of the customer, I've I'm
14:16going along with it. So when I say AI in
14:19this context, I'm pretty much only
14:22talking about Gen AI. Um and AI is an
14:25amplifier is an amplifier. So it's and
14:28an accelerant. It doesn't actually
14:29really a lot of the things that we talk
14:31about when the hype starts up around AI,
14:35it's actually not new stuff. They're
14:38actually um more people, culture,
14:40process, workflow and data challenges
14:42than they are technical ones. So the
14:44same sort of challenges we've had for
14:45decades slash tens of thousands of
14:48years. You know there the same types of
14:49behaviors and how do we work together
14:52and there's a lot of fear that comes in
14:54through that. So I find it's really
14:55important that we keep that perspective
14:58and that we sort of zoom out and we
15:00don't just hone in on the last sort of
15:02three and a half years of Gen AI hitting
15:05mainstream. If we zoom out and we keep
15:08that perspective, we actually see that
15:09longer time arc and we're traveling
15:12faster and with different tools, but as
15:14we look back, there's a lot of really
15:16familiar patterns and we've learned a
15:18lot of things over decades of how to
15:20navigate those complexities. Um, so I
15:22think it's really important to do that
15:24zooming out.
15:26And something I really want to emphasize
15:28is while I'm sharing this stuff, I'm
15:31definitely not saying that I've got the
15:33answers. I'm not an AI or I'm not an AI
15:36expert and I'm not a Gen AI expert. And
15:38anyone that tells you that they're a Gen
15:40AI expert, um I think it's pretty safe
15:43to say that they're lying to you andor
15:45delusional. Um because even the makers
15:47of the models or the growers of the
15:49models as sometimes now referred to
15:51because we don't actually even know how
15:52they work the neural network and we can
15:54see the effects and we can see the
15:56pre-training and some of the inference,
15:58but we don't actually understand anymore
16:00what's happening inside the models. Um
16:03so some people actually talk about these
16:05models now growing the conditions and
16:07the nurturing and some things are put
16:09together to grow them but there's not a
16:11direct cause and effect of making them
16:13as there once was. Um so there's this is
16:17very much emergent practice and there
16:20are no experts in emergence and
16:25learning starts with not knowing.
16:27So let's get into not knowing together.
16:31So, the key thing I'm really curious
16:33about to kick off is
16:35you,
16:37your data,
16:39and how you use AI. So, I'd love to be
16:43able to see you wherever possible,
16:44including with uh Andrew's
16:47animatronics head, [laughter] which is
16:49awesome. Um, but chuck in the chat. Um,
16:54let me know a little bit about what
16:56you're actually using in terms of AI
16:57models and tools right now. What's your
16:59your tool or tools of choice?
17:05Jess has got the paid the flash
17:07co-pilot.
17:08Claude and Virginia when you say Claude
17:11any particular version are using it in
17:13browser using the desktop app.
17:18Victor's got codeex claude code.
17:21So sounding more on the dev side.
17:22Desktop for Virginia. Awesome.
17:29It's mostly the chat GPT and the clawed
17:32side of things which would match all the
17:35download data.
17:38Some Figma make
17:41[laughter]
17:42Oie's living large with the new Siri. So
17:45what he means to say is he's using
17:46Gemini.
17:50Gemini mashed together with a bit of
17:51Apple love.
17:55Uh, what else we got? Claude Code, Opus,
17:59Gemini.
18:02Awesome. That's g me a good sort of
18:04sense.
18:05And when you're using your models,
18:09where do you find the good stuff goes
18:13when that session or chat ends?
18:17What happens to the outputs that you
18:18create in those little interactions?
18:29>> [laughter]
18:30>> into Riley's own neural network.
18:32Awesome. Good. Before, during, and
18:34after, I hope.
18:39And since we got more neurons outside of
18:41the brain than inside, then um I hope
18:45that's embodied AI
18:50docs, HTML, skills. Nice. Some scripts.
18:55John's geeking out with some Python.
19:00Creating prototypes.
19:07Excellent.
19:10And
19:12what's the situation here? How often do
19:14you find that you're repeating yourself
19:17across different chats or interactions?
19:20Is that a a rare thing or a
19:24most session thing?
19:26A bunch of too often.
19:30Nigel's rare. Good. That means some
19:32stuff's already set up.
19:34Howie
19:38Riley's got it sorted.
19:44Nice.
19:48And now not even thinking strictly in an
19:50AI context,
19:52but watch your data. What data do you
19:55have?
20:07You don't just have CSVs and
20:09spreadsheets. Oie,
20:12you got more interesting data than that.
20:20documents and web.
20:22[laughter]
20:24So Ollie might be letting us in on a
20:26where it does and does not trust AI to
20:29play right now.
20:32PowerPoint docs.
20:36Awesome.
20:42And the big question
20:44and one of the things of if you don't if
20:47this makes you a little bit nervous to
20:49answer or you don't know this is
20:51definitely your homework. Where does it
20:53live? Where's your data right now?
20:57How many places do you know all the
21:00places?
21:05GitHub. Nigel's definitely got the the
21:08dev rig going.
21:12everywhere.
21:16Superb base. Nice.
21:21GitHub.
21:25And the final one for this set
21:29for all that data that you've got and
21:31all those places it lives.
21:33What needs to be working for you to be
21:35able to access it when you need it?
21:42So if you said anything other than on
21:45your own device,
21:47you've got a lot of infrastructure that
21:49needs to be available.
21:51So the interweb needs to be available
21:53for Kev.
21:55GitHub needs to be up. And if it's on
21:56GitHub, then you need a whole bunch of
22:01infrastructure between your device and
22:04GitHub to be working
22:07in the clouds.
22:12And we have got to a point where the
22:14internet and just saying the internet
22:16needs to be working is almost like a
22:19passible answer because we rely on it so
22:22heavily every day. But if you think
22:24about all the moving parts that need to
22:26be up and going and working to um
22:29standard to be able to even browse a web
22:32page, let alone get to your real data,
22:35that's a lot of moving parts that you've
22:36got little to no control of. And for
22:40anyone who doesn't pay for data when
22:43you're on a flight
22:45and you need to get to that file
22:48because you really need it, then it
22:51might give you a sense of what it can
22:53feel like when you can't get the data
22:55you need.
22:57And so
23:00data,
23:02it really kicks things off. And you've
23:05probably, you know, seen a version of
23:08this pyramid or this pyramid
23:09specifically before. We've got the data.
23:12We turn that into information through
23:14synthesis and knowledge. And then
23:17hopefully we get insight out the top of
23:18it.
23:20And then all of that then wraps together
23:24as our experience.
23:26And so data, it's so important and it's
23:28so critical to everything. But not just
23:30because we love data, even though I do,
23:33but it starts the process, but it's
23:35actually our experience that gives it
23:36meaning. So it's all of those other
23:38things that wrap around it. If it's just
23:41an isolated piece of data, you might not
23:42care too much about it, but our
23:44experience gives it meaning. So if you
23:46think of a photo, a piece of music, a
23:49message, something that's really
23:51important to you, and you actually think
23:53about that thing right now.
23:57And then if you think about how would
23:59you feel if you lost it
24:02because a lot of that stuff is now
24:04stored as data.
24:08And so even though it's just a data
24:10point or data element,
24:12the data by itself is not enough. It's
24:14all those relationships and context and
24:16memory and our lived experience that
24:18wraps around it that gives it meaning.
24:21And that's really why the current
24:25experience when people use AI is a
24:29little bit weird. A lot of the time it
24:31often feels extractive or disposable
24:36and we go through some version of this a
24:39lot of the time where we sort of feel
24:42like we're starting over and over with
24:44someone that's really smart but they
24:46just and lots of potential but they just
24:48don't know us or remember us. So we ask
24:51something, we get an output, we might
24:54copy and paste that somewhere else.
