WEBVTT
00:00:02.080 --> 00:00:13.199
A US vs China AI Cold War is starting, and most business leaders have no idea they may already be participants.
00:00:13.599 --> 00:00:18.559
And it's moving much faster than hardly anyone can keep up with.
00:00:18.719 --> 00:00:28.320
About two weeks ago, Reuters reported that China is considering locking down its most advanced AI models, keeping its best from the rest of the world.
00:00:28.559 --> 00:00:37.600
Then, just yesterday, the Trump administration is signaling it may restrict Chinese AI models inside the United States.
00:00:37.759 --> 00:00:45.119
Yeah, read that again, because both sides are now moving to wall off AI access within days of each other.
00:00:45.359 --> 00:00:50.799
And here's why two governments suddenly care so much about who runs which model.
00:00:51.359 --> 00:00:58.240
Well, that's because this was never really about who has the most powerful AI or who has the best benchmarks.
00:00:58.399 --> 00:01:05.760
It's actually about economic growth, geopolitical power, and who sets the standards the rest of the world builds on.
00:01:06.000 --> 00:01:15.599
That's because whoever leads AI doesn't just win a tech race towards superintelligence, they gain leverage over every other country's future.
00:01:15.920 --> 00:01:22.560
And that's the fight your business may now be caught in the middle of, whether you signed up for it or not.
00:01:22.879 --> 00:01:24.400
So let's dig in.
00:01:24.640 --> 00:01:26.879
Here is the big picture.
00:01:27.599 --> 00:01:32.079
The AI power is flipping right in front of our very eyes.
00:01:32.239 --> 00:01:41.359
That's because for three plus years, the US has been pretty far ahead, at least when it came to the open source Chinese models.
00:01:41.519 --> 00:01:46.719
But now China is actually closing the gap with open weight models.
00:01:46.959 --> 00:01:53.280
Uh, yeah, closing the gap between the best that anthropic and open AI have to offer.
00:01:53.439 --> 00:02:00.159
But now both governments are treating their frontier models as strategic weapons worth potentially restricting.
00:02:00.319 --> 00:02:15.439
And US businesses that have been chasing models and maybe using these open source Chinese models because they were very capable, now might accidentally inherit some serious geopolitical deployment risk.
00:02:15.680 --> 00:02:23.120
So on today's show, you'll learn how these Chinese models became cheaper, closer, and harder to ignore.
00:02:23.280 --> 00:02:33.439
You're gonna know why a trillion parameter, yes, trillion with a T, a trillion parameter open model no longer means that you can simply run it on your computer.
00:02:33.680 --> 00:02:46.560
You're gonna know why Microsoft testing Chinese models could signal the enterprise AI future, and you're gonna know what leaders should deploy, document, and avoid before restrictions may harden.
00:02:46.719 --> 00:02:47.840
All right, let's get into it.
00:02:47.919 --> 00:02:49.199
Welcome to Everyday AI.
00:02:49.360 --> 00:02:53.280
My name's Jordan Wilson, and we do this every day, and it's yours.
00:02:53.439 --> 00:03:05.759
This is your unedited, unscripted daily live stream podcast and free daily newsletter, helping business leaders like you and me keep sense of all of these developments because, yeah, they're happening at warped speed.
00:03:05.919 --> 00:03:07.599
I tell you what matters, what doesn't.
00:03:07.680 --> 00:03:12.000
You take that information, and you're the smartest person in AI at your company.
00:03:12.159 --> 00:03:17.759
So it starts here with a podcast, but make sure to go to our website at your everydayai.com.
00:03:17.840 --> 00:03:21.840
We're going to be recapping the highlights from today's show and a whole lot more.
00:03:22.080 --> 00:03:24.319
Uh, let's get straight into it.
00:03:24.560 --> 00:03:26.400
What the heck has happened?
00:03:26.639 --> 00:03:49.039
I mean, my gosh, as someone that's been doing this everyday AI thing for three and a half years, uh, the last like five or six weeks, at least when it comes to uh the US versus China tensions and even what uh each country is pushing out, aside from just AI acceleration is at an all-time high.
00:03:49.199 --> 00:03:54.080
The battle between these two nations couldn't be uh any higher.
00:03:54.639 --> 00:04:08.479
So, uh, like I said, in the past two weeks, we saw reports first that uh Beijing is looking at curbing oversea access to its top AI models.
00:04:08.639 --> 00:04:17.680
And then we got a report just yesterday saying the Trump administration may also ban Chinese AI models, and that's big news.
00:04:17.839 --> 00:04:34.560
That's because we've seen a lot of reports, I've talked to people personally, uh, that are moving entire enterprises off, you know, maybe like an open AI or Anthropic or Google models, uh, right, and on to other Chinese models.
00:04:34.720 --> 00:04:51.759
I think especially when we saw GLM 5.2 from ZAI, which was technically a more affordable uh, you know, version of some of the Frontier models, it wasn't quite yet punching uh at the state-of-the-art um kind of AI class, but it was getting close.
00:04:51.920 --> 00:05:06.879
That's when it kind of started, but it's really snowballed just the past week with models like Kimi K3, uh, which is now just under the uh Fable 5 and uh GPT-5.6 soul class.
00:05:07.040 --> 00:05:16.000
Um, and then uh Quinn, their new 3.8 model, which we don't have full benchmarks on yet, uh, but it's released, which is weird, right?
00:05:16.079 --> 00:05:21.600
Normally companies don't release models and not release benchmarks, but it's live and it's seemingly really good.
00:05:21.680 --> 00:05:24.879
And it also may be entering uh that top tier.
00:05:25.040 --> 00:05:35.040
So in that upper echelon now, we might have models from anthropic, openai, and two open weight Chinese companies.
00:05:35.199 --> 00:05:37.279
So, in theory, these are models, right?
00:05:37.439 --> 00:05:52.480
Again, you can't really run them on your computer, but these are models that large, large enterprises that have the means to do it, they can run all of these things locally, again, assuming they have a couple uh servers to throw on.
00:05:53.279 --> 00:06:00.240
So here's why AI leadership now decides the global superpower status.
00:06:00.399 --> 00:06:02.480
It's three things, uh, right?
00:06:03.519 --> 00:06:08.560
It's the economy, power, and this multi-access.
00:06:08.959 --> 00:06:13.279
Okay, so here's what each of those boils down to.
00:06:13.439 --> 00:06:25.680
Whoever automates thinking work fastest compounds growth over every rival nation, and that helps whoever is actually winning AI have a leg up on the economy.
00:06:26.000 --> 00:06:34.720
Power, I mean, aside from you know, these uh systems, these AI systems, because that's what there are, they're much more than models, right?
00:06:34.800 --> 00:06:37.360
They're being used in military use cases.
00:06:37.439 --> 00:06:40.800
And I've been talking about this since the very beginning, right?
00:06:40.959 --> 00:06:51.600
Before we had models that should have even been touching the battlefield, I've said the future of AI is definitely uh it is going to become the new oil, it is going to become the new gold.
00:06:51.759 --> 00:06:56.079
And I think that most people that hang out on the bleeding edge understand that to be true.
00:06:56.160 --> 00:07:01.120
Maybe the rest of the world is we'll see that in a year or two, but that is the reality.
00:07:01.360 --> 00:07:17.920
If you control AI, if you control uh, you know, if you are in the lead toward AGI or artificial superintelligence, that means that your country will wield a or yield a power over every other country, and there may not be much that anyone can do about it.
00:07:18.079 --> 00:07:29.519
I mean, when you think about things like being able to uh, you know, put cyber attacks and being able to, in theory, take down entire countries' power grids, their banking systems, right?
00:07:29.680 --> 00:07:30.480
That's what we're at.
00:07:30.560 --> 00:07:35.279
It's things that uh physical weapons, you know, would try to do, uh, right.
00:07:35.360 --> 00:07:39.920
But this is something that could, in theory, be launched autonomously at scale.
00:07:40.079 --> 00:07:47.519
That's the down and the ugly side of AI, but that's ultimately anything as powerful as artificial intelligence in these models that we have now.
00:07:47.759 --> 00:07:56.319
You have to think it's much more about, you know, it's about much more than helping us all write better emails or helping us triage our days better.
00:07:56.480 --> 00:08:10.480
There is a bigger and sometimes badder purpose between but um behind what nations, especially the uh nations jostling for power at the top of the global uh pyramid.
00:08:10.639 --> 00:08:12.240
Uh, this is what they want it for.
00:08:12.399 --> 00:08:15.920
And then last but not least, it is this um this parody.
