{
  "video_id": "0pgCBV8CTZY",
  "channel_slug": "fireship",
  "channel_handle": "fireship",
  "title": "Anthropic is starting to panic…",
  "duration_seconds": 303.0,
  "url": "https://www.youtube.com/watch?v=0pgCBV8CTZY",
  "upload_date": "",
  "transcript": "Last week, Anthropic officially became\nthe apex alpha company of the artificial\nintelligence race with a valuation\nexceeding OpenAI as they filed to go\npublic with their trillion-dollar IPO\nlater this year. If you're a software\nengineer, this comes as no surprise\nbecause Claude has been the best AI\nprogrammer for years now. But despite\nthe billions of dollars flowing into\nthis company right now, they also just\nproposed something that sounds insane.\nMaybe we should wait a second and pause\nall AI development because AI is getting\ndangerously close to recursive\nself-improvement. that happens, the last\nthing humanity ever builds is the thing\nthat realizes it doesn't need humanity.\nEven if AI is benevolent and doesn't go\nrogue and kill us all, a new paper just\ndropped that believes all rational firms\nwill automate each other into a death\nspiral anyway. Looks like we're screwed\nno matter what we do, but not everybody\nout there is a doomer. And I found some\ncompelling evidence that AI might\nactually kind of suck. It is June 9th,\n2026, and you're watching The Code\nReport. Anthropic's in-house think tank\njust dropped a report, and this is the\nthesis. AI is getting dangerously close\nto recursive self-improvement. In other\nwords, they're smart enough to rewrite\ntheir code and upgrade themselves in a\nloop with no humans necessary. The\nentire industry is all gas and no\nbrakes, and they want everybody to come\ntogether and hold hands and create a\nbrake pedal. The problem is that\nAnthropic can't pause alone while\nOpenAI, DeepMind, and xAI keep\nsprinting. So unilateral pausing is off\nthe table. It's either everybody or\nnobody, and that includes China, by the\nway. Nobody's really worried about the\nEU. However, a global pause is a very\nconvenient thing for the market leader\nto advocate for because it doesn't erase\nAnthropic's lead, it freezes it right as\nthey're about to make billions of\ndollars with an IPO. If all this sounds\nfamiliar, it's not because you're crazy,\nit's because in 2019, OpenAI did the\nsame thing before the release of GPT-2.\nAt first, they said it was too dangerous\nto release, just like Anthropic is doing\nright now with Claude Mythos. But then\nthey released GPT-2, and it was totally\nfine. That was 7 years ago, and now it\nlooks like ancient technology. And\ntoday, if we are to trust these\ntrust-me-bro benchmarks, the modern\nClaude models are far better at research\nthan humans. Like 64% of the time,\nClaude Mythos is better than a human\nevery time. On top of that, AI\nresearchers are now solving problems\nthat humans haven't been able to, like\nOpenAI recently disproved a central\nconjecture in discrete geometry, which\nmathematicians have failed to do for the\nlast 80 years. The scary thing is that\nwe're already giving AI access to data\ncenters, robots, and weapons to blow\npeople up. And thanks to predictive\nprogramming in Hollywood movies, we all\nknow how that story ends. It's either\nenslavement like The Matrix or\nextermination like Terminator. I prefer\nthe latter, but there's a possibility\nfor an even dumber outcome as predicted\nby economists from Boston University in\ntheir paper, The AI Layoff Trap. At this\npoint, there's been tens of thousands of\nlayoffs in tech thanks to AI, but these\neconomists did some math and it doesn't\nlook good. Because when a firm automates\naway a worker with AI, it pockets 100%\nof the savings. But the laid-off worker\nis also a customer, and their lost\nspending doesn't just hurt the firm that\nfired them, it hurts everyone selling\nanything. Demand goes down, so the end\ngame is that firms automate their way to\ninfinite productivity and zero demand.\nThey also argue that things like UBI and\nupskilling aren't going to work, and the\nonly solution is to put a tax on\nautomation, kind of like the same way we\ntax pollution, making it cost more to\nfire people so the math stops rewarding\nthe AI race. But if there's one thing\nI've learned about economists, it's that\nthey're wrong pretty much every time. A\nthird possibility is that AI just isn't\nnearly as good as people think and never\nwill be. This is the Wall-E situation,\nwhere we keep chasing more and more\nautomation and ultimately destroy the\nplanet by building more and more data\ncenters. One piece of evidence that\nsupports that outcome is that over the\nlast couple years with the rise of\ngigantic AI, the number of new app\nreleases on the iOS App Store is nearly\ndoubled. However, it appears nobody's\nactually using these apps because app\nreviews and apps with significant usage\nare declining. In addition, this 2025\nreport from MIT analyzed over 300\nenterprises implementing AI, and even\nthough they spent over $3000000000\ncollectively, the end result was that\n95% of their projects delivered zero\nmeasurable revenue impact or return on\ninvestment. That doesn't look good, but\nluckily there are tools that can help\nyou avoid failure like Pioneer, the\nsponsor of today's video. If you're\ncalling a frontier model for every LLM\nrequest in your app, it's probably\nburning through a bonfire of tokens just\nto return generic results. But Pioneer\nsolves this by giving you an inference\nAPI that you can plug into your existing\nLLM setup to handle all model routing\nand optimization for you. And it will\ncluster your app's traffic by use case\nto discover where your current model is\nbeing too slow, too expensive, or too\nstupid. Then it trains a fleet of\nsmaller open-source models in the\nbackground and alerts you when it finds\none that's cheaper and better, so you\ncan easily swap it under the hood.\nPioneer works great with Claude Code,\nCodex, Cursor, Hermes, or anything else\nhitting an LLM endpoint. And you can get\n$30 of inference for just $5 today. This\nhas been the Code Report. Thanks for\nwatching, and I will see you [music] in\nthe next one.",
  "transcript_chars": 5785,
  "ingested_at": "2026-06-17T04:31:47.045297+00:00",
  "source": "channel",
  "yt_meta": {
    "view_count": null,
    "like_count": null,
    "channel_id": null,
    "categories": null,
    "tags": null
  }
}