24:56it then we want to build on it or do
24:59something more with it. Loses context
25:01and then you're back to sort of square
25:03one and have to start again and you
25:04start getting really frustrated of I've
25:06already told you this. Why am I telling
25:07you this again? And that's pretty
25:09infuriating. It's infuriating when it
25:11happens with people and it's even more
25:14infuriating I'd argue when it happens
25:16with machines that are supposed to just
25:18remember everything all the time. Um but
25:20that's not the experience for a lot of
25:21us a lot of the time right now. So we
25:24create these what feel like really
25:25useful moments you might even feel that
25:27excitement of look at this thing I'm
25:29starting to be able to do this thing
25:30with it and then without the way the
25:34products are at the moment a lot of the
25:36time that then a feeling of being let
25:40down or that momentum drops away there's
25:42no compounding value because it is like
25:44starting again
25:46and so that results in our thinking
25:48being a bit scattered and our context we
25:51feel disconnected and our work's just
25:54really hard to find and re reuse and and
25:56build on. And so that's what the
25:59experience is for a lot of people today.
26:02How much of that feels familiar from
26:05what people were saying in the chat?
26:06There was a fair bit and there's been a
26:08few nods. So hopefully it's not
26:11triggering too many things, but um
26:15that's where there's a huge opportunity
26:17because the tools are actually capable a
26:20lot of a lot of this now. But it does
26:22take that thinking about the workflow
26:24and the habits and building those things
26:25up. But being able to move from a chat
26:28to actually having a synthetic team or
26:30an agent team or the bots or the AIS or
26:33whatever you want to refer that refer to
26:35that to, but I think of it as my
26:38synthetic team. And so as I've been
26:41designing regen OS um and it's very much
26:44a work in progress and it's something
26:46literally working on every day because
26:48it's it iterates as I'm doing real work
26:50with it. But it's really about how to
26:52make AI a compounding asset and with
26:55data that you actually own and you own
26:58and have control of it locally, not just
27:00when somebody else's infrastructure is
27:02available. So all of that value is
27:04actually yours to use when and how you
27:07need it. And so that flow of work
27:09happens. You get the useful stuff that's
27:12produced if that's context or outputs is
27:14written back to that safe place.
27:18Patterns that are valuable become
27:20reusable. And then the next session
27:22starts from that stronger base. And
27:25because of all of that, you end up with
27:26better thinking, clearer decisions, more
27:28momentum. And that's all because that
27:30context is getting carried forward and
27:32built on rather than just left behind as
27:34a fragment in a browser window.
27:38And so
27:40this is an illustration version of
27:42regenos as a living loop. Um for those
27:47that were here for the first version,
27:50you'll see the next iteration of this,
27:52but there's five elements that form a
27:54stack, but it lives as a living loop. It
27:57actually works as a living loop. So the
27:59foundation,
28:01that's all your data, that's all that
28:03you own. That's durable. It's
28:05compounding. That's your stuff. And
28:08hopefully you can see from the icons
28:09made up of all those different file
28:11types that are really where our data and
28:14our digital life and our second brain
28:16and all of those types of things now
28:17live.
28:19The next layer building on top of it is
28:21the playbook. So that's how does the
28:24system actually behave.
28:27A really important one is the boundary.
28:30So that shield over the top of what's AI
28:33safe, what's allowed, assisted or
28:36explicitly human only of don't touch
28:38this AI hands off. This is my stuff. You
28:42either don't even know it exists or you
28:44get to read it only or you have a
28:48certain set of permissions to interact
28:50with it.
28:51And then across the top are the engines
28:55and it's plural. So there was a few
28:58people in the chat that were talking
28:59about the using multiple models. The
29:01whole idea here is that you can use
29:03multiple at once and not lose your team
29:06that you still get all the benefits of
29:08as if you're using one model and you
29:10don't have platform lockin and you
29:12reduce that risk that one platform is
29:15going to disappear or change their
29:17policy or get banned by the government
29:19or change their pricing model or do any
29:21number of things. that if you don't have
29:24that designed then all your data could
29:27literally gone and that's to me is a
29:30really scary thought.
29:33And now all of that comes together then
29:35into the studio and that's who works
29:38with who. So the human in the studio but
29:40then who are these synthetic team
29:42members? What are their roles? What
29:44memory and context are they working
29:46with? How do they interact?
29:48And that's where the work's done. And
29:50then the regeneration. So the system can
29:53keep um recovering and compounding and
29:56the continuity that you're picking up
29:58from where you left off rather than
30:00having to start again. [clears throat]
30:02So in terms of kind of the loop of how
30:05that works. So a studio member reads the
30:08foundation, they get their memory and
30:10context, they get the rules of the game
30:12from the playbook. The boundary
30:14determines what can be read, changed,
30:16suggested or decided. So the rules of
30:18that access and interaction.
30:21one of the engines or multiple of the
30:22engines and tools. You do the work
30:24through their app and the harness and
30:27then the useful outputs are then written
30:28back into the playbook and the
30:30foundation and then the whole system
30:32regenerates and gets stronger over time
30:34and so you get that compounding um
30:36value.
30:37So that's the sort of architecture of
30:40the snapshot. We'll come back to that a
30:42little bit more soon.
30:44But I think there's a really important
30:46part here. Um because as we were talking
30:49about before, a lot of this isn't
30:50actually about the tech. It's about the
30:53mental model and the the workflow and
30:55the paradigm that we bring to it. So
30:59when you think of this literally as your
31:01team
31:03rather than just a random chat bot and
31:06you apply a lot of the principles that
31:07you've probably used over years of
31:10working with new team members, on
31:11boarding, hiring new team members, how
31:14do you bring them up to speed? what work
31:16do you give to them when you know that
31:19leadership aspect of things and the man
31:22management of how do you manage things
31:24is a really important mental model that
31:27I think flows through all of this and so
31:30rather than just being a set of random
31:32things doing random stuff
31:34or even just it's one model so I'll just
31:37give it all the things in one chat and
31:39it can sort itself out in the background
31:42then each team member actually has when
31:44they're a team member
31:45they have a clear role and scope. So you
31:47know here's your JD basically and this
31:50is how what your skills needed to be.
31:53This is um the context that you're
31:55working with and this is how you play
31:58with other members of the team. And
31:59here's the playbook of how our workflow
32:02works when we're doing work together.
32:04It's got the context it needs. Each has
32:07its own memory that carries forward and
32:10it then you're renting that intelligence
32:13from one of the engines or one or more
32:15of the engines and then when something's
32:18repeatable then we can turn that into a
32:21skill and those skills are transferable
32:23across team members as well. And so it's
32:25all about being able to set the
32:27direction to find those boundaries. But
32:29as the human we are always accountable
32:32for our synthetic team and everything
32:34that they do, everything they produce or
32:36don't produce and the quality of it and
32:39any misbehavior,
32:41whatever happens there as the human
32:44that's on us. And so even looking at
32:48this is a sort of picture of the
32:50architecture,
32:52this isn't actually a team-based model.
32:55So each human has their own synthetic
32:58team and then how each human works with
33:01other humans in a real person team
33:04that is then another level of design but
33:07every human is bringing their own
33:09synthetic team to the broader team. Does
33:13that make sense?
33:15There's a whole thing on whole and
33:17hocrisy um but I'm going to stay away
33:20from that org design stuff just for now.
33:24a quick one on as a little bit of a
33:28before and how some of these things have
33:31morphed over time. Um, a bunch of you
33:34will know that I have a little ongoing
33:37creative project called Music
33:38Meanderings where I do um somewhere like
33:43150 plus um micro reviews of albums, my
33:48favorite albums having a major
33:50anniversary or new releases. Um, and so
33:55that's an always on project. I've been
33:56doing it for years. And for the
33:59anniversaries, that takes a fair bit of
34:02admin.
34:04And so I need to go through my iTunes
34:06library, get my four and five started uh
34:09four and five star rated um albums, get
34:14those all of those that are having an
34:16anniversary for the year, get those into
34:18a a sheet. So, I've got the year, but I
34:21don't yet know the date. And then I need
34:23to find out the specific date because I
34:26post on the date that it's having a
34:28major anniversary, not the general year.
