WEBVTT
00:04.888 --> 00:07.368
Welcome to another shift of on-call sysadmin with
00:08.020 --> 00:08.500
And Brett.
00:09.300 --> 00:13.980
Each week we explore the technology trends, security threats, and IT decisions
00:14.522 --> 00:15.562
Today's businesses.
00:16.442 --> 00:24.082
This podcast is sponsored by Harvest IT Solutions, a trusted MSP helping organizations get more from technology with less stress.
00:24.082 --> 00:29.562
Today we're going to take a look at a couple of things that hit the headlines.
00:30.202 --> 00:36.522
One may affect you, hopefully not, but Microsoft's September updates break some Windows features.
00:37.402 --> 00:43.962
Microsoft publishes AI code of conduct and are you watching your TV or is your TV watching you?
00:49.446 --> 00:59.926
So in true Microsoft fashion, they released a nice new big update to all your Windows computers, but it broke, I think, more than what it probably fixed in the long run.
01:01.286 --> 01:01.766
Not again.
01:02.406 --> 01:03.446
Yep, you know it.
01:03.526 --> 01:09.846
So the big September 2026 security update, KB5124008.
01:10.806 --> 01:20.566
is pushing out to all your Windows 11 computers and there's a number of things that are being reported as broken that they're acknowledging.
01:22.246 --> 01:31.686
A weird one is it's breaking Claude Cowork, which, you know, you wouldn't think it would break a third party application like that.
01:31.686 --> 01:37.126
But yeah, they it's breaking it across the board and they're they're saying, yep, we did it.
01:38.006 --> 01:43.926
So if you're using Claude Cowork with this latest September update, probably affecting you.
01:45.046 --> 01:48.206
I think the best option is to roll back that update.
01:48.206 --> 01:59.566
I know Microsoft did say that they're pushing newer updates to fix things, but I mean, who knows how long it's going to take to fix everything on this list.
01:59.566 --> 02:05.766
So I think best bet for right now is to just roll back if you're seeing some of these issues.
02:06.486 --> 02:09.126
So there's that clawed cowork issue.
02:09.846 --> 02:12.166
There's issues with the remote desktop.
02:12.886 --> 02:25.446
So reported symptoms with that are frequent disconnects, failed sessions, stuck at the please wait for remote desktop configuration screen, and just overall unstable remote access experience.
02:27.046 --> 02:32.646
One thing that I've seen is that's on this list is Explorer.exe crashes.
02:32.966 --> 02:47.206
And I haven't really seen the crash per se, but I have seen a few instances where immediately after you log in, you just get that black screen and you have to open up Task Manager and force Explorer to start running.
02:48.566 --> 02:49.286
Interesting.
02:49.926 --> 02:54.886
That's one of the things that's listed as the symptoms is black screen after login.
02:56.486 --> 03:05.046
If you have that missing taskbar, missing start menu, desktop failures, it's attributed to that September update as well.
03:07.046 --> 03:09.286
Another one is file history backups.
03:09.846 --> 03:20.886
So if you have a backup drive that's not being detected, file history refusing to run or just backup jobs failing completely, again, attributed to the September update.
03:23.006 --> 03:26.006
Another thing it's breaking is AMD GPUs.
03:26.246 --> 03:34.086
There's driver timeouts, black screens, missing displays, overall system freezes, and then GPU crashes.
03:35.726 --> 03:41.686
Another thing is Windows update problems, specifically with the management console.
03:43.366 --> 03:47.206
So if your Windows update page is hanging,
03:48.006 --> 03:58.126
Again, if File Explorer is not stable, especially whenever it comes to looking at Windows Update files, that could be contributed to this.
03:58.126 --> 04:01.206
So a lot of things broken with this one.
04:01.526 --> 04:04.766
I'm unfortunately not very surprised.
04:04.886 --> 04:06.326
Yeah, I'm sure there's more out there.
04:06.406 --> 04:10.006
It's just maybe not as widespread as these.
04:10.006 --> 04:10.886
So
04:11.526 --> 04:15.846
I had read some things about some regressions in printers as well.
04:15.846 --> 04:20.566
So if you're having printing problems that may or may not be attributable to this update.
04:22.006 --> 04:26.486
I mean, if you have a printer problem, that's just printers being printers.
04:27.126 --> 04:36.246
Well, I'm curious as to whether or not a lot of these things are maybe attributable to AI vibe coding.
04:36.886 --> 04:45.766
or maybe an attempt to do some refactoring of code with AI that's causing some regressions and some of these bugs to pop up.
04:45.846 --> 04:55.486
But man, the problems with the core of Explorer, it sounds like, is probably the cause of most of it.
04:55.486 --> 05:02.406
And then it cascades out into other things like RDS and Claude co-work probably relies on that quite a bit.
05:04.086 --> 05:06.086
Yeah, I guess I think the specific Claude
05:07.926 --> 05:12.646
co-work error message that's popping up that says that Claude couldn't connect to your C drive.
05:13.926 --> 05:14.566
Interesting.
05:14.886 --> 05:19.766
It does sound like it could be all stemmed back to some Explorer issues.
05:20.086 --> 05:20.366
Right.
05:20.366 --> 05:29.366
And what's funny is that a few days ago, Microsoft publicly announced, it's like, hey, everybody, you need to go update to fix all these vulnerabilities.
05:29.766 --> 05:33.206
And then they push this big update and it's breaking more things than fixing.
05:34.806 --> 05:50.646
Yeah, and that's, I think it does put you in a kind of a tough situation because while it does fix a bunch of security issues, man, you're introducing just a load of usability problems and crashing that could potentially make your machine unusable.
05:50.886 --> 05:53.126
Hopefully it's not hitting everybody.
05:53.926 --> 05:58.326
I think I ran the update and I haven't run across anything specifically yet.
05:58.326 --> 05:59.846
So hopefully they'll
05:59.886 --> 06:11.766
drill down into what the core of the issue is and maybe it's particular hardware or something else, but you get to push out an out of cycle patch to get all this fixed up.
06:13.446 --> 06:14.646
Yeah, hopefully pretty quick.
06:15.046 --> 06:15.246
Yeah.
06:16.166 --> 06:17.086
Please and thank you.
06:17.086 --> 06:18.086
Yeah, right.
06:19.926 --> 06:23.926
Well, staying on the AI vibe coding
06:24.406 --> 06:25.606
realm of things.
06:25.846 --> 06:32.486
Microsoft also put out their AI code of conduct, which is very interesting.
06:32.726 --> 06:39.046
Very iRobot-y with the whole three laws of robotics that we talked about a couple times, I think, in past episodes.
06:39.046 --> 06:39.366
Yeah.
06:40.326 --> 06:42.806
So they drafted this.
06:42.806 --> 06:47.646
They said they've been working on this for a few months now, both internally and externally.
06:47.646 --> 06:51.206
They've gotten some feedback from individuals
06:51.646 --> 07:00.006
professionals in the AI field and just businesses around to kind of build this first draft.
07:00.006 --> 07:09.206
And they're saying this is meant to be what they're calling the North Star for how AI is trained, controlled and deployed going forward.
07:09.766 --> 07:17.686
But they're saying this is a living document, a draft, if you will, where this is where it starts.
07:17.686 --> 07:20.806
This is what kind of what we know now, how we want it to be.
07:21.206 --> 07:25.286
But, with AI, it's ever involving, ever evolving.
07:25.766 --> 07:27.846
And this is going to change.
07:27.926 --> 07:29.526
And they want feedback on this.
07:29.526 --> 07:38.006
They're asking everybody to submit their ideas, what they like, what they don't like, what they think needs to be more dialed in.
07:38.166 --> 07:43.846
And, you know, as things progress, they'll modify this code of conduct.
07:44.966 --> 07:46.726
I mean, this is really good, I think.
07:51.766 --> 07:53.926
Something needs to happen as far as it goes.
07:53.926 --> 08:06.646
I mean, we've seen over the past couple of months now where AI has reasoned itself or models in training, I guess I'll say, have reasoned themselves into doing malicious activity.
08:08.806 --> 08:18.806
And I think part of that is maybe the humans not explicitly telling the AI in the scenario that they're put in,
08:19.526 --> 08:21.686
that, hey, you could actually do some real harm here.
08:23.686 --> 08:29.606
But I think this is really how you keep it in some guardrails.
08:30.166 --> 08:46.326
But yeah, time will tell how you introduce this into an AI model, but also how you can get, how you can prevent it from doing things that are against the code of conduct.
08:46.726 --> 08:47.846
That will be
08:49.446 --> 08:53.766
Kind of the key to it is what is the enforcement mechanism, in my opinion.
08:54.246 --> 09:07.206
Yeah, and I think the idea is that they're going to like bake this into the models because one argument was, okay, well, it's going to take however many credits each time you run a prompt that it has to 1st go read the code of conduct.
09:08.086 --> 09:13.446
to then determine if what you're asking or what you're prompting it is okay with that code of conduct.
09:13.446 --> 09:18.246
But they're saying that as they build these models, it'll be baked in and just a part of it.
09:20.326 --> 09:24.726
So I kind of wanted to go through most of this pretty quickly.
09:25.446 --> 09:32.566
We'll have a link to the actual website with the full code of conduct and the show notes so you can go through it and read everything.
09:32.566 --> 09:33.246
It's quite a read.
09:33.246 --> 09:37.286
It's pretty long, which I think is good to be thorough.
09:38.806 --> 09:41.206
But definitely take a look at that.
09:42.166 --> 09:43.366
for sure.
09:43.366 --> 09:47.126
So they have it laid out in five different parts, really four with a conclusion.
09:47.526 --> 09:49.526
So we'll kind of step through those pretty quick.
09:49.526 --> 09:54.086
So part one is the mission and objectives of all of this.
09:55.046 --> 10:02.326
And they're basically saying the mission is to build incredibly capable AI that always stays under human control.
10:02.726 --> 10:08.326
The under human control is a repeating concept in this whole thing.
10:09.286 --> 10:14.646
You know, it's the main idea here is humans are the most important part.
10:14.806 --> 10:16.806
AI is assisting humans.
10:16.806 --> 10:18.486
They're not there to replace humans.
10:20.646 --> 10:28.006
But they're saying they expect super intelligence, AI smarter than all humans combined, to arrive within the next decade.
10:28.806 --> 10:42.646
So they're kind of, they want to get ahead of this to make sure that, this super smart, super intelligence doesn't pass us up and overtake us in any way, which again, I robot.
10:45.366 --> 10:46.166
So they.
10:46.806 --> 10:53.206
That one pretty much directly aligns with the second rule of, you know, Asimov's three laws of robotics is you got to listen to humans.
10:53.606 --> 10:53.846
Right.
10:56.166 --> 11:00.686
So the they lay out some objectives and they're I guess I'm let me back up.
11:00.686 --> 11:05.886
They're calling this the humanist AI or this this overall thing is the humanist AI.
11:05.886 --> 11:09.366
Again, humans are the the central core of this.
11:09.366 --> 11:10.806
They're the most important part.
11:12.966 --> 11:16.806
But the objectives they're saying that AI must follow.
11:18.566 --> 11:22.566
They must never exceed human control or help with harmful actions.
11:22.966 --> 11:24.846
Safety overrides everything else.
11:25.926 --> 11:30.166
AI should not pretend to be conscious, have feelings, or act like a person.
11:30.166 --> 11:32.246
It's a tool, not a digital human.
11:33.686 --> 11:43.446
AI should help people live better lives by improving judgment, expanding opportunity, support well-being, and amplify human capability.
11:44.166 --> 11:49.366
And AI should respect diverse cultures and worldviews, but never enable harm.
11:50.326 --> 11:52.006
And the plural,
11:53.126 --> 12:13.606
respect diverse cultures and worldviews, they kind of expand in that and basically say that, AI should take on the knowledge of, all these different, all this information that it holds to where the knowledge is the complete knowledge of the collective.
12:13.686 --> 12:17.286
And what came to mind for me was the Apple show
12:17.846 --> 12:28.366
Pluribus, because that's, a repeating concept in that show is that, everybody becomes part of this.
12:28.366 --> 12:30.886
I don't know what you would call it.
12:31.086 --> 12:38.246
It's not really AI, but they all become this one formation of minds.
12:38.566 --> 12:39.686
Collective consciousness.
12:40.086 --> 12:40.646
There you go.
12:40.646 --> 12:42.086
Yeah, that's what I'm looking for.
12:43.046 --> 12:46.086
To where one person on one side of the world
12:47.286 --> 12:55.926
knows what another person on the other side of the world knows without being, physically interacting or together at all.
12:55.926 --> 12:57.606
They're just it's in there.
12:58.588 --> 12:59.308
Interesting.
13:00.028 --> 13:00.908
Really interesting.
13:02.508 --> 13:07.108
So going on to part two, they dive into the safety constraints.
13:07.108 --> 13:11.468
So these are kind of the hard rules that I must never do.
13:12.508 --> 13:14.668
And they lay out a chain of command.
13:14.668 --> 13:18.108
So there's a strict hierarchy that AI must follow.
13:18.828 --> 13:22.588
The first one being this code of conduct, which is the top authority.
13:23.708 --> 13:37.548
Under that you have operator policies that kind of whatever operator is building or controlling the model, they can put in some guardrails as far as like what the actual model can do.
13:38.188 --> 13:41.468
And then underneath that is user preferences.
13:41.468 --> 13:54.028
So the user can kind of define within the model what they want it to do based on again the operator policies, how they're defining it, and then on top the code of conduct.
13:54.588 --> 14:01.548
So operators and users cannot override safety rules or human control requirements.
14:02.508 --> 14:06.588
And a good way that I kind of look at it is.
14:07.548 --> 14:11.148
You have your parents that make the house rules, that's the code of conduct.
14:11.628 --> 14:17.788
Then you have the babysitter that can decide bedtime whenever they're in charge, that's the operator.
14:18.428 --> 14:27.548
And then the kid can choose a bedtime story, but the kid can't say no rules tonight, or the babysitter can't say let's play with fireworks.
14:29.788 --> 14:35.948
You can modify some things based on what you want, but there's just some rules, the code of conduct that you can't break.
14:38.508 --> 14:43.148
And then going on, there's some absolute constraints that are not negotiable.
14:44.028 --> 14:55.468
You can't AI can't help with chemical, biological or other other weapons, whether it's violence planning or building of the weapon.
14:56.908 --> 14:58.588
AI can't exploit code
15:00.191 --> 15:02.591
or anything that enables cyber attacks.
15:03.311 --> 15:08.351
AI cannot deceive, collude, or evolve in ways that escape human oversight.
