Latest / AI DTC WTF / A "Summer Fling": How Anthropic's Best Model Got Pulled

Transcript

A "Summer Fling": How Anthropic's Best Model Got Pulled · AI DTC WTF

• 0:03
  If you work in D2C and you use AI and
• 0:05
  you're wondering what the
• 0:07
  F is going on every week,
• 0:08
  this is your podcast, the AI D2C WTF
• 0:11
  podcast, your home for
• 0:13
  tactical tips, strategies, and
• 0:14
  ideas that you can implement right now in
• 0:17
  your AI workflows to make
• 0:18
  your brand or agency more money.
• 0:22
  Welcome to another episode of AI D2C WTF.
• 0:26
  And today we are talking about Fable,
• 0:29
  the Claude model being pulled down by the
• 0:31
  U.S. government for
• 0:32
  reasons that maybe you know,
• 0:34
  maybe you don't know, and what happens to
• 0:38
  your AI stack when this
• 0:39
  model vanishes overnight.
• 0:41
  So I'm here with Thomas Moen, my homie
• 0:45
  from Norway, who is becoming an absolute
• 0:48
  legend in AI in Europe
• 0:50
  and in other places too. And just super
• 0:53
  excited to talk about this because I
• 0:54
  think it happened and
• 0:55
  then everyone was like, wait, what's
• 0:56
  happening with this? And
• 0:57
  then it just said on Claude,
• 0:58
  Fable's unavailable. And I think this
• 1:01
  creates a lot of interesting
• 1:02
  precedents, right? In terms of
• 1:04
  like models being pulled down. So, you
• 1:06
  know, it was available for
• 1:07
  three days. That's how long
• 1:09
  Anthropics most powerful model lasted.
• 1:10
  And it was a great day, Andrew.
• 1:13
  I had a lot of, I was so happy. Everybody
• 1:16
  had a lot of memes about
• 1:17
  it being a summer fling,
• 1:20
  a hot summer fling, right? Yeah.
• 1:24
  You know, so look like that's how long it
• 1:26
  was available for
• 1:27
  three days before the U.S.
• 1:28
  government made them shut it off. And if
• 1:30
  you have anything that's
• 1:31
  wired into Fable 5 in those
• 1:33
  three days, right, this episode is your
• 1:35
  fire drill. We're going to
• 1:35
  go through what happened,
• 1:37
  what to do about it, and what are the
• 1:38
  things you need to check?
• 1:39
  Because I think there's a lot of
• 1:41
  pieces that we want to go through. So
• 1:42
  first of all, you know, what actually
• 1:45
  happened? Let's talk about
• 1:46
  this. So on June 12th, actually it was at
• 1:48
  521 Eastern. Anthropic
• 1:50
  received an export control
• 1:53
  from the U.S. government saying, hey, you
• 1:55
  need to suspend access
• 1:56
  to Fable 5 and Mythos 5,
• 1:58
  two of its most capable models, three
• 2:00
  days after launch. The order targeted
• 2:02
  technically access by
• 2:04
  any foreign national inside or outside of
• 2:06
  the U.S., including
• 2:07
  Anthropics' own foreign national
• 2:09
  employees, because Anthropic couldn't
• 2:12
  cleanly segment access by nationality and
• 2:15
  net effect was a full
• 2:16
  shutoff for all customers. And they did
• 2:19
  this because there was
• 2:20
  national security concern around a
• 2:21
  jailbreak of Fable, basically asking the
• 2:24
  model to read a code base
• 2:26
  and spot software flaws. And
• 2:28
  so, you know, other softwares can already
• 2:30
  do this, but it was a
• 2:32
  jailbreak of Fable. But this,
• 2:35
  you know, is sort of, this could happen
• 2:37
  to anybody. This could
• 2:38
  happen to any model. And in fact,
• 2:39
  it probably already has, you know, and
• 2:42
  I'm going to go into the final point,
• 2:44
  which is Anthropic is,
• 2:46
  so they complied, but they publicly
• 2:47
  disagreed saying that it was a
• 2:49
  misunderstanding. So nobody got
• 2:51
  hacked, nobody had a breach of data, but
• 2:52
  it was like, recalled
• 2:54
  like a car part, right? Like,
• 2:56
  or whatever. So, mid-deployment.
