What Really Happened?
Andrej Karpathy joined Anthropic on May 19, 2026. This news matters for the whole AI world. Karpathy is not just a researcher. He is a true icon in his field.
Many people think this is only about talent wars. But the real story is much more interesting. Karpathy did not come for money or fame.
His new role is very specific. He will work in the pre-training team. This is the stage where basic AI model abilities are built.
Anthropic said clearly what he will do. Karpathy will build a new team. This team will use Claude to speed up pre-training research.
What is Pre-training?
Pre-training is the stage of training a large language model. It is a very costly and complex process. Companies spend tens of millions of dollars on it.
During pre-training, the model learns language basics. It learns grammar, facts, and word relationships. This is the foundation for later skills.
Every small decision at this stage matters a lot. Even a tiny improvement can bring huge savings. That is why Anthropic wants to automate this research.
The Self-Improvement Loop
Karpathy has worked on AI research automation for a long time. After leaving OpenAI, he made a project called Auto Research. It was a simple but working prototype.
Auto Research is a system where AI improves itself. It works in four simple steps. Each step is easy to understand.
Step one: AI suggests a change. It is like a hypothesis in science. It says: „What if we do this differently?”
Step two: AI tests that change. It runs a short experiment. It checks what happens after the code change.
Step three: AI looks at the results. It compares scores before and after the change. If the result is better, the change stays. If worse, the system undoes it.
Step four: if the change works, the system keeps it. Then it starts the whole cycle again. This is a simple but very effective loop.
What Were the Auto Research Results?
Karpathy ran this system on his home computer. The system worked for about two days. During that time, it ran 700 experiments.
Out of those experiments, about 20 gave real improvements. Each improvement could be added to the previous ones. This gave a growing effect.
The most important result was shorter training time. A model the size of GPT2 trained 11% faster. From over two hours to about 1.8 hours.
This may not sound like a big change. But at the scale of large data centers, it is huge savings. Millions of dollars and weeks of time.
Karpathy himself was surprised that his first prototype worked so well. He said he was „mildly surprised.” This shows how much potential this method has.
Why Did Karpathy Choose Anthropic?
Karpathy had many options to choose from. He could go back to OpenAI, to Tesla, or start his own company. But he chose Anthropic. Why?
First, Karpathy believes in Anthropic’s vision. The company has a clear goal: safe and useful AI. This fits his own beliefs.
Second, the technical challenge was very important to him. Anthropic said: „Use Claude to improve Claude.” This sounds like a perfect task for someone like Karpathy.
Third, Karpathy wanted to return to the research front. He said before that being outside a lab weakens your sense of direction. People lose touch with what is really happening.
Karpathy once said he felt more on the side of humanity than labs. But now he changed his mind. He believes the next few years will be crucial for AI development.
What Do the Sources Say?
According to news reports, Karpathy will not be the company’s face. He will work in the „engine room” — right in the technical center. This confirms the focus is on real work, not marketing.
Anthropic could have hired Karpathy for a representative role. But they did not. They gave him a specific technical task. This says a lot about their strategy.
The company AI w Biznesie has watched such market moves for a long time. They know that automating AI research is the future. They help businesses put similar solutions to work on a smaller scale.
The Bigger Picture: A Debate in the AI Industry
Not everyone in the industry believes in AI research automation. There are two main groups. One supports the self-improvement loop. The other prefers a different approach.
The first group includes people from Anthropic and parts of OpenAI. They believe language models can improve themselves. You just need to give them tools and the right tasks.
The second group includes Demis Hassabis from Google DeepMind. He believes more in world models. These are systems that understand physics, images, and sounds.
Sergey Brin, a Google co-founder, has a different view than Demis. Brin returned to Google to work on coding AI agents. He wants Google to not fall behind.
This debate is not just academic. It decides where companies invest billions of dollars. And it decides which approach may lead to a breakthrough.
What Does This Mean for the Rest of the Market?
If Anthropic’s approach works, everything will change very fast. Models will improve themselves. Costs will go down, and abilities will go up.
Jack Clark from Anthropic said something very important. He believes by 2028 we have a 60% chance of fully automatic AI research. That means AI will create better AI without human help.
This may sound like science fiction. But Karpathy already built a prototype that works. These are not just promises — they are real results.
The company AI w Biznesie helps clients get ready for such a future. They put in place systems that automate simple research tasks. This gets companies ready for bigger changes.
What This Means for Business
Automating AI research is not just a technical curiosity. It is a real competitive advantage. Companies that understand this early will win.
Imagine that your AI system finds better ways to work on its own. You do not have to wait for a person to come up with a new idea. The system does it automatically, 24 hours a day.
Small companies can start with simple versions of such systems. Karpathy showed that even a home computer is enough for first tests. You do not need to invest millions right away.
As the technology grows, the benefits will grow too. Systems will find better and better improvements. They will work faster and cheaper.
How to Get Started?
First, understand what you are automating. It is not about replacing people. It is about speeding up simple tasks. Testing ideas, collecting data, checking results.
Second, choose the right tools. You do not have to build everything from scratch. There are ready-made frameworks you can adjust to your needs.
Third, start with small experiments. Like Karpathy — run a simple loop on a small model. See what works and what does not. Learn from your mistakes.
The company AI w Biznesie offers ready-made solutions for small and medium businesses. They help put research automation in place without big risks. It is a safe way to learn about this technology.
Summary and Conclusions
Andrej Karpathy did not join Anthropic for money or fame. He joined to work on something truly important. He wants to build a system where AI improves itself.
His Auto Research project showed this is possible. Even a simple prototype on a home computer gives real results. Imagine what you can do with billions of dollars of computing power.
In the next 6-12 months, we will see if this strategy works. If it does, the AI world will change beyond recognition. Models will get better, cheaper, and faster.
For businesses, this means one thing: you must be ready for change. Research automation is not the future — it is already here. Companies that understand this early will gain a huge advantage.
The company AI w Biznesie has helped clients put such solutions to work for years. They offer support at every stage — from first tests to full automation. They are a partner who will help you get ready for the changes ahead.
Watch what market leaders are doing. Karpathy, Anthropic, Google, and others show the direction. The rest must adapt to the new reality. The time to act is now.
No responses yet