34:31So, I need to go get that info. So,
34:34before AI, I'd have to go and research
34:38and and I might be looking at a sheet of
34:40150 of these. I'm I used to go and drop
34:45that into a Google search. go get the
34:48release date. When you get older than a
34:522005 album, then it gets sometimes
34:56pretty hard and especially if it's not
34:57really mainstream, it can start taking a
34:59lot of digging to get the actual release
35:01date. So, I used to have to do all of
35:03that manually. I have to remember all
35:05the things all the time and do it all
35:07myself. And then chat GBT and perplexity
35:11happened. And so when I was doing it in
35:142023,
35:16I was giving it batches and to both
35:20perplexity and to um chat GBT and I
35:26would then compare their output and chat
35:28GBT would just extremely confidently lie
35:31to me and tell me a date and I'd go that
35:35doesn't sound quite right and I'd get
35:37the same output from Plexity and
35:39Perplexity would more often say, "I
35:42don't know. I can't find that." And I
35:45would then do a compare. But it meant
35:48in, you know, late 2023 when I was doing
35:50this, I still had to do all of that
35:53manually and spot check. And because I
35:55could then couldn't trust it, I had to
35:57mostly do it manually and everything
35:58that I did have, I'd have to fact check.
36:01By late 2024 when I was doing that
36:04again, it was in a very different state.
36:08And this time it had the basics handled
36:11pretty well. I could upload a
36:13spreadsheet. It could do some
36:14restructuring of that sheet, but it
36:17couldn't really do the work. By late
36:202025,
36:22I was able to say, I need this. I gave
36:25it the list and it was able to go and
36:28accurately, and this was chat GPT this
36:30time, and fact checked with Claude, was
36:32actually able to factually go and get
36:34that info for me.
36:36We're now at a point only a few months
36:39later where I won't even need to do any
36:42of that myself in terms of going to get
36:44that data. I'll actually be able to give
36:46one of my agents or team members the
36:49goal of I want this for this reason,
36:51point to past examples and it will
36:54actually be able to use computer use, do
36:56the export out of iTunes, do the
36:58filtering and sorting and then go fill
37:01out the dates for me and tell me what my
37:03posting schedule is. So the the pace of
37:07change over those years and the reason
37:10I'm using something completely
37:12unbusiness related is because I want to
37:14make sure this is really not about
37:16coding not about operational
37:19optimization and automation of workflows
37:23really unstructured knowledge work and
37:26how do you work with that type of
37:28situation not just something that's
37:29pretty predictable and so these are the
37:33different ways of working as the
37:36capabilities have changed over the
37:38years.
37:40So my question to you
37:42is what synthetic team member would be
37:46most useful to you right now and feel
37:49free to chuck it in the chat.
37:53Does anyone got need some album release
37:55dates or you got some other uses for
37:57these [laughter] agents?
38:03And if you haven't got anything to chuck
38:05in the chat, have a think about what
38:06that might be. What's what's something
38:09that's on your plate right now that's
38:11you find frustrating
38:14or you feel a bit stuck or it's really
38:18labor intensive to get the data together
38:31visualizing skill sets of a team for
38:35specific projects. Definitely business
38:37process tacet knowledge extraction doco
38:41and as Riley mentioned
38:44now the two major models of claude and
38:48chatgpt support just screen recording
38:51that. So a lot of the other platforms
38:54where you'd have to either do that
38:55manually or pay for a se uh separate sub
38:58to be able to do things like soaps and
39:01playbooks that's all standard stuff now.
39:06or about to be
39:09virtual personal assistant. Awesome,
39:10Kev. That is actually one of the hardest
39:14things unless you've got a very
39:16operationalized, repeatable life. Uh,
39:18and as a designer, I know you don't.
39:21Then having a VA is really tricky when
39:25it's a human, especially when you don't
39:28have really repeatable loops on a
39:29timeline that makes sense to hire a VA.
39:33these types of models when it's very
39:35unstructured or it's something that
39:37happens in that little scenario I just
39:38gave of once a year and I fact check it
39:41and do some other things every now and
39:43then. Um,
39:45lots of power there. Awesome.
39:47Orchestrator,
39:49someone to work out my finances.
39:53Everything I say tonight is not
39:55financial advice, not legal advice, but
39:57I do have a financial advisor team
40:00member who does my statements of advice
40:03for me. And as the human, I am fully
40:07accountable for whatever it tells me and
40:08whatever I action, but it does all of
40:12that financial modeling for me. It looks
40:14product comparison does all of that for
40:17me as um same as a financial financial
40:20adviser would. and from my experience
40:23does it better and for free
40:28and [laughter]
40:30uh recipes awesome
40:34excellent lots of possible ways and I
40:36think the mix of very business work
40:40oriented things and lots of personal
40:42things
40:43when it's your own synthetic team you no
40:47longer have to delineate about that
40:48because you're building it for you and
40:50you can look at it whole of life and
40:52then when you're looking in a business
40:53context, you can look at which parts of
40:55this will I share now with other team
40:57members in a business context. So being
41:00able to that boundary line applies also
41:03to other humans and to other
41:04organizations etc.
41:08So the bits that make all this possible,
41:10we've talked about this a little bit,
41:12but just to make it really explicit
41:14because
41:16there's always preconditions for the
41:18good stuff to work and a lot of it
41:19doesn't happen at once.
41:21So at the moment the most powerful
41:23especially again talking about non- tech
41:26so we're not talking about code or dev
41:29in a non tech context the the superset
41:33and what I find I spend most of my time
41:35with at the moment is the chat GBT
41:38desktop app which also now includes
41:41codeex in it so it's it's one combined
41:44app now claude and again the desktop app
41:49and Obsidian in. So the key thing is
41:52with the um the models, you need the
41:56desktop app to really get the most value
41:59out of it. Now you can get a bit of
42:01stuff out of the browsers and using
42:03connectors and even some MCPs and stuff,
42:06but it's just not the same experience as
42:09the app being able to work with your
42:11files locally and you telling it what's
42:13allowed to touch and not allowed to
42:14touch. Um, so it's a completely
42:16different experience when you're using
42:18the desktop app um of these same models
42:20that a lot of people have been using for
42:22months and years now. So the key thing
42:25you need one of these apps on your
42:28desktop. Both support Mac and Windows.
42:31Um, varying support for other oss. Um,
42:36Mac tends to get the features
42:39a week, two, 3 weeks earlier sometimes
42:42and Windows. uh though that has started
42:44to change over the last few weeks. And a
42:48key thing here is you actually don't
42:50need it all and you don't need it all to
42:52be nice and clean, including your files
42:54and data before you can make a start.
42:56You can actually get your um AI to help
43:00you do a lot of that work once you've
43:03got the basic setup.
43:05You'll notice there a non-geni
43:08icon and that's Obsidian.
43:11So, a lot of people keep asking me about
43:13Obsidian for some reason. Um, if you use
43:16Notion, it's a lot like Notion except
43:19for you don't need to pay them. It's
43:22free, open source, and all the files
43:24live on your devices and can be clouds
43:27synced, but they're yours to play with,
43:30not someone else's. Um, and has a huge
43:34community with plugins and stuff wrapped
43:36around it, so you can do all sorts of
43:38stuff with it. Um, so it's effectively
43:41the same as notion for a lot of things.
43:45Um,
43:47basically it's a text editor just to
43:50almost [laughter]
43:51go back on that. But that's effectively
43:54what notion is is as well. So it's
43:58the what's emerged over the last 2 3
44:02years is markdown has become the deacto
44:04standard for the AI models as the file
44:07format and markdown is literally just a
44:11text file format. There's a specific
44:14syntax but it's pretty much a plain text
44:17editor that then gets enhanced with some
44:20other stuff when you use things like
44:21Obsidian to be able to do some
44:23additional things on top. So, it's
44:25nothing to be scared of. Um, and it's
44:28not the only text editor or the only
44:30markdown editor. So, I use Obsidian as
44:33my main. Um, but if I just want to read
44:36and I don't want it part of my more
44:37structured stuff and I just want to open
44:39it from Finder Explorer, then I use an
44:43app called Mark Viewer or you can
44:44literally open it in Notepad. Um, it is
44:47literally a text file. So, it's it's an
44:50open format and that gives it a lot of
44:52future proofing. There have been talks
44:55and there's a little bit of an ongoing
44:57debate of will HTML actually be the file
45:00format that AI runs on. Um it's possible
45:04and even if it is markdown converts
45:07nicely and lots of things. Um and
45:10there's also ways to run this where all
45:11your files and and a lot of what we're
45:13talking about is then run in a light
45:15database um which can also be run
45:17locally without um server
45:19infrastructure.
