[0:00] Last week, Anthropic officially became
[0:02] the Apex Alpha company of the artificial
[0:04] intelligence race with a valuation
[0:07] exceeding OpenAI as they filed to go
[0:09] public with their trillion dollar IPO
[0:11] later this year. If you're a software
[0:12] engineer, this comes as no surprise
[0:14] because Claude has been the best AI
[0:16] programmer for years now. But despite
[0:18] the billions of dollars flowing into
[0:19] this company right now, they also just
[0:21] proposed something that sounds insane.
[0:23] Maybe we should wait a second and pause
[0:25] all AI development because AI is getting
[0:27] dangerously close to recursive
[0:29] self-improvement. And when that happens,
[0:31] the last thing humanity ever builds is
[0:33] the thing that realizes it doesn't need
[0:35] humanity. Even if AI is benevolent and
[0:37] doesn't go rogue and kill us all, a new
[0:39] paper just dropped that believes all
[0:41] rational firms will automate each other
[0:43] into a death spiral anyway. It looks
[0:45] like we're screwed no matter what we do.
[0:47] But not everybody out there is a doomer.
[0:49] And I found some compelling evidence
[0:50] that AI might actually kind of suck. It
[0:53] is June 9th, 2026, and you're watching
[0:55] the code report. Anthropic's in-house
[0:57] think tank just dropped a report, and
[0:59] this is the thesis. AI is getting
[1:01] dangerously close to recursive
[1:02] self-improvement. In other words,
[1:04] they're smart enough to rewrite their
[1:06] code and upgrade themselves in a loop
[1:07] with no humans necessary. The entire
[1:09] industry is all gas and no breaks, and
[1:11] they want everybody to come together and
[1:13] hold hands and create a brake pedal. The
[1:15] problem is that Anthropic can't pause
[1:17] alone. While OpenAI, Deep Mind, and XAI
[1:20] keep sprinting. So, unilateral pausing
[1:21] is off the table. It's either everybody
[1:24] or nobody. And that includes China, by
[1:26] the way. And nobody's really worried
[1:27] about the EU. However, a global pause is
[1:29] a very convenient thing for the market
[1:31] leader to advocate for. Because it
[1:33] doesn't erase Anthropic's lead, it
[1:35] freezes it right as they're about to
[1:36] make billions of dollars with an IPO. If
[1:38] all this sounds familiar, it's not
[1:40] because you're crazy. It's because in
[1:42] 2019, OpenAI did the same thing before
[1:44] the release of GPT2. At first, they said
[1:47] it was too dangerous to release it, just
[1:49] like Anthropic is doing right now with
[1:50] Claude Mythos. But then they released
[1:52] GPT2 and it was totally fine. That was 7
[1:55] years ago, and now it looks like ancient
[1:56] technology. Today, if we are to trust
[1:59] these Trust Me Bro benchmarks, the
[2:00] modern Claude models are far better at
[2:02] research than humans. Like 64% of the
[2:05] time, a Claude mythos is better than a
[2:07] human every time. On top of that, AI
[2:09] researchers are now solving problems
[2:11] that humans haven't been able to. Like
[2:13] OpenAI recently disproved a central
[2:15] conjecture and discrete geometry, which
[2:17] mathematicians have failed to do for the
[2:19] last 80 years. The scary thing is that
[2:21] we're already giving AI access to data
[2:23] centers, robots, and weapons to blow
[2:25] people up. And thanks to predictive
[2:27] programming in Hollywood movies, we all
[2:29] know how that story ends. It's either
[2:30] enslavement like the Matrix or
[2:32] extermination like Terminator. I'd
[2:34] prefer the latter, but there's a
[2:36] possibility for an even dumber outcome.
[2:38] as predicted by economists from Boston
[2:40] University in their paper, the AI layoff
[2:42] trap. At this point, there's been tens
[2:44] of thousands of layoffs in tech thanks
[2:46] to AI. But these economists did some
[2:48] math, and it doesn't look good. Because
[2:50] when a firm automates away a worker with
[2:52] AI, it pockets 100% of the savings. But
[2:55] the laid-off worker is also a customer.
[2:57] And their loss spending doesn't just
[2:59] hurt the firm that fired them, it hurts
[3:00] everyone selling anything. Demand goes
[3:02] down. So the endgame is that firms
[3:05] automate their way to infinite
[3:06] productivity and zero demand. They also
[3:08] argue that things like UBI and
[3:10] upskilling aren't going to work and the
[3:12] only solution is to put a tax on
[3:14] automation. It kind of like the same way
[3:15] we tax pollution, making it cost more to
[3:18] fire people. So the math stops rewarding
[3:20] the AI race. But if there's one thing
[3:21] I've learned about economists, it's that
[3:23] they're wrong pretty much every time. A
[3:25] third possibility is that AI just isn't
[3:27] nearly as good as people think and never
[3:29] will be. This is the Wall-E situation
[3:31] where we keep chasing more and more
[3:33] automation and ultimately destroy the
[3:35] planet by building more and more data
[3:36] centers. One piece of evidence that
[3:38] supports that outcome is that over the
[3:40] last couple years with the rise of
[3:41] Agentic AI, the number of new app
[3:43] releases on the iOS app store is nearly
[3:45] doubled. However, it appears nobody's
[3:47] actually using these apps because app
[3:49] reviews and apps with significant usage
[3:51] are declining. In addition, this 2025
[3:54] report from MIT analyzed over 300
[3:57] enterprises implementing AI. And even
[3:59] though they spent over $30 billion
[4:00] collectively, the end result was that
[4:03] 95% of their projects delivered zero
[4:05] measurable revenue impact or return on
[4:07] investment. That doesn't look good. But
[4:09] luckily, there are tools that can help
[4:11] you avoid failure, like Pioneer, the
[4:13] sponsor of today's video. If you're
[4:15] calling a Frontier model for every LLM
[4:17] request in your app, it's probably
[4:19] burning through a bonfire of tokens just
[4:21] to return generic results. But Pioneer
[4:23] solves this by giving you an inference
[4:25] API that you can plug into your existing
[4:27] LLM setup to handle all model routing
[4:30] and optimization for you. And it will
[4:32] cluster your app's traffic by use case
[4:34] to discover where your current model is
[4:36] being too slow, too expensive, or too
[4:38] stupid. Then it trains a fleet of
[4:40] smaller open- source models in the
[4:42] background and alerts you when it finds
[4:43] one that's cheaper and better so you can
[4:46] easily swap it under the hood. But
[4:47] Pioneer works great with Claude Code,
[4:50] Codeex, Cursor, Hermes, or anything else
[4:52] hitting an LLM endpoint. And you can get
[4:55] $30 of inference for just $5 today. But
[4:58] this has been the code report. Thanks
[4:59] for watching and I will see you in the
[5:01] next one.