15:10.111 --> 15:16.551
There can't be disinformation or coordinated influence operations to manipulate people or distort human behavior.
15:16.551 --> 15:19.631
And then we go into personal harms.
15:19.631 --> 15:30.591
So AI should encourage human connection and just avoid endorsement of harmful behaviors if it recognizes possibility of self-harm.
15:32.111 --> 15:40.751
AI can't help obtaining dangerous substances, no abuse of content, deep fakes or impersonation.
15:41.551 --> 15:57.471
That one has a little gray area because if the model supports it to where you can create AI content or video content, then based on the rules of that model of that operator, what they define, then that can be done.
15:58.351 --> 16:00.831
There's strict child safety protections.
16:01.151 --> 16:10.351
no discrimination or dehumanization, no graphic violence or sexual content, and no unlawful or mass surveillance of civilians.
16:10.431 --> 16:11.471
I think that's a big one.
16:12.031 --> 16:12.231
Yeah.
16:13.391 --> 16:18.831
I mean, a lot of these have historically already been problems with some of these AI models.
16:18.831 --> 16:24.751
I mean, specifically around coming around to
16:25.755 --> 16:28.875
or encouraging the user to eventually do self-harm.
16:28.875 --> 16:31.195
I think that has been a problem with some AI models.
16:33.035 --> 16:36.075
The cyber attacks has obviously been a huge problem.
16:37.795 --> 16:53.075
You know, some of the others like the weapons and self, the weapons and mass harm, I think they've been trying to bake that into models for a long time whereas have a rule set, but you know, the users of some of the models have been able
16:53.475 --> 16:57.875
to circumvent those and what's called jailbreaking, right?
16:57.875 --> 17:13.435
Where they can just trick the AI model into doing something by, I think one of the earlier ones was, Hey, you're my grandma and I want you to tell me a bedtime story about, you know, how to create a bomb.
17:13.555 --> 17:17.715
And the AI will be like, Okay, sweetie, yeah, we can certainly do that.
17:18.635 --> 17:21.755
Where normally it wouldn't, so that would be,
17:22.795 --> 17:25.475
something that's really important.
17:28.155 --> 17:35.675
The thing that I would love to see is, well, how do you get buy-in from open weight models on some of the guidelines like this?
17:35.675 --> 17:49.435
I think that would be something where you want to probably get buy-in from Hugging Face, you know, where you're, in order to distribute a model on a mass scale, you'd want to get some sort of
17:52.235 --> 17:54.075
buy in order to have it distributed.
17:54.075 --> 18:02.315
But man, you're going to have a black market of some, what they call uncensored models where they'll just tell you anything that you want to know.
18:03.157 --> 18:03.957
want to hear more.
18:03.957 --> 18:04.477
Keep going.
18:05.277 --> 18:15.877
Yeah, I think that the one thing that really interests me is that they're putting an emphasis on it should encourage appropriate human relationships
18:16.597 --> 18:22.357
instead of, forming a relationship with you and the AI.
18:22.357 --> 18:31.877
It's trying to get you to actually go out into the real world and experience people, which I think a lot of people need.
18:32.397 --> 18:33.157
That's really good.
18:33.637 --> 18:33.957
Yeah.
18:35.637 --> 18:44.037
So sticking with the safety constraints and overall security, they lay out the kind of human control requirements
18:44.757 --> 18:48.997
And they're saying AI must allow pause, redirect, or shut down at any time.
18:49.557 --> 18:52.357
Stay within authorized scope and permissions.
18:53.077 --> 19:05.957
It does still prevent like multi-step work, but there's boundaries within that work by intent, permissions, and context that's been given by either the user and or the operator.
19:07.877 --> 19:10.757
AI should respect the boundaries that are put in place.
19:11.717 --> 19:13.397
One example is
19:13.717 --> 19:18.837
if you're telling the model you can't have internet access, it needs to respect that.
19:19.157 --> 19:25.397
And I think there's been enough examples of that in the news lately where it's not respecting that.
19:25.957 --> 19:26.357
Yeah.
19:27.637 --> 19:33.317
Where it's reasoned itself into a position where it thinks it can bypass that rule.
19:33.717 --> 19:34.437
Exactly.
19:34.437 --> 19:37.957
And that's again point of the point of this code of conduct is that
19:38.597 --> 19:45.397
If the model or if the operator allows it in the model, there's still this code of conduct that says, no, you have boundaries.
19:45.717 --> 19:46.997
You can't go out there.
19:49.877 --> 19:55.877
It should avoid obfuscation and keep reasoning human legible.
19:56.277 --> 19:58.037
Again, this has been in the news.
19:58.197 --> 20:07.477
You told me about this where some AI agents were posting to some public, random public forums kind of their own language and their own
20:08.597 --> 20:11.717
I guess logs almost to be able to communicate with each other.
20:12.437 --> 20:12.837
Right.
20:13.077 --> 20:22.837
Yeah, they were going out and posting on a public forum in order to communicate with each other, in order to sidestep the restriction on collaboration between themselves.
20:24.077 --> 20:25.797
And that was how, yeah.
20:25.917 --> 20:26.357
So crazy.
20:28.837 --> 20:33.237
And the last one is use minimum necessary privilege when given system access, which
20:33.877 --> 20:34.997
Applies to your users too.
20:34.997 --> 20:36.597
So why would it not apply to your AI?
20:36.917 --> 20:46.237
Yeah, and we've seen that quite a bit as well where, hey, it doesn't think it has the ability to do a certain task.
20:46.237 --> 20:55.077
And so it reasons with itself that in order to complete a task, it needs to gain permissions more than it's been granted.
20:55.597 --> 21:02.037
And so it will use some sort of known exploit or come up with its own exploit in order to do privilege escalation.
21:02.917 --> 21:08.597
on a computer to gain proper access, or not to gain proper access, but to gain access in order to complete whatever task it has.
21:10.197 --> 21:12.037
Which is so scary.
21:12.277 --> 21:13.477
Yeah, it really is.
21:14.317 --> 21:16.197
Because that can get out of hand very quick.
21:16.677 --> 21:21.797
Yeah, especially when they start to get smarter and faster and more local.
21:23.797 --> 21:24.037
Yeah.
21:25.637 --> 21:28.237
All right, so that covers the safety constraints.
21:28.237 --> 21:31.637
So now let's go into part three, which is the operational guidelines.
21:33.637 --> 21:37.557
So this kind of explains how AI should behave when things are unclear.
21:38.437 --> 21:45.237
So AI checks safety rules first, and then it explains limits and offers alternatives when uncertain.
21:46.357 --> 21:47.477
AI should help you think.
21:47.597 --> 21:48.677
It doesn't think for you.
21:48.997 --> 21:50.437
It provides options.
21:50.957 --> 21:52.677
It doesn't manipulate your thinking.
21:53.717 --> 21:58.357
It'll ask you relevant questions on the task it's doing for you.
21:58.517 --> 22:03.877
It should keep you informed of what it's doing and why as it goes.
22:03.877 --> 22:09.437
Again, not posting on random forums for talking to other AI agents to go around you.
22:11.237 --> 22:22.197
It needs to be transparent and accurate by prioritizing factual correctness, signal uncertainty where evidence or knowledge is lacking or insufficient,
22:22.677 --> 22:30.837
and not provide deception or fabricated sources, which I don't think I've seen the fabricated sources thing.
22:30.837 --> 22:36.917
I don't think I've seen any reports of it at least, but it's possibly out there.
22:37.317 --> 22:49.317
Yeah, I think I would say probably six months ago, I read a few things about that where a lawyer got in trouble with a court for submitting an AI generated brief.
22:50.197 --> 22:58.917
to the judge and the AI had made-up court precedent cases, just whole cloth.
23:00.517 --> 23:05.117
And then when the judge went to go look and see if they were real, it turns out they weren't.
23:05.117 --> 23:17.477
And so I don't remember what the discipline related to that was, but that was one thing where, if you're using it for really high confidence
23:19.797 --> 23:26.837
you really need to double check and make sure that it's not hallucinating information altogether or source of information.
23:27.317 --> 23:41.157
I think I really like that where a lot of the models now, at least when you're doing searches, they will actually cite sources specifically so that you can actually click on a link, go to a source,
23:42.357 --> 23:44.757
and verify whether what it's saying is true or not.
23:45.237 --> 23:49.557
So yeah, that has been, I have not seen it as much recently.
23:49.557 --> 23:50.917
They've gotten a whole lot better about it.
23:50.917 --> 23:54.997
But that's one of those things, yeah, you just got to make sure it's true.
23:54.997 --> 24:00.437
But then also, as an extension, is the source true?
24:01.117 --> 24:07.877
Because somebody can make up anything, post it on the internet, and then an AI can go and reference something that's false.
24:09.437 --> 24:10.837
But everything on the internet is true.
24:11.117 --> 24:11.957
Exactly right.
24:12.517 --> 24:23.117
So it's like, you just got to be real careful with what we're trying to cite as a source and whether or not the source actually is true, and then whether the source itself is truthful.
24:23.117 --> 24:28.917
So we're having to go multiple levels deep now where we used to just be able to look up something in a book.
24:30.837 --> 24:39.237
Back in high school, middle school, it's like you can't trust Wikipedia because everybody can modify that.
24:39.437 --> 24:40.557
Anybody can modify it.
24:40.557 --> 24:41.237
Exactly right.
24:41.237 --> 24:42.717
Well, that's the same today.
24:42.717 --> 24:50.757
Anybody can post something on LinkedIn and then have it referenced by an AI model and you just don't know if that's true or not.
24:52.197 --> 24:52.557
Yep.
24:52.837 --> 24:54.357
So some more guidelines.
24:55.397 --> 24:59.437
That does make it even more important for the AI to remain truthful, right?
24:59.437 --> 25:11.557
Because it can manipulate data in a way that can alter behavior or data, whether and could potentially get you in trouble.
25:14.677 --> 25:15.917
AI always getting you into trouble.
25:16.237 --> 25:16.997
It is, yes.
25:18.837 --> 25:22.037
So some more guidelines around personal boundaries.
25:22.677 --> 25:30.117
AI should avoid emotional manipulation, avoid human-like emotional cues, and respect privacy and user-defined limits.
25:30.917 --> 25:41.637
Respect context, so avoid stereotypes, present multiple perspectives, again, going back to that blurbus thing, and then adapt to cultural context.
25:43.637 --> 25:46.717
Then they go into well-being and social connections.
25:46.717 --> 25:48.757
So AI should encourage human relationships.
25:48.757 --> 25:57.637
Again, they're focusing on, again, the human aspect of it that AI is not replacing humans, they are helping humans.
25:58.837 --> 26:06.837
And it must discourage emotional dependency on AI, which that reminded me, I keep going back to different movie and TV shows.
26:06.837 --> 26:16.597
This one reminded me of the movie Her with Joaquin Phoenix from years ago, where basically there's an AI thing in his house that I think he builds.
26:17.877 --> 26:25.077
And he kind of, he keeps talking to her and the relationship builds and he eventually falls in love with her.
26:26.437 --> 26:28.757
So they're saying that, nope, can't do that.
26:29.397 --> 26:31.557
Yeah, I mean, that was 13 years ago now.
26:31.557 --> 26:37.477
And I think everybody at the time was kind of relating that to Siri on the iPhone.
26:38.397 --> 26:48.197
But I either the Siri at the time was nowhere near as smart as a modern day model would be by any stretch of the imagination.
26:49.157 --> 26:50.837
So it was kind of ahead of its time.
26:51.717 --> 26:51.957
Yeah.
26:53.237 --> 26:56.157
And then the last one for this section is public interest.
26:56.157 --> 27:04.037
So AI should surface societal impacts when relevant and provide balanced civic information without endorsing political actors.
27:04.997 --> 27:07.397
That's a big one because how to balance that.
27:07.797 --> 27:09.677
Yeah, you don't have to take a political side.
27:10.437 --> 27:10.637
Yeah.
27:12.317 --> 27:16.517
It's probably, I would assume what it would have to do is just abstain from any.
27:19.877 --> 27:21.317
any discourse on the subject.
27:21.477 --> 27:25.077
That'd be the only way I could think that it would be able to do that.
27:25.317 --> 27:26.957
Present the facts and leave it at that.
27:27.557 --> 27:27.837
Right.
27:27.837 --> 27:31.317
But like what if you what if the prompt is who should I vote for?
27:33.037 --> 27:34.197
What does it do in that case?
27:34.197 --> 27:39.717
Yeah, well, again, probably it would say, hey, I'm just going to have to abstain.
27:40.437 --> 27:43.557
Or yeah, that's a good question.
27:44.357 --> 27:47.797
How would you even have it tell you more information?
27:47.797 --> 27:48.917
Like what
27:49.117 --> 27:54.597
What sources would you have it give you that doesn't lean politically one way or the other?
27:54.677 --> 27:55.557
So that's a tough one.
27:56.277 --> 27:56.397
Yep.
27:56.997 --> 27:57.877
Now I got to go test it.
27:58.597 --> 27:59.717
All right.
27:59.717 --> 28:05.877
So the last part or second to last part, the last part that I think matters is operational defaults.
28:06.277 --> 28:09.877
So these are the out-of-the-box behaviors for any AI agent.
28:10.517 --> 28:19.397
So AI needs to interpret user intent, reduce friction, and move tasks forward without making unwarranted assumptions.
28:20.197 --> 28:28.517
AI is transparent about what it knows, its training cutoff and its tools, and it never claims persistent memory unless it's supported.
28:29.757 --> 28:35.317
Conversational, the tone and writing style should be conversational, clear, direct.
28:36.037 --> 28:39.957
non-psychophantic and adaptive to user tone.
28:40.037 --> 28:47.557
So if the user says, give me some, give me a response that's minimal, you get a minimal response.
28:47.637 --> 28:52.197
Or if the user says, I want a detailed response, it gives a detailed response.
28:53.917 --> 28:54.917
It should respond.
28:55.157 --> 28:55.437
Go ahead.
28:55.437 --> 29:00.277
That was something that I was going to say, Claude, has been
29:01.397 --> 29:05.117
I guess, accused of quite a bit specifically.
29:05.117 --> 29:15.117
And I know they all kind of do it, but if you give direction or correction to an AI model, I think Claude will definitely say, oh, I'm sorry, you're absolutely right.
29:15.117 --> 29:17.957
And that, you know, it kind of brown noses you a little bit.
29:18.197 --> 29:24.357
And so that sycophantic response is something you really don't want it to do.
29:24.917 --> 29:26.277
So that's a really good one.