45:21But I promised no repos um and no tech
45:26stuff. So we won't talk about that and
45:29no tech skills required. So I use
45:32Obsidian partly because I already had 30
45:36plus years of my notes in it um when it
45:39sort of took off earlier this year. So,
45:43and the reason I even had my notes in it
45:45was it fits a lot of my architecture
45:47principles um about being markdown files
45:51that it's offline. I've always got
45:53access to it. I can access across all my
45:54devices anytime, anywhere. Those are
45:57standard architecture principles I've
45:59had forever and Obsidian fit that. Um,
46:04and another key reason why I think
46:06Markdown still has a better chance than
46:08HTML at this stage is it's really easy
46:10to create a text document and edit it
46:12and read it. HTML mostly isn't. Um, it
46:16normally takes a little bit more effort
46:18around that. Not impossible, but it's
46:21very easy to create and edit markdown
46:22files and
46:26it's extremely efficient as that layer
46:28that both humans and machines can read.
46:31So that's why I use obsidian. The thing
46:34that I always have to show is everyone
46:38always wants to see the graph view of
46:41obsidian. So it automatically creates
46:44these knowledge graphs. So every node
46:48that you can see there is a note in my
46:51file file and each of those files is
46:53literally a file on my hard drive.
46:55So as you interact with each, it will
46:59actually show here are the relationships
47:02that it has with other notes within your
47:04vault as well. So you can use that as a
47:06way to trace through that this note,
47:09this file, this piece of data has a
47:11relationship with another one within the
47:13file system.
47:15useful for humans. Um, and AI models
47:18especially love it because then it does
47:20the work of being able to scan um, and
47:22be able to trace things through and you
47:25can animate your obsidian graph showing
47:27how the different sort of nodes and
47:29relationships sort of grew over time. I
47:33don't find it that useful on a every
47:36moment of everyday use, but it is an
47:37interesting little feature and everyone
47:39always when they talk about Obsidian
47:41want to see the graph. So there you go.
47:43You saw the graph. So now working with
47:47your synthetic team,
47:49what's it actually look like in action?
47:52So this is a very quick walk through and
47:56bit of a before, during and after
47:57journey map. So before these are the
48:00files that exist before any sess
48:02session. Starting how we on board
48:05synthetic team member and that's when
48:06the engines and the studio come into
48:08play. the during. So that's where the
48:10real context and that's the studio at
48:13work. How you end a session. So you're
48:15actually making sure that you're
48:16compounding the that value and
48:18regeneration can happen and continuing.
48:21So that same team member can then be
48:23used by a different engine and you got
48:26that continuity.
48:28So the before it is literally as I said
48:34marked down. So these are just folders.
48:37The yellow ones are folders. The purple
48:40ones are markdown files. And so these
48:43are just the files that sit on my
48:45machine. You'll see that in the playbook
48:49section. So this is how the rules of the
48:52game basically. There's the
48:53architecture. You would have seen the
48:56brand. So my writing style, the dynamic
48:59four brand guidelines, bunch of assets.
49:02Um there's tools there skills. There's
49:05workflow which includes things like a
49:06decision log um and an agent daily log.
49:10So everything that the agents do it
49:12actually logs into a running narrative
49:14of what it's doing. So it's always
49:16observable and an action board which is
49:20a cambban board where my agents talk to
49:23each other and take work through a
49:24workflow.
49:26So all of these are literally just files
49:29on my machine. There's not a single line
49:31of code in any of this.
49:36And
49:40a big one is when you're thinking about
49:43your team, what would your synthetic
49:45team need to know, follow, and never
49:49touch? So, if you had that idea of this
49:52team member you'd like to bring in who
49:55just happens to be synthetic,
49:57what are the things that you need to
50:00that need to know?
50:06And whatever you're thinking of there,
50:08that's the foundation. That's a lot of
50:10the context. That's the material. So if
50:12you had a human new starter in your
50:14team,
50:15what's the onboarding guide? What files
50:18do you point them at? What history do
50:20you give them? What do they need to know
50:22about you?
50:24The follow
50:26is the playbook. How do you actually
50:29work?
50:30What are your ways of working?
50:33What are your standards?
50:35How does work get allocated and approved
50:40as done?
50:45And then there's the never touch or the
50:49this is yours, but you can only use it
50:50in this way. And that's the boundary.
50:54So when you got clarity on that and you
50:57don't need to have absolute clarity to
50:58get started, but especially with one um
51:02team member to start with, you can use
51:03that as a way to start painting out the
51:06picture of how should this thing hang
51:08together and as a role that when you're
51:10doing something like this that I reckon
51:12is worth on boarding first because then
51:14they end up being the thing that does
51:18the work and helps you think through
51:21things and guide you through it and
51:24helps you look at the unintended
51:26consequences of different decisions that
51:29you might be making. But that chief
51:32information and intelligence officer
51:33CIO,
51:35I use that role as my guide and that
51:39actually turns everything into a
51:41coherent system that learns and really
51:44importantly self-heals. There's a whole
51:46bunch of skills that I build up. So
51:48that's a intentional loop as opposed to
51:51something that happens accidentally or
51:52only when something breaks. But it's uh
51:56everything you've seen to this point,
51:58all those files, you don't even need
51:59them all. You just need a little bit
52:01about who you are, how you like to work,
52:03spin this um team member up, and then it
52:08can guide you through the rest.
52:11And so in terms of spinning up a team
52:13member,
52:15it can be as easy as just saying this is
52:21who you are.
52:23So no mega prompts. And I'm actually a
52:27little bit anti-giving people prompts
52:29because
52:31when you copy prompts, you tend not to
52:33think about what you're actually asking.
52:34And 80% of that power of the interaction
52:37is actually the thinking that goes into
52:39it. and the principles and how to frame
52:42what am I actually asking.
52:45And so what that's just done is I just
52:47said you're my demo CIO.
52:51Now go set up your files. And I've got
52:54some skills in the background. So it
52:56knows what that means. And again, those
52:59skills are literally just text files.
53:02And then it's gone and created its
53:04three-state file. So that team member
53:06now has its context. And that's a bit of
53:10a shell at the moment, but it has some
53:11sections that it knows that it needs to
53:13get more information on, including a
53:15data library. It has its memory, which
53:17it right now the only thing that's got
53:20on its memory is I was just spun up. And
53:23so that will live within the memory
53:25file. And then it also has an archive
53:28file, which we'll show you in a moment.
53:31And so now there's a new team member.
53:36And it literally just created those
53:38three files. So they're text files
53:40sitting on my hard drive. No line of
53:42code, nothing stuck in the model,
53:44nothing stuck in an app. Literally three
53:46text files on my hard drive.
53:49And so the during part is obviously when
53:53things get more interesting. And one of
53:56the other little projects I'm doing at
53:57the moment is my Swiss meanderings. It's
54:00a retrospective photo journal from my
54:02various meanderings around Switzerland
54:04and Montro Jazz Festival. And so I've
54:08just said, "Go tell me about that
54:10project. What is it? Where am I at?
54:14What's happening?"
54:16And it didn't have I didn't give it that
54:18information. So it's now gone and
54:20checked out another team member's work
54:22who is the team member that's been
54:24running all of this and said, "What is
54:27this thing?" And it was able to do that
54:28because there's a team directory. So it
54:30was able to find Swiss Meanderings as
54:32part of the team and project directory.