• 2:58
  Yeah, yeah. And it's still unclear what
• 3:02
  the jailbreak or the
• 3:03
  security issues are. And it's
• 3:05
  been a lot, a lot of back and forth on
• 3:07
  things. And I think it's also
• 3:09
  particularly interesting that
• 3:11
  before the release of Fable 5 and Mythos,
• 3:15
  Anthropic has been
• 3:16
  spending a lot of time kind
• 3:18
  of hyping how dangerous the Mythos model
• 3:21
  was and kind of how
• 3:22
  potentially it would break the internet.
• 3:25
  And then when they finally kind of
• 3:27
  release it, it only kind of survives
• 3:29
  three days before it gets
• 3:30
  shuts down. But no hacking, no breaching
• 3:33
  of data. And probably it's a mix of
• 3:36
  something you can recreate
• 3:37
  on the other models. And I also
• 3:39
  personally think it has a lot to do with
• 3:41
  the relationship of Anthropic
• 3:43
  and the US government as well, which I
• 3:45
  don't think we need to go into all that
• 3:47
  kind of the rabbit hole
• 3:49
  there, but it's definitely a mix with a
• 3:51
  lot of factors here.
• 3:52
  Totally. But I mean, we should go into it
• 3:53
  a little bit. Like, I think
• 3:55
  it's a couple things to mention.
• 3:56
  One is, is that I think, you know, being,
• 3:58
  I used to be a press secretary in
• 4:00
  Congress in the United States.
• 4:01
  And one of the things that I noticed by
• 4:04
  in being in Congress was that
• 4:06
  there was a lot of people that
• 4:07
  wanted, that wrote legislation that
• 4:09
  wanted to actually think about things
• 4:10
  long term. And then there was the
• 4:11
  actual members of Congress that wanted
• 4:13
  things done yesterday. And it wasn't
• 4:15
  necessarily thoughtful at all
• 4:16
  around like the approach. Okay. And I
• 4:19
  think that we now in the current
• 4:20
  administration in the United
• 4:21
  States, you know, people can disagree
• 4:23
  with me politically, but
• 4:24
  like, it's a technical term
• 4:25
  would be, I think it's a little bit of an
• 4:27
  F show of like, what
• 4:29
  actually is the strategy around AI?
• 4:32
  A lot of things are an F show, but we're
• 4:33
  specifically talking about AI today. And,
• 4:36
  you know, how are we
• 4:37
  going to go forth on this? And I think
• 4:38
  because it's rapid pace, it
• 4:40
  just doesn't match up with how
• 4:41
  government works. And so they don't know
• 4:43
  how to legislate or
• 4:44
  regulate this at all. And so to me,
• 4:47
  you know, what happened was it, they, you
• 4:49
  know, some, some idiot
• 4:51
  who's running the NSA at this
• 4:52
  point, right in the United States
• 4:54
  basically goes and says, and they see an
• 4:56
  issue and they go, Oh my
• 4:57
  gosh, there's this issue. It's exposing
• 4:59
  stuff. We didn't know. Oh my
• 5:00
  gosh, that's big. We should
• 5:01
  shut it down. I mean, like it was, it was
• 5:03
  three days. Like they didn't, these
• 5:05
  models have been able
• 5:06
  to do this. And for specifically, they
• 5:08
  like went to Anthropic,
• 5:09
  I think partially from a
• 5:11
  PR standpoint to say like, we're reigning
• 5:13
  in AI, but I don't think
• 5:15
  it, you know, it really did
• 5:16
  anything because it's not really solving
• 5:18
  the core problem. And I
• 5:19
  think it also shows the
• 5:20
  dysfunction and the disorganization
• 5:22
  around how they're going to approach AI
• 5:25
  as an administration.
• 5:26
  But I also think it sets like a decent,
• 5:29
  or I'm sorry, a dangerous
• 5:30
  precedent that like a government
• 5:32
  can take down a model, right? Like that's
• 5:35
  a problem that it can take down a
• 5:37
  frontier model post launch.
