Elon Says We're in the Singularity. Here's Why.

Back in March, Andrej Karpathy ran a small experiment. He is one of the founders of OpenAI and the former head of AI at Tesla. He took his own small AI project and let an AI work on it alone for about 2 days.

The AI tried about 700 changes by itself. It kept about 20. Together, those changes made his AI learn about 11% faster.

He posted the result, and 87 minutes later Elon Musk replied with five words: "We are in the Singularity."

So why would one small speed up get that reaction? Because when AI improves AI, every improvement brings the next one sooner. I think how this loop works is way more interesting than the hype or the doom.

What the words mean

Normally, technology only gets better when people make it better. AI has worked that way too. Recursive self-improvement is when the AI starts doing part of that improving itself.

In plain words, it is AI that gets better at building better AI. It writes better code, the instructions that tell a computer what to do. Or it finds better ways for AI to learn. Then the improved AI is even better at improving AI. Recursive just means the loop feeds back into itself.

That one idea decides how fast the AI on your phone gets smarter and cheaper, and how soon it shows up in things like new medicines.

The idea is not new. In 1965, the mathematician I.J. Good wrote that a machine smarter than us could design even better machines. In his words: "There would then unquestionably be an intelligence explosion." For about 60 years, that was an idea on paper. Now it is showing up in actual numbers.

The weird savings account

The best way to picture this is compound interest. Put $100 in a savings account that pays 10% a year. In year two you earn $11 instead of $10, because now you are earning interest on your interest. That is normal progress, with humans improving AI at a set pace.

Now picture a weird account where the interest rate itself goes up every year, from 10% to 20% to 30%. After 3 years, the normal account has about $133. The weird one has about $172, and the gap keeps growing from there.

That weird account is recursive self-improvement. The thing being improved is also the thing doing the improving.

That is what makes AI different from electricity or the combustion engine. Both changed the world. Neither one ever got better on its own. AI can make itself better. So I think what is coming is bigger than both of them combined.

Every loop needs a judge

How does a machine know whether its own change is any good? It uses a judge.

First, the AI tries something, like a move in a game or a change to some code. Then an automatic judge scores it. Did it win the game? Did the code run faster? The AI keeps the winners and tries again, millions of times. The faster the judge, the faster the whole loop spins.

The cleanest picture of this is DeepMind's AlphaGo Zero from 2017. It learned the board game Go knowing only the rules, by playing against itself 4.9 million times. After 3 days, it beat the version of AlphaGo that had beaten world champion Lee Sedol, 100 games to zero.

Karpathy's experiment is the same loop pointed at AI itself. Training is the long process where an AI learns from examples. His AI kept trying changes to how his project trains. The judge was basically a stopwatch.

To me, this is the moment we went from humans tuning the AI to AI tuning the AI. And Karpathy said: "All LLM frontier labs will do this. It's the final boss battle."

He is saying every major AI lab is going to run this loop on its own AI. Any goal you can measure can be handed to a swarm of AIs, as long as you have a quick way to test it.

The loop is already running

Google built an AI called AlphaEvolve that runs that same loop. Gemini, the AI behind Google's chatbot, suggests changes to code. An automatic checker scores every version, and the best ones get kept and improved again.

Then Google pointed it at Gemini itself. AlphaEvolve sped up one piece of the code that trains Gemini by 23%, and that cut Gemini's total training time by 1%. Gemini is the exact AI that AlphaEvolve runs on.

1% sounds tiny. Remember the savings account. That 1% is a deposit the next version gets to build on.

At Anthropic, the company behind Claude, more than 80% of the code it adds to its own software is now written by Claude. Before early 2025, that number was in the low single digits. In about 15 months, the AI went from writing almost none of its company's code to writing most of it.

OpenAI says that last month it hit its goal of an automated AI research intern. That is an AI that can carry out research tasks under human direction, including tasks that would take a skilled researcher a few days.

The night they announced it, I posted: "We're totally in the singularity. Every week is insane." The singularity is the point where progress moves faster than people can keep up. The pieces of the loop are already here.

Measuring the curve

A research group called METR tests how long a task an AI can finish on its own, about half the time. They size each task by how long it takes a human expert.

In early 2023, the best AI handled tasks that take a person about 4 minutes. By this spring, it was at least 16 hours. That is two full work days. AI went from answering quick questions to handling a couple of days of work.

And the speed up is speeding up. In March of 2025, METR said this number was doubling about every 7 months. Since 2023, it has been doubling about every 4 months.

That is the weird savings account, where the interest rate itself keeps climbing. Most of us think in straight lines, so this is the part people miss. A loop like this curves upward, and it curves faster every year.

Four brakes

If this thing compounds that fast, why hasn't it exploded already? Because the loop has brakes. This is where the savings account comparison stops working. A bank pays you interest for free. This loop has to pay for every single round.

The first brake is computing power, because every new idea has to be tested.

The second brake is electricity. The International Energy Agency expects data centers, the giant buildings full of computers that run AI, to more than double their electricity use by 2030.

The third brake is the real world. An AI can come up with a new drug idea, but that drug still has to be tested in real cells and real people. As Anthropic put it, more intelligence can't learn what a drug does over decades of use.

The fourth brake is the judge. The loop runs fastest where there is a quick automatic score, like code speed or a math answer. And people still choose which problems the AI works on.

So the speed up is real, but it hasn't gone vertical yet. The biggest brakes are physical things, like chip factories and power plants. That is why I look at AI, robots and energy as one flywheel, a wheel that gets easier to spin the faster it goes. As robots and cheap energy take those brakes off, the loop spins faster, and every turn makes the next one easier.

The price of a thought

If you don't work in AI, why should you care? Because of what this does to the price of intelligence.

Epoch AI, a group that tracks AI trends, found the same level of AI ability has been getting about 13 times cheaper every year since 2023. For one science test, Epoch found a 725-fold price drop in under 18 months and put it like this: "It is like the sticker price on a new car falling from $50,000 to $69." Brain power that cost you a dollar last year costs you about 8 cents this year.

Cheap intelligence can be pointed at problems we have been stuck on for decades. Google DeepMind's AlphaFold predicted the shape of virtually all 200 million proteins that scientists have identified. Proteins are the tiny machines doing the work inside your body. Knowing their shape helps scientists design medicines that fit them.

Put the two together and you get a loop that keeps making AI smarter while the price keeps crashing. My guess is that for you, that means better medicines and cheaper energy showing up years sooner than anyone planned.

Where it leads

Sam Altman, the CEO of OpenAI, called this stage "a larval version of recursive self-improvement." Larval means the baby stage, like a caterpillar before it becomes a butterfly.

Dario Amodei, the CEO of Anthropic, describes where it is heading as "a country of geniuses in a data center." He means millions of AI copies, each smarter than a Nobel Prize winner, working way faster than any person. His bet is that this squeezes the next 50 to 100 years of progress in biology into 5 to 10 years.

As for me, I'm with Elon on this one. The loop has already started. My projection is that by the end of 2027, AI does about 80% of digital tasks at or above the level of the top 20% of people. And by 2030, the world looks dramatically different.

The impossible becomes inevitable faster than critics expect, but slower than I thought. Recursive self-improvement is the reason it keeps beating the critics.

So think back to that Karpathy post. One AI, working alone for about two days, made his AI learn about 11% faster. That was one person's small project, and Elon answered it with five words. Now Google, Anthropic and OpenAI are running that same loop on the AI you use every day.

The next time the AI on your phone suddenly gets smarter and cheaper, that is the loop paying interest.