Выпуск
AI Is Now Improving Itself. We're Running Out of Time?
AsianDadEnergy's Podcast · 17 ноября 2026 г. · 12 мин
Something rather extraordinary may be happening inside the world’s leading AI laboratories. AI models are increasingly being used to build, train, evaluate, and improve the next generation of AI models. And apparently, they’re getting pretty good at it. This process is known as recursive self-improvement , and it could fundamentally change the trajectory of artificial intelligence. In fact, according to a recent report from METR, an independent organization that evaluates AI capabilities, the rate of AI progress at Anthropic has accelerated dramatically over the past year. If these claims are accurate, then the implications are potentially enormous. Some researchers and industry observers believe we could be just one or two years away from Artificial General Intelligence (AGI), or even Artificial Superintelligence (ASI). In other words, we might be approaching a point where machines can outperform humans across virtually every intellectual endeavor. By 2027. Now, I’m not going to lie. My spidey senses are tingling. The idea that Claude is about to transform into some omniscient machine god in the next 12 months sounds suspiciously like something a company preparing for an enormous IPO would want investors to believe. But I also don’t want my natural tech cynicism to get the better of me. Because when you actually examine the mechanics of recursive self-improvement, there are some very compelling reasons to believe that AI progress could accelerate substantially. And there are also some very important reasons to believe that this acceleration will eventually hit a wall. Let’s unpack this. 1. AI Is Starting to Accelerate Its Own Development First, let’s talk about a rather interesting report published by METR. METR is an independent organization best known for its Time Horizon benchmark, which measures how long AI systems can independently work on complex technical tasks. Think software engineering, programming, and other technically demanding activities. A team from METR spent several months working alongside Anthropic to evaluate the capabilities and development of its AI models. One of their findings was particularly interesting. Based on Anthropic’s internal benchmarks, the rate of AI capability improvement had roughly doubled over the preceding year. In practical terms, AI was gaining in one year what previously took almost two years of progress according to their own measuring standards. Even more interestingly, the METR team attributed approximately 80% of this acceleration to AI helping accelerate AI research itself. Think about that for a moment. We’re no longer talking exclusively about humans using AI to write code, answer questions, or automate office work. We’re talking about AI systems helping researchers develop better AI systems. And that brings us to the concept of recursive self-improvement. Recursive self-improvement is the process by which an AI system helps improve its own capabilities, creating increasingly powerful versions of itself. The theoretical feedback loop looks something like this: * Humans develop an AI model. * That model helps researchers improve the next generation of AI. * The improved model becomes even better at AI research. * That model helps develop an even more capable successor. * Repeat. In theory, every generation becomes better at building the next generation. And if this feedback loop becomes sufficiently powerful, AI development could enter a period of explosive acceleration. This is the basic idea behind an intelligence explosion. 2. An Anthropic Whistleblower Makes Some Extraordinary Claims Around the same time as the METR report, a former Anthropic employee named Jacob Coxon made some rather extraordinary claims.<