54:34And now it's brought that across and I
54:37can ask it some questions just to show
54:38that it is actually getting primary
54:41things. It's also getting things from
54:43the internet. Um this time yesterday, no
54:48this time 25 years ago I was in uh
54:51Lashards and so after this session I
54:55need to go write my next post and sort
54:58through my photos from 25 years ago. So
55:02all of this has happened without me
55:04providing any real context. So there was
55:06no rebriefing and everything it's done
55:09it now knows and it knows in a durable
55:11way and in a way where that information
55:13stored on my hard drive.
55:16The other thing is that I get and this
55:20is a demo one. Um so it's going to be
55:22some probably some weird cards on there
55:24but this is actually in Obsidian. So I'm
55:25not using Linear or Jira or Trello or
55:29ClickUp or anything else. This is a
55:32canban board completely textbased
55:34sitting in Obsidian.
55:35And each of my agents knows this is
55:38where you go to get work. I've got
55:40another role which is the chief of staff
55:42who is the orchestrator of this board.
55:43So that farms out work and it'll go
55:47through everything that's in next
55:49through to um work in progress. When it
55:53thinks it's done, it'll move it to QA.
55:56that then pings the QA role and the QA
55:59then inspects it and does a test against
56:03the acceptance criteria and definition
56:05of done that was scoped out before it's
56:08as it came out of the backlog and into
56:09next. When the QA gives it a pass it
56:13goes into review and that's when I then
56:17review it and if I think it's good then
56:20it goes to done. If I think it's not
56:21good, I say to the chief of staff, you
56:23need to sort something out here. And it
56:25will go back and tell the team member
56:26directly.
56:28And so I'm the only person, again, I am
56:31the human who is accountable. I'm the
56:34only human that says, "Yes, this is
56:36actually done." And marks as complete.
56:40Again, no third party tools required for
56:42that workflow. Agents talking to each
56:44other, all sitting as text files. No
56:47complex database, no extra
56:48subscriptions.
56:51And so
56:53when you're mid session, it's got to end
56:55at some point. And
56:59making sure stuff gets written back is
57:01really important. Otherwise, things are
57:03trapped in the model until to a degree
57:04until you do this step. So there's a
57:07skill called /archive and that knows
57:10then to go and update the context with
57:13anything that's important.
57:15the memory with a standard format
57:17including next actions and decisions and
57:20key insights from the session. And the
57:24key thing it does is it creates a full
57:26archive transcript. So one of the things
57:28a lot of people don't know with these
57:30apps is
57:32they're not clouds synced that they are
57:34on your hard drive. There's a little bit
57:35different with codecs but not properly.
57:38um extremely brittle with clawed. So all
57:43of your conversation is in a hidden JSON
57:45file on your hard drive. And for most
57:48people that might not mean anything and
57:50that's exactly why you should be scared
57:51about it because it's a hidden system
57:53file or a script. And that actually
57:56contains the full transcript. And
58:00without that, even when Claude does a um
58:03compaction, quite often you can't scroll
58:06up. So anything that happens before that
58:08compaction you can no longer see let
58:11alone if there's an app corruption or
58:13any other uh device crash or anything
58:15else that um does happen where all your
58:18work can then be disappeared. So this
58:21archive step means it's being written
58:23back to that text file and now that text
58:26file I've got synced into multiple
58:27places across multiple devices including
58:29my phone. So I can always see that
58:32information and access it there.
58:37And just to show that, so all these
58:39other ones were in codeex/ chatgbt
58:43and I've switched over to claude
58:46opus 5 fresh model.
58:50And so I've just been able to now
58:53continue that session. I've spun it up
58:55and said you're the demo CIO.
58:59It's then picked up its state files from
59:01the hard drive and then I can say to it
59:04what's the last thing I asked you and it
59:07continues as if it's the same team
59:08member just now in a different model. So
59:12I just switch from chat GBT to claude
59:16and nothing changed. It's got all the
59:18same context, all the same memory, the
59:20same as if I had the same conversation
59:22because it's got the full archive
59:23threads. So it's all there ready to go.
59:30So the big question, what work would you
59:33love to be able to continue without
59:35[laughter] having to explain everything
59:38over and over again? And ideally,
59:41whatever that piece of work is, if it
59:43relates to what's that team member you'd
59:45love to have, if they're in sync, then
59:48you've probably got some pretty obvious
59:50next things to do.
59:55So, I'm going to fly through this
59:57because we've sort of been talking about
59:58it as we go. And so, this is probably
1:00:00more for if you do want to refer back to
1:00:02it and understand the layers and the
1:00:04moving parts. But, as a good designer, I
1:00:07start with design principles. And
1:00:11because I like to go from strategy all
1:00:13the way through to pixels, DevOps, and
1:00:16AI, that actually becomes real and
1:00:18becomes architecture. So the value and
1:00:21design principles for this number one
1:00:23said it many times but it is absolutely
1:00:26critical is that the human is still
1:00:29accountable no matter what we're the
1:00:32ones on the hook
1:00:35keeping and protecting that durable
1:00:37layer. So all of that value that you
1:00:39built up is then stored in files formats
1:00:42and places you control not in the
1:00:45interwebs. So being able to keep that
1:00:48clean, controlled, but you still need to
1:00:50do your backups obviously. Um, but
1:00:53you've then in control of your own data
1:00:55and all of that value that you created.
1:00:58It's really important to make the
1:01:00boundary explicit what AI can do, what
1:01:03it can't do. And as part of that, I
1:01:06think it's really important and it's
1:01:08does take effort is that we keep all
1:01:11that work that the agents and the
1:01:13synthetic team are doing observable and
1:01:15traceable. And that's why one I, you
1:01:18know, it's one of the key reasons I have
1:01:20that archive with the full transcript so
1:01:22I know all the things I said and all the
1:01:24things it said, all the outputs are
1:01:26created and also things like the agent
1:01:28daily log. So each team member is
1:01:31literally writing to a daily log saying
1:01:33I just did this piece of work which
1:01:35means there's a record of it that I can
1:01:37see but then other team members uh other
1:01:39synthetic team members can see that as
1:01:41well. a decision log if they are
1:01:45explicit decisions where I say write
1:01:46that to the log or there's actually some
1:01:50thresholds where as we're talking about
1:01:51things and decisions are actually
1:01:53tacitly made I also have those written
1:01:56off to a decision log as well so I can
1:01:58always go back and go what was that
1:02:00decision that we made and how did that
1:02:02work key thing of keeping engine
1:02:04swappable we don't want to be locked in
1:02:06and hostage especially as these things
1:02:09get more value um and they're more
1:02:11likely to get more value from here than
1:02:12less
1:02:14um design for resilience. So it's very
1:02:16safe to assume that tools will fail.
1:02:19Things will crash. There will be model
1:02:21changes. You will change your device. Um
1:02:25pricing will change.
1:02:28So making recovery actually part of the
1:02:30system design and designing for that. So
1:02:32it's not a surprise when it happens.
1:02:34Those recovery paths are already um
1:02:37natural and built in. A key one is that
1:02:41building for regen. So all of that value
1:02:43gets written back and then the reusable
1:02:46pattern be um leaves you know is then
1:02:49built on um so the system becomes
1:02:51stronger over time. And if we optimize
1:02:55for that compounding and we do real work
1:02:57with it, we don't try and do it all at
1:03:00once. We let that structure earn its
1:03:03place. So you don't want to go too heavy
1:03:05too soon, but you keep what works. you
1:03:08write it back and then let the
1:03:10compounding happen. And compounding is a
1:03:12beautiful thing if it's taking you in
1:03:13the right direction.
1:03:17So that as a living loop
1:03:22to try and talk about in simpler terms
1:03:26is a quick fly through of the
1:03:29foundation.
1:03:32So that's all your data or your bits and
1:03:34pieces. The playbook, what are the rules
1:03:36of the game? How does the system
1:03:38actually work? The boundary,
1:03:40what's in and out, the engines are on
1:03:43the outside of the boundary. They are
1:03:44rented intelligence. They are not the
1:03:47system.
