WEBVTT

1
00:00:00.140 --> 00:00:11.180
On Friday, February 27th, Sam Altman went on national television and said he agrees with Anthropic's red lines on autonomous weapons and mass surveillance. He said he shares those same principles.

2
00:00:11.220 --> 00:00:21.850
He said he trusts Anthropic. On Saturday, February 28th, he signed a deal that doesn't include those red lines and posted about it like it was a win for safety. This isn't a story about the government.

3
00:00:22.030 --> 00:00:31.680
The government was clear about what they wanted the whole time. This is a story about a nonprofit company that will say anything in the moment to secure a deal. So what happened?

4
00:00:31.920 --> 00:00:43.120
The US Department of War offered the same deal to every major AI company. They wanted total unfettered access to all AI models for all lawful uses. There were no compromises, no carve-outs.

5
00:00:43.220 --> 00:00:54.460
The, the deal was consistent, transparent, and it was the same terms for everyone. It wasn't a surprise what the government wanted. That's an important note. There was no special deal for any one AI company.

6
00:00:54.670 --> 00:01:07.300
Anthropic said no to two specific things. They wanted an agreement that their contract would restrict surveillance of citizens and that they would not be used in fully autonomous weapons. They wanted a human in the loop.

7
00:01:07.340 --> 00:01:18.560
Anthropic wanted specific language in the contract for those two specific things. They'd been in classified networks since June 2024. That's over 18 months. They weren't new to the government.

8
00:01:18.570 --> 00:01:26.820
They weren't refusing to work with the military. They just didn't want their model to be used in life or death situations without a person with human judgment there.

9
00:01:26.860 --> 00:01:37.520
They also didn't want their model to be used to surveil people. They drew two lines and held them. Anthropic calls out why this is in their statement after negotiations went sour.

10
00:01:37.580 --> 00:01:44.220
They said, quote, "Some uses are also simply outside the bounds of what today's technology can safely and reliably do."

11
00:01:44.460 --> 00:01:57.440
Anthropic's CEO even mentions that, quote, "Frontier AI systems are simply not reliable enough to power fully autonomous weapons." These negotiations have been going on for weeks or even months. Then Friday comes along.

12
00:01:57.820 --> 00:02:05.400
Sam Altman goes on CNBC's Squawk Box. He says he agrees with Anthropic's red lines. He says he shares those principles.

13
00:02:05.440 --> 00:02:14.080
He says the Pentagon shouldn't be threatening the Defense Production Act against AI companies or any private company for that matter. He says he trusts Anthropic as a company.

14
00:02:14.660 --> 00:02:27.660
He specifically says, quote, "I think we share those red lines with Anthropic." That was Friday Sam Altman. Apparently, Saturday Sam Altman is a different person. On Saturday, we get the news that OpenAI signs the deal.

15
00:02:27.760 --> 00:02:37.720
Sam posts about safety principles, technical safeguards, and mutual agreement. The framework is still all awful use. These are the same terms Anthropic refused.

16
00:02:37.760 --> 00:02:48.549
Sam tweets, quote, "AI safety and wide distribution of benefits are the core of our mission." Something's not adding up. The timeline here is the story. Friday, he says he agrees with their principles.

17
00:02:48.750 --> 00:03:01.860
Saturday, he signs a deal without those principles. Same weekend, same man. His principles lasted a whole 24 hours. Now, let's talk about what that means. OpenAI wants you to believe that they're a safety-first company.

18
00:03:02.260 --> 00:03:12.740
They want you to believe that they still have humanity's best interests in mind. But when safety had a price tag, they signed the deal anyway. Then they dressed it up like it was the same thing as having principles.

19
00:03:12.980 --> 00:03:24.120
Instead of fighting for AI safety, they talked the talk and then cashed the check. I believe that when someone tells you who they are, you should believe them. OpenAI told us who they are in this situation.

20
00:03:24.400 --> 00:03:36.940
Sam told us what he believed on Friday. OpenAI showed us what they'll do on Saturday. He wasn't ambushed on CNBC. He volunteered those positions. He said he shares Anthropic's red lines. He said he trusts them.

21
00:03:37.360 --> 00:03:41.700
He said the government shouldn't be using emergency wartime powers against AI companies.

22
00:03:41.980 --> 00:03:52.400
But when it came time to actually stand up for those beliefs, he and his company signed a deal that didn't include those red lines and principles but then posted about it like it was a safety milestone.

23
00:03:52.940 --> 00:04:01.600
This wasn't a pivot based on new information. It was a less than 24-hour gap between what the CEO says he believes and what his company does. The deal was the same for everyone.

24
00:04:01.980 --> 00:04:11.500
This is what makes the contradiction undeniable and jarring. The government didn't offer OpenAI a better deal. They didn't change the terms. Every company got the same contract.

25
00:04:11.910 --> 00:04:21.000
Anthropic looked at it and said, "We can't really agree on these two things." OpenAI looked at the same contract, told the public, "Yeah, we agree and stand with Anthropic," yet they signed it.

26
00:04:21.220 --> 00:04:29.780
Both companies made a choice. OpenAI's safety mechanisms are the government's safety mechanisms. Read Sam's post carefully.

27
00:04:29.840 --> 00:04:39.400
He talks about the technical safeguards and principles that the DOW agrees with and reflects in law and policy. That's not OpenAI negotiating new protections.

