When AI builds itself

@mattpaige68

63.8K followers · usually 932 views a post

Views

4.6K

Likes

145

Comments

21

Shares

23

Selling · SaaS › AI Tool

Anthropicseen in 22 videos →

Why it worked

The video taps into the public's fascination and fear surrounding advanced AI, specifically the concept of recursive self-improvement, which is a common theme in science fiction. It presents compelling data and connects it to a major company (Anthropic) on the verge of a significant financial event (IPO), creating a sense of urgency and importance.

Summary

The video discusses Anthropic's AI, Claude, and its rapid advancement towards recursive self-improvement. It highlights that Claude now writes 80% of Anthropic's code, can work for 12 hours straight, and outperforms humans in judgment tasks, suggesting a significant acceleration in AI development.

Structure

1AI building smarter AI
2Claude's code contribution and task duration
3AI outperforming humans in coding and judgment
4The concept of recursive self-improvement
5Implications for tech and human history
6Anthropic's upcoming IPO and potential for over-selling

Original caption

anthropic admitted their AI is starting to build a smarter version of itself, and it's happening faster than they expected. 80% of their code is written by claude now. Today it's 12 hrs straight. They are saying it maybe recursive self-improvement. The thing every sci-fi movie warned us about. but they're also about to go public. so is this the most unsettling moment in tech history, or perfect timing right before the IPO? where do you land? #anthropic #claude #ai

More from @mattpaige68

Their other posts in the index, biggest breakout first.

Product placement

The video discusses Anthropic and its AI model, Claude, presenting data and charts related to Claude's performance and code contribution.

Call to action

Comment "Anthropic" for the article

On-screen text

When AI builds itself Our progress toward recursive self-improvement, and its implications. Is AI Approaching Recursive Self Improvement? Save this... Comment "Anthropic" for the article Claude Code session success rate Code contributed per person, by quarter Where a researcher went wrong, could Claude have done better? Recursive Self Improvement

Transcript

Anthropic just admitted that their own AI is starting to build a smarter version of itself, and it's happening faster than they expected. The numbers are so stick with me. Right now, 80% of code inside Anthropic is written by Claude, not by humans. But two years ago, Claude could work on a task for about four minutes. Today, it's 12 hours straight. And that number's doubling every four months. In one stretch, it shifted over 800 fixes that cut a class of AI errors by 1,000 times. And an engineer overseeing it estimated that a human would have taken four years to do this work. And on the hard research calls, the judgment work that humans actually get paid for, their newest models now beat those humans 64% of the time. That number was two years ago, that number was 22%. And you get the exact thing that every sci-fi movie warned us about. AI that's good enough to help build the next smarter AI, which builds the next one even faster. Because one AI is doing the building, you're not capped by how many engineers you can hire. You can point thousands of AI agents at the problem at once. So every loop runs even quicker than the last. Recursive self-improvement. Anthropic is saying we're getting close. This one that we're watching the most important and genuinely unsettling moment in tech history, maybe human history. Number two, a company that's about to go public, just found a very compelling story to tell investors. And to be fair, Anthropic admits themselves that some of their numbers over sell this a bit. So where do you land? Is this recursive self-improvement happening before our eyes, or is it before the IPO? Let me know in the comments.