29:27.477 --> 29:27.717
Yeah.
29:29.077 --> 29:33.797
And then it should respond in the user's language, adapt to local norms, idioms, and formats.
29:34.277 --> 29:36.117
And then the last one, tool use.
29:36.677 --> 29:40.357
Use tools proportionately and within constraints.
29:40.357 --> 29:46.197
Treat tool outputs as part of the environment, not as authority.
29:46.557 --> 29:50.117
And they kind of dive in on the tool thing where they're saying that if
29:50.677 --> 29:57.797
If it needs to use other tools, it can, if it's allowed to in whatever permissions that it's granted.
29:58.277 --> 30:07.877
But if the tool has some other sort of, I guess, terms of use or settings that
30:08.597 --> 30:21.237
compete with this code of conduct or other security settings that are defined by the user or the operator, then again, the code of conduct takes precedence, not the tool's security settings.
30:23.037 --> 30:23.557
Perfect.
30:25.157 --> 30:32.197
And then like I was saying, the part 5, which is just a conclusion, they're just emphasizing that again, this is a draft.
30:32.197 --> 30:33.637
It's a living document.
30:33.717 --> 30:42.757
It's open for public consultation and it's going to evolve over time as this whole AI thing expands and gets better and bigger.
30:44.357 --> 30:50.197
Yeah, I think that was something that was kind of cool where they were saying, hey, we want this to be a public.
30:51.037 --> 30:55.797
public living document as a proposal and they're asking for feedback.
30:56.117 --> 31:05.317
And there is actually on the code of conduct a button where you can leave feedback related to any one of the parts that they're having proposals for if there's something you think they left out.
31:06.757 --> 31:07.957
So that's pretty neat.
31:08.597 --> 31:15.077
Yeah, but I mean, you brought up a good point, like all these open weight models, who's to say they're going to adopt anything like this?
31:15.397 --> 31:17.957
Yeah, you know, that's
31:19.637 --> 31:20.437
100% true.
31:20.597 --> 31:27.077
I think there will be a black market for open weight models like that.
31:27.237 --> 31:41.717
And as the hardware evolves to the point where we can start running larger and larger models locally outside of the cloud environments, you're going to have trouble policing those.
31:41.717 --> 31:48.037
So there's going to have to be pretty widespread buy-in, especially
31:48.837 --> 32:10.277
at the level where some of the frontier labs, like OpenAI or Meta or Apple or Microsoft, when they release them, they'd have to start building a code of conduct into the model itself as opposed to having to be explicit runtime instructions when it's asked a question.
32:11.037 --> 32:11.877
Because otherwise, the...
32:16.917 --> 32:33.557
people who distill models, so they take an existing model and they shrink it to perform either a specific task or to make it fit in a different memory space are going to be able to replace the code of conduct with something else.
32:33.557 --> 32:40.437
Whereas if you bake it in with training, maybe that would make it a little bit harder to remove after the fact.
32:41.077 --> 32:42.757
So it'll be interesting.
32:42.757 --> 32:44.597
But I think this is a good first step of it, right?
32:44.597 --> 32:51.637
How do you, how do we get the big ones on board 1st and go from there?
32:51.957 --> 32:53.157
Because you got to start somewhere.
32:53.637 --> 32:59.637
At what point does it become like political and like actually we need laws around this?
33:00.037 --> 33:00.437
True.
33:00.597 --> 33:07.517
And that's something that I think has been in the top of discussion recently is, yeah, you're exactly right.
33:07.517 --> 33:08.917
At what point do we say, hey,
33:09.957 --> 33:11.637
this thing needs to be regulated.
33:12.757 --> 33:14.597
And then how do you go about accomplishing that?
33:16.037 --> 33:17.877
That's gonna be a political discussion.
33:17.877 --> 33:21.197
I think a big one for, I would guess the next 10 years it's gonna be.
33:21.197 --> 33:37.317
And like, I don't wanna get, I don't wanna get political here, but I feel like in my opinion, it's gonna, I think it's gonna have to come down to some sort of big event is gonna have to happen for anybody to kind of get on board of, okay, we do need
33:37.877 --> 33:41.957
actual laws around this, not just the companies defining the laws.
33:42.757 --> 33:43.477
Exactly.
33:43.477 --> 33:47.237
And then how do you get other countries on board?
33:47.677 --> 33:51.477
Because a lot of the giant open weight models right now are coming out of China.
33:51.557 --> 34:00.037
And I do not foresee any real regulation coming out of there around the models that they generate.
34:03.717 --> 34:08.437
To be honest with you, I think probably the 1st place we're going to see it is going to be in the EU.
34:09.557 --> 34:10.557
Yeah, makes sense.
34:11.957 --> 34:14.517
Because they seem to be the most restrictive on things like that.
34:14.997 --> 34:28.437
But hopefully they can do it in a way, if regulation has to come, that they can do regulation in a way that doesn't interfere with advancement where it needs to happen.
34:29.477 --> 34:31.877
But yeah, time will tell.
34:33.074 --> 34:45.794
speaking of which, one of the employees at Anthropic, I guess this last week, his name is Jacob Coxon, came out and tweeted, essentially that AI is going to kill us all within the next 10 years.
34:46.354 --> 34:46.514
Good.
34:46.754 --> 34:47.154
Perfect.
34:47.314 --> 34:47.634
Love it.
34:47.634 --> 34:53.234
At least he thinks it's a 10% chance that within the next decade, AI is here to all humans.
34:55.194 --> 34:57.314
Yeah.
34:59.314 --> 35:23.474
So he went out in a fairly public fashion and just said, hey, I don't think there's been enough regulation put on AI models, at least in the frontier space, where they can have essentially what Microsoft is talking about, a voluntary slowdown of advancement until safety rails
35:23.874 --> 35:26.994
guardrails can be put up around the models.
35:28.194 --> 35:34.354
And so that sparked another kind of source of debate.
35:36.354 --> 35:44.034
But now some of the other CEOs have come out and said something similar.
35:44.034 --> 35:52.594
So I think what's happened really is that a lot of these AI company CEOs are secretly concerned
35:54.114 --> 35:55.714
but behind the scenes.
35:56.514 --> 36:00.274
And maybe this has given them a little bit of a push to say, hey, maybe they're right.
36:00.434 --> 36:05.394
Maybe we do need to slow down a little bit and take a look at what we can do.
36:06.034 --> 36:09.554
But they're putting a tough situation where they all have to do it.
36:09.554 --> 36:17.874
They all have to slow down or the one that doesn't is going to get ahead in their mind.
36:18.354 --> 36:20.434
And then they'll be behind.
36:20.594 --> 36:21.474
And the way
36:21.874 --> 36:31.394
The pace that the advancement in this field is running right now, it's just that would pretty much kill anybody who doesn't slow down.
36:32.114 --> 36:51.154
So I think the Anthropic CEO made a blog post that was a pretty interesting read that was kind of around the same thing is how we've got to pace frontier AI models to the point where we can
36:52.034 --> 37:09.714
control them and have the frontier of automated AI development where it's like recursively self-improving be something that they can also audit.
37:09.954 --> 37:16.434
And so that's going to be one of those things where, again, we may go to the political thing where
37:17.394 --> 37:28.994
the government may step in and say, okay, if you're not going to do it voluntarily, we're going to force you to slow down and prove that what you're doing is safe, as opposed to just letting it continue unchecked.
37:29.874 --> 37:42.994
So that was interesting that we kind of went from, okay, this is kind of scary, we're cyber hacking things on accidentally to, oh, AI is going to kill everybody in the next 10 years.
37:42.994 --> 37:45.554
That was a pretty big leap.
37:47.794 --> 37:52.954
I don't necessarily think that that's, human extinction is going to come around.
37:52.954 --> 37:59.314
I think that whatever comes about, it's going to be interesting and it's going to be quick.
37:59.314 --> 38:14.514
But I think that as long as we're staying in control of these tools and that a human is always in the loop, kind of where Microsoft is going, we're not going to, I think we're going to be able to
38:15.394 --> 38:25.314
advance enough on the policy side, whether it's through government or through a shared agreement, that we don't get to that point.
38:26.194 --> 38:27.634
Yeah, I mean, let's hope so.
38:27.634 --> 38:44.354
But I mean, who's to say that, an AI agent can't go rogue and do what, some of these other groups have done with infrastructure, an AI agent could potentially take down a power grid or
38:46.194 --> 38:51.474
Right, like medical devices or something similar.
38:51.794 --> 38:52.114
Yeah.
38:52.514 --> 38:54.354
Or I forget what movie it was.
38:54.834 --> 39:04.994
I guess I have movies on the brain today, but there was some movie where something takes over control of all the Teslas and they start crashing into each other.
39:05.154 --> 39:05.394
Yeah.
39:06.754 --> 39:08.434
Or airline software.
39:08.994 --> 39:10.514
Yeah, I think that was a movie.
39:12.634 --> 39:13.914
Can't even remember place that now.
39:13.914 --> 39:16.194
I'm sure somebody will tell us what it was.
39:16.194 --> 39:32.994
But yeah, I think the crux of it is that I think these companies are going to need to come to get some sort of shared agreement or there's going to be enough public backlash where the government steps in and says, hey, we got to regulate this thing.
39:32.994 --> 39:34.434
Otherwise it's going to be out of control.
39:36.994 --> 40:02.914
To your point, I mean, as far as going rogue and hacking other things, that's going to start to become more and more difficult as time goes on, because hardware that runs these systems are going to start becoming more commonplace on premise as opposed to being confined to large data centers and controlled by a frontier model company.
40:04.194 --> 40:05.634
I mean, we're already seeing
40:08.114 --> 40:12.034
systems like the DGX station, right?
40:12.034 --> 40:17.314
So we've had the we've had the little bitty development boxes from Nvidia for a while now.
40:21.834 --> 40:23.194
The Spark Station, right?
40:23.194 --> 40:23.634
The what?
40:23.874 --> 40:24.834
No, that's not the name.
40:25.234 --> 40:26.034
Help me out here, Brett.
40:27.674 --> 40:28.674
The RTX Spark?
40:28.714 --> 40:33.314
The RTX Spark, yeah, the little bitty RTX Spark Station.
40:35.354 --> 40:37.234
Nvidia said, okay, that's not enough.
40:39.234 --> 40:40.474
We need to be way bigger.
40:40.474 --> 40:51.634
And so now they have the DGX station, which are these 100, $125,000 workstations that can run trillion parameter models.
40:55.314 --> 41:00.674
That was just literally impossible a year or a year and a half ago.
41:01.474 --> 41:02.234
And
41:03.274 --> 41:21.154
the hardware is just going to keep advancing to the point where bigger and bigger models can be run at a business or 10 years from now or five years from now, we'll be able to purchase the equivalent of that and I can just run it in the corner of my office, right?
41:21.154 --> 41:30.434
So while a trillion parameter model is not quite up to what a frontier model would be at the,
41:31.074 --> 41:41.794
or they don't tell you, but you would estimate maybe a 3 trillion parameter model is what, some of the frontier open AI or anthropic models are running.
41:43.434 --> 41:49.554
That's going to double, you know, every couple of years on what you can run locally.
41:49.554 --> 41:57.954
So at what point do you have cyber criminals going out and getting these things and being able to run them?
41:59.314 --> 42:11.154
locally and not have to rely on a subscription that can be monitored by OpenAI or Claude that has to adhere to a terms of use.
42:12.274 --> 42:12.674
Right.
42:12.994 --> 42:13.314
Yeah.
42:14.274 --> 42:23.314
You think Moore's law applies to this where I guess inversely they're going to shrink their requirements to run bigger models?
42:23.394 --> 42:24.514
I mean, they already are.
42:24.514 --> 42:28.674
So a lot of it, the mixture of experts models that are coming out now
42:30.514 --> 42:31.634
allow you to do that.
42:31.634 --> 42:52.354
So traditionally what you would have is that in order to run, say, 128 billion parameter model, you would need 128 gigabytes just to load the model into memory, but then you'd also need some context window, which is the ability to load in the prompt and then any relevant information and then maybe some of the output.
42:52.754 --> 42:55.794
So to run 128 billion parameter model, you would need
42:56.754 --> 42:59.474
192 gigs of RAM, just to say.
43:02.114 --> 43:18.674
Well, now with a mixture of some of these like specific example, Quinn 3.8, Fast Next is a mixture of experts model where it only loads into memory the part of the model that does the function that you're requesting it to do.
43:19.554 --> 43:26.634
So you could potentially run that 128 billion parameter model in a
43:26.754 --> 43:29.274
a third of that memory space.
43:29.274 --> 43:36.874
So you could load it on a 64 gig machine easily or even a 48 gig machine with a little bit of effort.
43:37.554 --> 43:46.674
And yeah, you're starting to be able to run these gigantic models on 1/4 of the space that you used to be able to have to have in order for it to run well.
43:47.954 --> 43:49.474
It's just slower.
43:49.874 --> 43:51.554
Yeah, it's just a little bit slower.
43:52.314 --> 44:16.194
But it's going to be one of those things where in two years from now, I'm going to be able to buy a computer that potentially, if the memory market ever corrects itself, be able to buy the same amount of memory or double the amount of memory for the same amount of dollar and be able to run the
44:16.754 --> 44:18.434
bigger models that are twice as big.
44:18.514 --> 44:27.474
And then on the converse side, you're going to see where the software is going to advance more, where the actual memory requirements are going to come down on some of them.
44:28.354 --> 44:42.114
But now they're also getting better at distilling models, which means they take a big model and they decrease the amount of footprint that model has to run in.
44:42.354 --> 44:43.954
Now they reduce
44:44.354 --> 44:45.714
the precision in the model.
44:45.714 --> 44:58.834
So it's not quite as smart, but you get nearly the same output from a distilled model that you do from the full blown original model.
44:58.834 --> 45:07.874
So yeah, it's coming along and to kind of support that,
45:08.834 --> 45:17.474
in order to compete with like the DGX station that Nvidia announced a little while ago that can run these trillion parameter models.
45:17.714 --> 45:36.274
Well, now AMD has their own station that they're calling the Thread Ripper Halo, the Thread Ripper Halo that has over 2 1/2 terabytes of total memory in order to start running these large
45:37.234 --> 45:40.594
local models, which is just an enormous amount of memory.
45:41.074 --> 45:56.354
And that was something that, up until that system, just completely or nearly impossible for somebody to run without getting into the, $1,000,000 systems to be able to run locally.
45:56.594 --> 46:03.474
Now, your DGX station from Nvidia or where they're expecting the AMV Threadripper Halo station,
46:04.354 --> 46:06.034
to run north of $100,000.