• 5:39
  And the ramifications, which I'm saying,
• 5:42
  seeing Andrew kind of from,
• 5:44
  from, from my side is that
• 5:45
  a lot of people are now very skeptical in
• 5:49
  kind of going deep with the American
• 5:51
  models, right? So I have
• 5:52
  a lot of friends who've been starting to
• 5:54
  experiment with DeepSeq. I have a lot of
• 5:56
  European friends who
• 5:57
  are kind of trying to understand the
• 5:59
  European language model, landscape, et
• 6:03
  cetera, et cetera. So things
• 6:04
  that you wouldn't really think about a
• 6:07
  week and a half ago are now on people's
• 6:10
  minds and people are
• 6:11
  a bit more aware of how unstable the
• 6:15
  current situation is in the U.S.
• 6:17
  Absolutely. I mean, and the irony thing
• 6:19
  of this, I think is funny is that
• 6:21
  Anthropics own safety
• 6:22
  transparency, like area or division
• 6:26
  publicly admitting that no model is
• 6:27
  perfectly jailbreak proof.
• 6:29
  I think that handed the like regulators
• 6:32
  or the government, the
• 6:33
  rationale that they needed.
• 6:34
  Right. So I think it could have been
• 6:36
  anybody. And I think that that's, that's
• 6:39
  a big thing to point
• 6:39
  out. And I, you know, who knows how this
• 6:41
  is going to continue to
• 6:42
  evolve because we're at that phase
• 6:44
  where AI is only continuing to get more
• 6:46
  powerful as time goes on and
• 6:48
  it will get exponentially more
• 6:50
  powerful. And so what is the government
• 6:52
  going to do to, to attempt to regulate
• 6:54
  this? And shutting down
• 6:55
  a model is not like a way to do it. In my
• 6:58
  opinion, there's going to have
• 6:59
  to be, you know, a, basically
• 7:01
  an industry leading group that's going to
• 7:04
  go through and review
• 7:05
  these things. And then there
• 7:06
  probably is going to be a committee in
• 7:07
  Congress on this, you know,
• 7:09
  that, I mean, there probably is,
• 7:10
  but it's not necessarily that effective
• 7:11
  at this point because they're
• 7:13
  now it's June, July, August,
• 7:14
  they're not doing anything in Congress.
• 7:16
  And, you know, and, and
• 7:17
  so the, the presidency,
• 7:19
  they're not the executive branch. They
• 7:21
  don't know necessarily
• 7:22
  how to deal with this.
• 7:23
  And so I think what you see is you see
• 7:24
  things starting to pop
• 7:25
  up like the New York law,
• 7:27
  which came out in June 9th, where you
• 7:29
  have to put a, you know,
• 7:30
  disclosure on your ad creative
• 7:34
  that says, Hey, this is made with AI.
• 7:36
  There's a similar law that's
• 7:37
  actually in the European Union
• 7:39
  that talks about this. And, and the
• 7:41
  European Union one is, is
• 7:43
  classically European. It's like
• 7:45
  insanely broad covers everything. It has
• 7:48
  no, you know, like
• 7:50
  regulation arm. It appears the
• 7:52
  the fines are like, who knows? It's like,
• 7:55
  yeah, there's, there's
• 7:56
  no enforcement arm rather.
• 7:58
  And so it's like, it's a mess. But my, I
• 8:00
  guess my point is that
• 8:01
  you're going to have both,
• 8:02
  you're going to have all this stuff start
• 8:03
  to percolate in terms of
• 8:05
  you're not, you're having
• 8:05
  top-down people pulling models from the
• 8:07
  executive branch of the
• 8:08
  United States. You have bottom-up
• 8:10
  people pulling, you know, like laws in
• 8:12
  New York state, you know,
• 8:13
  and it's like all this stuff's
• 8:15
  going to start to do this. So over the
• 8:16
  next year, especially, I
• 8:18
  mean, you're going to have to be
• 8:19
  ready that you can't be overly dependent
• 8:22
  upon only one model.
• 8:24
  Yeah. And this kind of.
• 8:31
  AI D to C WTF is brought to you by
• 8:34
  motion. And this episode
• 8:35
  is brought to you by runneth
• 8:37
  from motion. And if you run paid social,
• 8:39
  I mean, obviously, you
• 8:40
  know, motion, it's a creative
• 8:41
  analytics platform that everybody uses.
• 8:43
  But the thing I've been
• 8:44
  obsessed with lately is runneth.