00:08:16.079 --> 00:08:23.600
There is this uh parody that's happening right now because right now, no nation leads in every access, right?
00:08:23.680 --> 00:08:28.800
When it comes to model capability, uh available compute, cost, adoption, and deployment.
00:08:28.879 --> 00:08:29.360
Uh, right.
00:08:29.439 --> 00:08:33.039
And I think that's one of the things that uh China is really working on.
00:08:33.200 --> 00:08:36.879
And just FY, I as I'm talking about this, I'm obviously right.
00:08:37.039 --> 00:08:39.039
Um, I'm based in the US.
00:08:39.200 --> 00:08:42.559
Um, most of the people and companies I work with are based in the US.
00:08:42.639 --> 00:08:46.879
So I'm obviously coming at it from that perspective, if you couldn't tell already.
00:08:47.200 --> 00:08:51.440
So one of the biggest questions is like, why?
00:08:51.679 --> 00:08:57.440
Why does China have they been coming with this open source or open weight approach?
00:08:57.519 --> 00:08:59.919
Um, and the US just really hasn't.
00:09:00.080 --> 00:09:06.960
Well, first, uh, you know, some recent models from the US have done fairly okay um on the open source scheme.
00:09:07.120 --> 00:09:10.799
So uh Thinking Machines Labs, new inkling model, really good.
00:09:10.960 --> 00:09:18.639
Um, you know, NVIDIA's uh open models fairly good, but no one's been able to touch the uh Chinese models.
00:09:18.799 --> 00:09:33.279
Part of that is because of distillation, which we'll get to, but it always gets to why would people always ask why would China put out these open source models that, you know, at least six months ago, you know, you could download them on very powerful consumer hardware and run them.
00:09:33.360 --> 00:09:34.879
And everyone was always confused.
00:09:34.960 --> 00:09:36.799
And I think that there's a couple of reasons.
00:09:36.960 --> 00:09:50.240
Uh, but one, it China wants to be the default on what the rest of the world builds on, because if so, that makes the all the other services that you might need to run those models more valuable.
00:09:50.480 --> 00:09:53.919
And every enterprise, here's the thing: it is a competition.
00:09:54.080 --> 00:10:02.960
Every enterprise that switches from US models drains the revenue funding, uh, essentially Silicon Valley's next training runs, right?
00:10:03.039 --> 00:10:08.240
So if you take the fuel out of the car, the car can no longer run.
00:10:08.559 --> 00:10:14.799
Uh, also, huge models make enterprises rent Chinese compatible hosting tools and support.
00:10:14.960 --> 00:10:21.919
And then China gains adoption and they they weaken US's pricing power and they keep the leverage.
00:10:22.080 --> 00:10:30.960
The other thing that most people don't talk about is controlling the narrative, something that China is obviously uh very concerned about and has been concerned about for many decades.
00:10:31.120 --> 00:10:39.519
But uh, these Chinese models, uh, right, studies have shown that they avoid sensitive topics that Beijing does not want to discuss globally.
00:10:39.840 --> 00:10:52.480
So people are just sometimes copying and pasting whatever uh an AI model spits out, and they're maybe sending it to colleagues, they're sending it to clients, or in many cases, they're just putting it on the internet, um, right?
00:10:52.639 --> 00:10:57.679
And then large language models start to regurgitate this, and this becomes part of the training data.
00:10:57.840 --> 00:11:07.200
So uh this is a way, and you know, it this gets um, you know, put out in schools, media, government, documents, everything.
00:11:07.840 --> 00:11:13.679
So, um, and you also have people using these models to distill and create other models, right?
00:11:13.759 --> 00:11:24.080
As an example, uh, I believe cursors, uh, Kimmy, uh, I believe cursors models, uh, their first ones that uh were based off of Kimmy's open source models.
00:11:24.799 --> 00:11:26.879
So, why does that matter?
00:11:27.039 --> 00:11:29.120
Why does China want to control the narrative?
00:11:29.279 --> 00:11:36.879
Well, uh, a Stanford study even found that China origin models answered political questions less directly.
00:11:37.039 --> 00:11:53.120
So it's it's not like you know these models are going to say something that you know overtly slams the US or overtly, you know, puts uh you know, Chinese uh, you know, morals and ethics on a pedestal.
00:11:53.200 --> 00:11:55.600
That's not what I'm saying, but it's just the nuance.
00:11:55.759 --> 00:11:59.039
It's the you know, describing things in a slightly different way.
00:11:59.200 --> 00:12:02.480
And you know, you know, if you think of the game of telephone, right?
00:12:02.639 --> 00:12:19.200
Each time that happens, you know, each time someone just blindly copies and pastes something on the internet, and then the next round of frontier models get trained on that information, and it just starts to weaken and distill maybe certain talking points that Beijing would rather not be out there.
00:12:19.360 --> 00:12:30.960
Uh, so there's it's much more than just about controlling the um, you know, what the rest of the world builds on and you know, maybe sucking uh the US's uh power supply dry.
00:12:31.039 --> 00:12:33.360
It's also about controlling the narrative.
00:12:33.759 --> 00:12:36.960
All right, so let's talk about some of the more recent models.
00:12:37.120 --> 00:12:46.000
So I think this all started uh earlier this summer or late spring with uh ZAI's GLM 5.2.
00:12:46.159 --> 00:12:59.679
Uh so that is not nearly on the same tier as the most recent ones from this past week, and that's Moonshot's Kimi K3 and Alibaba's uh Alibaba's Quen 3.8 Max.
00:12:59.919 --> 00:13:06.399
And they essentially, on some things, undercut uh US pricing, especially GLM 5.2.
00:13:06.720 --> 00:13:10.159
And in many cases, they have been good enough, right?
00:13:10.399 --> 00:13:21.679
We've literally read uh stories where uh, you know, I would say more uh tech forward or AI native companies, but large ones essentially took their clawed spend, right?
00:13:21.759 --> 00:13:24.559
Because Anthropic's models are the most expensive.
00:13:24.720 --> 00:13:35.840
And there was a period, uh, right, where uh Anthropic's Mythos V Fable 5 came out uh before OpenAI released their uh competing model in GBD 5.6.
00:13:35.919 --> 00:13:43.200
So there was this time and period where Anthropic had a lead in the quote unquote model wars, but it was just ridonculously expensive.
00:13:43.360 --> 00:14:01.440
And everyone's like, wait, we could use a model like GLM5.2, which isn't that far behind their OPIS class model, and we could get like 95% uh of the power of like an Opus model for a fraction of the cost.
00:14:02.000 --> 00:14:12.559
Um, so they were just saying they're trying to undercut US pricing with good enough models, not necessarily by being the number one model in the world, and that's kind of where we stand now.
00:14:12.720 --> 00:14:15.039
And I I said this on a show earlier.
00:14:15.120 --> 00:14:20.080
Um, I think previously Chinese models were like six months behind.
00:14:20.240 --> 00:14:22.960
Um, now it's like one to two months.
00:14:23.120 --> 00:14:28.159
Uh, part of that is I think their distillation efforts and their architecture under, right?
00:14:28.240 --> 00:14:29.360
It's not just distillation.
00:14:29.519 --> 00:14:39.120
Obviously, uh, these labs have some of the most talented engineers in the world, so it's a combination, I think, of you know, number one, their distillation uh efforts have increased.
00:14:39.279 --> 00:14:45.200
Uh, number two, some of their architecture that they're putting out is truly good and novel and making a difference.
00:14:45.360 --> 00:15:01.600
And well, um, number three, which you can't overlook, is the recent uh kind of US sanctions uh that have been handed down largely because of anthropic to all model providers here in the US, which is delaying.
00:15:01.679 --> 00:15:10.960
We've seen reports uh maybe like a 30-day, 45-day or more, whereas these companies would have been pushing these models out to the public a little faster.
00:15:11.120 --> 00:15:20.879
But now that we live in permission slip AI land, uh, you know, it's a combination of those three things that have kind of closed this gap that was much wider before.
00:15:21.120 --> 00:15:22.960
But it's the benchmarks.
00:15:23.039 --> 00:15:24.639
We do have to talk about the benchmarks.
00:15:24.799 --> 00:15:25.360
All right.
00:15:25.759 --> 00:15:43.039
So if we look at the artificial analysis index, which we talk about, uh, sorry, the artificial analysis intelligence index, which we do talk about a lot on this show, you know, now all of a sudden you have Kimmy K3 uh in this same uh tier, right?
00:15:43.200 --> 00:15:47.519
Almost as Claude Fable 5 and GPT 5.6 soul.
