WEBVTT
00:00:01.280 --> 00:00:15.359
Anthropic finally responded as the last AI lab out in not supporting open source AI in the viral open weights and American AI leadership paperslash coalition.
00:00:15.519 --> 00:00:25.039
That's because Anthropic's CEO, Dario Mati, just published a letter defending the company's decision to not join the other players.
00:00:25.280 --> 00:00:32.799
And I think it's actually more about Anthropic's IPO than it is about open source.
00:00:32.960 --> 00:00:35.039
But here's the setup and the backdrop.
00:00:35.280 --> 00:00:42.880
Nvidia and Microsoft put out a letter four days ago defending open models, and basically everyone signed it.
00:00:43.200 --> 00:00:46.560
Meta, OpenAI, Google, IBM, everyone.
00:00:46.719 --> 00:00:51.119
And Anthropic was the only major tech player that said no.
00:00:51.759 --> 00:01:12.480
And late yesterday, Dario published Anthropic's Answer in a blog post, and it opens by insisting they've never wanted to ban open models, which is technically true in writing, but not really truly in line with their previous statements, warning about how dangerous open models are, and there should be government mandates for them.
00:01:12.719 --> 00:01:15.439
Now, here's the part that nobody's talking about.
00:01:15.599 --> 00:01:20.719
Why Anthropic is really trying to get DC lawmakers scared of Chinese open models?
00:01:20.959 --> 00:01:22.879
It's one number, 80%.
00:01:23.599 --> 00:01:34.879
Yeah, that's roughly 80% of Anthropic revenues just comes from businesses paying per token, while the other big players have a fraction of that.
00:01:35.120 --> 00:01:42.079
So free, downloadable Chinese open source models literally choke Anthropic's only real source of revenue.
00:01:42.239 --> 00:01:46.799
And Anthropic filed for an IPO targeting an October launch.
00:01:47.040 --> 00:01:54.159
Hence the reason why Anthropic zigged while every other tech company zagged when it came to supporting open source.
00:01:54.400 --> 00:02:06.560
So today we're breaking down not just what the letter says, but where it actually even contradicts Dario's own previous words and why I think this was written for Washington policymakers and not for you and me.
00:02:06.799 --> 00:02:08.080
All right, let's get into it.
00:02:08.240 --> 00:02:09.840
Here is the big picture.
00:02:10.159 --> 00:02:22.719
Anthropic is the only major AI lab that refused to defend open models in the open weights and American AI leadership, uh paperslash coalition that was put forth by Microsoft and NVIDIA.
00:02:23.039 --> 00:02:31.919
Essentially, they said, well, open source is good for the American economy, it's good for national security, and we're gonna talk a little bit more about that paper.
00:02:32.240 --> 00:02:33.520
But why?
00:02:33.840 --> 00:02:41.199
Well, Anthropic will tell you it's ultimately about safety and control, but I think it is that 80% number.
00:02:41.280 --> 00:02:55.199
And we're gonna look at, well, it might not work very well if Anthropic can't get this essentially to go in their favor, and that's because their IPO is potentially only weeks away in October.
00:02:55.919 --> 00:03:03.360
So on today's show, stick with me, and you're gonna learn why that 80% meter token revenue makes Anthropic uniquely exposed to open models.
00:03:03.439 --> 00:03:10.960
You're gonna know what the regulatory capture means when it comes to AI and why this letter just fits the pattern exactly.
00:03:11.120 --> 00:03:21.039
You're gonna know why the timing points at Washington policymakers, because there's an important date coming up in just a couple of days, and I don't think this letter was made for me and you.
00:03:21.199 --> 00:03:27.520
And you're gonna know what Dario said in 2025 that flatly contradicts today's letter.
00:03:27.759 --> 00:03:29.039
All right, let's get into it.
00:03:29.199 --> 00:03:30.240
Welcome to Everyday AI.
00:03:30.319 --> 00:03:31.840
If you're new here, my name's Jordan.
00:03:32.080 --> 00:03:33.280
We do this well every day.
00:03:33.439 --> 00:03:42.159
It's your unedited, unscripted live stream podcast and free daily newsletter helping everyday business leaders like you and me keep up with the nonstop AI avalanche.
00:03:42.319 --> 00:03:44.639
I tell you what's important, what's not.
00:03:44.800 --> 00:03:50.159
You use that decision, and all of a sudden, wow, you're the smartest person in AI at in your company.
00:03:50.400 --> 00:03:56.319
So it starts here, but please make sure to subscribe to the podcast and go to our website at your everydayai.com.
00:03:56.479 --> 00:03:57.759
All right, let's get into it.
00:03:58.000 --> 00:04:04.800
This has been all the talk on tech, Twitter, and you know, in the media, in anything AI, right?
00:04:04.879 --> 00:04:13.039
If you follow AI on a daily basis like I do, this conversation has been just grabbing every single headline.
00:04:13.120 --> 00:04:16.800
And it's extremely important and also a little bit nuanced.
00:04:16.959 --> 00:04:28.480
So today's show, I'm gonna try to go through it a little bit quicker, stick to just the facts, and my opinion, maybe I'll, you know, uh bring back an accidental hot take Tuesday for old time's sake.
00:04:28.560 --> 00:04:32.319
Uh, but make sure you go check out today's newsletter uh for more.
00:04:32.560 --> 00:04:40.639
So here's what actually uh Dario put in his published response uh on open models.
00:04:40.800 --> 00:04:57.519
So the core claim was that Anthropic has never advocated banning open weight models, and he rejected the argument that open models actually help defenders more than attackers, which was a core claim in the Microsoft slash NVIDIA uh paper that everyone kind of got behind.
00:04:57.600 --> 00:05:04.800
And Anthropic says it's better for chip controls uh to crack down on distillation and mandatory safety testing for all.
00:05:05.040 --> 00:05:10.000
And they deny that protecting Anthropics business ever motivated the position.
00:05:10.079 --> 00:05:11.199
Oh, that's a funny one.
00:05:11.360 --> 00:05:11.920
All right.
00:05:12.560 --> 00:05:14.800
So let's take a quick look uh at the letter.
00:05:14.879 --> 00:05:19.920
I'm not gonna read it, but I will kind of read uh the opening paragraph and go through some of the bullet points here.
00:05:20.079 --> 00:05:23.680
So uh this is from Dario Marty, uh, Anthropics CEO.
00:05:23.759 --> 00:05:27.839
Uh so yeah, this just came late last night, actually, or late afternoon.
00:05:28.000 --> 00:05:29.120
Uh, I'm so tired.
00:05:29.199 --> 00:05:30.240
Sometimes like 4 p.m.
00:05:30.480 --> 00:05:31.439
feels like late night.
00:05:31.600 --> 00:05:38.000
So it says over the last few days, there have been a lot of discussions about open weight models, especially those from China.
00:05:38.160 --> 00:05:40.000
Reports suggest that some U.S.
00:05:40.079 --> 00:05:48.240
officials are considering banning the use of Chinese open source models, uh, sorry, Chinese open weights models by US companies.
00:05:48.319 --> 00:05:59.519
In response, many tech companies have signed a letter supporting open weights models, and some people have even accused Anthropic of wanting to ban open source models as a means of protecting our business.
00:05:59.759 --> 00:06:12.639
Anyone who has read my past writing should know that I don't regard such bans as a useful measure, but let me state it clearly so that there is no doubt Anthropic has never advocated for a ban on open weights models.
00:06:12.800 --> 00:06:17.920
Uh, then he says open weights models that don't have dangerous capabilities are a public good.
00:06:18.079 --> 00:06:24.639
They don't cost anything besides the compute needed to run them, and they provide value to businesses, developers, and researchers.
00:06:24.800 --> 00:06:31.040
All right, so let me just take a pause, and I probably should just kind of talk very briefly about open source open weights models.
00:06:31.120 --> 00:06:35.439
So if you are brand new, maybe this is the first episode you've ever listened to, and you're like, what the heck is going on?
00:06:35.600 --> 00:06:36.480
Why does this matter?
00:06:36.639 --> 00:06:39.839
Well, a couple important points that Dario brought up here in the letter.
00:06:40.079 --> 00:06:52.480
Uh, number one, there has been this idea floated recently, and you know, we'll see uh, you know, maybe in 10, 20 years, how much of Anthropic's DC lobbying efforts maybe uh led to this, right?
00:06:52.560 --> 00:06:54.720
I'm a former journalist, I understand how this works.
00:06:54.879 --> 00:07:10.720
All the big labs have a lot of lobbyists, but uh the US government is actually considering a ban on American companies using Chinese open source models for you know what they're saying, uh competitiveness, national security concerns, etc.
00:07:11.360 --> 00:07:17.920
But essentially, open source or open weight models are models that you can well you can download and you can run them on your machine.
00:07:18.079 --> 00:07:23.759
Well, very small open source models or what open source models used to be, you know, like two years ago.
00:07:23.920 --> 00:07:28.319
Today's open weights Chinese models, you can't download those on your computer.
00:07:28.399 --> 00:07:32.879
You're literally going to need a spare uh data center to run the newest.
00:07:32.959 --> 00:07:40.720
And the ones that kind of set this off are uh Moonshots, Kimi K3, uh ZAI's GLM52.
00:07:41.120 --> 00:07:45.519
Um, and I think that eventually we'll be talking a little bit more about Quen 3.8.
00:07:45.680 --> 00:07:50.879
So essentially you have all these models that, well, if you're a huge enterprise, you can use them for free.
00:07:51.040 --> 00:07:59.360
Uh, but if you are using them um through a hosting provider, uh through an inference provider, they're way, way cheaper, right?
00:07:59.439 --> 00:08:00.079
For the most part.
00:08:00.240 --> 00:08:07.920
Not always, but usually these open source models are way cheaper uh if you can't download and host them or run them locally.
00:08:08.079 --> 00:08:08.560
Uh all right.
00:08:08.720 --> 00:08:14.720
So that's the big picture on uh what kind of Dario said in his letter, in all of that.
00:08:14.879 --> 00:08:19.279
So uh some of the more important bullet points here.
00:08:19.600 --> 00:08:41.679
Uh he said, My primary concern is the risk that authoritarian governments, not solely the Chinese uh Communist Party or CCP, although the CCP is clearly the most capable threat, build AI models that are more powerful than those built in the US and use them to achieve permanent military superiority or uh perpetrate incredibly deep repression of their own people.
00:08:41.840 --> 00:08:42.000
All right.
00:08:42.080 --> 00:08:49.200
So essentially he said the biggest concern is well, the Chinese government is going to use these for military superiority.
00:08:49.360 --> 00:08:59.919
Uh, number two, secondary concern, he says, is that powerful AI models may be misused to carry out cyber attacks or biological attacks, and those may have serious alignment issues.
00:09:00.159 --> 00:09:04.240
So, you know, I agree with both of those things, right?
00:09:04.320 --> 00:09:06.320
Um, I've been saying it all along.
00:09:06.480 --> 00:09:16.000
I think artificial intelligence, for the most part, it is uh gonna be more important maybe in five to ten years than you know what your traditional military might be looked at.
00:09:16.080 --> 00:09:28.799
It's gonna be more important than oil or gold or anything else, because uh the nation essentially that figures out artificial general intelligence and artificial superintelligence first has a legit, unfair advantage on the rest of the world.
00:09:28.879 --> 00:09:39.440
And you can go from a mid-tier global power to the uh global superpower if you control artificial general intelligence or artificial superintelligence.
00:09:39.759 --> 00:09:57.279
So, uh, kind of to address these concerns, what he does say in his letter, he says, number one, we should not sell powerful chips or chipmaking equipment to China, essentially saying, hey, we still have a lead, although it's maybe slimmer than it was before, but China can't catch us if they don't have our chips.