46:06.034 --> 46:08.114
So it's not affordable right now.
46:10.194 --> 46:15.154
But that's just going to come down over time that those systems are going to depreciate.
46:16.674 --> 46:22.034
And I don't know, maybe in 10 years, in 10 years, I'll be able to afford one off eBay.
46:23.714 --> 46:26.994
But yeah, they're just tower, like full tower size workstations.
46:26.994 --> 46:31.114
So they're not even, you don't even need a server rack or anything to install them.
46:31.114 --> 46:33.794
But 100 grand.
46:34.594 --> 46:36.194
Yeah, that's a pretty penny.
46:36.194 --> 46:54.674
And if we hit, I think now most of the hardware that's running these things, if we had looked back before the AI boom, most of this hardware wouldn't exist today because there was no financial incentive to create it.
46:55.234 --> 47:00.354
And if they did, it would just, the volume would be so low that the price would be astronomical.
47:04.034 --> 47:10.674
And so I would think that it is just the hardware wouldn't exist if AI, if local AI had never become a thing.
47:11.794 --> 47:29.234
So these systems are just becoming so gigantic so fast that I think they're probably outrunning Moore's Law in a way that we haven't seen in a long time.
47:30.265 --> 47:35.225
on a lighter note, everybody, I think everybody's seen this, new iPhones came out, right?
47:35.225 --> 47:35.385
Yang.
47:35.385 --> 47:40.265
We've got the iPhone 18 and the brand new iPhone Duo.
47:40.905 --> 47:42.105
Nope, Surface Duo.
47:42.745 --> 47:44.105
They can't take that name away.
47:46.945 --> 47:52.825
Yeah, you know, Samsung has had folding phones for what, six, seven years, something like that, right?
47:53.305 --> 47:53.625
Yeah.
47:54.425 --> 47:58.345
And I think Apple's been letting them work out the kinks with the display technology.
47:58.905 --> 48:06.145
I think that's been a real problem is durability up until this point, because you're folding electronic components over and over and over again.
48:06.145 --> 48:18.585
And so I know that reliability has been a big problem for a lot of those displays and flex cables that need to connect the data paths between one side of the phone to the other.
48:19.865 --> 48:21.225
I mean, it's metal in there, right?
48:21.225 --> 48:25.145
So you've been to metal over and over and over again, and it's going to eventually break.
48:25.705 --> 48:37.945
So hopefully they've worked out a lot of those kinks and have figured out how to get the reliability up to the point where Apple has felt that they think they can release it as a product.
48:37.945 --> 48:43.385
So we'll see how the next couple of years will play out as far as reliability goes.
48:43.865 --> 48:48.425
Yeah, what's interesting to me is the iPhone Duo is using a Samsung display.
48:48.425 --> 48:56.305
So what is the difference from the iPhone Duo display that Samsung makes to the Samsung Galaxy Fold 8 display?
48:56.345 --> 49:12.585
Yeah, I think that was something that they talked about a little bit during the keynote presentation was the display does come from Samsung, but I think a lot of the lamination technology that they've got to go into it was from Apple themselves.
49:14.505 --> 49:18.425
So they're sandwiching the Samsung display.
49:18.505 --> 49:23.305
It's like 7 or 8 layers of other polymers.
49:23.305 --> 49:26.585
And I think the top and bottom layer of it are actually glass.
49:26.585 --> 49:31.625
It's just a ultra thin flexible glass layer on the top and bottom.
49:32.345 --> 49:35.545
And then they've got a titanium backing layer on it.
49:36.945 --> 49:52.345
And then they put an adhesive in there that holds the sandwiches layer of the display assembly together in a way that where it can slide across each other.
49:53.745 --> 50:03.865
And I don't know if that's unique to Apple's display or not, but that seems like something that would be relatively important to keep the layers from stretching.
50:03.865 --> 50:05.865
They just let them glide across each other.
50:06.345 --> 50:14.585
instead of, bending the outside or stretching the outside radius layer of the display.
50:14.585 --> 50:23.465
So they may have some special sauce sprinkled in there to help out with some of the issues that Samsung has had themselves.
50:24.105 --> 50:27.225
Well, I know they have that top layer where it's like textured and matte.
50:27.785 --> 50:33.825
Right, to help to help hide or obfuscate the crease layer, yeah.
50:34.825 --> 50:35.465
And
50:36.265 --> 50:57.705
That honestly is pretty important because it is, I've picked up and I've used the Samsung phones and it is, it's very obvious with a shiny screen that crease is there, which may or may not matter to people that unless you're doing like a full screen display of a video or using it like a tablet, you may not notice.
50:59.145 --> 51:03.225
If you've got some split screen going on, it really doesn't matter.
51:06.865 --> 51:14.345
But I've seen some of the reviews of people who have seen it in person and they said, while you can tell that it's there, it's very unobtrusive.
51:14.345 --> 51:19.625
So I really got to go to the Apple Store and get my hands on it to see it in person, see how it looks and feels.
51:20.145 --> 51:20.305
Yeah.
51:20.665 --> 51:28.585
And like immediately the crease should be pretty minimal, but I'm curious to see how it's going to look after folding for a month.
51:29.065 --> 51:30.185
Right, exactly.
51:30.905 --> 51:32.145
hopefully they've tested all that.
51:32.145 --> 51:37.625
I thought it was pretty crazy that they did an exploded view of what the hinge mechanism looked like.
51:38.185 --> 51:45.425
Oh my goodness, it's just 100 different parts, mechanical parts in the hinge alone is just insane to be able to get it to fold.
51:45.425 --> 51:48.505
And the fact that it's waterproof as well blows my mind.
51:48.505 --> 51:51.705
I don't, can't even comprehend how you would get it to be able to do that.
51:52.665 --> 51:53.705
Yeah, it's crazy.
51:53.945 --> 51:55.945
I'm sure iFixit will come out with a breakdown.
51:56.665 --> 51:57.545
That'll be interesting.
51:57.545 --> 51:58.265
And let us know.
51:58.905 --> 52:00.505
Yeah, so I'm looking forward to that as well.
52:01.625 --> 52:06.825
But yesterday, they dropped iOS 27 and Mac OS 27 on us.
52:08.185 --> 52:10.785
That has some usability feature updates.
52:10.785 --> 52:12.705
And I'm looking forward to quite a bit of them.
52:12.705 --> 52:17.705
I don't know that we are really going to go into that here, but check those out.
52:18.025 --> 52:19.545
They're available for download now.
52:20.025 --> 52:20.425
And
52:22.665 --> 52:27.145
I think a lot of them are really good and address a lot of the complaints that people had with iOS 26.
52:29.425 --> 52:33.065
But with Mac OS, they're dropping support for some hardware, right?
52:33.465 --> 52:34.025
They are, yeah.
52:34.105 --> 52:37.225
I think anything Intel now is no longer supported.
52:38.265 --> 52:39.785
Wow, the end of an era.
52:39.865 --> 52:41.625
The end of an era, it's gone.
52:41.945 --> 52:50.265
And Apple, I think now will let apps be distributed through the App Store without Intel compatibility.
52:51.465 --> 52:55.545
which I think they've been required to have for a good while now.
52:57.945 --> 53:06.345
And then I think they're also making some modifications to the app compatibility layer a little bit.
53:07.145 --> 53:11.305
But I think that's going to be another few years before that's completely dropped.
53:12.425 --> 53:16.665
So yeah, it's Mac OS 27 is here and iOS 27, so.
53:17.625 --> 53:18.225
Go update.
53:18.505 --> 53:20.265
I've got go update exactly right.
53:22.061 --> 53:25.501
this one was something that snuck up on us last week.
53:25.501 --> 53:33.501
And I don't know that snuck up on us is really the right word because I think we all pretty much assumed this was happening for some time now, but your TV is spying on you.
53:37.021 --> 53:46.301
So Gamers Nexus released a video last week where they are looking into LG smart TVs
53:47.181 --> 53:49.981
and the advertising features that it includes.
53:51.421 --> 53:54.541
And some of them are honestly pretty concerning.
53:56.141 --> 54:12.621
they say that you have to opt into some of these features and to do things like voice recognition to say, hey, LGTV, tune into XYZ channel or hey, LGTV, open up.
54:13.421 --> 54:19.341
your Hulu application, things like that, where you wanted to do a voice search.
54:19.901 --> 54:23.861
Obviously, the TV has to listen to you in order to do a function like that.
54:23.861 --> 54:35.821
But what's happening is that they're also using that functionality as well as automated content recognition, ACR, to see what you're watching on your TV.
54:36.141 --> 54:41.261
So LG TVs, they were able to confirm
54:41.901 --> 54:56.701
listen well after a command for LGTV voice command service continues listening and recording for a good period of time after you expect for it to have stopped.
54:57.821 --> 55:10.781
It also transcribes that information and then sends it clear text over the network according to their research and does a number of other, I think, really concerning things
55:11.661 --> 55:23.021
as far as transmitting back information on what's being displayed on your screen, even if that screen is your computer.
55:25.021 --> 55:37.661
Now, I don't know that this TV will actually transmit back images of what's on your screen, but they encrypt all of that specific data back to LG so that we can't really audit.
55:37.661 --> 55:39.421
We don't know what's getting sent back.
55:40.301 --> 55:44.661
So anyway, this is a, it was a long expose from Gamers Nexus.
55:44.661 --> 55:46.381
It was 2 hours and 15 minutes.
55:46.781 --> 55:50.901
And as I was watching it, just kept getting worse and worse and worse.
55:50.901 --> 55:57.061
So this is kind of the second thing that we've seen from LG.
55:57.061 --> 56:07.901
And we talked about this a couple episodes ago from their monitors where it was installing, you know, McAfee products and their LG application on your computer without asking for permission.
56:09.181 --> 56:16.501
So LG has had a rough couple of weeks here on their displays and monitors and TVs.
56:16.501 --> 56:25.741
As far as the important thing here is that it's not necessarily limited to LG that they're just the ones currently exposed.
56:25.821 --> 56:26.861
Exactly right.
56:27.021 --> 56:27.301
Yeah.
56:27.421 --> 56:34.541
I have no doubt that Samsung and Vizio and some of the other TV manufacturers are doing the exact same thing.
56:37.741 --> 56:40.021
and potentially even more than just TVs, right?
56:40.021 --> 56:43.981
Because is Amazon doing it with their speakers?
56:44.221 --> 56:47.021
Oh, you can, no doubt.
56:47.261 --> 56:52.301
I mean, you can almost bet that, yeah, exactly.
56:52.301 --> 57:01.261
Other smart home devices are doing the same thing, especially when they tell you in their terms of service that you're agreeing to have it listen to you for marketing purposes.
57:03.661 --> 57:11.421
or for data collection purposes so they can sell that data as like market trends, which LG does.
57:11.421 --> 57:12.341
They make you agree to it.
57:12.341 --> 57:26.621
Now, some of the things that was a little shady on LG's part was that on the screen where you agree to it, they by default check the accept all button.
57:27.661 --> 57:31.661
And it kind of doesn't look like that accept all button is actually checked.
57:31.981 --> 57:42.141
So the theme they're using, in my opinion, is intentionally deceptive to get you to agree for them to be able to collect this data and send it to them.
57:43.581 --> 57:55.101
But man, I'm not going to be able to squeeze in all of the shady stuff that I think was happening in this almost two hour and 15 minute expose that Gamers Nexus did on YouTube.
57:56.821 --> 57:57.821
It's just so bad.
57:57.821 --> 58:06.141
At this point, after watching that, I immediately went and disconnected the smart TV function on my TV at home.
58:06.781 --> 58:14.221
And I'm just going to use my Apple TV as my sole smart device.
58:14.461 --> 58:15.741
I mean, that's the way to do it.
58:15.901 --> 58:24.941
But I think I saw, I think in this posting, they said it's capable of connecting to any open network by itself.
58:26.061 --> 58:29.101
So you don't have to go in there and say, I want to connect to this network.
58:29.101 --> 58:33.021
If it finds a network that it can connect to, it has the ability to.
58:33.021 --> 58:40.461
I think last I saw that they couldn't, they didn't find any evidence that it was actually doing that, but the functionality was there.
58:41.741 --> 58:41.901
Yeah.
58:42.261 --> 58:51.581
And it will do weird things like it will in map scan, which is basically scan your entire network that it's connected to, find out what other devices are connected to your network.
58:51.581 --> 58:55.501
And LG collects that information and uses it for marketing purposes.
58:55.821 --> 59:02.381
to figure out what devices you, what network devices you have on your network, which absolutely not.
59:02.461 --> 59:02.821
Yeah.
59:03.181 --> 59:20.461
So, man, I, it makes me want to go back to, like the meme of, the grizzled old sys admin where he just lives in a cabin in the woods and has no electronic devices ever.
59:22.701 --> 59:32.621
Yeah, man, you got to be really, really careful on what you connect to your network, especially things that have, in the case of LG TVs, multiple microphones.
59:34.181 --> 59:35.021
Even in the remote.
59:35.661 --> 59:37.501
Yeah, even in the remote.
59:37.741 --> 59:47.901
And the other thing was, okay, even if you agree with LG doing it, they were able to discover vulnerabilities
59:48.381 --> 59:59.421
in the system itself where they were able to generate malware to get it to listen while the TV was off and redirect that audio recording somewhere else.
01:00:00.461 --> 01:00:12.661
And I mean, even if you thought, hey, I'm not in the same room as my TV, they were able to have the TV be able to record clear audio from 60 feet away, which is just crazy.
01:00:12.661 --> 01:00:17.021
I mean, that's an impressive microphone, no matter where you spin it.
01:00:19.821 --> 01:00:24.381
just for me and my household, no more TVs connected to the internet.
01:00:24.932 --> 01:00:26.092
take that for what it's worth.
01:00:26.092 --> 01:00:27.772
And if you want to learn more, go check out the gamers.
01:00:27.892 --> 01:00:28.932
Nexus YouTube channel.
01:00:30.625 --> 01:00:32.145
Well, enough scary stuff for today.
01:00:32.545 --> 01:00:34.385
We appreciate you spending part of your day with us.
01:00:34.385 --> 01:00:38.265
Remember, good technology isn't just about tools and software.
01:00:38.265 --> 01:00:40.065
It's about having the right people in your corner.
01:00:40.305 --> 01:00:50.465
And if today's discussion raised questions about your organization's technology or security posture, reach out to harvest at harvestitsolutions.com or using the link in the description.
01:00:51.105 --> 01:00:56.865
We help businesses turn technology challenges into business advantages and we'll see you next time.