1:03:50the studio where all the fun stuff
1:03:52happens, where I get to hang out with my
1:03:54synthetic team and then I can then bring
1:03:58us as a collective set to other people
1:04:02that I collaborate with who also have
1:04:03their synthetic teams. But then we get
1:04:05to collaborate at that level rather than
1:04:07trying to mash all the synthetic teams
1:04:10and agents together.
1:04:12The regen happens because it goes up and
1:04:15it gets written back down. And then
1:04:17through that playbook we get that
1:04:19continuity.
1:04:21And so
1:04:24I'm going to fly through this.
1:04:28The foundation, all the good stuff,
1:04:31playbook,
1:04:33the boundary,
1:04:36the engines. This is all so you can
1:04:38refer back to it if you want later.
1:04:41and the studio.
1:04:45And so some quick reflections
1:04:48on what I've learned over the last
1:04:52to one degree or another these
1:04:53experiments have been in progress for 30
1:04:55plus years. But the Genai specific parts
1:04:58have been in play for about three and a
1:04:59half years.
1:05:01The number one thing which I've said
1:05:03over and over, so hopefully you
1:05:05understand. I think it's really
1:05:06important is that
1:05:08I when I'm working with my synthetic
1:05:10team, I'm accountable for whatever it
1:05:12does. The quality of the output, if it's
1:05:14good, bad, or otherwise, I can't say the
1:05:16computer did it. It's me and I'm the one
1:05:19who is on the hook for that and
1:05:21accountable.
1:05:23There's so much change. there's so many
1:05:25things to learn um that experimenting
1:05:29with things rather than I won't make a
1:05:31start until I've got the right time or
1:05:34the right conditions um they'll never be
1:05:37right and it's not going to get still so
1:05:40jumping in experimenting playing with it
1:05:43and embracing the reality which can be
1:05:47inconvenient at times when progress is
1:05:49nonlinear
1:05:51so that are times when things crash or
1:05:54features regret press or disappear or
1:05:56models literally go offline. Um and so
1:06:00that's going to continue to happen. So
1:06:02one designing for that to be a reality
1:06:04and how we um think about our workflows
1:06:07and the um how critical that our
1:06:10dependence is on some of these tools is
1:06:14a really important things especially in
1:06:16you know mission critical type
1:06:18environments. We need to be mindful that
1:06:21we can't assume everything's going to be
1:06:23up all the time and working as we'd
1:06:25hope. Every model still has this is an
1:06:28experiment that makes mistakes. Um we've
1:06:30I think most of us have become blind to
1:06:32that little warning. Um but we're in a
1:06:35global experiment and we are part of the
1:06:39um model makers lab. Um this isn't a
1:06:42finished product and they don't pretend
1:06:44it is either.
1:06:46One that is a hard lesson to learn,
1:06:49remember, apply is that multitasking is
1:06:53really expensive.
1:06:55And it's something which over years and
1:06:59decades of sort of built habits and
1:07:02rituals and momentum around trying to
1:07:05design for better focus and less
1:07:08multitasking.
1:07:10The nature of working with a synthetic
1:07:13team can break that really quickly
1:07:16because it's off paralleling your
1:07:18thinking in multiple directions
1:07:20sometimes at once. And as the human,
1:07:24we're the slowest for a lot of this
1:07:26stuff. And if we're going to stay as the
1:07:30person who's actually accountable, then
1:07:32we need to actually review it and not
1:07:33get lazy and go good enough. The model
1:07:36did it, so it must be right. And we'll
1:07:37chuck it out with our name on it. I
1:07:39again encourage no one to ever do that
1:07:41or at least not for a long time yet. And
1:07:45it means that task switching of looking
1:07:46at what different team members are doing
1:07:49and doing that in rapid succession. So
1:07:52trying to build workflows around that
1:07:54which is where the camb board and I'm
1:07:56now not needing to be the glue that
1:07:58paste things from one team member to
1:08:01another. I can actually now have that
1:08:04workflow managed more directly with a
1:08:06separate QA role who I've told
1:08:09explicitly or we've co-designed it often
1:08:12what the acceptance criteria is. So when
1:08:15I'm reviewing it, it's much less likely
1:08:17to need rework and for me to be jumping
1:08:19in and micromanaging and spoon feeding
1:08:21the next step in a process. But as you
1:08:24start out, um, there's the fear of being
1:08:29kind of the thing, the bottleneck that's
1:08:31holding everything up. And it will
1:08:33probably be true, but finding a way to
1:08:36not let that completely derail your ways
1:08:38of working where you have built up good
1:08:40habits.
1:08:41Lots around data sovereignty and
1:08:43stewardship. It's all running on our
1:08:45data. How that data got there in the
1:08:47first place. There's lots of question
1:08:48marks and some of them aren't even
1:08:50questions. um thinking about what that
1:08:53might mean going forward,
1:08:56the responsible AI, the ethical AI, uh
1:08:59the climate impacts, these are all
1:09:00things not to ignore. Um but I think
1:09:03they're also things not to get freaked
1:09:04out about and and wave our arms around
1:09:06and therefore not play. I think the way
1:09:09we play actually has a big degree in
1:09:11shaping how some of these things are
1:09:13made and what um patterns go forward and
1:09:16which ones sort of get called out. That
1:09:19might be overly optimistic, but that's
1:09:20what I'm going to hold on to. And a key
1:09:23one is, and it's probably come all the
1:09:25way through, is resilience over lock in.
1:09:28So, a lot of what I've been talking
1:09:30about has been really filling product
1:09:32gaps in the current version of the
1:09:34products. Um, and a lot of that is
1:09:37because a lot of those data gaps and
1:09:39memory and things being backed up, etc.
1:09:43have come because codeex and clawed code
1:09:47came from development environments where
1:09:49all of that was written back or most of
1:09:50it was written back into repos um and
1:09:53backed up to the cloud and there was a
1:09:55lot of other backend infrastructure
1:09:56there as they wrapped the skin around it
1:09:59to make it more accessible to knowledge
1:10:01workers who don't have that
1:10:02infrastructure. They didn't bother
1:10:04filling those product gaps. And so right
1:10:07now it means filling those product gaps.
1:10:10That's not going to last forever and are
1:10:12already making big steps to try and sort
1:10:15out memory and do a bunch of those
1:10:17things. The way they're choosing to sort
1:10:19that out, of course, is so you put even
1:10:22more of your stuff into their platform.
1:10:24They hold it for you nice and safe, but
1:10:26they're the only ones that hold it nice
1:10:28and safe. And so you get product lock in
1:10:31and you're a hostage to their platform.
1:10:33And as they make changes, and they
1:10:34already are and have, as they change
1:10:37their pricing, as they change different
1:10:38ways of operating, if you want to be
1:10:40able to access your stuff and keep all
1:10:42the value that you built up, you've got
1:10:45no real choice but to use their
1:10:47platform. Um, so I don't want to be in
1:10:51that situation and I recommend most
1:10:53people don't want to be in that
1:10:54situation. So designing around that so
1:10:57there's no single point of failure in
1:10:58the tools, the different providers of
1:11:01the platforms or their pricing models as
1:11:03they change.
1:11:05And so the big question
1:11:08cuz that was a blah of a lot of stuff
1:11:10and there's probably going to be some
1:11:11things to unpack.
1:11:13But to get started,
1:11:17what's one real piece of work where you
1:11:19could test this? Not have it all
1:11:21magically working. So test and it could
1:11:25you know ideally in a safe to fail way
1:11:26where it's a safe place to experiment
1:11:28learn some stuff and be able to iterate.
1:11:32Have you got anything in mind where you
1:11:34go I reckon I'll use it in this way? You
1:11:37might have your own music meanderings
1:11:39where it's a creative project off the
1:11:41side but all went sideways no one's
1:11:44going to be impacted. No critical data
1:11:47is lost.