00:09:57.440 --> 00:10:06.559
Uh, right, previously uh they said uh Anthropic said that by essentially creating a chip ban, uh the US could get a 12 to 24 month uh lead in AI.
00:10:06.720 --> 00:10:08.480
I don't necessarily think that's true.
00:10:08.799 --> 00:10:12.240
The lead could extend from where it is now, which is like two months.
00:10:12.480 --> 00:10:18.799
It used to be about six months uh between the um US frontier models and the Chinese open source models.
00:10:18.879 --> 00:10:20.080
So maybe it could extend it.
00:10:20.159 --> 00:10:22.399
I don't necessarily think it would be 12 to 24 months.
00:10:22.559 --> 00:10:24.720
That's a little bit of lofty thinking, I would say.
00:10:24.879 --> 00:10:31.679
And then uh Dario also says we should crack down on industrial scale distillation uh distillation operations, right?
00:10:31.759 --> 00:10:34.799
So essentially how China is catching up, right?
00:10:34.879 --> 00:10:43.679
They do have some uh novel architecture and uh kind of ML um uh approaches, but for the most part, it's distillation, right?
00:10:43.759 --> 00:10:51.440
They're just copying uh, you know, millions of inputs and outputs from you know clawed models and other models to make their models smart.
00:10:51.600 --> 00:10:54.720
So they're just copying someone else's answers at a fraction of the cost.
00:10:54.879 --> 00:11:02.159
Uh and then uh Dario says all sufficiently capable models, open and closed, should go through mandatory safety testing.
00:11:02.320 --> 00:11:05.039
All right, there's more on the um the letter.
00:11:05.120 --> 00:11:05.919
It's not super long.
00:11:06.000 --> 00:11:07.039
I'm not gonna read it all.
00:11:07.120 --> 00:11:10.720
Uh, we'll make sure to put it in today's newsletter uh as well.
00:11:11.039 --> 00:11:25.200
So this is all essentially in response to the original uh kind of movement that uh NVIDIA and Microsoft put forth uh on July 24th, which was this past Friday.
00:11:25.360 --> 00:11:28.879
And this was called the open weights in American AI leadership.
00:11:29.039 --> 00:11:29.519
All right.
00:11:29.679 --> 00:11:36.720
So um essentially they circulated this, arguing that open models cut costs and avoid vendor locked in.
00:11:36.879 --> 00:11:47.919
So they said it's good for American businesses, it's good for uh competitiveness, and then they also said it's better for safety because they said that the core safety claim of having open models is that defenders, right?
00:11:48.000 --> 00:11:51.840
Because this whole thing comes down to cyber attacks versus defending your business, right?
00:11:52.000 --> 00:11:58.799
Uh, one of the big arguments is oh my gosh, if you have these, you know, these open source models everywhere, you can't keep track of them, right?
00:11:58.879 --> 00:12:17.840
And eventually, and I agree, probably in two years, I've said this, you'll be able to, you know, have a version or an open source model that is of the GPD56 soul fable five um tier of models, you'll be able to have something that powerful on a you know consumer-ish desktop device.
00:12:18.000 --> 00:12:24.480
So at that point, you know, yeah, there is going to be, you know, critical infrastructure at risk from cyber attacks.
00:12:24.639 --> 00:12:34.960
You know, think of you know, phishing scams and you know, people holding you know web systems for hostage uh times a thousand and compact it into a very tight time frame, right?
00:12:35.039 --> 00:12:36.559
Once that does happen, right?
00:12:36.799 --> 00:12:41.679
So if you're thinking of hospitals, school systems, power grids, right?
00:12:41.919 --> 00:12:47.600
Um, entire small nations, you know, city and state governments, right?
00:12:47.759 --> 00:12:50.159
All these places are going to be extremely at risk.
00:12:50.240 --> 00:12:50.879
And I agree with that.
00:12:50.960 --> 00:13:01.919
But essentially, Microsoft and NVIDIA are saying, well, the best way to defend against this is well, open models because you have more people contributing, people are going to find these vulnerabilities, and it collectively makes you know everyone stronger.
00:13:02.000 --> 00:13:05.039
So it gives everyone better defensive tools, all right.
00:13:05.120 --> 00:13:14.559
Um, and essentially everyone literally agreed with that from Amazon, OpenAI, Meta, Google, IBM, everyone except Anthropic.
00:13:15.360 --> 00:13:19.039
So my take on this is Dario's letter is not for us.
00:13:19.279 --> 00:13:22.240
This is literally for regulatory capture.
00:13:22.480 --> 00:13:26.960
This is to uh influence the policy making in uh DC.
00:13:27.039 --> 00:13:32.159
So there's been a lot of articles about, you know, just how big of an impact that Anthropic has in lobbying.
00:13:32.320 --> 00:13:38.159
And uh yeah, I think that this has to be the real reason here.
00:13:38.240 --> 00:13:41.279
Um, but I mean, my take on the actual letter, it's well written.
00:13:41.360 --> 00:13:46.799
There's great arguments in there, and it gener it genuinely engages with real risk, right?
00:13:47.200 --> 00:13:52.879
There's nothing in there uh that I'm looking at and I'm like, this is absolutely blasphemy, right?
00:13:52.960 --> 00:14:02.960
Like, aside from the fact that they say that they absolutely never, you know, called for a ban on open models, they have just without using the word ban, right?
00:14:03.120 --> 00:14:03.600
I don't know.
00:14:03.679 --> 00:14:13.200
If you say something, uh, you know, walks like a duck, quacks like a duck, is yellow like a duck, you know, is is in the jersey on mighty ducks, but you're like, well, it's not a duck.
00:14:13.360 --> 00:14:16.000
Well, yeah, that's essentially what you said.
00:14:16.080 --> 00:14:25.679
You just because you didn't say that animal right there is a duck, but you you described every single duck characteristic and used the word duck in describing the characteristics, but then you're like, Well, it's not a duck.
00:14:25.840 --> 00:14:27.360
Well, yeah, it is.
00:14:27.600 --> 00:14:28.000
All right.
00:14:28.159 --> 00:14:34.879
And but the important thing here is Dario's letter comes just days before an important federal deadline.
00:14:35.039 --> 00:14:38.159
And that is why the regulatory piece is very important.
00:14:38.399 --> 00:14:46.159
So President Trump signed essentially this uh AI uh order on June 2nd with a 60-day deadline.
00:14:46.240 --> 00:14:46.720
And guess what?
00:14:46.799 --> 00:14:49.600
That 60-day deadline is well, this week.
00:14:49.679 --> 00:14:54.960
Uh so the framework right now does not cover open weight models, right?
00:14:55.039 --> 00:15:04.639
So that's why there's all this kind of uh jockeying right now, and uh, you know, why you've had these stories come out in the past uh seven to ten days, right?
00:15:04.799 --> 00:15:13.600
Saying that the US is now maybe considering uh kind of throwing open source models, uh not in this executive order, but potentially just banning them.
00:15:13.759 --> 00:15:23.679
Because out of nowhere, uh right, you had you know, Moonshots, Kimi K3, I think especially come up and you know, essentially it is in that top tier.
00:15:23.759 --> 00:15:27.360
Whereas before the open source Chinese models were never in that tier.
00:15:27.440 --> 00:15:32.240
So that obviously takes away from um American business, number one, right?
00:15:32.399 --> 00:15:39.519
Because whether you know it or not, and I've been saying this all along, AI and large language models, right now they run the American economy.
00:15:39.679 --> 00:15:47.840
So when you have a Chinese open uh weights uh contender come in, and well, potentially you don't want to take all that spend, right?
00:15:47.919 --> 00:15:58.799
From uh uh you know, putting America first sort of thing, which is the uh kind of the Trump's uh uh the Trump people's uh position on this, right?
00:15:58.960 --> 00:16:04.639
Well, that takes away probably a lot of money that American businesses would be spending with other American businesses.
00:16:04.799 --> 00:16:21.039
So not only that, but you know, kind of the other uh uh chain of thought, haha uh AI joke accidentally made, is that you know, the more powerful these models become, well, they're just gonna become better at distilling versions of themselves from the other frontier labs.
00:16:21.200 --> 00:16:26.080
And you know, it seems like Anthropic, maybe more so than others, seemingly is struggling.
00:16:26.240 --> 00:16:30.720
Maybe they're just the ones complaining a little bit more because they technically have more to lose.
00:16:31.039 --> 00:16:38.480
But the Dario letter lands just five days uh before this uh kind of order gets locked in.
00:16:38.559 --> 00:16:49.759
So the order is voluntary and there's no licensing and no pre-clearance requirements, but essentially you're supposed to clear uh models with the government 30 days before release to you know give everyone a fair amount of time.
00:16:50.000 --> 00:16:55.519
So uh yeah, there's that pretty big thing that no one is talking about.
00:16:55.679 --> 00:16:58.000
But the real reason is the money, y'all.
00:16:58.320 --> 00:17:02.159
So these are according to public estimates, which are pretty close.
00:17:02.240 --> 00:17:12.720
A lot of it is from semi-analysis, which is kind of the uh the leader in the space in terms of uh looking at compute and its impact on the economy, the money in and the money out.
00:17:12.880 --> 00:17:21.519
So essentially, um, the best estimates are 80% of anthropics revenue comes from just selling tokens, right?
00:17:21.680 --> 00:17:26.240
Whereas its closest competitor, OpenAI, it's estimated that it's only 20%.
00:17:26.640 --> 00:17:31.119
Uh and then, you know, your other big tech companies, it's very small considered, right?
00:17:31.200 --> 00:17:41.599
Because they have you know dozens or hundreds of lines of revenue, like you know, Google, it's only about two percent, Microsoft two percent, Amazon one percent, and meta not even one percent yet.
00:17:41.759 --> 00:17:48.720
So essentially, Anthropic's entire business model is just selling tokens, right?
00:17:49.039 --> 00:17:56.960
And well, what if now all of a sudden, if you just stop buying those tokens, if your company stops buying them, right?
00:17:57.039 --> 00:18:13.519
If you're like, hey, we don't need right, we can get 95% of our AI use with a model like you know, GLM 5.2, and that one is a little bit smaller, and well, you know, we can invest, you know, uh six figures and getting this set up.
00:18:13.680 --> 00:18:17.839
We were spending six figures or seven figures in AI each month, right?
00:18:18.000 --> 00:18:20.079
Now all of a sudden we cut our cost by 95%.
00:18:20.480 --> 00:18:23.200
We have more control, we're running it locally, uh, etc.
00:18:23.440 --> 00:18:27.039
So you can see how this is in extreme.
00:18:27.440 --> 00:18:34.400
I am talking like potentially lights out uh for uh Anthropic's growth, right?
00:18:34.480 --> 00:18:40.319
Not for them as a company, but for their staggering growth that we've been reading about through 2024 and 2025.
00:18:40.400 --> 00:18:44.400
And you know, they've been a the main beneficiary of this token maxing.
00:18:44.480 --> 00:18:49.039
Uh, but now that token maxing is going away and it's uh going toward token efficiency.
00:18:49.200 --> 00:18:59.039
So not only, right, if you look at open router, uh, which kind of is a third-party tracker of you know who's spending uh tokens where, you know, it used to be anthropic.
00:18:59.119 --> 00:19:00.400
Anthropic used to dominate.
00:19:00.559 --> 00:19:02.400
Now uh I checked uh yesterday.
00:19:02.480 --> 00:19:03.920
I think they were like number seven.
00:19:04.000 --> 00:19:08.240
They're like the seventh most used lab in open routers, and they used to dominate, right?
00:19:08.319 --> 00:19:10.319
So people are coming to a couple of conclusions.
00:19:10.400 --> 00:19:15.920
Number one, uh yes, Anthropic still has probably the best models in the world, right?
00:19:16.079 --> 00:19:19.200
The combination of Opus V that just came out and Fable Five.