00:15:47.759 --> 00:15:54.159
So on this benchmark, which is probably the best single overall benchmarks because it is a conglomerate, right?
00:15:54.320 --> 00:16:02.480
Claude Fable 5 has a 60, GPT 5.6 soul has a 59, and Kimmy K3 has a 57, right?
00:16:02.559 --> 00:16:12.159
And there's usually always like at least a 10% drop off, uh, right, in terms of the AA uh index.
00:16:12.399 --> 00:16:16.480
Now it's like a 5% drop off, right?
00:16:16.720 --> 00:16:19.759
Which is not that big of a drop.
00:16:20.639 --> 00:16:31.039
So here's though where it's changed, especially the past like couple of models, uh, with ZAI's GLM 5.2.
00:16:31.279 --> 00:16:41.759
Um, this is something that, yeah, you could run it a slowed down, a watered-down version of that model if you had a super powerful uh consumer PC, right?
00:16:41.840 --> 00:16:57.360
Or um you couldn't actually run it at full speed, nothing like you could run online, but people out there that spent maybe way too much money on their uh consumer setups, they were able to run uh essentially watered down versions of GLM 5.2.
00:16:57.519 --> 00:16:58.720
Not anymore, right?
00:16:58.799 --> 00:17:05.599
Because now it seems like the next step that China is taking, they want to compete on the frontier, the frontier frontier.
00:17:05.759 --> 00:17:16.960
And as the US models, uh the state of the air models get more and more capable, uh, well, it gets harder for uh, you know, a small enough model to close that gap.
00:17:17.119 --> 00:17:22.480
Because now what we're seeing is these multiple trillion parameter models.
00:17:22.720 --> 00:17:24.400
So open source, right?
00:17:24.640 --> 00:17:33.359
Especially through 2025, generally meant okay, uh, you could get a quant version of this model, right?
00:17:33.519 --> 00:17:37.279
Which you know, uh only activates certain parameters.
00:17:37.359 --> 00:17:43.359
So think about two uh you know, four um uh context, right?
00:17:43.599 --> 00:17:47.920
The uh the GPT-4 family of models was like two trillion parameters.
00:17:48.079 --> 00:17:52.880
So now you have open source models that are bigger than that, um, right?
00:17:54.319 --> 00:17:56.160
Which is crazy.
00:17:56.319 --> 00:17:59.200
So these multiple trillion parameters, you can't run them.
00:17:59.359 --> 00:18:01.839
You literally need a small data center.
00:18:01.920 --> 00:18:10.559
Uh, or like I said, if you are an enterprise company that has access to compute, yeah, you're gonna need a whole rack of NVIDIA GPUs to actually run these things.
00:18:10.640 --> 00:18:16.079
So uh this does even change, I think, how most enterprise leaders should be viewing open source.
00:18:16.319 --> 00:18:18.319
It's almost like you should be saying, like, what kind?
00:18:18.480 --> 00:18:24.160
Like consumer open source or enterprise open source, because there really wasn't that distinction.
00:18:24.240 --> 00:18:26.640
There really wasn't that, you know, multiple tiers.
00:18:26.799 --> 00:18:40.000
Um, you know, because if someone chained together a couple Mac Studios a year ago, you could probably run a quant uh a quant version of these um, you know, Chinese open source models.
00:18:41.359 --> 00:18:47.759
But they're not exactly cheap anymore either.
00:18:47.920 --> 00:18:53.759
So with that distinction or moving away, um means they're not exactly cheap.
00:18:54.240 --> 00:18:56.799
And so for me, when I'm looking, right?
00:18:56.880 --> 00:18:59.279
If I'm a business leader, well, I am, right?
00:18:59.359 --> 00:19:06.160
But I'm not uh necessarily making enterprise decisions at Fortune 500 companies, although I do advise those type of companies.
00:19:06.319 --> 00:19:15.519
Um before, even two, three months ago, you say, Yeah, look at open source models as an alternative.
00:19:15.839 --> 00:19:19.279
Today, I don't really know why people should.
00:19:19.680 --> 00:19:24.240
And maybe that's largely because of what OpenAI has been able to do, um, right?
00:19:24.559 --> 00:19:34.319
In in terms of cost per intelligence index, uh, right, which I think is uh just as important as the overall intelligence from artificial analysis.
00:19:34.400 --> 00:19:40.079
But this essentially says, how much are you paying to get these tasks done?
00:19:40.240 --> 00:19:45.680
Uh right, because all of these benchmarks, right?
00:19:45.839 --> 00:19:51.119
Artificial analysis runs it and they say, here's how much it actually costs to get all of this work done.
00:19:51.279 --> 00:19:58.880
And what we've seen is, well, anthropic models, this is not one where you want to have a big lead, you want to have a bigger bar chart.
00:19:58.960 --> 00:19:59.839
No, that's bad.
00:20:00.000 --> 00:20:00.400
Uh, right.
00:20:00.559 --> 00:20:03.279
Anthropics models are ridiculously expensive.
00:20:03.440 --> 00:20:08.000
So, uh, as an example, it costs$2.75 per task.
00:20:08.079 --> 00:20:14.240
Um, whereas OpenAI's models um are much cheaper, up to a third cheaper.
00:20:14.400 --> 00:20:33.119
So their smaller version, GBT5.7, actually cost more um than GBD5.
00:20:34.640 --> 00:20:38.720
And the Quen costs more than the mid-tier.
00:20:39.279 --> 00:20:42.799
Um oh, sorry, I don't have this one on the benchmark here.
00:20:42.880 --> 00:20:54.240
Uh, but it right when you have a leading state-of-the-art model, um, and it's costing about the same or even less than the open source model.
00:20:54.319 --> 00:21:09.599
So it's like, okay, maybe there's uh couple hundred companies in the US that can actually go out and run this themselves without paying API prices, because if you are paying API prices, at this point, you would probably just should be using open AI.
00:21:09.839 --> 00:21:16.400
Um, because when it comes to price per task, which is what is ultimately gonna matter, they're winning, right?
00:21:16.559 --> 00:21:20.960
Or maybe you are looking at Grok in MetaMuse Spark, right?
00:21:21.119 --> 00:21:41.759
You remember on the show uh last week I said it was a really bad week for Anthropic with all of these new uh um open source models coming out, but then also with Grok and MetaMuse Spark uh 1.1 coming with some pretty good models that people weren't um expecting, especially on the coding and software engineering side.
00:21:41.920 --> 00:21:58.960
But similarly, uh artificial analysis has this uh this quadrant uh when it's cost per tax, uh cost per task, and the uh essentially intelligence or the artificial intelligence uh artificial analysis intelligence index score.
00:21:59.039 --> 00:22:10.240
So essentially you Be in the upper left hand corner, which means you have the smartest model at the cheapest cost, and none of the Chinese open source models are in that quadrant anymore.
00:22:10.400 --> 00:22:25.039
That obviously it resets as new models come out and the the the medians and the averages all change, but right now the three companies in there are OpenAI, Grok, and Meta.
00:22:25.680 --> 00:22:27.599
No open source Chinese models.
00:22:27.680 --> 00:22:49.200
Where this is actually a quadrant that they used to dominate, which is why I think six months in a year ago, it made a lot of sense for business leaders to be looking at Chinese open source models, especially when they were a little bit more tameable in terms of what you can do without having a multi-million dollar uh you know compute setup.
00:22:49.359 --> 00:22:51.519
Um, but it's just not the case anymore.
00:22:51.680 --> 00:23:11.519
So, for me, anyways, even though I know that the the war is going to rage on, if I'm making decisions, I'm looking at these charts and saying, well, there's maybe not that big of a reason for uh us to look at these models unless you do want or need that open nature, which I know some companies obviously do.
00:23:12.559 --> 00:23:28.799
But it's no longer just the cheap API prices that reveal the true cost of using these models, uh, because you have these aggressive Chinese prices that show their strategy is not just competing for money, right?
00:23:28.880 --> 00:23:40.640
Because there's other subsidies, cloud cross-selling competition, and also still these lower prices force enterprises to question the expensive closed model defaults.
00:23:41.440 --> 00:23:45.839
So I know that the models from the past week, right?
00:23:46.000 --> 00:23:58.160
Specifically Kimi K3 and Quen 3.8, they may not still fit in the traditional open source, cheap Chinese models to use paradigm, but there are still those models.
00:23:58.240 --> 00:24:02.319
They will still be uh continue to be developed, and I think that there will be a place for them.
00:24:02.559 --> 00:24:09.519
Um but it comes to distillation, we can't get to you know 20 plus minutes and not talk about distillation.