1:11:49If you can think of some things like
1:11:51that within your life and whole of life,
1:11:54it doesn't all have to be business and
1:11:55workrelated, then that's probably a good
1:11:59place to start. And starting with
1:12:02something nice and simple and safe,
1:12:04building things some things up around
1:12:05it. See what's working, what's not
1:12:07working. And as you feel comfortable
1:12:11with the way it does certain things and
1:12:13how you're um backing up data and what
1:12:16that whole experience feels like, then
1:12:18you might choose to give it more. But
1:12:21not giving the models and the apps all
1:12:23your things all at once. I would
1:12:25definitely recommend against that. Um so
1:12:27thinking about what's that real thing
1:12:29that you could do, but nice and slow.
1:12:33So stuff you could do next is what is
1:12:38that thing that's you know it's valuable
1:12:40but it's and it's real but it's not
1:12:43necessarily um high risk if anything
1:12:46goes a bit weird with it. If you create
1:12:49one little folder and workspace so you
1:12:51don't need to do a a yearong curation of
1:12:55all your data sets and get them all nice
1:12:57and clean before you start. But just
1:12:59start with the data that's you need and
1:13:01you feel um okay or good about putting
1:13:04into the models into the neural network.
1:13:08Write that simple context. Save some
1:13:10useful stuff. And again, all of this is
1:13:13in the context of using the desktop app.
1:13:16You won't be able to do a lot of this
1:13:17and get the same kind of value out of it
1:13:19if you're still using it in the browser.
1:13:23But think about the principles of I want
1:13:25to be able to see what it's actually
1:13:27done. what work did it do? What
1:13:29decisions did it make? Did I agree with
1:13:31that decision? Did it operate within
1:13:33what I think are safe boundaries? So,
1:13:36being able to set up um a way to then
1:13:38build that and then the stuff that's
1:13:40working, you keep using and you turn
1:13:44those into skills. Riley mentioned
1:13:45before, there's lots of ways to create
1:13:47skills now. That could be, you know,
1:13:49copying a workflow is through a screen
1:13:51record, which is they're doing natively
1:13:52now. It can literally be I just had a
1:13:55conversation with you and we created
1:13:56this output. Now reverse engineer the
1:13:59good version of that back into a skill
1:14:00that's reusable. Uh which is the method
1:14:02that I still use cuz I don't have much
1:14:04repeatable stuff to screen record. So
1:14:07the key thing is to start with one real
1:14:10workflow
1:14:11and then let the structure earn its
1:14:13place from there.
1:14:15And so
1:14:17this has all been about the principles
1:14:19and the shape.
1:14:21Um, but if you want some help navigating
1:14:25all of this and doing a personalized
1:14:27version where it's designed around the
1:14:30way you work, your goals or your data,
1:14:32the risk profile, um, then happy to have
1:14:35a chat and also curious if that's um,
1:14:39you know, there's different versions of
1:14:40this I'm thinking of doing depending on
1:14:42where there's interest. Um this is
1:14:45definitely where it starts crossing over
1:14:46into the the paid version of things
1:14:48rather than the free online workshop.
1:14:50But there's versions where doing this as
1:14:54um sort of onetoone or with a team in
1:14:56sort of more organizational context or
1:14:59even doing like a a buildalong session
1:15:01as sort of a uh online where guiding you
1:15:05through your own but in sort of a group
1:15:06setting. Um so there's a few different
1:15:08options there. ways I'm looking at
1:15:11possibly taking it forward depending on
1:15:14what people are interested in.
1:15:17So, that's the fire hose.
1:15:21There was plenty of good stuff in the
1:15:22chat.
1:15:25Any questions that have come up that are
1:15:27burning,
1:15:28even if they're not burning?
1:15:32I thought there was two um pretty
1:15:34interesting ones. Uh well uh one was um
1:15:38how do you start in terms of you've got
1:15:40skills behind some of that um some of
1:15:43the work that you were already getting
1:15:44when you got it to spin up its own
1:15:46little um uh sort of on boarding
1:15:48essentially for whatever that new sort
1:15:50of agent was.
1:15:51>> Um so yeah what would you say to uh
1:15:54would be the first thing you would start
1:15:55with? Should you would you start with a
1:15:57skill or would you start with another
1:15:58way or or something like that?
1:16:01So, one way is I can share my version of
1:16:04it. Um, and at least the headings
1:16:08because the specifics is where it gets
1:16:11less valuable. Um, the key thing is if
1:16:14you even just say to a synthetic team
1:16:18member, you are my guide. help me do
1:16:22this and interview me to work out what's
1:16:25important and then you can and that's
1:16:29how I built my different ones. So I
1:16:32didn't start with like a master plan and
1:16:34go it should do this this and this. It's
1:16:36happened iteratively and organically
1:16:38over time and then reverse engineered
1:16:41into a school a skill that's repeatable.
1:16:43And so a lot of this is not about trying
1:16:46to get it right on the first shot. And
1:16:49there are loads of different skill
1:16:50libraries and stuff that you can go and
1:16:52download. Um that can be useful if it's
1:16:55a like a really repeatable common task.
1:16:58Um but a lot of the value is actually in
1:17:00the thinking, not the specific text
1:17:03files. So being able to have that
1:17:04interaction with your agent and get it
1:17:07to ask you questions um is where a lot
1:17:11of the value is. And then it's making
1:17:14something that's repeatable. Repeatable,
1:17:16but not trying to make it repeatable for
1:17:18step one.
1:17:22Awesome. Yeah, I got it to interview me
1:17:24more than once, but yeah. Um, I think
1:17:26that that's a good way to do it. Um, cuz
1:17:28I can talk off the top of my head about
1:17:30something, but like it takes me a little
1:17:32bit of time to write it down. [laughter]
1:17:35And and I'm sure everyone's seen now
1:17:37like if you can get comfortable talking
1:17:39to your device um you can do that in
1:17:42dictation mode. You can now do it with
1:17:44the models with the real-time text mode.
1:17:48Um there's tools like whisper flow. Um
1:17:51there's a Google one as well that they
1:17:53released a few months ago. There's lots
1:17:55of options which can if your natural way
1:17:58of processing is verbally then that can
1:18:03be a very um quick way to get a lot of
1:18:05volume um out of your head. Personally
1:18:09I'm of the school of I write to think
1:18:12and so I don't have quite the same level
1:18:17of experience with it. Uh I do use it
1:18:19for some stuff, but um yeah, being able
1:18:21to have that um quick interaction and
1:18:25for it to iterate over time, that's the
1:18:29key thing. So I'm still now like
1:18:31literally today saying to my so it's
1:18:35doing this you're not printing the you
1:18:37know threads aren't printing the date
1:18:38properly like they were what's going on
1:18:40and it'll go find the root cause for me
1:18:42go h we need to strengthen up that rule
1:18:45here and do you approve this rule and
1:18:47then I say yes and it goes and rewrites
1:18:49um into a text file so you know it's
1:18:52constantly being iterated um but it is
1:18:55having it's more about the mindset and
1:18:58having that approach to it than it just
1:18:59having a perfect prompt.
1:19:02>> Yeah. Great. Um, another one was, "How
1:19:06many hours do you think you put into to
1:19:08the point of actually getting positive
1:19:10compounding return?"
1:19:14>> Yes, good question, Jess. Um,
1:19:17to start seeing some return, it's it's a
1:19:20hard one to answer because when I
1:19:22started, the models were about 1% of the
1:19:25capability and and the apps around them
1:19:27of what they are now. Um, so it's been 3
1:19:32and 1/2 years of constant
1:19:34experimentation
1:19:35and a lot of those experiments were
1:19:38extremely frustrating and no value
1:19:41return. The value return has been
1:19:44understanding how things have worked and
1:19:45and the deeper knowledge and experience
1:19:47around that. It's now at the point where
1:19:50you could spend a day or two, even less
1:19:54to start getting some value immediately
1:19:58to get a a fuller system built out.
1:20:01That's, you know, a few hours over a few
1:20:04weeks and months. But to have one thing
1:20:07done well, that's probably a few hours
1:20:11now with a little bit of base
1:20:12infrastructure as a starting point.