00:19:19.279 --> 00:19:24.400
Although I do think GPT-5-6 soul uh is is right there and on many important benchmarks.
00:19:24.480 --> 00:19:26.799
I do think uh GPT 5-6 soul is still better.
00:19:26.880 --> 00:19:51.039
But regardless, Anthropic has some of the best models in the world, but 80% of their revenue is just selling tokens, and now all of a sudden, over the past 10 days, you have Kimi K3, and you we're gonna be talking more about Quin38 and probably whenever GLM53 comes out, all of a sudden, yeah, Anthropic's you know that that hockey stick growth is not gonna look very hockey stick for much longer.
00:19:52.480 --> 00:20:02.720
So Anthropic built this massive business, and if I'm being honest, there's not a big moat, right?
00:20:02.960 --> 00:20:07.759
Google, open AI, and Microsoft are taking a very different approach, right?
00:20:07.839 --> 00:20:10.000
They're looking at the entire ecosystem.
00:20:10.240 --> 00:20:22.559
But right now, Anthropic's entire business is inference, and inference is just a commodity that's getting cheaper, mainly because of open source models and the technology itself.
00:20:22.720 --> 00:20:27.920
But some open models produce output roughly 10 times cheaper than Opus.
00:20:28.160 --> 00:20:37.039
So even if you're paying for it via the API, right, in some models and some certain tasks, it's just the same price, but in many instances, it is incredibly cheaper.
00:20:37.200 --> 00:20:39.519
And switching is easy, right?
00:20:39.680 --> 00:20:45.359
That's the difference between the ecosystem play versus the API charging you for tokens play, right?
00:20:45.440 --> 00:20:53.440
If a company, and you've seen plenty of stories, literally, people detail this at big multi-billion dollar companies where they're like, wait, we just cut our spend.
00:20:53.519 --> 00:20:56.559
You know, we were spending, you know, tens of millions of dollars a year.
00:20:56.720 --> 00:21:04.799
We just cut our spend by you know 80% just by now using a model router and you know, 80% of our queries, we can use open source models, right?
00:21:04.880 --> 00:21:11.200
Maybe they're still paying a little bit to you know, anthropic or open AI or Google or whatever for the actual just raw tokens.
00:21:11.359 --> 00:21:23.119
But switching is so easy to literally go in there and switch a couple endpoints, or you know, if you're using um, you know, a certain gateway, uh, you know, to go in there and just choose a new model and the drop down.
00:21:23.279 --> 00:21:31.119
And if you've been doing your due diligence and you know, doing your running your backups like I've been telling you to do for years, it's actually not that big of a disruption for your business.
00:21:31.279 --> 00:21:33.440
So switching is easier than ever.
00:21:33.759 --> 00:21:39.440
If you're just paying for APIs, switching out of uh an ecosystem, not easy, right?
00:21:39.519 --> 00:21:45.200
That's like I've always wished I could just, you know, switch out of using an iPhone or switch out of using a Mac.
00:21:45.440 --> 00:21:45.920
I can't.
00:21:46.079 --> 00:21:48.000
I am locked into that ecosystem.
00:21:48.160 --> 00:21:52.000
Even if the hardware isn't the best, even if the camera isn't the best, whatever.
00:21:52.079 --> 00:21:52.640
It doesn't matter.
00:21:52.720 --> 00:21:54.559
I can't get out of the Apple ecosystem.
00:21:54.640 --> 00:21:55.920
I am too locked in, right?
00:21:56.079 --> 00:21:58.160
That's the game that OpenAI has been playing.
00:21:58.240 --> 00:22:05.359
That's the game that Microsoft and Google has been playing and Infraopic has just been saying, Well, we're just going to sell tokens.
00:22:05.519 --> 00:22:12.160
Well, they're seeing and they're finding out that that may not work unless they can get some regulatory help.
00:22:12.400 --> 00:22:14.799
And that could just become their moat.
00:22:15.200 --> 00:22:18.000
So that is the regulatory capture here.
00:22:18.079 --> 00:22:22.000
So that's just winning in Washington instead of on an actual product.
00:22:22.160 --> 00:22:28.799
So essentially the incumbents, right, are backing these safety rules that only companies their size can afford.
00:22:28.880 --> 00:22:32.160
And the rules become the moat that the product just couldn't build, right?
00:22:32.240 --> 00:22:34.079
And this isn't just me saying this.
00:22:34.240 --> 00:22:36.960
Um, yeah, there's actually some very prominent names.
00:22:37.039 --> 00:22:40.160
Uh, you think I'm tough on anthropic sometimes.
00:22:40.319 --> 00:22:41.519
I'm definitely not.
00:22:41.680 --> 00:22:48.720
I actually think that this letter is pretty good, and there's very few things in there that I would argue against, uh, right.
00:22:48.799 --> 00:22:54.640
But White House AI czar David Sachs called it fear-based regulatory capture.
00:22:54.799 --> 00:22:59.359
I mean, he tore into this, and uh, interestingly enough, he actually called it, right?
00:22:59.519 --> 00:23:03.599
He essentially said, Yeah, Anthropic is gonna run uh a psyops on us.
00:23:03.680 --> 00:23:13.359
You know, he literally called this before uh Dario wrote this letter, and he said, you know, Anthropic's uh gonna claim that they never wanted to ban open models, although they did want to.
00:23:13.519 --> 00:23:16.799
And sure enough, that's literally what they did, right?
00:23:16.960 --> 00:23:24.160
Um, and also nearly 200 startups warned that banning open source models could hand Anthropic a monopoly.
00:23:24.559 --> 00:23:29.440
So let's get back to the timing because literally, this is what this is about.
00:23:29.680 --> 00:23:33.119
This is what the I don't know how many times you want me to tell this, right?
00:23:33.200 --> 00:23:40.720
This is the difference between a company, you know, being worth, you know, 500 billion and being worth three trillion dollars, right?
00:23:40.880 --> 00:23:44.559
In theory, anthropic could go either way in a year.
00:23:44.720 --> 00:23:50.480
Um, you know, and I actually wasn't even gonna do this show, uh, but I put out a tweet, yeah.
00:23:50.559 --> 00:23:52.240
I did, yeah, cringe, right?
00:23:52.400 --> 00:23:53.920
Uh, a couple of weeks ago.
00:23:54.160 --> 00:23:59.920
And it was actually a very high up person from Google that liked the tweet, right?
00:24:00.079 --> 00:24:10.480
And my tweet, I forgot exactly what it was, but it was something along the lines of like Anthropic is in deep trouble with these open models and with the shift from you know token maxing to token efficiency.
00:24:10.640 --> 00:24:19.359
And their hope is they want to go public as quickly as humanly possible before Wall Street and enterprises figure out what the heck is going on.
00:24:19.519 --> 00:24:25.279
And I was actually surprised that this very high-up ranking person uh at Google liked the tweet.
00:24:25.359 --> 00:24:30.880
And the thing is, they're not really a I mean, they are a direct competitor with anthropic, but they're also an investor.
00:24:31.039 --> 00:24:35.680
Um, and you know, likes are private, so don't think you can go in there and I'm not gonna say who this person was.
00:24:35.839 --> 00:24:48.720
But when I said, wait, this person is uh essentially agreeing with what I'm saying here, and that's why I'm like, no, I am a hundred percent um confident that this is what is happening.
00:24:48.799 --> 00:24:50.640
This isn't just some like some hunch, right?
00:24:50.720 --> 00:24:54.960
And I've talked to plenty of people off the record about this at the big labs.
00:24:55.039 --> 00:24:57.200
Um, and they've all pretty much agreed.
00:24:57.279 --> 00:25:01.920
So Enthropic, here's the timing Enthropic filed their confidential S1 June 1st.
00:25:02.000 --> 00:25:08.079
And, you know, according to reports, uh, they're looking at maybe listing as a public company as soon as October.
00:25:08.240 --> 00:25:15.359
So public investors will pay for that durable pricing power, which obviously the open models straight up erase.
00:25:15.759 --> 00:25:23.039
And you know, these restrictions that are landing here soon, uh, right, would remove the cheapest substitutes.
00:25:23.119 --> 00:25:28.000
So the safety concern can be sincere, but also can be commercially self-serving all at once.
00:25:28.160 --> 00:25:28.960
I think that's what's happening.
00:25:29.039 --> 00:25:43.680
That's why I'm like, you read the letter, yeah, most people would agree with most parts of it, but it can, you know, be a you know, a good letter that you can kind of agree with many parts of it, but also be incredibly self-serving at the same time.
00:25:44.480 --> 00:25:45.119
All right.
00:25:46.000 --> 00:25:50.720
So anthropics numbers, you gotta take a look at them.
00:25:50.880 --> 00:25:53.759
So obviously they're not public, but there is a couple of websites.
00:25:53.920 --> 00:26:00.240
One is called Notice, um, that essentially tracks, and this is, you know, all you know, in algorithm estimates.
00:26:00.319 --> 00:26:12.079
It's not, you know, actually secondary market prices, but you know, it essentially tracks, you know, the OpenAI, the Anthropics, all these companies that have yet to go public, it tracks what their stocks would roughly be.
00:26:12.400 --> 00:26:20.319
And Anthropic has had their worst seven-week span ever in a seven-week span that included a model release, right?
00:26:20.400 --> 00:26:26.559
Because there was one drought uh, you know, where their stock went down a little bit, um, but they just didn't have any releases.
00:26:26.720 --> 00:26:34.640
So essentially, in the history of the company, this is their worst period when it comes to their, you know, pre-IPO stock price.
00:26:34.880 --> 00:26:38.480
Um, and they had great models that ship during this time window, right?
00:26:38.640 --> 00:26:41.359
Fable five, sonnet five, opus five.
00:26:41.440 --> 00:26:48.880
Uh, so it, you know, not a huge drop, but it fell, you know, about 8% over seven weeks, their worst record on stretch.
00:26:49.200 --> 00:26:49.680
Why?
00:26:49.920 --> 00:26:50.240
Right?
00:26:50.480 --> 00:26:55.839
Makes no sense because they've been hockey sticking uh essentially since, you know, 2024.
00:26:56.319 --> 00:27:04.720
Well, I think early investors and people who are very closely attached to AI are starting to understand what's going on here.
00:27:04.880 --> 00:27:15.599
It is the combination of the reckoning going from token maxing to token efficiency to now, all of a sudden, people are starting to care more about the cost per tax, uh, cost per task, right?
00:27:15.680 --> 00:27:23.119
So they're not just solely looking at a benchmark and then you know, fully going full send token maxing on you know the best model.
00:27:23.279 --> 00:27:29.039
They're instead saying, what's the best model that gets the job done at a good cost, right?
00:27:29.200 --> 00:27:32.880
The best example is you know OpenAI's GPT 5.6.
00:27:33.519 --> 00:27:44.720
Their models might have one or two uh points lower on a benchmark like artificial analysis, but at sometimes at a third of the cost, right?
00:27:45.119 --> 00:27:50.559
So, yeah, there's just some bad things that are happening when it comes to token economics with anthropic.
00:27:50.640 --> 00:28:04.160
And they are literally hoping to go public ASAP as soon as humanly possible, uh, before essentially the rest of the uh it insiders wake up to see what's go actually going on.
00:28:04.480 --> 00:28:29.599
So let's inspect a little bit more about some of the reasons I think what Dario said don't exactly add up because he actually wrote something six weeks ago um that kind of sound like a ban to me, yet in today's uh sorry, in yesterday's uh you know letter um saying why they didn't join this coalition with Microsoft and NVIDIA and everyone else.
00:28:29.759 --> 00:28:35.279
Well, they're like, Well, we're not banning open source models, we just can't fully support them in the way that it's written.
00:28:36.000 --> 00:28:36.319
All right.
00:28:36.400 --> 00:28:42.000
So Dario had an essay called The Policy on the AI Exponential published on his personal website.