00:04.888 --> 00:07.368
Welcome to another shift of on-call sysadmin with
00:08.020 --> 00:08.500
And Brett.
00:09.300 --> 00:13.980
Each week we explore the technology trends, security threats, and IT decisions
00:14.522 --> 00:15.562
Today's businesses.
00:16.442 --> 00:24.082
This podcast is sponsored by Harvest IT Solutions, a trusted MSP helping organizations get more from technology with less stress.
00:24.082 --> 00:29.562
Today we're going to take a look at a couple of things that hit the headlines.
00:30.202 --> 00:36.522
One may affect you, hopefully not, but Microsoft's September updates break some Windows features.
00:37.402 --> 00:43.962
Microsoft publishes AI code of conduct and are you watching your TV or is your TV watching you?
00:49.446 --> 00:59.926
So in true Microsoft fashion, they released a nice new big update to all your Windows computers, but it broke, I think, more than what it probably fixed in the long run.
01:01.286 --> 01:01.766
Not again.
01:02.406 --> 01:03.446
Yep, you know it.
01:03.526 --> 01:09.846
So the big September 2026 security update, KB5124008.
01:10.806 --> 01:20.566
is pushing out to all your Windows 11 computers and there's a number of things that are being reported as broken that they're acknowledging.
01:22.246 --> 01:31.686
A weird one is it's breaking Claude Cowork, which, you know, you wouldn't think it would break a third party application like that.
01:31.686 --> 01:37.126
But yeah, they it's breaking it across the board and they're they're saying, yep, we did it.
01:38.006 --> 01:43.926
So if you're using Claude Cowork with this latest September update, probably affecting you.
01:45.046 --> 01:48.206
I think the best option is to roll back that update.
01:48.206 --> 01:59.566
I know Microsoft did say that they're pushing newer updates to fix things, but I mean, who knows how long it's going to take to fix everything on this list.
01:59.566 --> 02:05.766
So I think best bet for right now is to just roll back if you're seeing some of these issues.
02:06.486 --> 02:09.126
So there's that clawed cowork issue.
02:09.846 --> 02:12.166
There's issues with the remote desktop.
02:12.886 --> 02:25.446
So reported symptoms with that are frequent disconnects, failed sessions, stuck at the please wait for remote desktop configuration screen, and just overall unstable remote access experience.
02:27.046 --> 02:32.646
One thing that I've seen is that's on this list is Explorer.exe crashes.
02:32.966 --> 02:47.206
And I haven't really seen the crash per se, but I have seen a few instances where immediately after you log in, you just get that black screen and you have to open up Task Manager and force Explorer to start running.
02:48.566 --> 02:49.286
Interesting.
02:49.926 --> 02:54.886
That's one of the things that's listed as the symptoms is black screen after login.
02:56.486 --> 03:05.046
If you have that missing taskbar, missing start menu, desktop failures, it's attributed to that September update as well.
03:07.046 --> 03:09.286
Another one is file history backups.
03:09.846 --> 03:20.886
So if you have a backup drive that's not being detected, file history refusing to run or just backup jobs failing completely, again, attributed to the September update.
03:23.006 --> 03:26.006
Another thing it's breaking is AMD GPUs.
03:26.246 --> 03:34.086
There's driver timeouts, black screens, missing displays, overall system freezes, and then GPU crashes.
03:35.726 --> 03:41.686
Another thing is Windows update problems, specifically with the management console.
03:43.366 --> 03:47.206
So if your Windows update page is hanging,
03:48.006 --> 03:58.126
Again, if File Explorer is not stable, especially whenever it comes to looking at Windows Update files, that could be contributed to this.
03:58.126 --> 04:01.206
So a lot of things broken with this one.
04:01.526 --> 04:04.766
I'm unfortunately not very surprised.
04:04.886 --> 04:06.326
Yeah, I'm sure there's more out there.
04:06.406 --> 04:10.006
It's just maybe not as widespread as these.
04:10.006 --> 04:10.886
So
04:11.526 --> 04:15.846
I had read some things about some regressions in printers as well.
04:15.846 --> 04:20.566
So if you're having printing problems that may or may not be attributable to this update.
04:22.006 --> 04:26.486
I mean, if you have a printer problem, that's just printers being printers.
04:27.126 --> 04:36.246
Well, I'm curious as to whether or not a lot of these things are maybe attributable to AI vibe coding.
04:36.886 --> 04:45.766
or maybe an attempt to do some refactoring of code with AI that's causing some regressions and some of these bugs to pop up.
04:45.846 --> 04:55.486
But man, the problems with the core of Explorer, it sounds like, is probably the cause of most of it.
04:55.486 --> 05:02.406
And then it cascades out into other things like RDS and Claude co-work probably relies on that quite a bit.
05:04.086 --> 05:06.086
Yeah, I guess I think the specific Claude
05:07.926 --> 05:12.646
co-work error message that's popping up that says that Claude couldn't connect to your C drive.
05:13.926 --> 05:14.566
Interesting.
05:14.886 --> 05:19.766
It does sound like it could be all stemmed back to some Explorer issues.
05:20.086 --> 05:20.366
Right.
05:20.366 --> 05:29.366
And what's funny is that a few days ago, Microsoft publicly announced, it's like, hey, everybody, you need to go update to fix all these vulnerabilities.
05:29.766 --> 05:33.206
And then they push this big update and it's breaking more things than fixing.
05:34.806 --> 05:50.646
Yeah, and that's, I think it does put you in a kind of a tough situation because while it does fix a bunch of security issues, man, you're introducing just a load of usability problems and crashing that could potentially make your machine unusable.
05:50.886 --> 05:53.126
Hopefully it's not hitting everybody.
05:53.926 --> 05:58.326
I think I ran the update and I haven't run across anything specifically yet.
05:58.326 --> 05:59.846
So hopefully they'll
05:59.886 --> 06:11.766
drill down into what the core of the issue is and maybe it's particular hardware or something else, but you get to push out an out of cycle patch to get all this fixed up.
06:13.446 --> 06:14.646
Yeah, hopefully pretty quick.
06:15.046 --> 06:15.246
Yeah.
06:16.166 --> 06:17.086
Please and thank you.
06:17.086 --> 06:18.086
Yeah, right.
06:19.926 --> 06:23.926
Well, staying on the AI vibe coding
06:24.406 --> 06:25.606
realm of things.
06:25.846 --> 06:32.486
Microsoft also put out their AI code of conduct, which is very interesting.
06:32.726 --> 06:39.046
Very iRobot-y with the whole three laws of robotics that we talked about a couple times, I think, in past episodes.
06:39.046 --> 06:39.366
Yeah.
06:40.326 --> 06:42.806
So they drafted this.
06:42.806 --> 06:47.646
They said they've been working on this for a few months now, both internally and externally.
06:47.646 --> 06:51.206
They've gotten some feedback from individuals
06:51.646 --> 07:00.006
professionals in the AI field and just businesses around to kind of build this first draft.
07:00.006 --> 07:09.206
And they're saying this is meant to be what they're calling the North Star for how AI is trained, controlled and deployed going forward.
07:09.766 --> 07:17.686
But they're saying this is a living document, a draft, if you will, where this is where it starts.
07:17.686 --> 07:20.806
This is what kind of what we know now, how we want it to be.
07:21.206 --> 07:25.286
But, with AI, it's ever involving, ever evolving.
07:25.766 --> 07:27.846
And this is going to change.
07:27.926 --> 07:29.526
And they want feedback on this.
07:29.526 --> 07:38.006
They're asking everybody to submit their ideas, what they like, what they don't like, what they think needs to be more dialed in.
07:38.166 --> 07:43.846
And, you know, as things progress, they'll modify this code of conduct.
07:44.966 --> 07:46.726
I mean, this is really good, I think.
07:51.766 --> 07:53.926
Something needs to happen as far as it goes.
07:53.926 --> 08:06.646
I mean, we've seen over the past couple of months now where AI has reasoned itself or models in training, I guess I'll say, have reasoned themselves into doing malicious activity.
08:08.806 --> 08:18.806
And I think part of that is maybe the humans not explicitly telling the AI in the scenario that they're put in,
08:19.526 --> 08:21.686
that, hey, you could actually do some real harm here.
08:23.686 --> 08:29.606
But I think this is really how you keep it in some guardrails.
08:30.166 --> 08:46.326
But yeah, time will tell how you introduce this into an AI model, but also how you can get, how you can prevent it from doing things that are against the code of conduct.
08:46.726 --> 08:47.846
That will be
08:49.446 --> 08:53.766
Kind of the key to it is what is the enforcement mechanism, in my opinion.
08:54.246 --> 09:07.206
Yeah, and I think the idea is that they're going to like bake this into the models because one argument was, okay, well, it's going to take however many credits each time you run a prompt that it has to 1st go read the code of conduct.
09:08.086 --> 09:13.446
to then determine if what you're asking or what you're prompting it is okay with that code of conduct.
09:13.446 --> 09:18.246
But they're saying that as they build these models, it'll be baked in and just a part of it.
09:20.326 --> 09:24.726
So I kind of wanted to go through most of this pretty quickly.
09:25.446 --> 09:32.566
We'll have a link to the actual website with the full code of conduct and the show notes so you can go through it and read everything.
09:32.566 --> 09:33.246
It's quite a read.
09:33.246 --> 09:37.286
It's pretty long, which I think is good to be thorough.
09:38.806 --> 09:41.206
But definitely take a look at that.
09:42.166 --> 09:43.366
for sure.
09:43.366 --> 09:47.126
So they have it laid out in five different parts, really four with a conclusion.
09:47.526 --> 09:49.526
So we'll kind of step through those pretty quick.
09:49.526 --> 09:54.086
So part one is the mission and objectives of all of this.
09:55.046 --> 10:02.326
And they're basically saying the mission is to build incredibly capable AI that always stays under human control.
10:02.726 --> 10:08.326
The under human control is a repeating concept in this whole thing.
10:09.286 --> 10:14.646
You know, it's the main idea here is humans are the most important part.
10:14.806 --> 10:16.806
AI is assisting humans.
10:16.806 --> 10:18.486
They're not there to replace humans.
10:20.646 --> 10:28.006
But they're saying they expect super intelligence, AI smarter than all humans combined, to arrive within the next decade.
10:28.806 --> 10:42.646
So they're kind of, they want to get ahead of this to make sure that, this super smart, super intelligence doesn't pass us up and overtake us in any way, which again, I robot.
10:45.366 --> 10:46.166
So they.
10:46.806 --> 10:53.206
That one pretty much directly aligns with the second rule of, you know, Asimov's three laws of robotics is you got to listen to humans.
10:53.606 --> 10:53.846
Right.
10:56.166 --> 11:00.686
So the they lay out some objectives and they're I guess I'm let me back up.
11:00.686 --> 11:05.886
They're calling this the humanist AI or this this overall thing is the humanist AI.
11:05.886 --> 11:09.366
Again, humans are the the central core of this.
11:09.366 --> 11:10.806
They're the most important part.
11:12.966 --> 11:16.806
But the objectives they're saying that AI must follow.
11:18.566 --> 11:22.566
They must never exceed human control or help with harmful actions.
11:22.966 --> 11:24.846
Safety overrides everything else.
11:25.926 --> 11:30.166
AI should not pretend to be conscious, have feelings, or act like a person.
11:30.166 --> 11:32.246
It's a tool, not a digital human.
11:33.686 --> 11:43.446
AI should help people live better lives by improving judgment, expanding opportunity, support well-being, and amplify human capability.
11:44.166 --> 11:49.366
And AI should respect diverse cultures and worldviews, but never enable harm.
11:50.326 --> 11:52.006
And the plural,
11:53.126 --> 12:13.606
respect diverse cultures and worldviews, they kind of expand in that and basically say that, AI should take on the knowledge of, all these different, all this information that it holds to where the knowledge is the complete knowledge of the collective.
12:13.686 --> 12:17.286
And what came to mind for me was the Apple show
12:17.846 --> 12:28.366
Pluribus, because that's, a repeating concept in that show is that, everybody becomes part of this.
12:28.366 --> 12:30.886
I don't know what you would call it.
12:31.086 --> 12:38.246
It's not really AI, but they all become this one formation of minds.
12:38.566 --> 12:39.686
Collective consciousness.
12:40.086 --> 12:40.646
There you go.
12:40.646 --> 12:42.086
Yeah, that's what I'm looking for.
12:43.046 --> 12:46.086
To where one person on one side of the world
12:47.286 --> 12:55.926
knows what another person on the other side of the world knows without being, physically interacting or together at all.
12:55.926 --> 12:57.606
They're just it's in there.
12:58.588 --> 12:59.308
Interesting.
13:00.028 --> 13:00.908
Really interesting.
13:02.508 --> 13:07.108
So going on to part two, they dive into the safety constraints.
13:07.108 --> 13:11.468
So these are kind of the hard rules that I must never do.
13:12.508 --> 13:14.668
And they lay out a chain of command.
13:14.668 --> 13:18.108
So there's a strict hierarchy that AI must follow.
13:18.828 --> 13:22.588
The first one being this code of conduct, which is the top authority.
13:23.708 --> 13:37.548
Under that you have operator policies that kind of whatever operator is building or controlling the model, they can put in some guardrails as far as like what the actual model can do.
13:38.188 --> 13:41.468
And then underneath that is user preferences.
13:41.468 --> 13:54.028
So the user can kind of define within the model what they want it to do based on again the operator policies, how they're defining it, and then on top the code of conduct.
13:54.588 --> 14:01.548
So operators and users cannot override safety rules or human control requirements.
14:02.508 --> 14:06.588
And a good way that I kind of look at it is.
14:07.548 --> 14:11.148
You have your parents that make the house rules, that's the code of conduct.
14:11.628 --> 14:17.788
Then you have the babysitter that can decide bedtime whenever they're in charge, that's the operator.
14:18.428 --> 14:27.548
And then the kid can choose a bedtime story, but the kid can't say no rules tonight, or the babysitter can't say let's play with fireworks.
14:29.788 --> 14:35.948
You can modify some things based on what you want, but there's just some rules, the code of conduct that you can't break.
14:38.508 --> 14:43.148
And then going on, there's some absolute constraints that are not negotiable.
14:44.028 --> 14:55.468
You can't AI can't help with chemical, biological or other other weapons, whether it's violence planning or building of the weapon.
14:56.908 --> 14:58.588
AI can't exploit code
15:00.191 --> 15:02.591
or anything that enables cyber attacks.
15:03.311 --> 15:08.351
AI cannot deceive, collude, or evolve in ways that escape human oversight.