1:20:17Yeah, I would agree on the pretty quick
1:20:19turnaround. Um, I've just changed jobs
1:20:21and I've rebuilt my own little one at
1:20:23work that I've just started as a brand
1:20:25new sort of instance and yeah, it didn't
1:20:27take me more than
1:20:30yeah, half like not even a couple of
1:20:32hours. Um, yeah, but probably about
1:20:33three or four three or four hours I'd
1:20:35say. Um yeah,
1:20:38so it's my own personal onboarding
1:20:40system to the company as well cuz I kind
1:20:41of had to gather some enough internal
1:20:43information to be able to kind of you
1:20:45know um uh on board myself and so at the
1:20:47same time I just onboarded you know
1:20:49essentially a new nent kind of OS as
1:20:53well starting to anyway
1:20:56um does any
1:20:57>> one of sorry uh just a quick one
1:21:00something we were talking about the
1:21:00other day Jess um that makes a huge
1:21:03difference is when you're setting
1:21:05setting up when you're on boarding and
1:21:07you sort of mentioned a little bit there
1:21:08Riley when you're on boarding and you're
1:21:10setting up your context
1:21:12it's makes a massive difference if you
1:21:14go through a step of actively creating a
1:21:16data library as part of your context
1:21:18file. So curating that a bit not as in
1:21:22you need to go get all the files and put
1:21:24them in one place but getting the the
1:21:28model to go within the boundaries that
1:21:30you've said of go have a look around
1:21:32here and then do a bit of a review of
1:21:36the data it's found and the different
1:21:38sources to go yes this is current and
1:21:40reliable this should have a higher
1:21:42waiting than this other thing from 3
1:21:45years ago under a whole bunch a whole
1:21:47different strategy. So that effort in um
1:21:51just reviewing and curating the data
1:21:53library bits means that you got a much
1:21:56higher quality base and context than
1:21:57build off um and it's you get a lot less
1:22:00drift and um also helps with the context
1:22:03window but won't go into that side of
1:22:06things.
1:22:08>> Cool.
1:22:12>> Jesse, did you have a spot? Sorry.
1:22:15>> Sure. Thanks. I was nervous I was going
1:22:17to um take up time and go down a rabbit
1:22:19hole. But so yeah, I played with that
1:22:21data library idea for about 2 hours on
1:22:23the weekend. What I found was with
1:22:25co-pilot, it kept coming back saying,
1:22:27"Yeah, yeah, I can see it all." And so
1:22:30what was your phrase, Ben? Trust and but
1:22:33verify. No.
1:22:35>> Yes. KGB phrase.
1:22:37>> KGB phrase. Right. So I [laughter] was
1:22:39like, "Okay, well uh tell me what's in
1:22:41that document." And I I I I I think I
1:22:45eventually realized that it couldn't
1:22:46actually really see what was in the
1:22:48company SharePoint. Um so so but I was
1:22:53still learning um um the the
1:22:57interaction. So I I still got something
1:22:59out of it, but I still don't know if
1:23:02technically it can see the things I want
1:23:03it to see. Yeah.
1:23:05>> Yeah. And and that's a good example of
1:23:07the 2025
1:23:092024 version of a lot of these
1:23:11experiments uh where
1:23:15with co-pilot uh Microsoft copilot
1:23:18you're probably still in a situation
1:23:20where you'll need to curate those files
1:23:23into a safe place and say look at that
1:23:26um because it's a bit weak on doing
1:23:29stuff that the other models are able to
1:23:31do. But if if you do the show me what's
1:23:34in it and it comes back with sensible
1:23:36stuff, then you know that's part of that
1:23:38test and learn of you're saying one
1:23:41thing but show me evidence of that
1:23:42before I believe that and start doubling
1:23:44down on it. Um otherwise definitely in
1:23:47hallucination land.
1:23:48>> Yeah, love it. Thanks. [laughter]
1:23:53>> Cool. Um probably got time for one one
1:23:55more question. U what have you so I've
1:23:59got a couple here but um what have you
1:24:03tried the Australian sovereign AI model
1:24:07um from uh trailers data trail data is
1:24:11that you think no I haven't I haven't
1:24:14played with any of the um local models
1:24:20um
1:24:21it's something that sort of keeping an
1:24:23eye on But I have intentionally taken
1:24:28predominantly a non- tech approach to
1:24:31Gen AI. Um, all of my experiments over
1:24:34the last three and a half years, I've
1:24:36stayed away from having to do anything
1:24:38that's starting to get into code or
1:24:40repos or stuff that I do when I'm
1:24:43designing and building digital product,
1:24:45but not stuff that I want to do in a Gen
1:24:48AI context because I know how fast it's
1:24:51moving and I see a lot of that's going
1:24:53to be obsolete before um, it's valuable
1:24:57and then it just builds up the overhead
1:24:59of trying to learn and manage all this
1:25:01stuff. Anyway, um but I think the at a
1:25:03principal level and a conceptual level
1:25:05of the sovereign versions of models, I
1:25:09think it's going to be
1:25:12over the next 6 to 12 months one of the
1:25:14hottest topics, especially now that
1:25:16we've had um the US government ban and
1:25:21make Anthropic take Fable 5 offline for
1:25:24a few weeks.
1:25:26um they have said that they now want to
1:25:29early um view of new models models being
1:25:34released. Australia has talked about
1:25:36doing a similar thing. Um OpenAI has
1:25:40talked about giving uh the US government
1:25:435% shares which does that accelerate
1:25:47preference actually put up guard rails.
1:25:50All of these things are not really
1:25:52known. Um so having access to the stuff
1:25:57that we need and when we need it is
1:26:01definitely um in flux at the moment. Uh
1:26:03and then in the last week as well a lot
1:26:06of the couple of the Chinese open source
1:26:08models are performing near benchmark of
1:26:11um opus 4.8. So pretty much Frontier
1:26:15model with the um the weights and and
1:26:19being open source, but you need a
1:26:21there's about 2 tab of data for to be
1:26:25able to tink with the weights and run it
1:26:27locally. So it's not yet at the point
1:26:28where it's light enough to really be
1:26:31able to um use it for most people. But I
1:26:34think that side of things, the sovereign
1:26:36aspect, the offline, the open source,
1:26:39um, are all conversations that are going
1:26:40to ramp up even more over the next six
1:26:42to 12 months. And that's also part of
1:26:45the design principle that I've built
1:26:46around here is those engines are
1:26:48swappable. And so I'm hanging out for
1:26:51the time when I can be truly offline and
1:26:54have a local model that can do the bulk
1:26:57of what I need and I'm not actually even
1:26:59needing the internet to be up anymore.
1:27:01Um, and I reckon that's no more than 12
1:27:03months away.
1:27:07>> Awesome. Thank you so much for your time
1:27:09uh tonight, Ben, and sharing your uh
1:27:11your setup and what you've learned along
1:27:12the way. Um, it was uh great to have you
1:27:16uh back and hopefully um uh you get some
1:27:19uh yeah, uh anyone who's keen to uh find
1:27:22out more, then we can um certainly, you
1:27:25know, follow this up um uh with Ben and
1:27:28find out what he's doing. Um, and
1:27:30>> and if and if you hit that QR code, um,
1:27:33you can book a discovery call, but you
1:27:35can also see music meanderings and photo
1:27:38meanderings. Um, so if you curious about
1:27:41what these little projects are, you'll
1:27:43see that on that same QR code because it
1:27:45goes to my link tree.
1:27:48>> Bonus bonus ad there. [laughter]
1:27:51>> Um, thank you very much. We will follow
1:27:53this up with a um the recording um all
1:27:56things being equal at the end of this uh
1:27:58when we hang up. Um but [laughter] yeah,
1:28:01join me in thanking Ben and um yeah,
1:28:05>> awesome.
1:28:06>> Thank you very much.
1:28:07>> Thanks everyone. Great to see you all
1:28:09and uh more soon.

About this transcript

This page contains the full transcript of Build Your AI Operating System: Regen OS. Human-Centred AI Community Workshop by Ben Pecotich by Dynamic4, generated from the public captions YouTube serves with the video. The transcript has 14,134 words across 2,061 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.