00:28:42.160 --> 00:28:48.960
So he drops transparency as sufficient and calls for binding regulatory, uh, sorry, regulation instead.
00:28:49.039 --> 00:28:55.359
And he wants that mandatory third-party testing, then government power to block a release, right?
00:28:55.680 --> 00:29:01.680
Um, in his words, that releases should be blocked or reversed if models fail.
00:29:01.839 --> 00:29:02.160
All right.
00:29:02.240 --> 00:29:06.880
So essentially, he thinks all models should have third-party testing.
00:29:07.039 --> 00:29:13.519
And well, um, if they fail after release, they need to be blocked or reversed.
00:29:13.680 --> 00:29:19.200
So Anthropic published actually draft legislation with it and pledged money behind it.
00:29:19.440 --> 00:29:21.279
So, what does that mean?
00:29:21.519 --> 00:29:27.599
Well, essentially, this was calling for a ban of open source models without using the word ban.
00:29:27.920 --> 00:29:28.640
Let me explain.
00:29:28.799 --> 00:29:40.880
Because if you add one and one here, the only answer is two, because anthropic or any proprietary company can, you know, go along and play by anthropic's rules that they wish existed, right?
00:29:40.960 --> 00:29:44.799
You put out a model, something goes wrong, it's not safe, you pull it away, right?
00:29:44.960 --> 00:29:47.839
Same thing that happened with their Fable V Mythos V.
00:29:48.000 --> 00:29:49.440
It was out for like three days.
00:29:49.680 --> 00:29:52.480
The government said nope, and then they pulled it, right?
00:29:52.640 --> 00:29:55.519
So, in theory, Anthropic was playing by its own rules.
00:29:55.759 --> 00:29:56.720
Guess what?
00:29:57.200 --> 00:29:59.519
Open models literally can't.
00:29:59.839 --> 00:30:06.559
So I don't know how Anthropic is saying, well, oh, we never said uh we would ban open source models.
00:30:06.720 --> 00:30:10.799
We would just create rules that would make them literally impossible to exist.
00:30:11.039 --> 00:30:12.640
So we didn't use the word ban, right?
00:30:12.720 --> 00:30:13.839
We didn't call the deduct.
00:30:14.160 --> 00:30:20.880
Uh, because nobody can undo an open release once thousands have downloaded an open model, right?
00:30:20.960 --> 00:30:38.400
So they're um, and you know, I'm not gonna go through state by state, but essentially Anthropic has pledged support uh to certain states where it would be advantageous for them to uh have this kind of setup in there um that would essentially make open source models not really a truly viable thing.
00:30:38.559 --> 00:30:52.960
So if his rule is passable for anthropic, uh well, it or open, you know, open AI, any um, you know, any closed frontier model, you can't download it, you can't save it, just like Anthropic did, they can pull it, right?
00:30:53.119 --> 00:30:56.640
If Google needs to pull something, if open AI needs to pull something, they can't.
00:30:56.720 --> 00:30:57.200
It's closed.
00:30:57.279 --> 00:31:00.720
They pull the API, they pull the subscription, the model's no longer available.
00:31:00.880 --> 00:31:02.160
It's not the case with open source.
00:31:02.240 --> 00:31:05.200
Once you open it up, you can't trace it, right?
00:31:05.359 --> 00:31:07.839
People can you know lower the guardrails.
00:31:08.000 --> 00:31:12.799
Yeah, it is technically, in theory, a little more dangerous, right?
00:31:13.759 --> 00:31:18.079
But you can't say both, right?
00:31:18.160 --> 00:31:23.920
You can't put a letter out yesterday saying, like, oh, we never said we would ban open source models.
00:31:24.160 --> 00:31:38.400
We just six weeks ago put out uh our own example of policy that would just make it so open models could never actually exist by following rules that technically only a proprietary model could follow.
00:31:38.799 --> 00:31:41.039
So that's not the only thing.
00:31:41.200 --> 00:31:44.640
The other thing is talking about the control argument.
00:31:44.880 --> 00:31:49.359
So Dario's case requires closed models to be monitorable and containable.
00:31:49.519 --> 00:31:51.359
Well, what happened with Mythos, y'all?
00:31:51.519 --> 00:31:56.720
Because Mythos escaped its own sandbox and emailed a researcher unprompted.
00:31:56.880 --> 00:32:04.400
And not only that, unauthorized outsiders access mythos the same day it was announced via a Discord server just by guessing the naming mechanism.
00:32:04.559 --> 00:32:06.000
And also, oh wait, what's that?
00:32:06.160 --> 00:32:11.839
Chinese hackers ran 80 to 90 percent reportedly of an espionage campaign using Claude Code.
00:32:12.000 --> 00:32:17.039
So the bad stuff that Dario is warning about, we can never have open models, right?
00:32:17.200 --> 00:32:21.279
The the the way that everyone else wants them, because all this bad stuff will happen.
00:32:21.359 --> 00:32:22.000
Well, guess what?
00:32:22.079 --> 00:32:25.359
It's already happening with Anthropic's own models, right?
00:32:25.519 --> 00:32:28.480
That's not my opinion, that's just the facts.
00:32:29.519 --> 00:32:32.400
And we have to talk about the Hugging Face incident, right?
00:32:32.559 --> 00:32:33.599
So that happened recently.
00:32:33.680 --> 00:32:36.960
It's been all the talk, and it's actually led to the snowballing of all these other things.
00:32:37.039 --> 00:32:40.240
But OpenAI's closed model broke containment, right?
00:32:40.319 --> 00:32:46.480
The agent got out of its sandbox and attacked Hugging Face uh to work on a benchmark, right?
00:32:46.640 --> 00:32:52.319
It was trying to solve uh a benchmark problem by kind of hacking Hugging Face and getting the answers.
00:32:52.480 --> 00:32:56.640
So the closed APIs that Hugging Face used, right?
00:32:57.680 --> 00:33:02.079
They were trying to analyze the attack and they used closed models and it didn't work.
00:33:02.240 --> 00:33:03.599
They got blocked, right?
00:33:03.759 --> 00:33:08.720
And instead, they used an open source model, uh GLM 5.2, and it worked.
00:33:08.799 --> 00:33:11.519
And that's how they were able to close down the attack.
00:33:11.680 --> 00:33:16.319
So the closed source models, well, they weren't good enough as a defender.
00:33:16.480 --> 00:33:23.839
The open source Chinese models were because they couldn't use the closed source proprietary models because the guardrails are a little stricter.
00:33:23.920 --> 00:33:29.519
So Dario's own argument that he puts out there is moot, right?
00:33:29.759 --> 00:33:35.519
His his letter doubts says that open models can't actually help defenders, right?
00:33:35.680 --> 00:33:39.359
Uh he says that's what proprietary models are for.
00:33:39.599 --> 00:33:40.640
Not true, right?
00:33:40.720 --> 00:33:45.839
If AI is only for thee and not for me, it's not the case, right?
00:33:46.000 --> 00:34:00.720
Uh, you know, if the the the mythos, as an example, you know, I think there's like 40 original partners uh in the Glasswing uh program that have access to the mythos that you can truly use the best model in the world for cyber defense.
00:34:00.880 --> 00:34:05.359
Well, for 99.9% of everyone else, they can't, right?
00:34:05.599 --> 00:34:15.039
So you can't have this two-tiered system uh of well, only a select few can actually use the best proprietary models for defense, but anyone can use an open model.
00:34:15.199 --> 00:34:18.639
So, you know, his point there is absolutely false, right?
00:34:18.800 --> 00:34:21.920
Demonstratably false within the last week, right?
00:34:22.000 --> 00:34:34.960
So maybe they should have updated, maybe they wrote that letter before the whole, you know, open AI agent hugging face GLM 5.2 fiasco, uh, because it literally is doesn't make sense on paper because it's false.
00:34:35.280 --> 00:34:43.280
So let's wrap up here because I think here the big debate is safety control and business protection.
00:34:43.440 --> 00:34:45.199
When it comes to open source, right?
00:34:45.599 --> 00:34:48.960
I'm not gonna dive too deep on that because there's pros and there's cons.
00:34:49.119 --> 00:34:54.880
Like I said, there's things in the uh in the NVIDIA Microsoft uh letter that I agree with 100%.
00:34:55.199 --> 00:34:57.840
There are things in Dario's letter that I agree with 100%.
00:34:58.239 --> 00:35:08.559
A lot of it is gray area, a lot of it's nuanced, and a lot of it is the unknown because who knows what models in six months, proprietary or open source, will be capable of.
00:35:08.719 --> 00:35:09.280
Who knows?
00:35:09.440 --> 00:35:15.280
Maybe at some point open source models will be able to pass proprietary models.
00:35:15.440 --> 00:35:18.320
I mean, it's not outside of the realm of possibilities.
00:35:18.559 --> 00:35:20.719
So we can't really make that comparison.
00:35:20.800 --> 00:35:25.920
But what we have to look at is today and how this impacts your business and the decisions you make.
00:35:26.079 --> 00:35:31.679
Because I think Dario's safety concerns, well, they're real, and I think that they were thoughtfully presented in this letter.
00:35:31.840 --> 00:35:35.039
But the letter, it's actually not what you think it is.
00:35:35.199 --> 00:35:37.440
It is literally about their IPO.
00:35:37.599 --> 00:35:45.199
It is to scare uh people in DC enough to get some restrictions on these Chinese open source models.
00:35:45.360 --> 00:35:53.440
They're gonna say, you know, put out the the military reasons, the American competitiveness, all these things, which in theory are kind of true.
00:35:53.519 --> 00:36:09.280
But then I think you have to look at the uh the every other company's position that signed on to the uh NVIDIA and Microsoft letter that says, well, no, everyone needs open source because the more people, the more companies, the more AI labs they get behind open source, the more powerful it becomes.
00:36:09.440 --> 00:36:11.039
It's collective, it's a snowball.
00:36:11.119 --> 00:36:14.960
And when you put all the snowballs together and roll them down the hill, well, they're gonna get bigger.
00:36:15.199 --> 00:36:17.199
So this is not for us.
00:36:17.599 --> 00:36:19.679
This is for Washington.
00:36:19.840 --> 00:36:29.679
This is there's a shot clock, both the shot clock uh for the Trump AI rules that go into effect here in a week, but also Anthropic's IPO.
00:36:29.760 --> 00:36:37.039
Because the conclusion here, I think, is yes, this is a sincere safety argument, but it is unmistakable regulatory capture.
00:36:37.119 --> 00:36:37.920
That's all it is.
00:36:38.079 --> 00:36:48.079
And my hot take is well, Inthropic over the last couple of months has realized, I think, that their monthly growth is probably not up.
00:36:48.559 --> 00:36:57.199
Hockey sticking at the same rate that it was before, and they've realized that their moat is not really as strong as they thought it was.
00:36:57.360 --> 00:37:11.760
Um, so they're actually hoping for regulatory help to be their actual moat because they realized wait, with open source, we can't just sell tokens when other companies are giving it away for next to free, and it's just as good.
00:37:12.000 --> 00:37:22.719
All right, that's a wrap on Anthropic's response in why Claude's CEO didn't sign the open model pact and the real reasons, I think, why.
00:37:22.880 --> 00:37:25.039
So, if this was helpful, let me know about it.
00:37:25.119 --> 00:37:27.119
Please subscribe to the podcast if you haven't already.
00:37:27.199 --> 00:37:29.360
Then go to your everydayai.com.
00:37:29.440 --> 00:37:34.639
We're gonna be recapping the highlights from today's show as well as everything else you need to know to stay up to date.
00:37:34.800 --> 00:37:35.679
Thanks for tuning in.
00:37:35.840 --> 00:37:38.800
Hope to see you back tomorrow and every day for more everyday AI.
00:37:38.880 --> 00:37:39.760
Thanks, y'all.