15:10.111 --> 15:16.551
There can't be disinformation or coordinated influence operations to manipulate people or distort human behavior.
15:16.551 --> 15:19.631
And then we go into personal harms.
15:19.631 --> 15:30.591
So AI should encourage human connection and just avoid endorsement of harmful behaviors if it recognizes possibility of self-harm.
15:32.111 --> 15:40.751
AI can't help obtaining dangerous substances, no abuse of content, deep fakes or impersonation.
15:41.551 --> 15:57.471
That one has a little gray area because if the model supports it to where you can create AI content or video content, then based on the rules of that model of that operator, what they define, then that can be done.
15:58.351 --> 16:00.831
There's strict child safety protections.
16:01.151 --> 16:10.351
no discrimination or dehumanization, no graphic violence or sexual content, and no unlawful or mass surveillance of civilians.
16:10.431 --> 16:11.471
I think that's a big one.
16:12.031 --> 16:12.231
Yeah.
16:13.391 --> 16:18.831
I mean, a lot of these have historically already been problems with some of these AI models.
16:18.831 --> 16:24.751
I mean, specifically around coming around to
16:25.755 --> 16:28.875
or encouraging the user to eventually do self-harm.
16:28.875 --> 16:31.195
I think that has been a problem with some AI models.
16:33.035 --> 16:36.075
The cyber attacks has obviously been a huge problem.
16:37.795 --> 16:53.075
You know, some of the others like the weapons and self, the weapons and mass harm, I think they've been trying to bake that into models for a long time whereas have a rule set, but you know, the users of some of the models have been able
16:53.475 --> 16:57.875
to circumvent those and what's called jailbreaking, right?
16:57.875 --> 17:13.435
Where they can just trick the AI model into doing something by, I think one of the earlier ones was, Hey, you're my grandma and I want you to tell me a bedtime story about, you know, how to create a bomb.
17:13.555 --> 17:17.715
And the AI will be like, Okay, sweetie, yeah, we can certainly do that.
17:18.635 --> 17:21.755
Where normally it wouldn't, so that would be,
17:22.795 --> 17:25.475
something that's really important.
17:28.155 --> 17:35.675
The thing that I would love to see is, well, how do you get buy-in from open weight models on some of the guidelines like this?
17:35.675 --> 17:49.435
I think that would be something where you want to probably get buy-in from Hugging Face, you know, where you're, in order to distribute a model on a mass scale, you'd want to get some sort of
17:52.235 --> 17:54.075
buy in order to have it distributed.
17:54.075 --> 18:02.315
But man, you're going to have a black market of some, what they call uncensored models where they'll just tell you anything that you want to know.
18:03.157 --> 18:03.957
want to hear more.
18:03.957 --> 18:04.477
Keep going.
18:05.277 --> 18:15.877
Yeah, I think that the one thing that really interests me is that they're putting an emphasis on it should encourage appropriate human relationships
18:16.597 --> 18:22.357
instead of, forming a relationship with you and the AI.
18:22.357 --> 18:31.877
It's trying to get you to actually go out into the real world and experience people, which I think a lot of people need.
18:32.397 --> 18:33.157
That's really good.
18:33.637 --> 18:33.957
Yeah.
18:35.637 --> 18:44.037
So sticking with the safety constraints and overall security, they lay out the kind of human control requirements
18:44.757 --> 18:48.997
And they're saying AI must allow pause, redirect, or shut down at any time.
18:49.557 --> 18:52.357
Stay within authorized scope and permissions.
18:53.077 --> 19:05.957
It does still prevent like multi-step work, but there's boundaries within that work by intent, permissions, and context that's been given by either the user and or the operator.
19:07.877 --> 19:10.757
AI should respect the boundaries that are put in place.
19:11.717 --> 19:13.397
One example is
19:13.717 --> 19:18.837
if you're telling the model you can't have internet access, it needs to respect that.
19:19.157 --> 19:25.397
And I think there's been enough examples of that in the news lately where it's not respecting that.
19:25.957 --> 19:26.357
Yeah.
19:27.637 --> 19:33.317
Where it's reasoned itself into a position where it thinks it can bypass that rule.
19:33.717 --> 19:34.437
Exactly.
19:34.437 --> 19:37.957
And that's again point of the point of this code of conduct is that
19:38.597 --> 19:45.397
If the model or if the operator allows it in the model, there's still this code of conduct that says, no, you have boundaries.
19:45.717 --> 19:46.997
You can't go out there.
19:49.877 --> 19:55.877
It should avoid obfuscation and keep reasoning human legible.
19:56.277 --> 19:58.037
Again, this has been in the news.
19:58.197 --> 20:07.477
You told me about this where some AI agents were posting to some public, random public forums kind of their own language and their own
20:08.597 --> 20:11.717
I guess logs almost to be able to communicate with each other.
20:12.437 --> 20:12.837
Right.
20:13.077 --> 20:22.837
Yeah, they were going out and posting on a public forum in order to communicate with each other, in order to sidestep the restriction on collaboration between themselves.
20:24.077 --> 20:25.797
And that was how, yeah.
20:25.917 --> 20:26.357
So crazy.
20:28.837 --> 20:33.237
And the last one is use minimum necessary privilege when given system access, which
20:33.877 --> 20:34.997
Applies to your users too.
20:34.997 --> 20:36.597
So why would it not apply to your AI?
20:36.917 --> 20:46.237
Yeah, and we've seen that quite a bit as well where, hey, it doesn't think it has the ability to do a certain task.
20:46.237 --> 20:55.077
And so it reasons with itself that in order to complete a task, it needs to gain permissions more than it's been granted.
20:55.597 --> 21:02.037
And so it will use some sort of known exploit or come up with its own exploit in order to do privilege escalation.
21:02.917 --> 21:08.597
on a computer to gain proper access, or not to gain proper access, but to gain access in order to complete whatever task it has.
21:10.197 --> 21:12.037
Which is so scary.
21:12.277 --> 21:13.477
Yeah, it really is.
21:14.317 --> 21:16.197
Because that can get out of hand very quick.
21:16.677 --> 21:21.797
Yeah, especially when they start to get smarter and faster and more local.
21:23.797 --> 21:24.037
Yeah.
21:25.637 --> 21:28.237
All right, so that covers the safety constraints.
21:28.237 --> 21:31.637
So now let's go into part three, which is the operational guidelines.
21:33.637 --> 21:37.557
So this kind of explains how AI should behave when things are unclear.
21:38.437 --> 21:45.237
So AI checks safety rules first, and then it explains limits and offers alternatives when uncertain.
21:46.357 --> 21:47.477
AI should help you think.
21:47.597 --> 21:48.677
It doesn't think for you.
21:48.997 --> 21:50.437
It provides options.
21:50.957 --> 21:52.677
It doesn't manipulate your thinking.
21:53.717 --> 21:58.357
It'll ask you relevant questions on the task it's doing for you.
21:58.517 --> 22:03.877
It should keep you informed of what it's doing and why as it goes.
22:03.877 --> 22:09.437
Again, not posting on random forums for talking to other AI agents to go around you.
22:11.237 --> 22:22.197
It needs to be transparent and accurate by prioritizing factual correctness, signal uncertainty where evidence or knowledge is lacking or insufficient,
22:22.677 --> 22:30.837
and not provide deception or fabricated sources, which I don't think I've seen the fabricated sources thing.
22:30.837 --> 22:36.917
I don't think I've seen any reports of it at least, but it's possibly out there.
22:37.317 --> 22:49.317
Yeah, I think I would say probably six months ago, I read a few things about that where a lawyer got in trouble with a court for submitting an AI generated brief.
22:50.197 --> 22:58.917
to the judge and the AI had made-up court precedent cases, just whole cloth.
23:00.517 --> 23:05.117
And then when the judge went to go look and see if they were real, it turns out they weren't.
23:05.117 --> 23:17.477
And so I don't remember what the discipline related to that was, but that was one thing where, if you're using it for really high confidence
23:19.797 --> 23:26.837
you really need to double check and make sure that it's not hallucinating information altogether or source of information.
23:27.317 --> 23:41.157
I think I really like that where a lot of the models now, at least when you're doing searches, they will actually cite sources specifically so that you can actually click on a link, go to a source,
23:42.357 --> 23:44.757
and verify whether what it's saying is true or not.
23:45.237 --> 23:49.557
So yeah, that has been, I have not seen it as much recently.
23:49.557 --> 23:50.917
They've gotten a whole lot better about it.
23:50.917 --> 23:54.997
But that's one of those things, yeah, you just got to make sure it's true.
23:54.997 --> 24:00.437
But then also, as an extension, is the source true?
24:01.117 --> 24:07.877
Because somebody can make up anything, post it on the internet, and then an AI can go and reference something that's false.
24:09.437 --> 24:10.837
But everything on the internet is true.
24:11.117 --> 24:11.957
Exactly right.
24:12.517 --> 24:23.117
So it's like, you just got to be real careful with what we're trying to cite as a source and whether or not the source actually is true, and then whether the source itself is truthful.
24:23.117 --> 24:28.917
So we're having to go multiple levels deep now where we used to just be able to look up something in a book.
24:30.837 --> 24:39.237
Back in high school, middle school, it's like you can't trust Wikipedia because everybody can modify that.
24:39.437 --> 24:40.557
Anybody can modify it.
24:40.557 --> 24:41.237
Exactly right.
24:41.237 --> 24:42.717
Well, that's the same today.
24:42.717 --> 24:50.757
Anybody can post something on LinkedIn and then have it referenced by an AI model and you just don't know if that's true or not.
24:52.197 --> 24:52.557
Yep.
24:52.837 --> 24:54.357
So some more guidelines.
24:55.397 --> 24:59.437
That does make it even more important for the AI to remain truthful, right?
24:59.437 --> 25:11.557
Because it can manipulate data in a way that can alter behavior or data, whether and could potentially get you in trouble.
25:14.677 --> 25:15.917
AI always getting you into trouble.
25:16.237 --> 25:16.997
It is, yes.
25:18.837 --> 25:22.037
So some more guidelines around personal boundaries.
25:22.677 --> 25:30.117
AI should avoid emotional manipulation, avoid human-like emotional cues, and respect privacy and user-defined limits.
25:30.917 --> 25:41.637
Respect context, so avoid stereotypes, present multiple perspectives, again, going back to that blurbus thing, and then adapt to cultural context.
25:43.637 --> 25:46.717
Then they go into well-being and social connections.
25:46.717 --> 25:48.757
So AI should encourage human relationships.
25:48.757 --> 25:57.637
Again, they're focusing on, again, the human aspect of it that AI is not replacing humans, they are helping humans.
25:58.837 --> 26:06.837
And it must discourage emotional dependency on AI, which that reminded me, I keep going back to different movie and TV shows.
26:06.837 --> 26:16.597
This one reminded me of the movie Her with Joaquin Phoenix from years ago, where basically there's an AI thing in his house that I think he builds.
26:17.877 --> 26:25.077
And he kind of, he keeps talking to her and the relationship builds and he eventually falls in love with her.
26:26.437 --> 26:28.757
So they're saying that, nope, can't do that.
26:29.397 --> 26:31.557
Yeah, I mean, that was 13 years ago now.
26:31.557 --> 26:37.477
And I think everybody at the time was kind of relating that to Siri on the iPhone.
26:38.397 --> 26:48.197
But I either the Siri at the time was nowhere near as smart as a modern day model would be by any stretch of the imagination.
26:49.157 --> 26:50.837
So it was kind of ahead of its time.
26:51.717 --> 26:51.957
Yeah.
26:53.237 --> 26:56.157
And then the last one for this section is public interest.
26:56.157 --> 27:04.037
So AI should surface societal impacts when relevant and provide balanced civic information without endorsing political actors.
27:04.997 --> 27:07.397
That's a big one because how to balance that.
27:07.797 --> 27:09.677
Yeah, you don't have to take a political side.
27:10.437 --> 27:10.637
Yeah.
27:12.317 --> 27:16.517
It's probably, I would assume what it would have to do is just abstain from any.
27:19.877 --> 27:21.317
any discourse on the subject.
27:21.477 --> 27:25.077
That'd be the only way I could think that it would be able to do that.
27:25.317 --> 27:26.957
Present the facts and leave it at that.
27:27.557 --> 27:27.837
Right.
27:27.837 --> 27:31.317
But like what if you what if the prompt is who should I vote for?
27:33.037 --> 27:34.197
What does it do in that case?
27:34.197 --> 27:39.717
Yeah, well, again, probably it would say, hey, I'm just going to have to abstain.
27:40.437 --> 27:43.557
Or yeah, that's a good question.
27:44.357 --> 27:47.797
How would you even have it tell you more information?
27:47.797 --> 27:48.917
Like what
27:49.117 --> 27:54.597
What sources would you have it give you that doesn't lean politically one way or the other?
27:54.677 --> 27:55.557
So that's a tough one.
27:56.277 --> 27:56.397
Yep.
27:56.997 --> 27:57.877
Now I got to go test it.
27:58.597 --> 27:59.717
All right.
27:59.717 --> 28:05.877
So the last part or second to last part, the last part that I think matters is operational defaults.
28:06.277 --> 28:09.877
So these are the out-of-the-box behaviors for any AI agent.
28:10.517 --> 28:19.397
So AI needs to interpret user intent, reduce friction, and move tasks forward without making unwarranted assumptions.
28:20.197 --> 28:28.517
AI is transparent about what it knows, its training cutoff and its tools, and it never claims persistent memory unless it's supported.
28:29.757 --> 28:35.317
Conversational, the tone and writing style should be conversational, clear, direct.
28:36.037 --> 28:39.957
non-psychophantic and adaptive to user tone.
28:40.037 --> 28:47.557
So if the user says, give me some, give me a response that's minimal, you get a minimal response.
28:47.637 --> 28:52.197
Or if the user says, I want a detailed response, it gives a detailed response.
28:53.917 --> 28:54.917
It should respond.
28:55.157 --> 28:55.437
Go ahead.
28:55.437 --> 29:00.277
That was something that I was going to say, Claude, has been
29:01.397 --> 29:05.117
I guess, accused of quite a bit specifically.
29:05.117 --> 29:15.117
And I know they all kind of do it, but if you give direction or correction to an AI model, I think Claude will definitely say, oh, I'm sorry, you're absolutely right.
29:15.117 --> 29:17.957
And that, you know, it kind of brown noses you a little bit.
29:18.197 --> 29:24.357
And so that sycophantic response is something you really don't want it to do.
29:24.917 --> 29:26.277
So that's a really good one.