00:24:09.599 --> 00:24:23.519
So distillation is essentially where, for the most part, Chinese companies, you know, take it's just like they copy the questions and they copy the answers for a lack of a better term from the big AI model provider.
00:24:23.599 --> 00:24:31.039
So it's just like they copy the answers on the tests, they take all the hard work from the American companies and use it to train their own models, right?
00:24:31.119 --> 00:24:37.440
And that accelerates the progress, but it you still can't explain uh you know how they get there.
00:24:37.599 --> 00:24:48.640
Um, just through distillation, that's not enough because they have had some great um advancements, uh, the Chinese companies in architecture and just their overall execution.
00:24:48.960 --> 00:25:04.000
So AI model distillation isn't universally illegal, it's highly frowned upon, and it is um, you know, um sparking some backlash politically um and economically with export controls.
00:25:04.160 --> 00:25:08.240
I think we're gonna continue to see those ramped up now here for the rest of 2026.
00:25:08.400 --> 00:25:17.519
But you know, these alleged Chinese practices do violate contracts and have led to major US national security actions.
00:25:17.839 --> 00:25:21.920
So, what is the US doing to stop this model distillation?
00:25:22.079 --> 00:25:35.200
Well, they're trying to cluster traffic and you know, detect and you know, catch large-scale uh offenders and you know block them, but it's pretty hard because it's they're just playing whack-a-mole.
00:25:35.359 --> 00:25:43.119
Uh, a US House committee did, though, just unanimously back a bill to sanction foreign actors for extracting US models.
00:25:43.359 --> 00:25:48.960
But uh laws in the US are a slow-moving machine, right?
00:25:49.119 --> 00:25:57.119
So uh it could be many months or multiple quarters or even more than a year before something like that ever becomes law.
00:25:57.359 --> 00:26:02.880
But these published open weights cannot be recalled, so businesses can still adapt them, anyways, right?
00:26:03.039 --> 00:26:08.799
So when models go open source or open weight, it's kind of like the cats out of the bag at that point.
00:26:09.440 --> 00:26:28.960
So the US can try to restrict access, but that could potentially backfire because Washington can, yeah, they can pressure uh, you know, via export controls or otherwise, the chips, the cloud access, procurement, sanctions, and all these other things, but the restrictions may just protect the technology while raising the cost for American companies.
00:26:29.200 --> 00:26:47.839
Uh, like I said, if you um are restricting these models, but companies already have the weights, you may just be keeping money ultimately um from other ecosystems that could power the American um enterprise, right?
00:26:49.279 --> 00:26:54.480
So I think the writing though is kind of on the wall.
00:26:54.640 --> 00:27:09.839
Um, and I think it's actually Microsoft's the uh some of their recent actions that show I think what we might be looking at when it comes to how should large enterprises be looking at or using these Chinese models.
00:27:10.079 --> 00:27:18.000
So uh reportedly Microsoft was evaluating Deep Seek for cheaper copilot co-work model routing.
00:27:18.160 --> 00:27:21.200
So, not for all of Microsoft Copilot, right?
00:27:21.279 --> 00:27:22.880
And this is just according to reports.
00:27:22.960 --> 00:27:25.920
I don't believe Microsoft has confirmed anything yet.
00:27:26.240 --> 00:27:34.160
And they were looking at this for a backup because what they found, uh, Microsoft that the co-pilot co-work was actually very cost intensive.
00:27:34.319 --> 00:27:43.680
And they did start to bill it um, you know, uh at a token rate and no longer just including usage uh like they did when it was in beta.
00:27:43.839 --> 00:27:48.240
Um that is telling you everything that you need to know.
00:27:48.480 --> 00:28:00.000
They haven't switched to it, but Microsoft, yes, that Microsoft, the one that has huge stakes um in OpenAI and Anthropic.
00:28:00.160 --> 00:28:03.279
They are looking at a model like DeepSeq.
00:28:03.920 --> 00:28:15.519
For all the, you know, the time that DeepSeek's name has gotten dragged through the mud for, you know, some alleged shady practices, they're still looking at them to use them.
00:28:15.599 --> 00:28:23.759
And I think ultimately, uh, you have to look at your balance sheet, you have to look at the dollars and the cents and make it make sense.
00:28:24.000 --> 00:28:26.000
And in this case, Microsoft, right?
00:28:26.079 --> 00:28:37.519
And the other thing, again, if you're a large company and you can uh, you know, essentially download the weights to these models and you can fine-tune them and make sure that they perform up to a certain standard.
00:28:37.680 --> 00:28:46.640
But I think the future workloads, like in this Microsoft example, could just be routed by cost, capability, security, geography, and regulation, right?
00:28:46.799 --> 00:28:57.039
I've been a huge advocate over the years for who knows, maybe we'll get there eventually, but you know, the mixture of models, much different than a mixture of experts, right?
00:28:57.119 --> 00:29:03.920
That's using a single model, and then the model kind of calls on the different experts and the different parameters in this dense model, right?
00:29:04.079 --> 00:29:07.519
Mixture of models is similar to model routing, but it's just using, right?
00:29:07.599 --> 00:29:19.920
I think Perplexities model council, uh, Microsoft has something similar, but it's well, when you put a prompt out there and there's just a router, and it might send a simple prompt to one open source model, right?
00:29:20.000 --> 00:29:20.799
That's a good example.
00:29:20.960 --> 00:29:28.720
Or it might send a complex prompt to a hundred different models and then an orchestrator model to go in and collect all the information.
00:29:28.880 --> 00:29:33.039
And maybe half of those hundred models are open source.
00:29:33.119 --> 00:29:41.200
Uh, but I think if anything, the shift though does weaken loyalty to one model and rewards flexible architecture.
00:29:41.359 --> 00:29:45.599
I do think that is probably the future where enterprise leaders need to be focusing on.
00:29:45.759 --> 00:29:48.079
So as we wrap up, let's talk about that.
00:29:48.400 --> 00:29:52.640
What is coming next and what business leaders should be doing now?
00:29:52.799 --> 00:29:54.079
Well, I would do this.
00:29:54.240 --> 00:29:57.599
I'm gonna say expect controlled openness.
00:29:57.759 --> 00:29:58.319
All right.
00:29:58.480 --> 00:30:12.880
Whether it is China placing export on their open models, uh, which I know kind of uh goes against the uh very reason they put them out there in the first place, but that's another uh another topic for another episode.
00:30:12.960 --> 00:30:14.480
But I would expect controlled openness.
00:30:14.559 --> 00:30:38.480
So either China is going to restrict probably the US uh from using their maybe uh most frontier models, uh, and or uh the US may also uh restrict the usage of some of these Chinese open models as well, especially as now these models are getting more and more capable, uh more and more autonomous.
00:30:38.640 --> 00:30:44.240
And when it comes to cyber, things are getting a little bit scary-ish, right?
00:30:44.400 --> 00:30:50.240
Uh, I I think in six months, uh a year, that's when things are gonna be getting like real scary.
00:30:50.319 --> 00:31:05.119
Uh, in in terms of these autonomous models, as they can technically get smaller, faster, more capable, and well, more open source, that's when uh these exports are going to uh or these export controls um are going to come in.
00:31:05.279 --> 00:31:11.039
So expect both nations to probably just share older models, but maybe guard their best.
00:31:11.279 --> 00:31:22.000
Uh, but what you need to do is inventory where your current Chinese open source models are running and then route those tasks by cost, security, and policy risk.
00:31:22.079 --> 00:31:34.640
And then you need to document your sourcing and be ready to swap out a model or to uh you know go to your plan B uh when and if those government restrictions harden.
00:31:34.799 --> 00:31:37.680
All right, I hope this episode was helpful.
00:31:37.839 --> 00:31:54.960
But like I said, whether you know it or not, there's a good chance if you're using open source models that part of your plan might need to be modified pretty soon because the war between the US and China when it comes to AI is heating up.
00:31:55.039 --> 00:31:58.079
So I would expect a lot of back and forth racket.
00:31:58.319 --> 00:31:59.839
So don't get stuck in the middle.
00:32:00.000 --> 00:32:02.640
Don't waste hours every single day toiling over it.
00:32:02.880 --> 00:32:11.119
Just tune in to us here on the show and the newsletter at your everydayai.com because we're always going to keep you up to date.
00:32:11.279 --> 00:32:12.319
I hope this one was helpful.
00:32:12.480 --> 00:32:13.119
Thank you for tuning in.
00:32:13.279 --> 00:32:16.319
Hope to see you back tomorrow and every day for more everyday AI.
00:32:16.400 --> 00:32:17.279
Thanks, y'all.