00:00:01.280 --> 00:00:15.359
Anthropic finally responded as the last AI lab out in not supporting open source AI in the viral open weights and American AI leadership paperslash coalition.
00:00:15.519 --> 00:00:25.039
That's because Anthropic's CEO, Dario Mati, just published a letter defending the company's decision to not join the other players.
00:00:25.280 --> 00:00:32.799
And I think it's actually more about Anthropic's IPO than it is about open source.
00:00:32.960 --> 00:00:35.039
But here's the setup and the backdrop.
00:00:35.280 --> 00:00:42.880
Nvidia and Microsoft put out a letter four days ago defending open models, and basically everyone signed it.
00:00:43.200 --> 00:00:46.560
Meta, OpenAI, Google, IBM, everyone.
00:00:46.719 --> 00:00:51.119
And Anthropic was the only major tech player that said no.
00:00:51.759 --> 00:01:12.480
And late yesterday, Dario published Anthropic's Answer in a blog post, and it opens by insisting they've never wanted to ban open models, which is technically true in writing, but not really truly in line with their previous statements, warning about how dangerous open models are, and there should be government mandates for them.
00:01:12.719 --> 00:01:15.439
Now, here's the part that nobody's talking about.
00:01:15.599 --> 00:01:20.719
Why Anthropic is really trying to get DC lawmakers scared of Chinese open models?
00:01:20.959 --> 00:01:22.879
It's one number, 80%.
00:01:23.599 --> 00:01:34.879
Yeah, that's roughly 80% of Anthropic revenues just comes from businesses paying per token, while the other big players have a fraction of that.
00:01:35.120 --> 00:01:42.079
So free, downloadable Chinese open source models literally choke Anthropic's only real source of revenue.
00:01:42.239 --> 00:01:46.799
And Anthropic filed for an IPO targeting an October launch.
00:01:47.040 --> 00:01:54.159
Hence the reason why Anthropic zigged while every other tech company zagged when it came to supporting open source.
00:01:54.400 --> 00:02:06.560
So today we're breaking down not just what the letter says, but where it actually even contradicts Dario's own previous words and why I think this was written for Washington policymakers and not for you and me.
00:02:06.799 --> 00:02:08.080
All right, let's get into it.
00:02:08.240 --> 00:02:09.840
Here is the big picture.
00:02:10.159 --> 00:02:22.719
Anthropic is the only major AI lab that refused to defend open models in the open weights and American AI leadership, uh paperslash coalition that was put forth by Microsoft and NVIDIA.
00:02:23.039 --> 00:02:31.919
Essentially, they said, well, open source is good for the American economy, it's good for national security, and we're gonna talk a little bit more about that paper.
00:02:32.240 --> 00:02:33.520
But why?
00:02:33.840 --> 00:02:41.199
Well, Anthropic will tell you it's ultimately about safety and control, but I think it is that 80% number.
00:02:41.280 --> 00:02:55.199
And we're gonna look at, well, it might not work very well if Anthropic can't get this essentially to go in their favor, and that's because their IPO is potentially only weeks away in October.
00:02:55.919 --> 00:03:03.360
So on today's show, stick with me, and you're gonna learn why that 80% meter token revenue makes Anthropic uniquely exposed to open models.
00:03:03.439 --> 00:03:10.960
You're gonna know what the regulatory capture means when it comes to AI and why this letter just fits the pattern exactly.
00:03:11.120 --> 00:03:21.039
You're gonna know why the timing points at Washington policymakers, because there's an important date coming up in just a couple of days, and I don't think this letter was made for me and you.
00:03:21.199 --> 00:03:27.520
And you're gonna know what Dario said in 2025 that flatly contradicts today's letter.
00:03:27.759 --> 00:03:29.039
All right, let's get into it.
00:03:29.199 --> 00:03:30.240
Welcome to Everyday AI.
00:03:30.319 --> 00:03:31.840
If you're new here, my name's Jordan.
00:03:32.080 --> 00:03:33.280
We do this well every day.
00:03:33.439 --> 00:03:42.159
It's your unedited, unscripted live stream podcast and free daily newsletter helping everyday business leaders like you and me keep up with the nonstop AI avalanche.
00:03:42.319 --> 00:03:44.639
I tell you what's important, what's not.
00:03:44.800 --> 00:03:50.159
You use that decision, and all of a sudden, wow, you're the smartest person in AI at in your company.
00:03:50.400 --> 00:03:56.319
So it starts here, but please make sure to subscribe to the podcast and go to our website at your everydayai.com.
00:03:56.479 --> 00:03:57.759
All right, let's get into it.
00:03:58.000 --> 00:04:04.800
This has been all the talk on tech, Twitter, and you know, in the media, in anything AI, right?
00:04:04.879 --> 00:04:13.039
If you follow AI on a daily basis like I do, this conversation has been just grabbing every single headline.
00:04:13.120 --> 00:04:16.800
And it's extremely important and also a little bit nuanced.
00:04:16.959 --> 00:04:28.480
So today's show, I'm gonna try to go through it a little bit quicker, stick to just the facts, and my opinion, maybe I'll, you know, uh bring back an accidental hot take Tuesday for old time's sake.
00:04:28.560 --> 00:04:32.319
Uh, but make sure you go check out today's newsletter uh for more.
00:04:32.560 --> 00:04:40.639
So here's what actually uh Dario put in his published response uh on open models.
00:04:40.800 --> 00:04:57.519
So the core claim was that Anthropic has never advocated banning open weight models, and he rejected the argument that open models actually help defenders more than attackers, which was a core claim in the Microsoft slash NVIDIA uh paper that everyone kind of got behind.
00:04:57.600 --> 00:05:04.800
And Anthropic says it's better for chip controls uh to crack down on distillation and mandatory safety testing for all.
00:05:05.040 --> 00:05:10.000
And they deny that protecting Anthropics business ever motivated the position.
00:05:10.079 --> 00:05:11.199
Oh, that's a funny one.
00:05:11.360 --> 00:05:11.920
All right.
00:05:12.560 --> 00:05:14.800
So let's take a quick look uh at the letter.
00:05:14.879 --> 00:05:19.920
I'm not gonna read it, but I will kind of read uh the opening paragraph and go through some of the bullet points here.
00:05:20.079 --> 00:05:23.680
So uh this is from Dario Marty, uh, Anthropics CEO.
00:05:23.759 --> 00:05:27.839
Uh so yeah, this just came late last night, actually, or late afternoon.
00:05:28.000 --> 00:05:29.120
Uh, I'm so tired.
00:05:29.199 --> 00:05:30.240
Sometimes like 4 p.m.
00:05:30.480 --> 00:05:31.439
feels like late night.
00:05:31.600 --> 00:05:38.000
So it says over the last few days, there have been a lot of discussions about open weight models, especially those from China.
00:05:38.160 --> 00:05:40.000
Reports suggest that some U.S.
00:05:40.079 --> 00:05:48.240
officials are considering banning the use of Chinese open source models, uh, sorry, Chinese open weights models by US companies.
00:05:48.319 --> 00:05:59.519
In response, many tech companies have signed a letter supporting open weights models, and some people have even accused Anthropic of wanting to ban open source models as a means of protecting our business.
00:05:59.759 --> 00:06:12.639
Anyone who has read my past writing should know that I don't regard such bans as a useful measure, but let me state it clearly so that there is no doubt Anthropic has never advocated for a ban on open weights models.
00:06:12.800 --> 00:06:17.920
Uh, then he says open weights models that don't have dangerous capabilities are a public good.
00:06:18.079 --> 00:06:24.639
They don't cost anything besides the compute needed to run them, and they provide value to businesses, developers, and researchers.
00:06:24.800 --> 00:06:31.040
All right, so let me just take a pause, and I probably should just kind of talk very briefly about open source open weights models.
00:06:31.120 --> 00:06:35.439
So if you are brand new, maybe this is the first episode you've ever listened to, and you're like, what the heck is going on?
00:06:35.600 --> 00:06:36.480
Why does this matter?
00:06:36.639 --> 00:06:39.839
Well, a couple important points that Dario brought up here in the letter.
00:06:40.079 --> 00:06:52.480
Uh, number one, there has been this idea floated recently, and you know, we'll see uh, you know, maybe in 10, 20 years, how much of Anthropic's DC lobbying efforts maybe uh led to this, right?
00:06:52.560 --> 00:06:54.720
I'm a former journalist, I understand how this works.
00:06:54.879 --> 00:07:10.720
All the big labs have a lot of lobbyists, but uh the US government is actually considering a ban on American companies using Chinese open source models for you know what they're saying, uh competitiveness, national security concerns, etc.
00:07:11.360 --> 00:07:17.920
But essentially, open source or open weight models are models that you can well you can download and you can run them on your machine.
00:07:18.079 --> 00:07:23.759
Well, very small open source models or what open source models used to be, you know, like two years ago.
00:07:23.920 --> 00:07:28.319
Today's open weights Chinese models, you can't download those on your computer.
00:07:28.399 --> 00:07:32.879
You're literally going to need a spare uh data center to run the newest.
00:07:32.959 --> 00:07:40.720
And the ones that kind of set this off are uh Moonshots, Kimi K3, uh ZAI's GLM52.
00:07:41.120 --> 00:07:45.519
Um, and I think that eventually we'll be talking a little bit more about Quen 3.8.
00:07:45.680 --> 00:07:50.879
So essentially you have all these models that, well, if you're a huge enterprise, you can use them for free.
00:07:51.040 --> 00:07:59.360
Uh, but if you are using them um through a hosting provider, uh through an inference provider, they're way, way cheaper, right?
00:07:59.439 --> 00:08:00.079
For the most part.
00:08:00.240 --> 00:08:07.920
Not always, but usually these open source models are way cheaper uh if you can't download and host them or run them locally.
00:08:08.079 --> 00:08:08.560
Uh all right.
00:08:08.720 --> 00:08:14.720
So that's the big picture on uh what kind of Dario said in his letter, in all of that.
00:08:14.879 --> 00:08:19.279
So uh some of the more important bullet points here.
00:08:19.600 --> 00:08:41.679
Uh he said, My primary concern is the risk that authoritarian governments, not solely the Chinese uh Communist Party or CCP, although the CCP is clearly the most capable threat, build AI models that are more powerful than those built in the US and use them to achieve permanent military superiority or uh perpetrate incredibly deep repression of their own people.
00:08:41.840 --> 00:08:42.000
All right.
00:08:42.080 --> 00:08:49.200
So essentially he said the biggest concern is well, the Chinese government is going to use these for military superiority.
00:08:49.360 --> 00:08:59.919
Uh, number two, secondary concern, he says, is that powerful AI models may be misused to carry out cyber attacks or biological attacks, and those may have serious alignment issues.
00:09:00.159 --> 00:09:04.240
So, you know, I agree with both of those things, right?
00:09:04.320 --> 00:09:06.320
Um, I've been saying it all along.
00:09:06.480 --> 00:09:16.000
I think artificial intelligence, for the most part, it is uh gonna be more important maybe in five to ten years than you know what your traditional military might be looked at.
00:09:16.080 --> 00:09:28.799
It's gonna be more important than oil or gold or anything else, because uh the nation essentially that figures out artificial general intelligence and artificial superintelligence first has a legit, unfair advantage on the rest of the world.
00:09:28.879 --> 00:09:39.440
And you can go from a mid-tier global power to the uh global superpower if you control artificial general intelligence or artificial superintelligence.
00:09:39.759 --> 00:09:57.279
So, uh, kind of to address these concerns, what he does say in his letter, he says, number one, we should not sell powerful chips or chipmaking equipment to China, essentially saying, hey, we still have a lead, although it's maybe slimmer than it was before, but China can't catch us if they don't have our chips.