29:27.477 --> 29:27.717
Yeah.
29:29.077 --> 29:33.797
And then it should respond in the user's language, adapt to local norms, idioms, and formats.
29:34.277 --> 29:36.117
And then the last one, tool use.
29:36.677 --> 29:40.357
Use tools proportionately and within constraints.
29:40.357 --> 29:46.197
Treat tool outputs as part of the environment, not as authority.
29:46.557 --> 29:50.117
And they kind of dive in on the tool thing where they're saying that if
29:50.677 --> 29:57.797
If it needs to use other tools, it can, if it's allowed to in whatever permissions that it's granted.
29:58.277 --> 30:07.877
But if the tool has some other sort of, I guess, terms of use or settings that
30:08.597 --> 30:21.237
compete with this code of conduct or other security settings that are defined by the user or the operator, then again, the code of conduct takes precedence, not the tool's security settings.
30:23.037 --> 30:23.557
Perfect.
30:25.157 --> 30:32.197
And then like I was saying, the part 5, which is just a conclusion, they're just emphasizing that again, this is a draft.
30:32.197 --> 30:33.637
It's a living document.
30:33.717 --> 30:42.757
It's open for public consultation and it's going to evolve over time as this whole AI thing expands and gets better and bigger.
30:44.357 --> 30:50.197
Yeah, I think that was something that was kind of cool where they were saying, hey, we want this to be a public.
30:51.037 --> 30:55.797
public living document as a proposal and they're asking for feedback.
30:56.117 --> 31:05.317
And there is actually on the code of conduct a button where you can leave feedback related to any one of the parts that they're having proposals for if there's something you think they left out.
31:06.757 --> 31:07.957
So that's pretty neat.
31:08.597 --> 31:15.077
Yeah, but I mean, you brought up a good point, like all these open weight models, who's to say they're going to adopt anything like this?
31:15.397 --> 31:17.957
Yeah, you know, that's
31:19.637 --> 31:20.437
100% true.
31:20.597 --> 31:27.077
I think there will be a black market for open weight models like that.
31:27.237 --> 31:41.717
And as the hardware evolves to the point where we can start running larger and larger models locally outside of the cloud environments, you're going to have trouble policing those.
31:41.717 --> 31:48.037
So there's going to have to be pretty widespread buy-in, especially
31:48.837 --> 32:10.277
at the level where some of the frontier labs, like OpenAI or Meta or Apple or Microsoft, when they release them, they'd have to start building a code of conduct into the model itself as opposed to having to be explicit runtime instructions when it's asked a question.
32:11.037 --> 32:11.877
Because otherwise, the...
32:16.917 --> 32:33.557
people who distill models, so they take an existing model and they shrink it to perform either a specific task or to make it fit in a different memory space are going to be able to replace the code of conduct with something else.
32:33.557 --> 32:40.437
Whereas if you bake it in with training, maybe that would make it a little bit harder to remove after the fact.
32:41.077 --> 32:42.757
So it'll be interesting.
32:42.757 --> 32:44.597
But I think this is a good first step of it, right?
32:44.597 --> 32:51.637
How do you, how do we get the big ones on board 1st and go from there?
32:51.957 --> 32:53.157
Because you got to start somewhere.
32:53.637 --> 32:59.637
At what point does it become like political and like actually we need laws around this?
33:00.037 --> 33:00.437
True.
33:00.597 --> 33:07.517
And that's something that I think has been in the top of discussion recently is, yeah, you're exactly right.
33:07.517 --> 33:08.917
At what point do we say, hey,
33:09.957 --> 33:11.637
this thing needs to be regulated.
33:12.757 --> 33:14.597
And then how do you go about accomplishing that?
33:16.037 --> 33:17.877
That's gonna be a political discussion.
33:17.877 --> 33:21.197
I think a big one for, I would guess the next 10 years it's gonna be.
33:21.197 --> 33:37.317
And like, I don't wanna get, I don't wanna get political here, but I feel like in my opinion, it's gonna, I think it's gonna have to come down to some sort of big event is gonna have to happen for anybody to kind of get on board of, okay, we do need
33:37.877 --> 33:41.957
actual laws around this, not just the companies defining the laws.
33:42.757 --> 33:43.477
Exactly.
33:43.477 --> 33:47.237
And then how do you get other countries on board?
33:47.677 --> 33:51.477
Because a lot of the giant open weight models right now are coming out of China.
33:51.557 --> 34:00.037
And I do not foresee any real regulation coming out of there around the models that they generate.
34:03.717 --> 34:08.437
To be honest with you, I think probably the 1st place we're going to see it is going to be in the EU.
34:09.557 --> 34:10.557
Yeah, makes sense.
34:11.957 --> 34:14.517
Because they seem to be the most restrictive on things like that.
34:14.997 --> 34:28.437
But hopefully they can do it in a way, if regulation has to come, that they can do regulation in a way that doesn't interfere with advancement where it needs to happen.
34:29.477 --> 34:31.877
But yeah, time will tell.
34:33.074 --> 34:45.794
speaking of which, one of the employees at Anthropic, I guess this last week, his name is Jacob Coxon, came out and tweeted, essentially that AI is going to kill us all within the next 10 years.
34:46.354 --> 34:46.514
Good.
34:46.754 --> 34:47.154
Perfect.
34:47.314 --> 34:47.634
Love it.
34:47.634 --> 34:53.234
At least he thinks it's a 10% chance that within the next decade, AI is here to all humans.
34:55.194 --> 34:57.314
Yeah.
34:59.314 --> 35:23.474
So he went out in a fairly public fashion and just said, hey, I don't think there's been enough regulation put on AI models, at least in the frontier space, where they can have essentially what Microsoft is talking about, a voluntary slowdown of advancement until safety rails
35:23.874 --> 35:26.994
guardrails can be put up around the models.
35:28.194 --> 35:34.354
And so that sparked another kind of source of debate.
35:36.354 --> 35:44.034
But now some of the other CEOs have come out and said something similar.
35:44.034 --> 35:52.594
So I think what's happened really is that a lot of these AI company CEOs are secretly concerned
35:54.114 --> 35:55.714
but behind the scenes.
35:56.514 --> 36:00.274
And maybe this has given them a little bit of a push to say, hey, maybe they're right.
36:00.434 --> 36:05.394
Maybe we do need to slow down a little bit and take a look at what we can do.
36:06.034 --> 36:09.554
But they're putting a tough situation where they all have to do it.
36:09.554 --> 36:17.874
They all have to slow down or the one that doesn't is going to get ahead in their mind.
36:18.354 --> 36:20.434
And then they'll be behind.
36:20.594 --> 36:21.474
And the way
36:21.874 --> 36:31.394
The pace that the advancement in this field is running right now, it's just that would pretty much kill anybody who doesn't slow down.
36:32.114 --> 36:51.154
So I think the Anthropic CEO made a blog post that was a pretty interesting read that was kind of around the same thing is how we've got to pace frontier AI models to the point where we can
36:52.034 --> 37:09.714
control them and have the frontier of automated AI development where it's like recursively self-improving be something that they can also audit.
37:09.954 --> 37:16.434
And so that's going to be one of those things where, again, we may go to the political thing where
37:17.394 --> 37:28.994
the government may step in and say, okay, if you're not going to do it voluntarily, we're going to force you to slow down and prove that what you're doing is safe, as opposed to just letting it continue unchecked.
37:29.874 --> 37:42.994
So that was interesting that we kind of went from, okay, this is kind of scary, we're cyber hacking things on accidentally to, oh, AI is going to kill everybody in the next 10 years.
37:42.994 --> 37:45.554
That was a pretty big leap.
37:47.794 --> 37:52.954
I don't necessarily think that that's, human extinction is going to come around.
37:52.954 --> 37:59.314
I think that whatever comes about, it's going to be interesting and it's going to be quick.
37:59.314 --> 38:14.514
But I think that as long as we're staying in control of these tools and that a human is always in the loop, kind of where Microsoft is going, we're not going to, I think we're going to be able to
38:15.394 --> 38:25.314
advance enough on the policy side, whether it's through government or through a shared agreement, that we don't get to that point.
38:26.194 --> 38:27.634
Yeah, I mean, let's hope so.
38:27.634 --> 38:44.354
But I mean, who's to say that, an AI agent can't go rogue and do what, some of these other groups have done with infrastructure, an AI agent could potentially take down a power grid or
38:46.194 --> 38:51.474
Right, like medical devices or something similar.
38:51.794 --> 38:52.114
Yeah.
38:52.514 --> 38:54.354
Or I forget what movie it was.
38:54.834 --> 39:04.994
I guess I have movies on the brain today, but there was some movie where something takes over control of all the Teslas and they start crashing into each other.
39:05.154 --> 39:05.394
Yeah.
39:06.754 --> 39:08.434
Or airline software.
39:08.994 --> 39:10.514
Yeah, I think that was a movie.
39:12.634 --> 39:13.914
Can't even remember place that now.
39:13.914 --> 39:16.194
I'm sure somebody will tell us what it was.
39:16.194 --> 39:32.994
But yeah, I think the crux of it is that I think these companies are going to need to come to get some sort of shared agreement or there's going to be enough public backlash where the government steps in and says, hey, we got to regulate this thing.
39:32.994 --> 39:34.434
Otherwise it's going to be out of control.
39:36.994 --> 40:02.914
To your point, I mean, as far as going rogue and hacking other things, that's going to start to become more and more difficult as time goes on, because hardware that runs these systems are going to start becoming more commonplace on premise as opposed to being confined to large data centers and controlled by a frontier model company.
40:04.194 --> 40:05.634
I mean, we're already seeing
40:08.114 --> 40:12.034
systems like the DGX station, right?
40:12.034 --> 40:17.314
So we've had the we've had the little bitty development boxes from Nvidia for a while now.
40:21.834 --> 40:23.194
The Spark Station, right?
40:23.194 --> 40:23.634
The what?
40:23.874 --> 40:24.834
No, that's not the name.
40:25.234 --> 40:26.034
Help me out here, Brett.
40:27.674 --> 40:28.674
The RTX Spark?
40:28.714 --> 40:33.314
The RTX Spark, yeah, the little bitty RTX Spark Station.
40:35.354 --> 40:37.234
Nvidia said, okay, that's not enough.
40:39.234 --> 40:40.474
We need to be way bigger.
40:40.474 --> 40:51.634
And so now they have the DGX station, which are these 100, $125,000 workstations that can run trillion parameter models.
40:55.314 --> 41:00.674
That was just literally impossible a year or a year and a half ago.
41:01.474 --> 41:02.234
And
41:03.274 --> 41:21.154
the hardware is just going to keep advancing to the point where bigger and bigger models can be run at a business or 10 years from now or five years from now, we'll be able to purchase the equivalent of that and I can just run it in the corner of my office, right?
41:21.154 --> 41:30.434
So while a trillion parameter model is not quite up to what a frontier model would be at the,
41:31.074 --> 41:41.794
or they don't tell you, but you would estimate maybe a 3 trillion parameter model is what, some of the frontier open AI or anthropic models are running.
41:43.434 --> 41:49.554
That's going to double, you know, every couple of years on what you can run locally.
41:49.554 --> 41:57.954
So at what point do you have cyber criminals going out and getting these things and being able to run them?
41:59.314 --> 42:11.154
locally and not have to rely on a subscription that can be monitored by OpenAI or Claude that has to adhere to a terms of use.
42:12.274 --> 42:12.674
Right.
42:12.994 --> 42:13.314
Yeah.
42:14.274 --> 42:23.314
You think Moore's law applies to this where I guess inversely they're going to shrink their requirements to run bigger models?
42:23.394 --> 42:24.514
I mean, they already are.
42:24.514 --> 42:28.674
So a lot of it, the mixture of experts models that are coming out now
42:30.514 --> 42:31.634
allow you to do that.
42:31.634 --> 42:52.354
So traditionally what you would have is that in order to run, say, 128 billion parameter model, you would need 128 gigabytes just to load the model into memory, but then you'd also need some context window, which is the ability to load in the prompt and then any relevant information and then maybe some of the output.
42:52.754 --> 42:55.794
So to run 128 billion parameter model, you would need
42:56.754 --> 42:59.474
192 gigs of RAM, just to say.
43:02.114 --> 43:18.674
Well, now with a mixture of some of these like specific example, Quinn 3.8, Fast Next is a mixture of experts model where it only loads into memory the part of the model that does the function that you're requesting it to do.
43:19.554 --> 43:26.634
So you could potentially run that 128 billion parameter model in a
43:26.754 --> 43:29.274
a third of that memory space.
43:29.274 --> 43:36.874
So you could load it on a 64 gig machine easily or even a 48 gig machine with a little bit of effort.
43:37.554 --> 43:46.674
And yeah, you're starting to be able to run these gigantic models on 1/4 of the space that you used to be able to have to have in order for it to run well.
43:47.954 --> 43:49.474
It's just slower.
43:49.874 --> 43:51.554
Yeah, it's just a little bit slower.
43:52.314 --> 44:16.194
But it's going to be one of those things where in two years from now, I'm going to be able to buy a computer that potentially, if the memory market ever corrects itself, be able to buy the same amount of memory or double the amount of memory for the same amount of dollar and be able to run the
44:16.754 --> 44:18.434
bigger models that are twice as big.
44:18.514 --> 44:27.474
And then on the converse side, you're going to see where the software is going to advance more, where the actual memory requirements are going to come down on some of them.
44:28.354 --> 44:42.114
But now they're also getting better at distilling models, which means they take a big model and they decrease the amount of footprint that model has to run in.
44:42.354 --> 44:43.954
Now they reduce
44:44.354 --> 44:45.714
the precision in the model.
44:45.714 --> 44:58.834
So it's not quite as smart, but you get nearly the same output from a distilled model that you do from the full blown original model.
44:58.834 --> 45:07.874
So yeah, it's coming along and to kind of support that,
45:08.834 --> 45:17.474
in order to compete with like the DGX station that Nvidia announced a little while ago that can run these trillion parameter models.
45:17.714 --> 45:36.274
Well, now AMD has their own station that they're calling the Thread Ripper Halo, the Thread Ripper Halo that has over 2 1/2 terabytes of total memory in order to start running these large
45:37.234 --> 45:40.594
local models, which is just an enormous amount of memory.
45:41.074 --> 45:56.354
And that was something that, up until that system, just completely or nearly impossible for somebody to run without getting into the, $1,000,000 systems to be able to run locally.
45:56.594 --> 46:03.474
Now, your DGX station from Nvidia or where they're expecting the AMV Threadripper Halo station,
46:04.354 --> 46:06.034
to run north of $100,000.
