It's Beginning: AI Improves Itself

[00:00] Artificial intelligence companies are  now asking a question that used to   belong to science fiction writers and  slightly over-caffeinated bloggers:   what happens when AI begins to improve  AI? Because it is beginning to happen.

[00:16] I'm always looking for new products that  you might find useful. And today's sponsor   Consensus is one that you should definitely have  a look at because it solved a big problem for me.  

[00:28] My problem with using AI for research is  that if you ask a chatbot about science,   it'll happily invent a source, present it with  total confidence and sound completely convincing,  

[00:40] except the paper doesn't exist. This is why  Consensus works the other way around. It searches   220 million peer-reviewed papers first and then  uses AI to summarize what it actually found. That  

[00:56] way, every claim links to a real verifiable study.  I particularly like the deep search function,   which shows you the consensus across the whole  literature. You can see where studies agree,  

[01:09] where they conflict, and how the research has  changed over time. It also integrates with Claude.   Use the link in the description for a full week of  Consensus free. And now back to the science news.

[01:23] That AI improves itself is widely regarded  as the beginning of the singularity,   aka the intelligence explosion, after which  AI will see a rapid self-powered improvement  

[01:35] that will either turn the world into a  paradise or hell, depending on who you ask. Jimmy Ba, the cofounder of xAI, recently tweeted  that: “Recursive self-improvement loops likely do  

[01:48] live in the next 12 months.” And that “2026  is gonna be insane and likely the busiest   and most consequential year for the future of  our species.” Dario Amodei, CEO of Anthropic,  

[02:02] one of the world’s leading AI firms, also thinks  it’s the biggest thing to watch out for right now,   and I agree. [Dario Amodei:] “I think  the biggest thing to watch is this issue   of AI systems building AI systems. How that  goes, whether that goes one way or another,  

[02:19] that will determine, you know, whether  it's a few more years until we get there,   or if we have wonders and a great emergency  in front of us that we have to face.”

[02:33] Indeed, Anthropic just created an internal  think tank as a “fire drill” to prepare for   the intelligence explosion and to look  for “early warning signals for recursive  

[02:45] self-improvement”. They called it the “Anthropic  Institute”. It’s led by co-founder Jack Clark   who put the odds above sixty percent that  by twenty twenty-eight an AI system could  

[02:57] be told to make a better version of  itself, and then do it autonomously.  However, the closer we have come to AI  self-improvement, the more obvious it’s become   that it’s an incredibly vague and useless term.  One can argue that if an AI researcher uses AI  

[03:15] in their research, then that counts as a sort of  self-improvement of AI. Though seeing how things   are going with the scientific literature,  I think that’s an intelligence implosion.

[03:27] But I digress. My point is that there are many  ways AI can self-improve in some sense and yet   it’s not the sense that anyone had in mind  when they were thinking of the intelligence   explosion. It was supposed to be a runaway  effect, caused by AI improving their own code,  

[03:44] that leads to exponential or super  exponential intelligence increase. I already talked last year about the beginning of  self-improving AI with Large Language Models that  

[03:57] tune their own hyperparameters or, on the more  applied side, AI that has improved microchips   used for AI training. Self-improvement, yes,  but not the runaway effect we’re waiting for.

[04:11] But, these examples have multiplied in  the past months, and some of these AIs   do now indeed write new AI code. The clearest  recent example is probably Andrej Karpathy’s  

[04:25] AutoResearch. It is a small Python setup  where an artificial intelligence agent   modifies language-model training code, trains  the model for 5 minutes, checks whether the   result improved, keeps or discards the change,  and repeats. Meta’s HyperAgents work like this,  

[04:42] too. These are LLM-based research agents that  write code, run experiments, debug failures,   keep what works and toss what doesn’t. These are not yet models directly rewriting  

[04:55] their own brains. But it is artificial  intelligence developing new artificial   intelligence, albeit on a small scale. We are  indeed getting closer to the self-recursive part.

[05:07] METR, a nonprofit institute for Model Evaluation  and Threat Research is keeping track of this   closely. In November they asked whether we  have already reached the point where AI can  

[05:20] do AI research well enough to speed up the  next AI? To answer this question, they tested   GPT-5.1-Codex-Max on some research-engineering  tasks, like fixing a damaged small language  

[05:34] model. In November their answer was: we have  not yet reached the self-acceleration point. But that was in November, and at the  pace that things are going, that's   basically the stone age. Already in February, one  METR researcher published a model predicting that,  

[05:51] if current trends continue, AI could automate  more than 99 percent of AI research and   development around 2032. Personally, I think  that it will happen much sooner than that.

[06:03] So this is no longer just “someone  on the internet says the singularity   is near”. The people who work in the  field are now discussing when and how   the recursive loop is going to close  and how to prepare for that event.

[06:19] That said, so far, artificial intelligence has  not replaced scientists. It has merely automated   the part where we try 700 things that don’t  work and pretend this was the plan all along.

[06:33] Thanks for watching. See you tomorrow.