00:09:57.440 --> 00:10:06.559
Uh, right, previously uh they said uh Anthropic said that by essentially creating a chip ban, uh the US could get a 12 to 24 month uh lead in AI.
00:10:06.720 --> 00:10:08.480
I don't necessarily think that's true.
00:10:08.799 --> 00:10:12.240
The lead could extend from where it is now, which is like two months.
00:10:12.480 --> 00:10:18.799
It used to be about six months uh between the um US frontier models and the Chinese open source models.
00:10:18.879 --> 00:10:20.080
So maybe it could extend it.
00:10:20.159 --> 00:10:22.399
I don't necessarily think it would be 12 to 24 months.
00:10:22.559 --> 00:10:24.720
That's a little bit of lofty thinking, I would say.
00:10:24.879 --> 00:10:31.679
And then uh Dario also says we should crack down on industrial scale distillation uh distillation operations, right?
00:10:31.759 --> 00:10:34.799
So essentially how China is catching up, right?
00:10:34.879 --> 00:10:43.679
They do have some uh novel architecture and uh kind of ML um uh approaches, but for the most part, it's distillation, right?
00:10:43.759 --> 00:10:51.440
They're just copying uh, you know, millions of inputs and outputs from you know clawed models and other models to make their models smart.
00:10:51.600 --> 00:10:54.720
So they're just copying someone else's answers at a fraction of the cost.
00:10:54.879 --> 00:11:02.159
Uh and then uh Dario says all sufficiently capable models, open and closed, should go through mandatory safety testing.
00:11:02.320 --> 00:11:05.039
All right, there's more on the um the letter.
00:11:05.120 --> 00:11:05.919
It's not super long.
00:11:06.000 --> 00:11:07.039
I'm not gonna read it all.
00:11:07.120 --> 00:11:10.720
Uh, we'll make sure to put it in today's newsletter uh as well.
00:11:11.039 --> 00:11:25.200
So this is all essentially in response to the original uh kind of movement that uh NVIDIA and Microsoft put forth uh on July 24th, which was this past Friday.
00:11:25.360 --> 00:11:28.879
And this was called the open weights in American AI leadership.
00:11:29.039 --> 00:11:29.519
All right.
00:11:29.679 --> 00:11:36.720
So um essentially they circulated this, arguing that open models cut costs and avoid vendor locked in.
00:11:36.879 --> 00:11:47.919
So they said it's good for American businesses, it's good for uh competitiveness, and then they also said it's better for safety because they said that the core safety claim of having open models is that defenders, right?
00:11:48.000 --> 00:11:51.840
Because this whole thing comes down to cyber attacks versus defending your business, right?
00:11:52.000 --> 00:11:58.799
Uh, one of the big arguments is oh my gosh, if you have these, you know, these open source models everywhere, you can't keep track of them, right?
00:11:58.879 --> 00:12:17.840
And eventually, and I agree, probably in two years, I've said this, you'll be able to, you know, have a version or an open source model that is of the GPD56 soul fable five um tier of models, you'll be able to have something that powerful on a you know consumer-ish desktop device.
00:12:18.000 --> 00:12:24.480
So at that point, you know, yeah, there is going to be, you know, critical infrastructure at risk from cyber attacks.
00:12:24.639 --> 00:12:34.960
You know, think of you know, phishing scams and you know, people holding you know web systems for hostage uh times a thousand and compact it into a very tight time frame, right?
00:12:35.039 --> 00:12:36.559
Once that does happen, right?
00:12:36.799 --> 00:12:41.679
So if you're thinking of hospitals, school systems, power grids, right?
00:12:41.919 --> 00:12:47.600
Um, entire small nations, you know, city and state governments, right?
00:12:47.759 --> 00:12:50.159
All these places are going to be extremely at risk.
00:12:50.240 --> 00:12:50.879
And I agree with that.
00:12:50.960 --> 00:13:01.919
But essentially, Microsoft and NVIDIA are saying, well, the best way to defend against this is well, open models because you have more people contributing, people are going to find these vulnerabilities, and it collectively makes you know everyone stronger.
00:13:02.000 --> 00:13:05.039
So it gives everyone better defensive tools, all right.
00:13:05.120 --> 00:13:14.559
Um, and essentially everyone literally agreed with that from Amazon, OpenAI, Meta, Google, IBM, everyone except Anthropic.
00:13:15.360 --> 00:13:19.039
So my take on this is Dario's letter is not for us.
00:13:19.279 --> 00:13:22.240
This is literally for regulatory capture.
00:13:22.480 --> 00:13:26.960
This is to uh influence the policy making in uh DC.
00:13:27.039 --> 00:13:32.159
So there's been a lot of articles about, you know, just how big of an impact that Anthropic has in lobbying.
00:13:32.320 --> 00:13:38.159
And uh yeah, I think that this has to be the real reason here.
00:13:38.240 --> 00:13:41.279
Um, but I mean, my take on the actual letter, it's well written.
00:13:41.360 --> 00:13:46.799
There's great arguments in there, and it gener it genuinely engages with real risk, right?
00:13:47.200 --> 00:13:52.879
There's nothing in there uh that I'm looking at and I'm like, this is absolutely blasphemy, right?
00:13:52.960 --> 00:14:02.960
Like, aside from the fact that they say that they absolutely never, you know, called for a ban on open models, they have just without using the word ban, right?
00:14:03.120 --> 00:14:03.600
I don't know.
00:14:03.679 --> 00:14:13.200
If you say something, uh, you know, walks like a duck, quacks like a duck, is yellow like a duck, you know, is is in the jersey on mighty ducks, but you're like, well, it's not a duck.
00:14:13.360 --> 00:14:16.000
Well, yeah, that's essentially what you said.
00:14:16.080 --> 00:14:25.679
You just because you didn't say that animal right there is a duck, but you you described every single duck characteristic and used the word duck in describing the characteristics, but then you're like, Well, it's not a duck.
00:14:25.840 --> 00:14:27.360
Well, yeah, it is.
00:14:27.600 --> 00:14:28.000
All right.
00:14:28.159 --> 00:14:34.879
And but the important thing here is Dario's letter comes just days before an important federal deadline.
00:14:35.039 --> 00:14:38.159
And that is why the regulatory piece is very important.
00:14:38.399 --> 00:14:46.159
So President Trump signed essentially this uh AI uh order on June 2nd with a 60-day deadline.
00:14:46.240 --> 00:14:46.720
And guess what?
00:14:46.799 --> 00:14:49.600
That 60-day deadline is well, this week.
00:14:49.679 --> 00:14:54.960
Uh so the framework right now does not cover open weight models, right?
00:14:55.039 --> 00:15:04.639
So that's why there's all this kind of uh jockeying right now, and uh, you know, why you've had these stories come out in the past uh seven to ten days, right?
00:15:04.799 --> 00:15:13.600
Saying that the US is now maybe considering uh kind of throwing open source models, uh not in this executive order, but potentially just banning them.
00:15:13.759 --> 00:15:23.679
Because out of nowhere, uh right, you had you know, Moonshots, Kimi K3, I think especially come up and you know, essentially it is in that top tier.
00:15:23.759 --> 00:15:27.360
Whereas before the open source Chinese models were never in that tier.
00:15:27.440 --> 00:15:32.240
So that obviously takes away from um American business, number one, right?
00:15:32.399 --> 00:15:39.519
Because whether you know it or not, and I've been saying this all along, AI and large language models, right now they run the American economy.
00:15:39.679 --> 00:15:47.840
So when you have a Chinese open uh weights uh contender come in, and well, potentially you don't want to take all that spend, right?
00:15:47.919 --> 00:15:58.799
From uh uh you know, putting America first sort of thing, which is the uh kind of the Trump's uh uh the Trump people's uh position on this, right?
00:15:58.960 --> 00:16:04.639
Well, that takes away probably a lot of money that American businesses would be spending with other American businesses.
00:16:04.799 --> 00:16:21.039
So not only that, but you know, kind of the other uh uh chain of thought, haha uh AI joke accidentally made, is that you know, the more powerful these models become, well, they're just gonna become better at distilling versions of themselves from the other frontier labs.
00:16:21.200 --> 00:16:26.080
And you know, it seems like Anthropic, maybe more so than others, seemingly is struggling.
00:16:26.240 --> 00:16:30.720
Maybe they're just the ones complaining a little bit more because they technically have more to lose.
00:16:31.039 --> 00:16:38.480
But the Dario letter lands just five days uh before this uh kind of order gets locked in.
00:16:38.559 --> 00:16:49.759
So the order is voluntary and there's no licensing and no pre-clearance requirements, but essentially you're supposed to clear uh models with the government 30 days before release to you know give everyone a fair amount of time.
00:16:50.000 --> 00:16:55.519
So uh yeah, there's that pretty big thing that no one is talking about.
00:16:55.679 --> 00:16:58.000
But the real reason is the money, y'all.
00:16:58.320 --> 00:17:02.159
So these are according to public estimates, which are pretty close.
00:17:02.240 --> 00:17:12.720
A lot of it is from semi-analysis, which is kind of the uh the leader in the space in terms of uh looking at compute and its impact on the economy, the money in and the money out.
00:17:12.880 --> 00:17:21.519
So essentially, um, the best estimates are 80% of anthropics revenue comes from just selling tokens, right?
00:17:21.680 --> 00:17:26.240
Whereas its closest competitor, OpenAI, it's estimated that it's only 20%.
00:17:26.640 --> 00:17:31.119
Uh and then, you know, your other big tech companies, it's very small considered, right?
00:17:31.200 --> 00:17:41.599
Because they have you know dozens or hundreds of lines of revenue, like you know, Google, it's only about two percent, Microsoft two percent, Amazon one percent, and meta not even one percent yet.
00:17:41.759 --> 00:17:48.720
So essentially, Anthropic's entire business model is just selling tokens, right?
00:17:49.039 --> 00:17:56.960
And well, what if now all of a sudden, if you just stop buying those tokens, if your company stops buying them, right?
00:17:57.039 --> 00:18:13.519
If you're like, hey, we don't need right, we can get 95% of our AI use with a model like you know, GLM 5.2, and that one is a little bit smaller, and well, you know, we can invest, you know, uh six figures and getting this set up.
00:18:13.680 --> 00:18:17.839
We were spending six figures or seven figures in AI each month, right?
00:18:18.000 --> 00:18:20.079
Now all of a sudden we cut our cost by 95%.
00:18:20.480 --> 00:18:23.200
We have more control, we're running it locally, uh, etc.
00:18:23.440 --> 00:18:27.039
So you can see how this is in extreme.
00:18:27.440 --> 00:18:34.400
I am talking like potentially lights out uh for uh Anthropic's growth, right?
00:18:34.480 --> 00:18:40.319
Not for them as a company, but for their staggering growth that we've been reading about through 2024 and 2025.
00:18:40.400 --> 00:18:44.400
And you know, they've been a the main beneficiary of this token maxing.
00:18:44.480 --> 00:18:49.039
Uh, but now that token maxing is going away and it's uh going toward token efficiency.
00:18:49.200 --> 00:18:59.039
So not only, right, if you look at open router, uh, which kind of is a third-party tracker of you know who's spending uh tokens where, you know, it used to be anthropic.
00:18:59.119 --> 00:19:00.400
Anthropic used to dominate.
00:19:00.559 --> 00:19:02.400
Now uh I checked uh yesterday.
00:19:02.480 --> 00:19:03.920
I think they were like number seven.
00:19:04.000 --> 00:19:08.240
They're like the seventh most used lab in open routers, and they used to dominate, right?
00:19:08.319 --> 00:19:10.319
So people are coming to a couple of conclusions.
00:19:10.400 --> 00:19:15.920
Number one, uh yes, Anthropic still has probably the best models in the world, right?
00:19:16.079 --> 00:19:19.200
The combination of Opus V that just came out and Fable Five.