00:00:02.080 --> 00:00:13.199
A US vs China AI Cold War is starting, and most business leaders have no idea they may already be participants.
00:00:13.599 --> 00:00:18.559
And it's moving much faster than hardly anyone can keep up with.
00:00:18.719 --> 00:00:28.320
About two weeks ago, Reuters reported that China is considering locking down its most advanced AI models, keeping its best from the rest of the world.
00:00:28.559 --> 00:00:37.600
Then, just yesterday, the Trump administration is signaling it may restrict Chinese AI models inside the United States.
00:00:37.759 --> 00:00:45.119
Yeah, read that again, because both sides are now moving to wall off AI access within days of each other.
00:00:45.359 --> 00:00:50.799
And here's why two governments suddenly care so much about who runs which model.
00:00:51.359 --> 00:00:58.240
Well, that's because this was never really about who has the most powerful AI or who has the best benchmarks.
00:00:58.399 --> 00:01:05.760
It's actually about economic growth, geopolitical power, and who sets the standards the rest of the world builds on.
00:01:06.000 --> 00:01:15.599
That's because whoever leads AI doesn't just win a tech race towards superintelligence, they gain leverage over every other country's future.
00:01:15.920 --> 00:01:22.560
And that's the fight your business may now be caught in the middle of, whether you signed up for it or not.
00:01:22.879 --> 00:01:24.400
So let's dig in.
00:01:24.640 --> 00:01:26.879
Here is the big picture.
00:01:27.599 --> 00:01:32.079
The AI power is flipping right in front of our very eyes.
00:01:32.239 --> 00:01:41.359
That's because for three plus years, the US has been pretty far ahead, at least when it came to the open source Chinese models.
00:01:41.519 --> 00:01:46.719
But now China is actually closing the gap with open weight models.
00:01:46.959 --> 00:01:53.280
Uh, yeah, closing the gap between the best that anthropic and open AI have to offer.
00:01:53.439 --> 00:02:00.159
But now both governments are treating their frontier models as strategic weapons worth potentially restricting.
00:02:00.319 --> 00:02:15.439
And US businesses that have been chasing models and maybe using these open source Chinese models because they were very capable, now might accidentally inherit some serious geopolitical deployment risk.
00:02:15.680 --> 00:02:23.120
So on today's show, you'll learn how these Chinese models became cheaper, closer, and harder to ignore.
00:02:23.280 --> 00:02:33.439
You're gonna know why a trillion parameter, yes, trillion with a T, a trillion parameter open model no longer means that you can simply run it on your computer.
00:02:33.680 --> 00:02:46.560
You're gonna know why Microsoft testing Chinese models could signal the enterprise AI future, and you're gonna know what leaders should deploy, document, and avoid before restrictions may harden.
00:02:46.719 --> 00:02:47.840
All right, let's get into it.
00:02:47.919 --> 00:02:49.199
Welcome to Everyday AI.
00:02:49.360 --> 00:02:53.280
My name's Jordan Wilson, and we do this every day, and it's yours.
00:02:53.439 --> 00:03:05.759
This is your unedited, unscripted daily live stream podcast and free daily newsletter, helping business leaders like you and me keep sense of all of these developments because, yeah, they're happening at warped speed.
00:03:05.919 --> 00:03:07.599
I tell you what matters, what doesn't.
00:03:07.680 --> 00:03:12.000
You take that information, and you're the smartest person in AI at your company.
00:03:12.159 --> 00:03:17.759
So it starts here with a podcast, but make sure to go to our website at your everydayai.com.
00:03:17.840 --> 00:03:21.840
We're going to be recapping the highlights from today's show and a whole lot more.
00:03:22.080 --> 00:03:24.319
Uh, let's get straight into it.
00:03:24.560 --> 00:03:26.400
What the heck has happened?
00:03:26.639 --> 00:03:49.039
I mean, my gosh, as someone that's been doing this everyday AI thing for three and a half years, uh, the last like five or six weeks, at least when it comes to uh the US versus China tensions and even what uh each country is pushing out, aside from just AI acceleration is at an all-time high.
00:03:49.199 --> 00:03:54.080
The battle between these two nations couldn't be uh any higher.
00:03:54.639 --> 00:04:08.479
So, uh, like I said, in the past two weeks, we saw reports first that uh Beijing is looking at curbing oversea access to its top AI models.
00:04:08.639 --> 00:04:17.680
And then we got a report just yesterday saying the Trump administration may also ban Chinese AI models, and that's big news.
00:04:17.839 --> 00:04:34.560
That's because we've seen a lot of reports, I've talked to people personally, uh, that are moving entire enterprises off, you know, maybe like an open AI or Anthropic or Google models, uh, right, and on to other Chinese models.
00:04:34.720 --> 00:04:51.759
I think especially when we saw GLM 5.2 from ZAI, which was technically a more affordable uh, you know, version of some of the Frontier models, it wasn't quite yet punching uh at the state-of-the-art um kind of AI class, but it was getting close.
00:04:51.920 --> 00:05:06.879
That's when it kind of started, but it's really snowballed just the past week with models like Kimi K3, uh, which is now just under the uh Fable 5 and uh GPT-5.6 soul class.
00:05:07.040 --> 00:05:16.000
Um, and then uh Quinn, their new 3.8 model, which we don't have full benchmarks on yet, uh, but it's released, which is weird, right?
00:05:16.079 --> 00:05:21.600
Normally companies don't release models and not release benchmarks, but it's live and it's seemingly really good.
00:05:21.680 --> 00:05:24.879
And it also may be entering uh that top tier.
00:05:25.040 --> 00:05:35.040
So in that upper echelon now, we might have models from anthropic, openai, and two open weight Chinese companies.
00:05:35.199 --> 00:05:37.279
So, in theory, these are models, right?
00:05:37.439 --> 00:05:52.480
Again, you can't really run them on your computer, but these are models that large, large enterprises that have the means to do it, they can run all of these things locally, again, assuming they have a couple uh servers to throw on.
00:05:53.279 --> 00:06:00.240
So here's why AI leadership now decides the global superpower status.
00:06:00.399 --> 00:06:02.480
It's three things, uh, right?
00:06:03.519 --> 00:06:08.560
It's the economy, power, and this multi-access.
00:06:08.959 --> 00:06:13.279
Okay, so here's what each of those boils down to.
00:06:13.439 --> 00:06:25.680
Whoever automates thinking work fastest compounds growth over every rival nation, and that helps whoever is actually winning AI have a leg up on the economy.
00:06:26.000 --> 00:06:34.720
Power, I mean, aside from you know, these uh systems, these AI systems, because that's what there are, they're much more than models, right?
00:06:34.800 --> 00:06:37.360
They're being used in military use cases.
00:06:37.439 --> 00:06:40.800
And I've been talking about this since the very beginning, right?
00:06:40.959 --> 00:06:51.600
Before we had models that should have even been touching the battlefield, I've said the future of AI is definitely uh it is going to become the new oil, it is going to become the new gold.
00:06:51.759 --> 00:06:56.079
And I think that most people that hang out on the bleeding edge understand that to be true.
00:06:56.160 --> 00:07:01.120
Maybe the rest of the world is we'll see that in a year or two, but that is the reality.
00:07:01.360 --> 00:07:17.920
If you control AI, if you control uh, you know, if you are in the lead toward AGI or artificial superintelligence, that means that your country will wield a or yield a power over every other country, and there may not be much that anyone can do about it.
00:07:18.079 --> 00:07:29.519
I mean, when you think about things like being able to uh, you know, put cyber attacks and being able to, in theory, take down entire countries' power grids, their banking systems, right?
00:07:29.680 --> 00:07:30.480
That's what we're at.
00:07:30.560 --> 00:07:35.279
It's things that uh physical weapons, you know, would try to do, uh, right.
00:07:35.360 --> 00:07:39.920
But this is something that could, in theory, be launched autonomously at scale.
00:07:40.079 --> 00:07:47.519
That's the down and the ugly side of AI, but that's ultimately anything as powerful as artificial intelligence in these models that we have now.
00:07:47.759 --> 00:07:56.319
You have to think it's much more about, you know, it's about much more than helping us all write better emails or helping us triage our days better.
00:07:56.480 --> 00:08:10.480
There is a bigger and sometimes badder purpose between but um behind what nations, especially the uh nations jostling for power at the top of the global uh pyramid.
00:08:10.639 --> 00:08:12.240
Uh, this is what they want it for.
00:08:12.399 --> 00:08:15.920
And then last but not least, it is this um this parody.
