Anthropic is starting to panic…

[00:00] Last week, Anthropic officially became

[00:02] the Apex Alpha company of the artificial

[00:04] intelligence race with a valuation

[00:07] exceeding OpenAI as they filed to go

[00:09] public with their trillion dollar IPO

[00:11] later this year. If you're a software

[00:12] engineer, this comes as no surprise

[00:14] because Claude has been the best AI

[00:16] programmer for years now. But despite

[00:18] the billions of dollars flowing into

[00:19] this company right now, they also just

[00:21] proposed something that sounds insane.

[00:23] Maybe we should wait a second and pause

[00:25] all AI development because AI is getting

[00:27] dangerously close to recursive

[00:29] self-improvement. And when that happens,

[00:31] the last thing humanity ever builds is

[00:33] the thing that realizes it doesn't need

[00:35] humanity. Even if AI is benevolent and

[00:37] doesn't go rogue and kill us all, a new

[00:39] paper just dropped that believes all

[00:41] rational firms will automate each other

[00:43] into a death spiral anyway. It looks

[00:45] like we're screwed no matter what we do.

[00:47] But not everybody out there is a doomer.

[00:49] And I found some compelling evidence

[00:50] that AI might actually kind of suck. It

[00:53] is June 9th, 2026, and you're watching

[00:55] the code report. Anthropic's in-house

[00:57] think tank just dropped a report, and

[00:59] this is the thesis. AI is getting

[01:01] dangerously close to recursive

[01:02] self-improvement. In other words,

[01:04] they're smart enough to rewrite their

[01:06] code and upgrade themselves in a loop

[01:07] with no humans necessary. The entire

[01:09] industry is all gas and no breaks, and

[01:11] they want everybody to come together and

[01:13] hold hands and create a brake pedal. The

[01:15] problem is that Anthropic can't pause

[01:17] alone. While OpenAI, Deep Mind, and XAI

[01:20] keep sprinting. So, unilateral pausing

[01:21] is off the table. It's either everybody

[01:24] or nobody. And that includes China, by

[01:26] the way. And nobody's really worried

[01:27] about the EU. However, a global pause is

[01:29] a very convenient thing for the market

[01:31] leader to advocate for. Because it

[01:33] doesn't erase Anthropic's lead, it

[01:35] freezes it right as they're about to

[01:36] make billions of dollars with an IPO. If

[01:38] all this sounds familiar, it's not

[01:40] because you're crazy. It's because in

[01:42] 2019, OpenAI did the same thing before

[01:44] the release of GPT2. At first, they said

[01:47] it was too dangerous to release it, just

[01:49] like Anthropic is doing right now with

[01:50] Claude Mythos. But then they released

[01:52] GPT2 and it was totally fine. That was 7

[01:55] years ago, and now it looks like ancient

[01:56] technology. Today, if we are to trust

[01:59] these Trust Me Bro benchmarks, the

[02:00] modern Claude models are far better at

[02:02] research than humans. Like 64% of the

[02:05] time, a Claude mythos is better than a

[02:07] human every time. On top of that, AI

[02:09] researchers are now solving problems

[02:11] that humans haven't been able to. Like

[02:13] OpenAI recently disproved a central

[02:15] conjecture and discrete geometry, which

[02:17] mathematicians have failed to do for the

[02:19] last 80 years. The scary thing is that

[02:21] we're already giving AI access to data

[02:23] centers, robots, and weapons to blow

[02:25] people up. And thanks to predictive

[02:27] programming in Hollywood movies, we all

[02:29] know how that story ends. It's either

[02:30] enslavement like the Matrix or

[02:32] extermination like Terminator. I'd

[02:34] prefer the latter, but there's a

[02:36] possibility for an even dumber outcome.

[02:38] as predicted by economists from Boston

[02:40] University in their paper, the AI layoff

[02:42] trap. At this point, there's been tens

[02:44] of thousands of layoffs in tech thanks

[02:46] to AI. But these economists did some

[02:48] math, and it doesn't look good. Because

[02:50] when a firm automates away a worker with

[02:52] AI, it pockets 100% of the savings. But

[02:55] the laid-off worker is also a customer.

[02:57] And their loss spending doesn't just

[02:59] hurt the firm that fired them, it hurts

[03:00] everyone selling anything. Demand goes

[03:02] down. So the endgame is that firms

[03:05] automate their way to infinite

[03:06] productivity and zero demand. They also

[03:08] argue that things like UBI and

[03:10] upskilling aren't going to work and the

[03:12] only solution is to put a tax on

[03:14] automation. It kind of like the same way

[03:15] we tax pollution, making it cost more to

[03:18] fire people. So the math stops rewarding

[03:20] the AI race. But if there's one thing

[03:21] I've learned about economists, it's that

[03:23] they're wrong pretty much every time. A

[03:25] third possibility is that AI just isn't

[03:27] nearly as good as people think and never

[03:29] will be. This is the Wall-E situation

[03:31] where we keep chasing more and more

[03:33] automation and ultimately destroy the

[03:35] planet by building more and more data

[03:36] centers. One piece of evidence that

[03:38] supports that outcome is that over the

[03:40] last couple years with the rise of

[03:41] Agentic AI, the number of new app

[03:43] releases on the iOS app store is nearly

[03:45] doubled. However, it appears nobody's

[03:47] actually using these apps because app

[03:49] reviews and apps with significant usage

[03:51] are declining. In addition, this 2025

[03:54] report from MIT analyzed over 300

[03:57] enterprises implementing AI. And even

[03:59] though they spent over $30 billion

[04:00] collectively, the end result was that

[04:03] 95% of their projects delivered zero

[04:05] measurable revenue impact or return on

[04:07] investment. That doesn't look good. But

[04:09] luckily, there are tools that can help

[04:11] you avoid failure, like Pioneer, the

[04:13] sponsor of today's video. If you're

[04:15] calling a Frontier model for every LLM

[04:17] request in your app, it's probably

[04:19] burning through a bonfire of tokens just

[04:21] to return generic results. But Pioneer

[04:23] solves this by giving you an inference

[04:25] API that you can plug into your existing

[04:27] LLM setup to handle all model routing

[04:30] and optimization for you. And it will

[04:32] cluster your app's traffic by use case

[04:34] to discover where your current model is

[04:36] being too slow, too expensive, or too

[04:38] stupid. Then it trains a fleet of

[04:40] smaller open- source models in the

[04:42] background and alerts you when it finds

[04:43] one that's cheaper and better so you can

[04:46] easily swap it under the hood. But

[04:47] Pioneer works great with Claude Code,

[04:50] Codeex, Cursor, Hermes, or anything else

[04:52] hitting an LLM endpoint. And you can get

[04:55] $30 of inference for just $5 today. But

[04:58] this has been the code report. Thanks

[04:59] for watching and I will see you in the

[05:01] next one.