Diploma Be Gone: How the 'Wild Path' Leads to OpenAI and Anthropic

Forget everything you've been told about needing a PhD to work in the top tier of artificial intelligence. If once a ticket to the exclusive club of OpenAI or DeepMind was a stack of publications at top conferences like NeurIPS, today the rules have changed so much that yesterday's academics are increasingly giving way to those whom China calls "elüzi." This term means "wild path" or the path of a self-taught person, and it is precisely such people who are now becoming the backbone of teams creating next-generation models.

For a long time, it was believed that entry into the AI industry was protected by an insurmountable wall of academic credentials. You had to spend years studying advanced mathematics, writing papers and waiting for reviews, just to get a chance at an interview. But here's the paradox: the most impressive successes of recent years are not a victory of pure theory, but a triumph of scaling (Scaling Laws).

It turned out that to create GPT-4, you need not so much to invent a new type of neuron, but to possess phenomenal skills in infrastructure, data preparation and computation optimization. This is work for hackers and engineers, not theoreticians. Companies like OpenAI realized this before others.

In their staff, people are increasingly appearing who dropped out of university or never attended at all, but have spent years digging through open source code and participating in competitions on Kaggle. These "wild path" specialists possess a quality that Stanford graduates often lack—they know how to make code work under conditions of limited resources. In an era when the cost of training a model is measured in hundreds of millions of dollars, any engineer's mistake is too expensive.

Here those are valued who understand how "hardware" interacts with software at the deepest level. The paradigm shift occurred because AI has transformed from a scientific discipline into an industrial sector. Once aviation was also the domain of lone scientists, but today planes are built by huge factories.

In AI, the "factory" phase has arrived. To train a model like Claude 3, you don't need to rewrite the transformer architecture. You need to be able to distribute the load across thousands of graphics processors so they don't sit idle for a second.

That's why infrastructure engineers and data specialists have become the new rock stars, and their salaries have come in line with the incomes of leading researchers. Moreover, the development format itself has changed. OpenAI and its competitors are becoming increasingly closed.

They no longer publish detailed articles about their methods, fearing competition. In this atmosphere of secrecy, the academic skill of writing beautiful reports in LaTeX becomes useless. What becomes more important is the internal "kitchen": how exactly you clean data from noise and what tricks you use in reinforcement learning training (RLHF).

These skills are not taught in universities; they can only be acquired in practice, working on real projects. What does this mean for the future? We are witnessing a democratization of talent while simultaneously monopolizing resources.

The threshold for entry in terms of knowledge is lowered: you no longer need to spend seven years on a doctoral dissertation. However, demands on practical skills are growing exponentially. If you can show a working project that solves a specific optimization or data processing problem, your GitHub will mean more to a recruiter from Silicon Valley than any diploma.

The bottom line: the era of theorists in AI is being replaced by the era of builders. Will universities be able to adapt to a world where practice is years ahead of theory?