00:08:16.079 --> 00:08:23.600
There is this uh parody that's happening right now because right now, no nation leads in every access, right?
00:08:23.680 --> 00:08:28.800
When it comes to model capability, uh available compute, cost, adoption, and deployment.
00:08:28.879 --> 00:08:29.360
Uh, right.
00:08:29.439 --> 00:08:33.039
And I think that's one of the things that uh China is really working on.
00:08:33.200 --> 00:08:36.879
And just FY, I as I'm talking about this, I'm obviously right.
00:08:37.039 --> 00:08:39.039
Um, I'm based in the US.
00:08:39.200 --> 00:08:42.559
Um, most of the people and companies I work with are based in the US.
00:08:42.639 --> 00:08:46.879
So I'm obviously coming at it from that perspective, if you couldn't tell already.
00:08:47.200 --> 00:08:51.440
So one of the biggest questions is like, why?
00:08:51.679 --> 00:08:57.440
Why does China have they been coming with this open source or open weight approach?
00:08:57.519 --> 00:08:59.919
Um, and the US just really hasn't.
00:09:00.080 --> 00:09:06.960
Well, first, uh, you know, some recent models from the US have done fairly okay um on the open source scheme.
00:09:07.120 --> 00:09:10.799
So uh Thinking Machines Labs, new inkling model, really good.
00:09:10.960 --> 00:09:18.639
Um, you know, NVIDIA's uh open models fairly good, but no one's been able to touch the uh Chinese models.
00:09:18.799 --> 00:09:33.279
Part of that is because of distillation, which we'll get to, but it always gets to why would people always ask why would China put out these open source models that, you know, at least six months ago, you know, you could download them on very powerful consumer hardware and run them.
00:09:33.360 --> 00:09:34.879
And everyone was always confused.
00:09:34.960 --> 00:09:36.799
And I think that there's a couple of reasons.
00:09:36.960 --> 00:09:50.240
Uh, but one, it China wants to be the default on what the rest of the world builds on, because if so, that makes the all the other services that you might need to run those models more valuable.
00:09:50.480 --> 00:09:53.919
And every enterprise, here's the thing: it is a competition.
00:09:54.080 --> 00:10:02.960
Every enterprise that switches from US models drains the revenue funding, uh, essentially Silicon Valley's next training runs, right?
00:10:03.039 --> 00:10:08.240
So if you take the fuel out of the car, the car can no longer run.
00:10:08.559 --> 00:10:14.799
Uh, also, huge models make enterprises rent Chinese compatible hosting tools and support.
00:10:14.960 --> 00:10:21.919
And then China gains adoption and they they weaken US's pricing power and they keep the leverage.
00:10:22.080 --> 00:10:30.960
The other thing that most people don't talk about is controlling the narrative, something that China is obviously uh very concerned about and has been concerned about for many decades.
00:10:31.120 --> 00:10:39.519
But uh, these Chinese models, uh, right, studies have shown that they avoid sensitive topics that Beijing does not want to discuss globally.
00:10:39.840 --> 00:10:52.480
So people are just sometimes copying and pasting whatever uh an AI model spits out, and they're maybe sending it to colleagues, they're sending it to clients, or in many cases, they're just putting it on the internet, um, right?
00:10:52.639 --> 00:10:57.679
And then large language models start to regurgitate this, and this becomes part of the training data.
00:10:57.840 --> 00:11:07.200
So uh this is a way, and you know, it this gets um, you know, put out in schools, media, government, documents, everything.
00:11:07.840 --> 00:11:13.679
So, um, and you also have people using these models to distill and create other models, right?
00:11:13.759 --> 00:11:24.080
As an example, uh, I believe cursors, uh, Kimmy, uh, I believe cursors models, uh, their first ones that uh were based off of Kimmy's open source models.
00:11:24.799 --> 00:11:26.879
So, why does that matter?
00:11:27.039 --> 00:11:29.120
Why does China want to control the narrative?
00:11:29.279 --> 00:11:36.879
Well, uh, a Stanford study even found that China origin models answered political questions less directly.
00:11:37.039 --> 00:11:53.120
So it's it's not like you know these models are going to say something that you know overtly slams the US or overtly, you know, puts uh you know, Chinese uh, you know, morals and ethics on a pedestal.
00:11:53.200 --> 00:11:55.600
That's not what I'm saying, but it's just the nuance.
00:11:55.759 --> 00:11:59.039
It's the you know, describing things in a slightly different way.
00:11:59.200 --> 00:12:02.480
And you know, you know, if you think of the game of telephone, right?
00:12:02.639 --> 00:12:19.200
Each time that happens, you know, each time someone just blindly copies and pastes something on the internet, and then the next round of frontier models get trained on that information, and it just starts to weaken and distill maybe certain talking points that Beijing would rather not be out there.
00:12:19.360 --> 00:12:30.960
Uh, so there's it's much more than just about controlling the um, you know, what the rest of the world builds on and you know, maybe sucking uh the US's uh power supply dry.
00:12:31.039 --> 00:12:33.360
It's also about controlling the narrative.
00:12:33.759 --> 00:12:36.960
All right, so let's talk about some of the more recent models.
00:12:37.120 --> 00:12:46.000
So I think this all started uh earlier this summer or late spring with uh ZAI's GLM 5.2.
00:12:46.159 --> 00:12:59.679
Uh so that is not nearly on the same tier as the most recent ones from this past week, and that's Moonshot's Kimi K3 and Alibaba's uh Alibaba's Quen 3.8 Max.
00:12:59.919 --> 00:13:06.399
And they essentially, on some things, undercut uh US pricing, especially GLM 5.2.
00:13:06.720 --> 00:13:10.159
And in many cases, they have been good enough, right?
00:13:10.399 --> 00:13:21.679
We've literally read uh stories where uh, you know, I would say more uh tech forward or AI native companies, but large ones essentially took their clawed spend, right?
00:13:21.759 --> 00:13:24.559
Because Anthropic's models are the most expensive.
00:13:24.720 --> 00:13:35.840
And there was a period, uh, right, where uh Anthropic's Mythos V Fable 5 came out uh before OpenAI released their uh competing model in GBD 5.6.
00:13:35.919 --> 00:13:43.200
So there was this time and period where Anthropic had a lead in the quote unquote model wars, but it was just ridonculously expensive.
00:13:43.360 --> 00:14:01.440
And everyone's like, wait, we could use a model like GLM5.2, which isn't that far behind their OPIS class model, and we could get like 95% uh of the power of like an Opus model for a fraction of the cost.
00:14:02.000 --> 00:14:12.559
Um, so they were just saying they're trying to undercut US pricing with good enough models, not necessarily by being the number one model in the world, and that's kind of where we stand now.
00:14:12.720 --> 00:14:15.039
And I I said this on a show earlier.
00:14:15.120 --> 00:14:20.080
Um, I think previously Chinese models were like six months behind.
00:14:20.240 --> 00:14:22.960
Um, now it's like one to two months.
00:14:23.120 --> 00:14:28.159
Uh, part of that is I think their distillation efforts and their architecture under, right?
00:14:28.240 --> 00:14:29.360
It's not just distillation.
00:14:29.519 --> 00:14:39.120
Obviously, uh, these labs have some of the most talented engineers in the world, so it's a combination, I think, of you know, number one, their distillation uh efforts have increased.
00:14:39.279 --> 00:14:45.200
Uh, number two, some of their architecture that they're putting out is truly good and novel and making a difference.
00:14:45.360 --> 00:15:01.600
And well, um, number three, which you can't overlook, is the recent uh kind of US sanctions uh that have been handed down largely because of anthropic to all model providers here in the US, which is delaying.
00:15:01.679 --> 00:15:10.960
We've seen reports uh maybe like a 30-day, 45-day or more, whereas these companies would have been pushing these models out to the public a little faster.
00:15:11.120 --> 00:15:20.879
But now that we live in permission slip AI land, uh, you know, it's a combination of those three things that have kind of closed this gap that was much wider before.
00:15:21.120 --> 00:15:22.960
But it's the benchmarks.
00:15:23.039 --> 00:15:24.639
We do have to talk about the benchmarks.
00:15:24.799 --> 00:15:25.360
All right.
00:15:25.759 --> 00:15:43.039
So if we look at the artificial analysis index, which we talk about, uh, sorry, the artificial analysis intelligence index, which we do talk about a lot on this show, you know, now all of a sudden you have Kimmy K3 uh in this same uh tier, right?
00:15:43.200 --> 00:15:47.519
Almost as Claude Fable 5 and GPT 5.6 soul.