00:19:19.279 --> 00:19:24.400
Although I do think GPT-5-6 soul uh is is right there and on many important benchmarks.
00:19:24.480 --> 00:19:26.799
I do think uh GPT 5-6 soul is still better.
00:19:26.880 --> 00:19:51.039
But regardless, Anthropic has some of the best models in the world, but 80% of their revenue is just selling tokens, and now all of a sudden, over the past 10 days, you have Kimi K3, and you we're gonna be talking more about Quin38 and probably whenever GLM53 comes out, all of a sudden, yeah, Anthropic's you know that that hockey stick growth is not gonna look very hockey stick for much longer.
00:19:52.480 --> 00:20:02.720
So Anthropic built this massive business, and if I'm being honest, there's not a big moat, right?
00:20:02.960 --> 00:20:07.759
Google, open AI, and Microsoft are taking a very different approach, right?
00:20:07.839 --> 00:20:10.000
They're looking at the entire ecosystem.
00:20:10.240 --> 00:20:22.559
But right now, Anthropic's entire business is inference, and inference is just a commodity that's getting cheaper, mainly because of open source models and the technology itself.
00:20:22.720 --> 00:20:27.920
But some open models produce output roughly 10 times cheaper than Opus.
00:20:28.160 --> 00:20:37.039
So even if you're paying for it via the API, right, in some models and some certain tasks, it's just the same price, but in many instances, it is incredibly cheaper.
00:20:37.200 --> 00:20:39.519
And switching is easy, right?
00:20:39.680 --> 00:20:45.359
That's the difference between the ecosystem play versus the API charging you for tokens play, right?
00:20:45.440 --> 00:20:53.440
If a company, and you've seen plenty of stories, literally, people detail this at big multi-billion dollar companies where they're like, wait, we just cut our spend.
00:20:53.519 --> 00:20:56.559
You know, we were spending, you know, tens of millions of dollars a year.
00:20:56.720 --> 00:21:04.799
We just cut our spend by you know 80% just by now using a model router and you know, 80% of our queries, we can use open source models, right?
00:21:04.880 --> 00:21:11.200
Maybe they're still paying a little bit to you know, anthropic or open AI or Google or whatever for the actual just raw tokens.
00:21:11.359 --> 00:21:23.119
But switching is so easy to literally go in there and switch a couple endpoints, or you know, if you're using um, you know, a certain gateway, uh, you know, to go in there and just choose a new model and the drop down.
00:21:23.279 --> 00:21:31.119
And if you've been doing your due diligence and you know, doing your running your backups like I've been telling you to do for years, it's actually not that big of a disruption for your business.
00:21:31.279 --> 00:21:33.440
So switching is easier than ever.
00:21:33.759 --> 00:21:39.440
If you're just paying for APIs, switching out of uh an ecosystem, not easy, right?
00:21:39.519 --> 00:21:45.200
That's like I've always wished I could just, you know, switch out of using an iPhone or switch out of using a Mac.
00:21:45.440 --> 00:21:45.920
I can't.
00:21:46.079 --> 00:21:48.000
I am locked into that ecosystem.
00:21:48.160 --> 00:21:52.000
Even if the hardware isn't the best, even if the camera isn't the best, whatever.
00:21:52.079 --> 00:21:52.640
It doesn't matter.
00:21:52.720 --> 00:21:54.559
I can't get out of the Apple ecosystem.
00:21:54.640 --> 00:21:55.920
I am too locked in, right?
00:21:56.079 --> 00:21:58.160
That's the game that OpenAI has been playing.
00:21:58.240 --> 00:22:05.359
That's the game that Microsoft and Google has been playing and Infraopic has just been saying, Well, we're just going to sell tokens.
00:22:05.519 --> 00:22:12.160
Well, they're seeing and they're finding out that that may not work unless they can get some regulatory help.
00:22:12.400 --> 00:22:14.799
And that could just become their moat.
00:22:15.200 --> 00:22:18.000
So that is the regulatory capture here.
00:22:18.079 --> 00:22:22.000
So that's just winning in Washington instead of on an actual product.
00:22:22.160 --> 00:22:28.799
So essentially the incumbents, right, are backing these safety rules that only companies their size can afford.
00:22:28.880 --> 00:22:32.160
And the rules become the moat that the product just couldn't build, right?
00:22:32.240 --> 00:22:34.079
And this isn't just me saying this.
00:22:34.240 --> 00:22:36.960
Um, yeah, there's actually some very prominent names.
00:22:37.039 --> 00:22:40.160
Uh, you think I'm tough on anthropic sometimes.
00:22:40.319 --> 00:22:41.519
I'm definitely not.
00:22:41.680 --> 00:22:48.720
I actually think that this letter is pretty good, and there's very few things in there that I would argue against, uh, right.
00:22:48.799 --> 00:22:54.640
But White House AI czar David Sachs called it fear-based regulatory capture.
00:22:54.799 --> 00:22:59.359
I mean, he tore into this, and uh, interestingly enough, he actually called it, right?
00:22:59.519 --> 00:23:03.599
He essentially said, Yeah, Anthropic is gonna run uh a psyops on us.
00:23:03.680 --> 00:23:13.359
You know, he literally called this before uh Dario wrote this letter, and he said, you know, Anthropic's uh gonna claim that they never wanted to ban open models, although they did want to.
00:23:13.519 --> 00:23:16.799
And sure enough, that's literally what they did, right?
00:23:16.960 --> 00:23:24.160
Um, and also nearly 200 startups warned that banning open source models could hand Anthropic a monopoly.
00:23:24.559 --> 00:23:29.440
So let's get back to the timing because literally, this is what this is about.
00:23:29.680 --> 00:23:33.119
This is what the I don't know how many times you want me to tell this, right?
00:23:33.200 --> 00:23:40.720
This is the difference between a company, you know, being worth, you know, 500 billion and being worth three trillion dollars, right?
00:23:40.880 --> 00:23:44.559
In theory, anthropic could go either way in a year.
00:23:44.720 --> 00:23:50.480
Um, you know, and I actually wasn't even gonna do this show, uh, but I put out a tweet, yeah.
00:23:50.559 --> 00:23:52.240
I did, yeah, cringe, right?
00:23:52.400 --> 00:23:53.920
Uh, a couple of weeks ago.
00:23:54.160 --> 00:23:59.920
And it was actually a very high up person from Google that liked the tweet, right?
00:24:00.079 --> 00:24:10.480
And my tweet, I forgot exactly what it was, but it was something along the lines of like Anthropic is in deep trouble with these open models and with the shift from you know token maxing to token efficiency.
00:24:10.640 --> 00:24:19.359
And their hope is they want to go public as quickly as humanly possible before Wall Street and enterprises figure out what the heck is going on.
00:24:19.519 --> 00:24:25.279
And I was actually surprised that this very high-up ranking person uh at Google liked the tweet.
00:24:25.359 --> 00:24:30.880
And the thing is, they're not really a I mean, they are a direct competitor with anthropic, but they're also an investor.
00:24:31.039 --> 00:24:35.680
Um, and you know, likes are private, so don't think you can go in there and I'm not gonna say who this person was.
00:24:35.839 --> 00:24:48.720
But when I said, wait, this person is uh essentially agreeing with what I'm saying here, and that's why I'm like, no, I am a hundred percent um confident that this is what is happening.
00:24:48.799 --> 00:24:50.640
This isn't just some like some hunch, right?
00:24:50.720 --> 00:24:54.960
And I've talked to plenty of people off the record about this at the big labs.
00:24:55.039 --> 00:24:57.200
Um, and they've all pretty much agreed.
00:24:57.279 --> 00:25:01.920
So Enthropic, here's the timing Enthropic filed their confidential S1 June 1st.
00:25:02.000 --> 00:25:08.079
And, you know, according to reports, uh, they're looking at maybe listing as a public company as soon as October.
00:25:08.240 --> 00:25:15.359
So public investors will pay for that durable pricing power, which obviously the open models straight up erase.
00:25:15.759 --> 00:25:23.039
And you know, these restrictions that are landing here soon, uh, right, would remove the cheapest substitutes.
00:25:23.119 --> 00:25:28.000
So the safety concern can be sincere, but also can be commercially self-serving all at once.
00:25:28.160 --> 00:25:28.960
I think that's what's happening.
00:25:29.039 --> 00:25:43.680
That's why I'm like, you read the letter, yeah, most people would agree with most parts of it, but it can, you know, be a you know, a good letter that you can kind of agree with many parts of it, but also be incredibly self-serving at the same time.
00:25:44.480 --> 00:25:45.119
All right.
00:25:46.000 --> 00:25:50.720
So anthropics numbers, you gotta take a look at them.
00:25:50.880 --> 00:25:53.759
So obviously they're not public, but there is a couple of websites.
00:25:53.920 --> 00:26:00.240
One is called Notice, um, that essentially tracks, and this is, you know, all you know, in algorithm estimates.
00:26:00.319 --> 00:26:12.079
It's not, you know, actually secondary market prices, but you know, it essentially tracks, you know, the OpenAI, the Anthropics, all these companies that have yet to go public, it tracks what their stocks would roughly be.
00:26:12.400 --> 00:26:20.319
And Anthropic has had their worst seven-week span ever in a seven-week span that included a model release, right?
00:26:20.400 --> 00:26:26.559
Because there was one drought uh, you know, where their stock went down a little bit, um, but they just didn't have any releases.
00:26:26.720 --> 00:26:34.640
So essentially, in the history of the company, this is their worst period when it comes to their, you know, pre-IPO stock price.
00:26:34.880 --> 00:26:38.480
Um, and they had great models that ship during this time window, right?
00:26:38.640 --> 00:26:41.359
Fable five, sonnet five, opus five.
00:26:41.440 --> 00:26:48.880
Uh, so it, you know, not a huge drop, but it fell, you know, about 8% over seven weeks, their worst record on stretch.
00:26:49.200 --> 00:26:49.680
Why?
00:26:49.920 --> 00:26:50.240
Right?
00:26:50.480 --> 00:26:55.839
Makes no sense because they've been hockey sticking uh essentially since, you know, 2024.
00:26:56.319 --> 00:27:04.720
Well, I think early investors and people who are very closely attached to AI are starting to understand what's going on here.
00:27:04.880 --> 00:27:15.599
It is the combination of the reckoning going from token maxing to token efficiency to now, all of a sudden, people are starting to care more about the cost per tax, uh, cost per task, right?
00:27:15.680 --> 00:27:23.119
So they're not just solely looking at a benchmark and then you know, fully going full send token maxing on you know the best model.
00:27:23.279 --> 00:27:29.039
They're instead saying, what's the best model that gets the job done at a good cost, right?
00:27:29.200 --> 00:27:32.880
The best example is you know OpenAI's GPT 5.6.
00:27:33.519 --> 00:27:44.720
Their models might have one or two uh points lower on a benchmark like artificial analysis, but at sometimes at a third of the cost, right?
00:27:45.119 --> 00:27:50.559
So, yeah, there's just some bad things that are happening when it comes to token economics with anthropic.
00:27:50.640 --> 00:28:04.160
And they are literally hoping to go public ASAP as soon as humanly possible, uh, before essentially the rest of the uh it insiders wake up to see what's go actually going on.
00:28:04.480 --> 00:28:29.599
So let's inspect a little bit more about some of the reasons I think what Dario said don't exactly add up because he actually wrote something six weeks ago um that kind of sound like a ban to me, yet in today's uh sorry, in yesterday's uh you know letter um saying why they didn't join this coalition with Microsoft and NVIDIA and everyone else.
00:28:29.759 --> 00:28:35.279
Well, they're like, Well, we're not banning open source models, we just can't fully support them in the way that it's written.
00:28:36.000 --> 00:28:36.319
All right.
00:28:36.400 --> 00:28:42.000
So Dario had an essay called The Policy on the AI Exponential published on his personal website.