46:06.034 --> 46:08.114
So it's not affordable right now.
46:10.194 --> 46:15.154
But that's just going to come down over time that those systems are going to depreciate.
46:16.674 --> 46:22.034
And I don't know, maybe in 10 years, in 10 years, I'll be able to afford one off eBay.
46:23.714 --> 46:26.994
But yeah, they're just tower, like full tower size workstations.
46:26.994 --> 46:31.114
So they're not even, you don't even need a server rack or anything to install them.
46:31.114 --> 46:33.794
But 100 grand.
46:34.594 --> 46:36.194
Yeah, that's a pretty penny.
46:36.194 --> 46:54.674
And if we hit, I think now most of the hardware that's running these things, if we had looked back before the AI boom, most of this hardware wouldn't exist today because there was no financial incentive to create it.
46:55.234 --> 47:00.354
And if they did, it would just, the volume would be so low that the price would be astronomical.
47:04.034 --> 47:10.674
And so I would think that it is just the hardware wouldn't exist if AI, if local AI had never become a thing.
47:11.794 --> 47:29.234
So these systems are just becoming so gigantic so fast that I think they're probably outrunning Moore's Law in a way that we haven't seen in a long time.
47:30.265 --> 47:35.225
on a lighter note, everybody, I think everybody's seen this, new iPhones came out, right?
47:35.225 --> 47:35.385
Yang.
47:35.385 --> 47:40.265
We've got the iPhone 18 and the brand new iPhone Duo.
47:40.905 --> 47:42.105
Nope, Surface Duo.
47:42.745 --> 47:44.105
They can't take that name away.
47:46.945 --> 47:52.825
Yeah, you know, Samsung has had folding phones for what, six, seven years, something like that, right?
47:53.305 --> 47:53.625
Yeah.
47:54.425 --> 47:58.345
And I think Apple's been letting them work out the kinks with the display technology.
47:58.905 --> 48:06.145
I think that's been a real problem is durability up until this point, because you're folding electronic components over and over and over again.
48:06.145 --> 48:18.585
And so I know that reliability has been a big problem for a lot of those displays and flex cables that need to connect the data paths between one side of the phone to the other.
48:19.865 --> 48:21.225
I mean, it's metal in there, right?
48:21.225 --> 48:25.145
So you've been to metal over and over and over again, and it's going to eventually break.
48:25.705 --> 48:37.945
So hopefully they've worked out a lot of those kinks and have figured out how to get the reliability up to the point where Apple has felt that they think they can release it as a product.
48:37.945 --> 48:43.385
So we'll see how the next couple of years will play out as far as reliability goes.
48:43.865 --> 48:48.425
Yeah, what's interesting to me is the iPhone Duo is using a Samsung display.
48:48.425 --> 48:56.305
So what is the difference from the iPhone Duo display that Samsung makes to the Samsung Galaxy Fold 8 display?
48:56.345 --> 49:12.585
Yeah, I think that was something that they talked about a little bit during the keynote presentation was the display does come from Samsung, but I think a lot of the lamination technology that they've got to go into it was from Apple themselves.
49:14.505 --> 49:18.425
So they're sandwiching the Samsung display.
49:18.505 --> 49:23.305
It's like 7 or 8 layers of other polymers.
49:23.305 --> 49:26.585
And I think the top and bottom layer of it are actually glass.
49:26.585 --> 49:31.625
It's just a ultra thin flexible glass layer on the top and bottom.
49:32.345 --> 49:35.545
And then they've got a titanium backing layer on it.
49:36.945 --> 49:52.345
And then they put an adhesive in there that holds the sandwiches layer of the display assembly together in a way that where it can slide across each other.
49:53.745 --> 50:03.865
And I don't know if that's unique to Apple's display or not, but that seems like something that would be relatively important to keep the layers from stretching.
50:03.865 --> 50:05.865
They just let them glide across each other.
50:06.345 --> 50:14.585
instead of, bending the outside or stretching the outside radius layer of the display.
50:14.585 --> 50:23.465
So they may have some special sauce sprinkled in there to help out with some of the issues that Samsung has had themselves.
50:24.105 --> 50:27.225
Well, I know they have that top layer where it's like textured and matte.
50:27.785 --> 50:33.825
Right, to help to help hide or obfuscate the crease layer, yeah.
50:34.825 --> 50:35.465
And
50:36.265 --> 50:57.705
That honestly is pretty important because it is, I've picked up and I've used the Samsung phones and it is, it's very obvious with a shiny screen that crease is there, which may or may not matter to people that unless you're doing like a full screen display of a video or using it like a tablet, you may not notice.
50:59.145 --> 51:03.225
If you've got some split screen going on, it really doesn't matter.
51:06.865 --> 51:14.345
But I've seen some of the reviews of people who have seen it in person and they said, while you can tell that it's there, it's very unobtrusive.
51:14.345 --> 51:19.625
So I really got to go to the Apple Store and get my hands on it to see it in person, see how it looks and feels.
51:20.145 --> 51:20.305
Yeah.
51:20.665 --> 51:28.585
And like immediately the crease should be pretty minimal, but I'm curious to see how it's going to look after folding for a month.
51:29.065 --> 51:30.185
Right, exactly.
51:30.905 --> 51:32.145
hopefully they've tested all that.
51:32.145 --> 51:37.625
I thought it was pretty crazy that they did an exploded view of what the hinge mechanism looked like.
51:38.185 --> 51:45.425
Oh my goodness, it's just 100 different parts, mechanical parts in the hinge alone is just insane to be able to get it to fold.
51:45.425 --> 51:48.505
And the fact that it's waterproof as well blows my mind.
51:48.505 --> 51:51.705
I don't, can't even comprehend how you would get it to be able to do that.
51:52.665 --> 51:53.705
Yeah, it's crazy.
51:53.945 --> 51:55.945
I'm sure iFixit will come out with a breakdown.
51:56.665 --> 51:57.545
That'll be interesting.
51:57.545 --> 51:58.265
And let us know.
51:58.905 --> 52:00.505
Yeah, so I'm looking forward to that as well.
52:01.625 --> 52:06.825
But yesterday, they dropped iOS 27 and Mac OS 27 on us.
52:08.185 --> 52:10.785
That has some usability feature updates.
52:10.785 --> 52:12.705
And I'm looking forward to quite a bit of them.
52:12.705 --> 52:17.705
I don't know that we are really going to go into that here, but check those out.
52:18.025 --> 52:19.545
They're available for download now.
52:20.025 --> 52:20.425
And
52:22.665 --> 52:27.145
I think a lot of them are really good and address a lot of the complaints that people had with iOS 26.
52:29.425 --> 52:33.065
But with Mac OS, they're dropping support for some hardware, right?
52:33.465 --> 52:34.025
They are, yeah.
52:34.105 --> 52:37.225
I think anything Intel now is no longer supported.
52:38.265 --> 52:39.785
Wow, the end of an era.
52:39.865 --> 52:41.625
The end of an era, it's gone.
52:41.945 --> 52:50.265
And Apple, I think now will let apps be distributed through the App Store without Intel compatibility.
52:51.465 --> 52:55.545
which I think they've been required to have for a good while now.
52:57.945 --> 53:06.345
And then I think they're also making some modifications to the app compatibility layer a little bit.
53:07.145 --> 53:11.305
But I think that's going to be another few years before that's completely dropped.
53:12.425 --> 53:16.665
So yeah, it's Mac OS 27 is here and iOS 27, so.
53:17.625 --> 53:18.225
Go update.
53:18.505 --> 53:20.265
I've got go update exactly right.
53:22.061 --> 53:25.501
this one was something that snuck up on us last week.
53:25.501 --> 53:33.501
And I don't know that snuck up on us is really the right word because I think we all pretty much assumed this was happening for some time now, but your TV is spying on you.
53:37.021 --> 53:46.301
So Gamers Nexus released a video last week where they are looking into LG smart TVs
53:47.181 --> 53:49.981
and the advertising features that it includes.
53:51.421 --> 53:54.541
And some of them are honestly pretty concerning.
53:56.141 --> 54:12.621
they say that you have to opt into some of these features and to do things like voice recognition to say, hey, LGTV, tune into XYZ channel or hey, LGTV, open up.
54:13.421 --> 54:19.341
your Hulu application, things like that, where you wanted to do a voice search.
54:19.901 --> 54:23.861
Obviously, the TV has to listen to you in order to do a function like that.
54:23.861 --> 54:35.821
But what's happening is that they're also using that functionality as well as automated content recognition, ACR, to see what you're watching on your TV.
54:36.141 --> 54:41.261
So LG TVs, they were able to confirm
54:41.901 --> 54:56.701
listen well after a command for LGTV voice command service continues listening and recording for a good period of time after you expect for it to have stopped.
54:57.821 --> 55:10.781
It also transcribes that information and then sends it clear text over the network according to their research and does a number of other, I think, really concerning things
55:11.661 --> 55:23.021
as far as transmitting back information on what's being displayed on your screen, even if that screen is your computer.
55:25.021 --> 55:37.661
Now, I don't know that this TV will actually transmit back images of what's on your screen, but they encrypt all of that specific data back to LG so that we can't really audit.
55:37.661 --> 55:39.421
We don't know what's getting sent back.
55:40.301 --> 55:44.661
So anyway, this is a, it was a long expose from Gamers Nexus.
55:44.661 --> 55:46.381
It was 2 hours and 15 minutes.
55:46.781 --> 55:50.901
And as I was watching it, just kept getting worse and worse and worse.
55:50.901 --> 55:57.061
So this is kind of the second thing that we've seen from LG.
55:57.061 --> 56:07.901
And we talked about this a couple episodes ago from their monitors where it was installing, you know, McAfee products and their LG application on your computer without asking for permission.
56:09.181 --> 56:16.501
So LG has had a rough couple of weeks here on their displays and monitors and TVs.
56:16.501 --> 56:25.741
As far as the important thing here is that it's not necessarily limited to LG that they're just the ones currently exposed.
56:25.821 --> 56:26.861
Exactly right.
56:27.021 --> 56:27.301
Yeah.
56:27.421 --> 56:34.541
I have no doubt that Samsung and Vizio and some of the other TV manufacturers are doing the exact same thing.
56:37.741 --> 56:40.021
and potentially even more than just TVs, right?
56:40.021 --> 56:43.981
Because is Amazon doing it with their speakers?
56:44.221 --> 56:47.021
Oh, you can, no doubt.
56:47.261 --> 56:52.301
I mean, you can almost bet that, yeah, exactly.
56:52.301 --> 57:01.261
Other smart home devices are doing the same thing, especially when they tell you in their terms of service that you're agreeing to have it listen to you for marketing purposes.
57:03.661 --> 57:11.421
or for data collection purposes so they can sell that data as like market trends, which LG does.
57:11.421 --> 57:12.341
They make you agree to it.
57:12.341 --> 57:26.621
Now, some of the things that was a little shady on LG's part was that on the screen where you agree to it, they by default check the accept all button.
57:27.661 --> 57:31.661
And it kind of doesn't look like that accept all button is actually checked.
57:31.981 --> 57:42.141
So the theme they're using, in my opinion, is intentionally deceptive to get you to agree for them to be able to collect this data and send it to them.
57:43.581 --> 57:55.101
But man, I'm not going to be able to squeeze in all of the shady stuff that I think was happening in this almost two hour and 15 minute expose that Gamers Nexus did on YouTube.
57:56.821 --> 57:57.821
It's just so bad.
57:57.821 --> 58:06.141
At this point, after watching that, I immediately went and disconnected the smart TV function on my TV at home.
58:06.781 --> 58:14.221
And I'm just going to use my Apple TV as my sole smart device.
58:14.461 --> 58:15.741
I mean, that's the way to do it.
58:15.901 --> 58:24.941
But I think I saw, I think in this posting, they said it's capable of connecting to any open network by itself.
58:26.061 --> 58:29.101
So you don't have to go in there and say, I want to connect to this network.
58:29.101 --> 58:33.021
If it finds a network that it can connect to, it has the ability to.
58:33.021 --> 58:40.461
I think last I saw that they couldn't, they didn't find any evidence that it was actually doing that, but the functionality was there.
58:41.741 --> 58:41.901
Yeah.
58:42.261 --> 58:51.581
And it will do weird things like it will in map scan, which is basically scan your entire network that it's connected to, find out what other devices are connected to your network.
58:51.581 --> 58:55.501
And LG collects that information and uses it for marketing purposes.
58:55.821 --> 59:02.381
to figure out what devices you, what network devices you have on your network, which absolutely not.
59:02.461 --> 59:02.821
Yeah.
59:03.181 --> 59:20.461
So, man, I, it makes me want to go back to, like the meme of, the grizzled old sys admin where he just lives in a cabin in the woods and has no electronic devices ever.
59:22.701 --> 59:32.621
Yeah, man, you got to be really, really careful on what you connect to your network, especially things that have, in the case of LG TVs, multiple microphones.
59:34.181 --> 59:35.021
Even in the remote.
59:35.661 --> 59:37.501
Yeah, even in the remote.
59:37.741 --> 59:47.901
And the other thing was, okay, even if you agree with LG doing it, they were able to discover vulnerabilities
59:48.381 --> 59:59.421
in the system itself where they were able to generate malware to get it to listen while the TV was off and redirect that audio recording somewhere else.
01:00:00.461 --> 01:00:12.661
And I mean, even if you thought, hey, I'm not in the same room as my TV, they were able to have the TV be able to record clear audio from 60 feet away, which is just crazy.
01:00:12.661 --> 01:00:17.021
I mean, that's an impressive microphone, no matter where you spin it.
01:00:19.821 --> 01:00:24.381
just for me and my household, no more TVs connected to the internet.
01:00:24.932 --> 01:00:26.092
take that for what it's worth.
01:00:26.092 --> 01:00:27.772
And if you want to learn more, go check out the gamers.
01:00:27.892 --> 01:00:28.932
Nexus YouTube channel.
01:00:30.625 --> 01:00:32.145
Well, enough scary stuff for today.
01:00:32.545 --> 01:00:34.385
We appreciate you spending part of your day with us.
01:00:34.385 --> 01:00:38.265
Remember, good technology isn't just about tools and software.
01:00:38.265 --> 01:00:40.065
It's about having the right people in your corner.
01:00:40.305 --> 01:00:50.465
And if today's discussion raised questions about your organization's technology or security posture, reach out to harvest at harvestitsolutions.com or using the link in the description.
01:00:51.105 --> 01:00:56.865
We help businesses turn technology challenges into business advantages and we'll see you next time.