28
00:04:39.640 --> 00:04:47.320
That's just OpenAI pointing at things the government already does and framing it as their own safety framework. The Saturday post is PR, not policy.

29
00:04:47.680 --> 00:05:00.500
Sam's tweet mentions, quote, "prohibitions on domestic mass surveillance" and, quote, "human responsibility for the use of force," which are the exact issues Anthropic refused to compromise on, but Anthropic wanted those as contractual red lines.

30
00:05:00.820 --> 00:05:07.680
OpenAI's version references existing legal authorities, meaning the law's already on the books, not new enforceable commitments.

31
00:05:07.860 --> 00:05:16.060
There's a massive difference between we negotiated hard limits and we're trusting that current law covers it. One is a guardrail. The other is a press release.

32
00:05:16.340 --> 00:05:25.670
So when you're recommending their tools to clients, it's worth understanding what it truly means now to use OpenAI's platform and how that might look in the future.

33
00:05:25.720 --> 00:05:39.288
OpenAI may genuinely believe working within the system is the better strategy.There's an argument to be had that having a seat at the table, even with imperfect terms, gives you more influence over how AI gets used than refusing and getting shut out.

34
00:05:39.368 --> 00:05:50.687
But openness is not really a signal that we've seen from the current administration here, certainly not during these negotiations. But signing and then advocating internally isn't inherently dishonest.

35
00:05:51.148 --> 00:05:59.888
It's definitely an opportunity that Anthropic is now going to miss out on. Sam's Friday comments and Saturday's deal may not necessarily be contradictory in his mind.

36
00:05:59.988 --> 00:06:07.868
He might argue that he does agree with the principles, and he really does believe that the deal actually reflects those principles through existing law.

37
00:06:07.988 --> 00:06:18.708
That's not insane, but it's just a much weaker version of the protection Anthropic was demanding. He might see referencing legal authorities as sufficient where Anthropic sees it as incomplete.

38
00:06:18.788 --> 00:06:26.968
I can see how this topic could be debated. It's not something that's totally out of the realm of possibility, but it's the timeline between the two positions that's the problem.

39
00:06:27.428 --> 00:06:38.408
Anthropic said that these two things need to be contractual red lines because we don't trust that existing law is enough for technology this new. OpenAI said that existing law is fine, and we trust you guys.

40
00:06:38.648 --> 00:06:48.348
But when Sam told us on Friday that he agrees with Anthropic's position but then signs on Saturday a version of the contract that doesn't include that position, we got a glimpse of who he is.

41
00:06:48.408 --> 00:06:57.888
It's not possible to defend the deal and defend the principles. Now, one could frame this as pragmatism, but one can't frame it as sharing the same principles.

42
00:06:58.308 --> 00:07:07.588
Those are different things, and the twenty-four-hour timeline makes it very hard to argue otherwise. More than anything, it reflects a desire to maximize profit, not benefit humanity.

43
00:07:08.168 --> 00:07:19.928
We're seeing OpenAI move from nonprofit to profit at all costs, and the way Sam positioned his Saturday post matters. He didn't say we made a difficult compromise. He didn't say we wish the deal included more.

44
00:07:20.568 --> 00:07:33.248
He posted it as a win. He posted it as if safety principles were upheld and the OpenAI mission was still intact. If he genuinely shared Anthropic's concerns, the tone of that post is misleading at best.

45
00:07:33.388 --> 00:07:43.588
You don't celebrate a deal you think is missing important protections. Most AI companies say they care about safety. It's in every mission statement, every blog post, every keynote.

46
00:07:43.668 --> 00:07:46.458
But this was the week that it actually cost something.

47
00:07:47.148 --> 00:07:58.108
Safety and privacy for American citizens and a human in the loop for fully autonomous weapons or a two hundred million dollar contract, a government relationship, and a reputation in Washington.

48
00:07:58.728 --> 00:08:08.568
Anthropic paid the price required to have standards. OpenAI signed the deal wholesale and wrote a blog post about how safe they are. The real question this raises isn't about one contract.

49
00:08:08.928 --> 00:08:11.348
It's about what AI safety means going forward.

50
00:08:11.488 --> 00:08:23.808
If the most well-funded AI company in the world can have its CEO state his principles on camera and then sign a deal contradictory to those principles the next day, then safety is just a brand word. It means nothing.

51
00:08:24.058 --> 00:08:33.408
It's a badge they wear until it gets expensive, then you awkwardly cover it up or remove it from your website. Now, I'm not here to tell you who should or shouldn't work with the government. It's not my place.

52
00:08:33.628 --> 00:08:38.808
It's also not my place to tell you which AI company to use or who the good guys are in this situation.

53
00:08:39.048 --> 00:08:46.328
There are real decisions with real trade-offs and real-life consequences, and reasonable people land on different sides of that argument.

54
00:08:46.368 --> 00:08:56.548
But I am here to tell you that saying you share someone's principles on Friday and signing a deal without those principles on Saturday isn't a compromise. It's not pragmatism.

55
00:08:57.108 --> 00:09:05.128
It's a company telling you exactly how much their values are worth. OpenAI doesn't have a safety policy. They have a communications strategy.

56
00:09:05.158 --> 00:09:14.848
And this past weekend, for exactly twenty-four hours, Sam Altman let you see the difference. I'll see you in the next video. [outro music]