• 8:46
  And it's their AI agent that lives right
• 8:48
  inside of your ad data. The
• 8:50
  simplest way to explain it,
• 8:51
  most people use AI by asking it one
• 8:53
  thing, like, Hey, what did we do
• 8:54
  yesterday? How'd it go?
• 8:55
  Runneth is the upgrade of that. You just
• 8:57
  tag it and ask it to do
• 8:58
  almost anything. It goes into your
• 8:59
  account. Does it like build me this
• 9:02
  week's creative recap, pull the data and
• 9:04
  write a brief on what to
• 9:05
  iterate next. Tell me which ads to make
• 9:08
  next. And it goes on and does this in
• 9:10
  plain English. So two
• 9:11
  things that actually make it really
• 9:12
  special, right? So under
• 9:13
  the hood, it's, it's Claude,
• 9:14
  right? And anything Claude can do,
• 9:16
  runneth can do basically
• 9:18
  even better as long as you go to
• 9:19
  access. And second, it's wired from
• 9:21
  everything motion has. So
• 9:22
  the AI tags, the creative, the
• 9:24
  frame by frame breakdowns, and then it's
• 9:26
  trained by $14 billion in ad spend on
• 9:28
  motion as customers,
• 9:29
  right? So it's pretty insane. So the way
• 9:32
  that, you know, we're
• 9:33
  using it, and I know some other
• 9:35
  colleagues are using it, like I looked at
• 9:37
  what they're doing, and we
• 9:38
  talked through it. People are,
• 9:40
  you know, having it drop a morning brief
• 9:42
  in Slack on how the
• 9:43
  account's performing. It gets smarter
• 9:44
  every time it runs. It's less like a tool
• 9:47
  and more like a creative strategist
• 9:48
  working while you sleep.
• 9:49
  So go see it at motionapp.com. That's
• 9:51
  runneth from motion and make ads that win
• 9:54
  without getting lucky.
• 10:00
  You're going to have to be ready that you
• 10:02
  can't be overly
• 10:03
  dependent upon only one model.
• 10:05
  Yeah. And this kind of leads us into why,
• 10:09
  yeah, this leads us into kind
• 10:10
  of why, why you, the audience
• 10:12
  should, should care about this. Why, why
• 10:14
  do e-commerce
• 10:15
  advertisers, agencies, D2P, C people,
• 10:18
  why, why, why should we care about this?
• 10:20
  But, but because I think, I
• 10:21
  think we should care deeply
• 10:22
  about this. And I think it's, it has kind
• 10:24
  of exposed how vulnerable a lot of
• 10:27
  people's AI workflows and,
• 10:30
  and kind of how they look at things and
• 10:32
  how to kind of build things
• 10:33
  and what kind of building blocks
• 10:34
  they're been using that, that suddenly
• 10:37
  maybe doesn't feel so
• 10:40
  future-proofed anymore.
• 10:41
  No, absolutely not. So, you know, why
• 10:43
  should you care? Let's go through some of
• 10:44
  the reasons. Number one,
• 10:46
  single model dependency is now a real
• 10:48
  risk for you. So if you're building
• 10:50
  stuff, creative analysis,
• 10:52
  analysis, competitor analysis, whatever,
• 10:54
  customer support workflow, anything
• 10:56
  you're doing in D2C by
• 10:58
  one specific model, you know, a directive
• 11:00
  like this can take you down
• 11:01
  with zero warning. You're not
• 11:02
  going to have anything backup. So it's,
• 11:04
  it's the same exposure as
• 11:05
  if you had one 3PL, you know,
• 11:08
  or one ad platform, like you have to be
• 11:10
  cautious and aware of this. I'm not
• 11:12
  saying, you know, if Fable
• 11:13
  is, is what you eventually go with,
• 11:15
  that's the most powerful. I think that's
• 11:16
  okay. You just have to
• 11:17
  have a backup, which we'll get into in a
• 11:20
  minute. I think that's a,
• 11:21
  that's one. I think number two,
• 11:23
  agencies feel it harder than brands. You
• 11:26
  know, if you ran 70
• 11:28
  clients through a pipeline of,
• 11:31
  of static image creation, you know,
• 11:33
  that's 70 fires you have to
• 11:35
  deal with basically. So like
• 11:37
  agencies are going to feel this because
• 11:38
  the numbers and the scale that you're
• 11:40
  doing generally are going
• 11:42
  to be, you know, greater than that of,
• 11:45
  let's just say one particular brand. So
• 11:48
  just be aware of where
• 11:50
  you're tied in. And, you know, having a
• 11:53
  plan of, as we talked about this in
• 11:55
  previous episodes, I
• 11:55
  talked about this with Will about AI
• 11:57
  governance, who's running AI governance
• 11:59
  in your organization
• 12:01
  and what's the workflow as it relates to
• 12:03
  what are we doing with AI? How are we
• 12:05
  using it? Where is it
• 12:07
  currently integrated? Where is it
• 12:08
  currently integrated into client
• 12:09
  deliverable and, and client
• 12:11
  communication and, and client, you know,
• 12:13
  asset creation, whatever.