00:15:47.759 --> 00:15:54.159
So on this benchmark, which is probably the best single overall benchmarks because it is a conglomerate, right?
00:15:54.320 --> 00:16:02.480
Claude Fable 5 has a 60, GPT 5.6 soul has a 59, and Kimmy K3 has a 57, right?
00:16:02.559 --> 00:16:12.159
And there's usually always like at least a 10% drop off, uh, right, in terms of the AA uh index.
00:16:12.399 --> 00:16:16.480
Now it's like a 5% drop off, right?
00:16:16.720 --> 00:16:19.759
Which is not that big of a drop.
00:16:20.639 --> 00:16:31.039
So here's though where it's changed, especially the past like couple of models, uh, with ZAI's GLM 5.2.
00:16:31.279 --> 00:16:41.759
Um, this is something that, yeah, you could run it a slowed down, a watered-down version of that model if you had a super powerful uh consumer PC, right?
00:16:41.840 --> 00:16:57.360
Or um you couldn't actually run it at full speed, nothing like you could run online, but people out there that spent maybe way too much money on their uh consumer setups, they were able to run uh essentially watered down versions of GLM 5.2.
00:16:57.519 --> 00:16:58.720
Not anymore, right?
00:16:58.799 --> 00:17:05.599
Because now it seems like the next step that China is taking, they want to compete on the frontier, the frontier frontier.
00:17:05.759 --> 00:17:16.960
And as the US models, uh the state of the air models get more and more capable, uh, well, it gets harder for uh, you know, a small enough model to close that gap.
00:17:17.119 --> 00:17:22.480
Because now what we're seeing is these multiple trillion parameter models.
00:17:22.720 --> 00:17:24.400
So open source, right?
00:17:24.640 --> 00:17:33.359
Especially through 2025, generally meant okay, uh, you could get a quant version of this model, right?
00:17:33.519 --> 00:17:37.279
Which you know, uh only activates certain parameters.
00:17:37.359 --> 00:17:43.359
So think about two uh you know, four um uh context, right?
00:17:43.599 --> 00:17:47.920
The uh the GPT-4 family of models was like two trillion parameters.
00:17:48.079 --> 00:17:52.880
So now you have open source models that are bigger than that, um, right?
00:17:54.319 --> 00:17:56.160
Which is crazy.
00:17:56.319 --> 00:17:59.200
So these multiple trillion parameters, you can't run them.
00:17:59.359 --> 00:18:01.839
You literally need a small data center.
00:18:01.920 --> 00:18:10.559
Uh, or like I said, if you are an enterprise company that has access to compute, yeah, you're gonna need a whole rack of NVIDIA GPUs to actually run these things.
00:18:10.640 --> 00:18:16.079
So uh this does even change, I think, how most enterprise leaders should be viewing open source.
00:18:16.319 --> 00:18:18.319
It's almost like you should be saying, like, what kind?
00:18:18.480 --> 00:18:24.160
Like consumer open source or enterprise open source, because there really wasn't that distinction.
00:18:24.240 --> 00:18:26.640
There really wasn't that, you know, multiple tiers.
00:18:26.799 --> 00:18:40.000
Um, you know, because if someone chained together a couple Mac Studios a year ago, you could probably run a quant uh a quant version of these um, you know, Chinese open source models.
00:18:41.359 --> 00:18:47.759
But they're not exactly cheap anymore either.
00:18:47.920 --> 00:18:53.759
So with that distinction or moving away, um means they're not exactly cheap.
00:18:54.240 --> 00:18:56.799
And so for me, when I'm looking, right?
00:18:56.880 --> 00:18:59.279
If I'm a business leader, well, I am, right?
00:18:59.359 --> 00:19:06.160
But I'm not uh necessarily making enterprise decisions at Fortune 500 companies, although I do advise those type of companies.
00:19:06.319 --> 00:19:15.519
Um before, even two, three months ago, you say, Yeah, look at open source models as an alternative.
00:19:15.839 --> 00:19:19.279
Today, I don't really know why people should.
00:19:19.680 --> 00:19:24.240
And maybe that's largely because of what OpenAI has been able to do, um, right?
00:19:24.559 --> 00:19:34.319
In in terms of cost per intelligence index, uh, right, which I think is uh just as important as the overall intelligence from artificial analysis.
00:19:34.400 --> 00:19:40.079
But this essentially says, how much are you paying to get these tasks done?
00:19:40.240 --> 00:19:45.680
Uh right, because all of these benchmarks, right?
00:19:45.839 --> 00:19:51.119
Artificial analysis runs it and they say, here's how much it actually costs to get all of this work done.
00:19:51.279 --> 00:19:58.880
And what we've seen is, well, anthropic models, this is not one where you want to have a big lead, you want to have a bigger bar chart.
00:19:58.960 --> 00:19:59.839
No, that's bad.
00:20:00.000 --> 00:20:00.400
Uh, right.
00:20:00.559 --> 00:20:03.279
Anthropics models are ridiculously expensive.
00:20:03.440 --> 00:20:08.000
So, uh, as an example, it costs$2.75 per task.
00:20:08.079 --> 00:20:14.240
Um, whereas OpenAI's models um are much cheaper, up to a third cheaper.
00:20:14.400 --> 00:20:33.119
So their smaller version, GBT5.7, actually cost more um than GBD5.
00:20:34.640 --> 00:20:38.720
And the Quen costs more than the mid-tier.
00:20:39.279 --> 00:20:42.799
Um oh, sorry, I don't have this one on the benchmark here.
00:20:42.880 --> 00:20:54.240
Uh, but it right when you have a leading state-of-the-art model, um, and it's costing about the same or even less than the open source model.
00:20:54.319 --> 00:21:09.599
So it's like, okay, maybe there's uh couple hundred companies in the US that can actually go out and run this themselves without paying API prices, because if you are paying API prices, at this point, you would probably just should be using open AI.
00:21:09.839 --> 00:21:16.400
Um, because when it comes to price per task, which is what is ultimately gonna matter, they're winning, right?
00:21:16.559 --> 00:21:20.960
Or maybe you are looking at Grok in MetaMuse Spark, right?
00:21:21.119 --> 00:21:41.759
You remember on the show uh last week I said it was a really bad week for Anthropic with all of these new uh um open source models coming out, but then also with Grok and MetaMuse Spark uh 1.1 coming with some pretty good models that people weren't um expecting, especially on the coding and software engineering side.
00:21:41.920 --> 00:21:58.960
But similarly, uh artificial analysis has this uh this quadrant uh when it's cost per tax, uh cost per task, and the uh essentially intelligence or the artificial intelligence uh artificial analysis intelligence index score.
00:21:59.039 --> 00:22:10.240
So essentially you Be in the upper left hand corner, which means you have the smartest model at the cheapest cost, and none of the Chinese open source models are in that quadrant anymore.
00:22:10.400 --> 00:22:25.039
That obviously it resets as new models come out and the the the medians and the averages all change, but right now the three companies in there are OpenAI, Grok, and Meta.
00:22:25.680 --> 00:22:27.599
No open source Chinese models.
00:22:27.680 --> 00:22:49.200
Where this is actually a quadrant that they used to dominate, which is why I think six months in a year ago, it made a lot of sense for business leaders to be looking at Chinese open source models, especially when they were a little bit more tameable in terms of what you can do without having a multi-million dollar uh you know compute setup.
00:22:49.359 --> 00:22:51.519
Um, but it's just not the case anymore.
00:22:51.680 --> 00:23:11.519
So, for me, anyways, even though I know that the the war is going to rage on, if I'm making decisions, I'm looking at these charts and saying, well, there's maybe not that big of a reason for uh us to look at these models unless you do want or need that open nature, which I know some companies obviously do.
00:23:12.559 --> 00:23:28.799
But it's no longer just the cheap API prices that reveal the true cost of using these models, uh, because you have these aggressive Chinese prices that show their strategy is not just competing for money, right?
00:23:28.880 --> 00:23:40.640
Because there's other subsidies, cloud cross-selling competition, and also still these lower prices force enterprises to question the expensive closed model defaults.
00:23:41.440 --> 00:23:45.839
So I know that the models from the past week, right?
00:23:46.000 --> 00:23:58.160
Specifically Kimi K3 and Quen 3.8, they may not still fit in the traditional open source, cheap Chinese models to use paradigm, but there are still those models.
00:23:58.240 --> 00:24:02.319
They will still be uh continue to be developed, and I think that there will be a place for them.
00:24:02.559 --> 00:24:09.519
Um but it comes to distillation, we can't get to you know 20 plus minutes and not talk about distillation.