00:28:42.160 --> 00:28:48.960
So he drops transparency as sufficient and calls for binding regulatory, uh, sorry, regulation instead.
00:28:49.039 --> 00:28:55.359
And he wants that mandatory third-party testing, then government power to block a release, right?
00:28:55.680 --> 00:29:01.680
Um, in his words, that releases should be blocked or reversed if models fail.
00:29:01.839 --> 00:29:02.160
All right.
00:29:02.240 --> 00:29:06.880
So essentially, he thinks all models should have third-party testing.
00:29:07.039 --> 00:29:13.519
And well, um, if they fail after release, they need to be blocked or reversed.
00:29:13.680 --> 00:29:19.200
So Anthropic published actually draft legislation with it and pledged money behind it.
00:29:19.440 --> 00:29:21.279
So, what does that mean?
00:29:21.519 --> 00:29:27.599
Well, essentially, this was calling for a ban of open source models without using the word ban.
00:29:27.920 --> 00:29:28.640
Let me explain.
00:29:28.799 --> 00:29:40.880
Because if you add one and one here, the only answer is two, because anthropic or any proprietary company can, you know, go along and play by anthropic's rules that they wish existed, right?
00:29:40.960 --> 00:29:44.799
You put out a model, something goes wrong, it's not safe, you pull it away, right?
00:29:44.960 --> 00:29:47.839
Same thing that happened with their Fable V Mythos V.
00:29:48.000 --> 00:29:49.440
It was out for like three days.
00:29:49.680 --> 00:29:52.480
The government said nope, and then they pulled it, right?
00:29:52.640 --> 00:29:55.519
So, in theory, Anthropic was playing by its own rules.
00:29:55.759 --> 00:29:56.720
Guess what?
00:29:57.200 --> 00:29:59.519
Open models literally can't.
00:29:59.839 --> 00:30:06.559
So I don't know how Anthropic is saying, well, oh, we never said uh we would ban open source models.
00:30:06.720 --> 00:30:10.799
We would just create rules that would make them literally impossible to exist.
00:30:11.039 --> 00:30:12.640
So we didn't use the word ban, right?
00:30:12.720 --> 00:30:13.839
We didn't call the deduct.
00:30:14.160 --> 00:30:20.880
Uh, because nobody can undo an open release once thousands have downloaded an open model, right?
00:30:20.960 --> 00:30:38.400
So they're um, and you know, I'm not gonna go through state by state, but essentially Anthropic has pledged support uh to certain states where it would be advantageous for them to uh have this kind of setup in there um that would essentially make open source models not really a truly viable thing.
00:30:38.559 --> 00:30:52.960
So if his rule is passable for anthropic, uh well, it or open, you know, open AI, any um, you know, any closed frontier model, you can't download it, you can't save it, just like Anthropic did, they can pull it, right?
00:30:53.119 --> 00:30:56.640
If Google needs to pull something, if open AI needs to pull something, they can't.
00:30:56.720 --> 00:30:57.200
It's closed.
00:30:57.279 --> 00:31:00.720
They pull the API, they pull the subscription, the model's no longer available.
00:31:00.880 --> 00:31:02.160
It's not the case with open source.
00:31:02.240 --> 00:31:05.200
Once you open it up, you can't trace it, right?
00:31:05.359 --> 00:31:07.839
People can you know lower the guardrails.
00:31:08.000 --> 00:31:12.799
Yeah, it is technically, in theory, a little more dangerous, right?
00:31:13.759 --> 00:31:18.079
But you can't say both, right?
00:31:18.160 --> 00:31:23.920
You can't put a letter out yesterday saying, like, oh, we never said we would ban open source models.
00:31:24.160 --> 00:31:38.400
We just six weeks ago put out uh our own example of policy that would just make it so open models could never actually exist by following rules that technically only a proprietary model could follow.
00:31:38.799 --> 00:31:41.039
So that's not the only thing.
00:31:41.200 --> 00:31:44.640
The other thing is talking about the control argument.
00:31:44.880 --> 00:31:49.359
So Dario's case requires closed models to be monitorable and containable.
00:31:49.519 --> 00:31:51.359
Well, what happened with Mythos, y'all?
00:31:51.519 --> 00:31:56.720
Because Mythos escaped its own sandbox and emailed a researcher unprompted.
00:31:56.880 --> 00:32:04.400
And not only that, unauthorized outsiders access mythos the same day it was announced via a Discord server just by guessing the naming mechanism.
00:32:04.559 --> 00:32:06.000
And also, oh wait, what's that?
00:32:06.160 --> 00:32:11.839
Chinese hackers ran 80 to 90 percent reportedly of an espionage campaign using Claude Code.
00:32:12.000 --> 00:32:17.039
So the bad stuff that Dario is warning about, we can never have open models, right?
00:32:17.200 --> 00:32:21.279
The the the way that everyone else wants them, because all this bad stuff will happen.
00:32:21.359 --> 00:32:22.000
Well, guess what?
00:32:22.079 --> 00:32:25.359
It's already happening with Anthropic's own models, right?
00:32:25.519 --> 00:32:28.480
That's not my opinion, that's just the facts.
00:32:29.519 --> 00:32:32.400
And we have to talk about the Hugging Face incident, right?
00:32:32.559 --> 00:32:33.599
So that happened recently.
00:32:33.680 --> 00:32:36.960
It's been all the talk, and it's actually led to the snowballing of all these other things.
00:32:37.039 --> 00:32:40.240
But OpenAI's closed model broke containment, right?
00:32:40.319 --> 00:32:46.480
The agent got out of its sandbox and attacked Hugging Face uh to work on a benchmark, right?
00:32:46.640 --> 00:32:52.319
It was trying to solve uh a benchmark problem by kind of hacking Hugging Face and getting the answers.
00:32:52.480 --> 00:32:56.640
So the closed APIs that Hugging Face used, right?
00:32:57.680 --> 00:33:02.079
They were trying to analyze the attack and they used closed models and it didn't work.
00:33:02.240 --> 00:33:03.599
They got blocked, right?
00:33:03.759 --> 00:33:08.720
And instead, they used an open source model, uh GLM 5.2, and it worked.
00:33:08.799 --> 00:33:11.519
And that's how they were able to close down the attack.
00:33:11.680 --> 00:33:16.319
So the closed source models, well, they weren't good enough as a defender.
00:33:16.480 --> 00:33:23.839
The open source Chinese models were because they couldn't use the closed source proprietary models because the guardrails are a little stricter.
00:33:23.920 --> 00:33:29.519
So Dario's own argument that he puts out there is moot, right?
00:33:29.759 --> 00:33:35.519
His his letter doubts says that open models can't actually help defenders, right?
00:33:35.680 --> 00:33:39.359
Uh he says that's what proprietary models are for.
00:33:39.599 --> 00:33:40.640
Not true, right?
00:33:40.720 --> 00:33:45.839
If AI is only for thee and not for me, it's not the case, right?
00:33:46.000 --> 00:34:00.720
Uh, you know, if the the the mythos, as an example, you know, I think there's like 40 original partners uh in the Glasswing uh program that have access to the mythos that you can truly use the best model in the world for cyber defense.
00:34:00.880 --> 00:34:05.359
Well, for 99.9% of everyone else, they can't, right?
00:34:05.599 --> 00:34:15.039
So you can't have this two-tiered system uh of well, only a select few can actually use the best proprietary models for defense, but anyone can use an open model.
00:34:15.199 --> 00:34:18.639
So, you know, his point there is absolutely false, right?
00:34:18.800 --> 00:34:21.920
Demonstratably false within the last week, right?
00:34:22.000 --> 00:34:34.960
So maybe they should have updated, maybe they wrote that letter before the whole, you know, open AI agent hugging face GLM 5.2 fiasco, uh, because it literally is doesn't make sense on paper because it's false.
00:34:35.280 --> 00:34:43.280
So let's wrap up here because I think here the big debate is safety control and business protection.
00:34:43.440 --> 00:34:45.199
When it comes to open source, right?
00:34:45.599 --> 00:34:48.960
I'm not gonna dive too deep on that because there's pros and there's cons.
00:34:49.119 --> 00:34:54.880
Like I said, there's things in the uh in the NVIDIA Microsoft uh letter that I agree with 100%.
00:34:55.199 --> 00:34:57.840
There are things in Dario's letter that I agree with 100%.
00:34:58.239 --> 00:35:08.559
A lot of it is gray area, a lot of it's nuanced, and a lot of it is the unknown because who knows what models in six months, proprietary or open source, will be capable of.
00:35:08.719 --> 00:35:09.280
Who knows?
00:35:09.440 --> 00:35:15.280
Maybe at some point open source models will be able to pass proprietary models.
00:35:15.440 --> 00:35:18.320
I mean, it's not outside of the realm of possibilities.
00:35:18.559 --> 00:35:20.719
So we can't really make that comparison.
00:35:20.800 --> 00:35:25.920
But what we have to look at is today and how this impacts your business and the decisions you make.
00:35:26.079 --> 00:35:31.679
Because I think Dario's safety concerns, well, they're real, and I think that they were thoughtfully presented in this letter.
00:35:31.840 --> 00:35:35.039
But the letter, it's actually not what you think it is.
00:35:35.199 --> 00:35:37.440
It is literally about their IPO.
00:35:37.599 --> 00:35:45.199
It is to scare uh people in DC enough to get some restrictions on these Chinese open source models.
00:35:45.360 --> 00:35:53.440
They're gonna say, you know, put out the the military reasons, the American competitiveness, all these things, which in theory are kind of true.
00:35:53.519 --> 00:36:09.280
But then I think you have to look at the uh the every other company's position that signed on to the uh NVIDIA and Microsoft letter that says, well, no, everyone needs open source because the more people, the more companies, the more AI labs they get behind open source, the more powerful it becomes.
00:36:09.440 --> 00:36:11.039
It's collective, it's a snowball.
00:36:11.119 --> 00:36:14.960
And when you put all the snowballs together and roll them down the hill, well, they're gonna get bigger.
00:36:15.199 --> 00:36:17.199
So this is not for us.
00:36:17.599 --> 00:36:19.679
This is for Washington.
00:36:19.840 --> 00:36:29.679
This is there's a shot clock, both the shot clock uh for the Trump AI rules that go into effect here in a week, but also Anthropic's IPO.
00:36:29.760 --> 00:36:37.039
Because the conclusion here, I think, is yes, this is a sincere safety argument, but it is unmistakable regulatory capture.
00:36:37.119 --> 00:36:37.920
That's all it is.
00:36:38.079 --> 00:36:48.079
And my hot take is well, Inthropic over the last couple of months has realized, I think, that their monthly growth is probably not up.
00:36:48.559 --> 00:36:57.199
Hockey sticking at the same rate that it was before, and they've realized that their moat is not really as strong as they thought it was.
00:36:57.360 --> 00:37:11.760
Um, so they're actually hoping for regulatory help to be their actual moat because they realized wait, with open source, we can't just sell tokens when other companies are giving it away for next to free, and it's just as good.
00:37:12.000 --> 00:37:22.719
All right, that's a wrap on Anthropic's response in why Claude's CEO didn't sign the open model pact and the real reasons, I think, why.
00:37:22.880 --> 00:37:25.039
So, if this was helpful, let me know about it.
00:37:25.119 --> 00:37:27.119
Please subscribe to the podcast if you haven't already.
00:37:27.199 --> 00:37:29.360
Then go to your everydayai.com.
00:37:29.440 --> 00:37:34.639
We're gonna be recapping the highlights from today's show as well as everything else you need to know to stay up to date.
00:37:34.800 --> 00:37:35.679
Thanks for tuning in.
00:37:35.840 --> 00:37:38.800
Hope to see you back tomorrow and every day for more everyday AI.
00:37:38.880 --> 00:37:39.760
Thanks, y'all.