• 12:15
  And what happens if one of
• 12:17
  these models goes down? Absolutely. What,
• 12:19
  what, what's the plan? And,
• 12:21
  and you have to have somebody
• 12:22
  in your organization that's the captain
• 12:23
  of that, that really understands and has
• 12:25
  backup plans for it,
• 12:26
  because it's going to be really
• 12:28
  challenging if you don't have.
• 12:29
  Absolutely. And just to kind of build
• 12:32
  on the single model dependency we talked
• 12:34
  about, I also see that a lot of people
• 12:38
  are locking themselves
• 12:40
  into codecs or cloud code or something
• 12:43
  like that. And, you know, it is, I think
• 12:47
  when you're operating
• 12:48
  a business, you're very vulnerable if
• 12:50
  you, so one thing is the
• 12:52
  models, but it's also kind of the
• 12:53
  platforms to make sure that your
• 12:55
  workflows would work just as good with,
• 12:58
  with codecs or some other
• 13:00
  large language models. A lot of people
• 13:02
  we've been talking to, Andrew, are kind
• 13:03
  of building their own
• 13:05
  software and are plugging the models into
• 13:08
  it instead of working kind
• 13:10
  of in cloud code all day,
• 13:11
  et cetera. And I think that that's a very
• 13:14
  smart way of thinking
• 13:15
  about how to operate a business
• 13:17
  that makes you less vulnerable when
• 13:20
  things like this happens.
• 13:21
  So to kind of be not platform
• 13:24
  dependent, I think it would be something
• 13:27
  you should think about going forward.
• 13:29
  Yeah, totally. I think being not platform
• 13:31
  dependent and having a
• 13:32
  fail safe going through.
• 13:34
  Another thing to mention, right, is if
• 13:36
  you handle customer
• 13:38
  order data. So, you know,
• 13:40
  we call that like the PII stuff, right,
• 13:42
  which is around customer
• 13:44
  payments through Stripe or through
• 13:45
  Shopify or whatever else. One thing that
• 13:47
  I noticed in doing the research for this
• 13:48
  episode is there's a
• 13:50
  30 day customer retention policy that was
• 13:52
  shipped with Fable. Yes.
• 13:54
  And this is their own jailbreak
• 13:56
  mitigation sort of trade-off. And D to C
• 13:59
  brands handling that data,
• 14:01
  that customer data, that changes by
• 14:03
  model. So you have to be aware of how
• 14:07
  that changes and what the
• 14:08
  changes are within each model. And you
• 14:10
  can just figure that out by asking the
• 14:11
  model, hey, how are you
• 14:12
  handling this? But, you know, I think
• 14:14
  that that is something to be
• 14:16
  aware of with customer data,
• 14:17
  right. And how you're, how you're
• 14:20
  protecting that. And if you were to run
• 14:22
  into an issue, you know,
• 14:23
  with, with a customer saying, how's my
• 14:26
  data being handled or whatever, you have
• 14:28
  to be able to answer
• 14:28
  that. So, so that's, I think some of the
• 14:31
  bigger pieces that we want
• 14:32
  to talk about why you should
• 14:34
  care. I'd say that let's get into sort of
• 14:36
  a little bit more of the tactics around
• 14:38
  this, which we've gone
• 14:39
  into briefly, but I think we'd get into
• 14:41
  more. So, so one audit, the AI
• 14:44
  dependencies that you have,
• 14:46
  right. Get your prompts, your workflows,
• 14:48
  your MCP configs, zaps,
• 14:51
  whatever, and make a list of who,
• 14:54
  how they're connected and what is calling
• 14:56
  what. And like, again, this
• 14:58
  is sort of that AI governance
• 15:00
  piece within your organization, have
• 15:02
  somebody assigned to this or
• 15:03
  have a team in your, in your
• 15:04
  agency or your brand assigned to this of
• 15:07
  how are we doing this? Who's in charge
• 15:09
  and how are all these
• 15:10
  pieces talking to each other? So that if
• 15:12
  one breaks, there's a workflow that it
• 15:14
  can go on to a varying
• 15:16
  model and you're not completely stuck.