00:24:09.599 --> 00:24:23.519
So distillation is essentially where, for the most part, Chinese companies, you know, take it's just like they copy the questions and they copy the answers for a lack of a better term from the big AI model provider.
00:24:23.599 --> 00:24:31.039
So it's just like they copy the answers on the tests, they take all the hard work from the American companies and use it to train their own models, right?
00:24:31.119 --> 00:24:37.440
And that accelerates the progress, but it you still can't explain uh you know how they get there.
00:24:37.599 --> 00:24:48.640
Um, just through distillation, that's not enough because they have had some great um advancements, uh, the Chinese companies in architecture and just their overall execution.
00:24:48.960 --> 00:25:04.000
So AI model distillation isn't universally illegal, it's highly frowned upon, and it is um, you know, um sparking some backlash politically um and economically with export controls.
00:25:04.160 --> 00:25:08.240
I think we're gonna continue to see those ramped up now here for the rest of 2026.
00:25:08.400 --> 00:25:17.519
But you know, these alleged Chinese practices do violate contracts and have led to major US national security actions.
00:25:17.839 --> 00:25:21.920
So, what is the US doing to stop this model distillation?
00:25:22.079 --> 00:25:35.200
Well, they're trying to cluster traffic and you know, detect and you know, catch large-scale uh offenders and you know block them, but it's pretty hard because it's they're just playing whack-a-mole.
00:25:35.359 --> 00:25:43.119
Uh, a US House committee did, though, just unanimously back a bill to sanction foreign actors for extracting US models.
00:25:43.359 --> 00:25:48.960
But uh laws in the US are a slow-moving machine, right?
00:25:49.119 --> 00:25:57.119
So uh it could be many months or multiple quarters or even more than a year before something like that ever becomes law.
00:25:57.359 --> 00:26:02.880
But these published open weights cannot be recalled, so businesses can still adapt them, anyways, right?
00:26:03.039 --> 00:26:08.799
So when models go open source or open weight, it's kind of like the cats out of the bag at that point.
00:26:09.440 --> 00:26:28.960
So the US can try to restrict access, but that could potentially backfire because Washington can, yeah, they can pressure uh, you know, via export controls or otherwise, the chips, the cloud access, procurement, sanctions, and all these other things, but the restrictions may just protect the technology while raising the cost for American companies.
00:26:29.200 --> 00:26:47.839
Uh, like I said, if you um are restricting these models, but companies already have the weights, you may just be keeping money ultimately um from other ecosystems that could power the American um enterprise, right?
00:26:49.279 --> 00:26:54.480
So I think the writing though is kind of on the wall.
00:26:54.640 --> 00:27:09.839
Um, and I think it's actually Microsoft's the uh some of their recent actions that show I think what we might be looking at when it comes to how should large enterprises be looking at or using these Chinese models.
00:27:10.079 --> 00:27:18.000
So uh reportedly Microsoft was evaluating Deep Seek for cheaper copilot co-work model routing.
00:27:18.160 --> 00:27:21.200
So, not for all of Microsoft Copilot, right?
00:27:21.279 --> 00:27:22.880
And this is just according to reports.
00:27:22.960 --> 00:27:25.920
I don't believe Microsoft has confirmed anything yet.
00:27:26.240 --> 00:27:34.160
And they were looking at this for a backup because what they found, uh, Microsoft that the co-pilot co-work was actually very cost intensive.
00:27:34.319 --> 00:27:43.680
And they did start to bill it um, you know, uh at a token rate and no longer just including usage uh like they did when it was in beta.
00:27:43.839 --> 00:27:48.240
Um that is telling you everything that you need to know.
00:27:48.480 --> 00:28:00.000
They haven't switched to it, but Microsoft, yes, that Microsoft, the one that has huge stakes um in OpenAI and Anthropic.
00:28:00.160 --> 00:28:03.279
They are looking at a model like DeepSeq.
00:28:03.920 --> 00:28:15.519
For all the, you know, the time that DeepSeek's name has gotten dragged through the mud for, you know, some alleged shady practices, they're still looking at them to use them.
00:28:15.599 --> 00:28:23.759
And I think ultimately, uh, you have to look at your balance sheet, you have to look at the dollars and the cents and make it make sense.
00:28:24.000 --> 00:28:26.000
And in this case, Microsoft, right?
00:28:26.079 --> 00:28:37.519
And the other thing, again, if you're a large company and you can uh, you know, essentially download the weights to these models and you can fine-tune them and make sure that they perform up to a certain standard.
00:28:37.680 --> 00:28:46.640
But I think the future workloads, like in this Microsoft example, could just be routed by cost, capability, security, geography, and regulation, right?
00:28:46.799 --> 00:28:57.039
I've been a huge advocate over the years for who knows, maybe we'll get there eventually, but you know, the mixture of models, much different than a mixture of experts, right?
00:28:57.119 --> 00:29:03.920
That's using a single model, and then the model kind of calls on the different experts and the different parameters in this dense model, right?
00:29:04.079 --> 00:29:07.519
Mixture of models is similar to model routing, but it's just using, right?
00:29:07.599 --> 00:29:19.920
I think Perplexities model council, uh, Microsoft has something similar, but it's well, when you put a prompt out there and there's just a router, and it might send a simple prompt to one open source model, right?
00:29:20.000 --> 00:29:20.799
That's a good example.
00:29:20.960 --> 00:29:28.720
Or it might send a complex prompt to a hundred different models and then an orchestrator model to go in and collect all the information.
00:29:28.880 --> 00:29:33.039
And maybe half of those hundred models are open source.
00:29:33.119 --> 00:29:41.200
Uh, but I think if anything, the shift though does weaken loyalty to one model and rewards flexible architecture.
00:29:41.359 --> 00:29:45.599
I do think that is probably the future where enterprise leaders need to be focusing on.
00:29:45.759 --> 00:29:48.079
So as we wrap up, let's talk about that.
00:29:48.400 --> 00:29:52.640
What is coming next and what business leaders should be doing now?
00:29:52.799 --> 00:29:54.079
Well, I would do this.
00:29:54.240 --> 00:29:57.599
I'm gonna say expect controlled openness.
00:29:57.759 --> 00:29:58.319
All right.
00:29:58.480 --> 00:30:12.880
Whether it is China placing export on their open models, uh, which I know kind of uh goes against the uh very reason they put them out there in the first place, but that's another uh another topic for another episode.
00:30:12.960 --> 00:30:14.480
But I would expect controlled openness.
00:30:14.559 --> 00:30:38.480
So either China is going to restrict probably the US uh from using their maybe uh most frontier models, uh, and or uh the US may also uh restrict the usage of some of these Chinese open models as well, especially as now these models are getting more and more capable, uh more and more autonomous.
00:30:38.640 --> 00:30:44.240
And when it comes to cyber, things are getting a little bit scary-ish, right?
00:30:44.400 --> 00:30:50.240
Uh, I I think in six months, uh a year, that's when things are gonna be getting like real scary.
00:30:50.319 --> 00:31:05.119
Uh, in in terms of these autonomous models, as they can technically get smaller, faster, more capable, and well, more open source, that's when uh these exports are going to uh or these export controls um are going to come in.
00:31:05.279 --> 00:31:11.039
So expect both nations to probably just share older models, but maybe guard their best.
00:31:11.279 --> 00:31:22.000
Uh, but what you need to do is inventory where your current Chinese open source models are running and then route those tasks by cost, security, and policy risk.
00:31:22.079 --> 00:31:34.640
And then you need to document your sourcing and be ready to swap out a model or to uh you know go to your plan B uh when and if those government restrictions harden.
00:31:34.799 --> 00:31:37.680
All right, I hope this episode was helpful.
00:31:37.839 --> 00:31:54.960
But like I said, whether you know it or not, there's a good chance if you're using open source models that part of your plan might need to be modified pretty soon because the war between the US and China when it comes to AI is heating up.
00:31:55.039 --> 00:31:58.079
So I would expect a lot of back and forth racket.
00:31:58.319 --> 00:31:59.839
So don't get stuck in the middle.
00:32:00.000 --> 00:32:02.640
Don't waste hours every single day toiling over it.
00:32:02.880 --> 00:32:11.119
Just tune in to us here on the show and the newsletter at your everydayai.com because we're always going to keep you up to date.
00:32:11.279 --> 00:32:12.319
I hope this one was helpful.
00:32:12.480 --> 00:32:13.119
Thank you for tuning in.
00:32:13.279 --> 00:32:16.319
Hope to see you back tomorrow and every day for more everyday AI.
00:32:16.400 --> 00:32:17.279
Thanks, y'all.