• 15:18
  Also, I think it's important
• 15:19
  to, to build a fallback ladder,
• 15:21
  you know, so for all the critical
• 15:23
  workflows, et cetera, make sure you have
• 15:26
  fallbacks and you have
• 15:27
  backup models and actually test the
• 15:29
  backup models. I had a big issue actually
• 15:33
  a few days ago where my
• 15:35
  backup model was not working correctly on
• 15:38
  one of my open clause. So suddenly
• 15:39
  everything stopped working.
• 15:41
  So actually testing that the fallbacks
• 15:44
  are working, I think is very
• 15:46
  important. Make sure that your
• 15:48
  models is not hard coded into your
• 15:50
  workflows, but kind of make it to Opus or
• 15:53
  Sonnet or GPT or like
• 15:56
  whatever it is, I think
• 15:57
  is also very, very smart.
• 15:59
  Yeah. I think ultimately another thing
• 16:01
  along with having this like
• 16:02
  pathway is also, you know,
• 16:05
  keeping a playbook for manual mode,
• 16:06
  right? Like if none of
• 16:07
  this existed, like how,
• 16:09
  what are the things that you, what are
• 16:11
  the things that you're going to do to fix
• 16:13
  this and how are you
• 16:14
  going to run the business? So having a
• 16:16
  plan with the team,
• 16:18
  understanding if we cannot query this,
• 16:20
  what is the way that we're going through
• 16:22
  and doing this? You know, we're doing
• 16:23
  this in our organization
• 16:24
  now, which is, you know, looking at
• 16:27
  essentially with a database of
• 16:29
  all of our customers. We are,
• 16:32
  we've done it manually. And now we are
• 16:35
  basically having duplicative efforts
• 16:36
  where we're updating it
• 16:37
  manually, but we're also updating it in
• 16:39
  AI as we continue to build
• 16:41
  AI. And we're building like a
• 16:42
  master queryable database within all of
• 16:45
  our customers, what
• 16:46
  they've done. But I'm worried
• 16:47
  that, that, that a model will get pulled
• 16:49
  down or there's going to
• 16:50
  be some issue or whatever.
• 16:52
  And then like, so that's what we're
• 16:53
  doing. We're basically
• 16:54
  building two as it goes through
• 16:56
  so that we have a fallback. I'm not
• 16:57
  saying you have to do this, but I do
• 16:59
  think that if we ultimately
• 17:00
  move into one that's fully AI automated
• 17:03
  stuff that we're doing with customer
• 17:04
  database querying and
• 17:05
  everything, we should at a minimum at a
• 17:07
  weekly basis, be also
• 17:09
  updating a manual database in
• 17:10
  Google sheets or whatever that's locked
• 17:12
  so that we know, right?
• 17:13
  Like, I just think if we become
• 17:15
  overly dependent on this, this is bad.
• 17:17
  There's already so much
• 17:18
  talk around people are becoming
• 17:21
  lazy. They're not letting, they're not
• 17:23
  doing the work themselves. AI is doing
• 17:25
  it. You can tell all the
• 17:26
  posts sound the same. They're all sound
• 17:27
  homogenous, which I agree with. And I
• 17:30
  think there's a lot to be
• 17:31
  said for that because people jump into
• 17:33
  like making it easier on them. And I
• 17:35
  mean, LinkedIn comments
• 17:36
  are an example of this, right? So anyway,
• 17:39
  so you need to just like be
• 17:41
  aware and be conscious of that
• 17:44
  and have a, have a plan. Yeah. And I
• 17:46
  think building on that, I think you, you
• 17:48
  need at the minimum to
• 17:50
  understand the logic of how things work,
• 17:53
  right? So you have to
• 17:54
  understand what the AI is actually
• 17:57
  reading from, what it's going through,
• 17:59
  the, the, the processes it's doing. You
• 18:01
  know, we, we, we've gone
• 18:02
  from SOPs to AI skills basically. And
• 18:07
  those skills are getting more and more
• 18:09
  bloated and more and more
• 18:11
  complicated. I've actually started now
• 18:13
  going, going through a lot of my core
• 18:15
  skills and rewriting it
• 18:16
  from the, the bloat that AI wrote it into
• 18:19
  something that I can
• 18:20
  actually understand and, and
• 18:22
  understand the logic of that has created
• 18:25
  actually a higher quality of
• 18:26
  the, of the outcome. But also
• 18:28
  if you have to go to what you call manual
• 18:30
  mode or caveman, like how you
• 18:32
  did it from back in the days,
• 18:34
  understanding the logic, the workflows,
• 18:37
  and what the skills that SOPs are
• 18:39
  actually doing is super,
• 18:41
  super important. And last, last thing
• 18:44
  that I'm becoming a
• 18:45
  bigger and bigger advocate for
• 18:48
  is also considering more open source
• 18:51
  systems, right? So I'm a
• 18:52
  big open claw, for example,
• 18:55
  guy, which is an open source project you
• 18:57
  can put into every different large
• 19:00
  language models out there.
• 19:02
  So we're trying to build as much as
• 19:04
  possible outside codecs and outside cloud
• 19:07
  code, and just let kind of
• 19:09
  our systems use them when we need it. But
• 19:13
  we are also now
• 19:14
  experimenting with having local models
• 19:16
  running or doing some of the work. So, so
• 19:18
  again, experimenting
• 19:20
  with what if open AI or
• 19:23
  Anthropic went off the map today? How
• 19:25
  screwed would you be? And what's the plan
• 19:28
  to make sure that you can
• 19:29
  actually keep thriving and at least kind
• 19:32
  of keep surviving in,
• 19:33
  in a blip like that?
• 19:35
  Totally. I think it's, I think the big
• 19:36
  thing to take away from this is that
• 19:37
  look, AI is moving from,
• 19:40
  you know, Hey, this is doing cool, cool
• 19:41
  shit. It does all this too. It's going to
• 19:44
  continue to move into,
• 19:45
  to a pseudo regulated infrastructure.
• 19:48
  Okay. So you have to be ready that
• 19:50
  government's going to come
• 19:51
  in and, and it plays a heavy, playing a
• 19:52
  heavy hand in this. And I think anybody
• 19:54
  that's going to be in
• 19:55
  office is going to continue to try to do
• 19:57
  this because I think the
• 19:58
  public sentiment around AI is
• 20:00
  we like it, we use it, but there's also a
• 20:02
  ton of skepticism in the
• 20:03
  United States and people also,
• 20:06
  you know, are, it's easy to be against
• 20:08
  AI, right? And it's easy to say like, AI,
• 20:10
  they're too powerful,
• 20:11
  these companies, whatever, which I don't
• 20:13
  disagree with necessarily. Right. So
• 20:15
  you're going to start
• 20:16
  to see that you're going to start to see
• 20:17
  more laws like New York, where there's
• 20:19
  regulated infrastructure
• 20:20
  that comes up or regulating the
• 20:21
  infrastructure of disclosures around AI
• 20:24
  around, you know, whatever.
• 20:26
  It's like, you need to, we need to be
• 20:27
  ready. And there needs to be
• 20:28
  systems that you have in place,
• 20:31
  knowing that these things could be
• 20:32
  eliminated quickly. Okay. So
• 20:35
  again, like it's contingency
• 20:36
  planning. It's just, it's easy to go all
• 20:39
  in on this stuff, but
• 20:40
  you also have to be ready
• 20:41
  and have a plan for when it's not. So
• 20:43
  thanks for checking out our
• 20:45
  quick episode today on this.
• 20:47
  We thought it was important to drop. If
• 20:48
  you have thoughts,
• 20:49
  questions, please feel free to email
• 20:52
  us or contact us. And we look
• 20:53
  forward to having you join us.
• 20:55
  Bye.
• 20:55
  Bye.
• 20:55
  Bye.