🐔 Don’t trust “AI experts”.

The term “AI expert” is largely a marketing label, often applied to random individuals to drive SaaS sales, secure funding, obtain promotions, or increase ad revenue. Those things, as always, come in the way of real facts.

Measuring expertise precisely is challenging, but I would argue that 99.99% of those who self-identify as “AI experts”, along with 99% of those labeled as such by others, fall into the bottom 30% of the field in terms of true expertise.

As more people are called experts, the value of the term diminishes—a clear case of inflation. The biggest contributors to this inflation and its undesirable effects IMO are business “AI experts” that have barely ever touched a line of code.

An “AI expert” from 10 years ago is vastly different from one today. The field has evolved and grown so much that 1) the term “AI expert” has lost much of its meaning and 2) the requirements to qualify as an AI expert have accumulated beyond what a person can reasonably achieve.

I would call someone an “AI expert”, if (among other things, and in loosely decreasing order of difficulty):

• They spent >20 years, >70 hours a week, focused on AI
• They can build a GPT3.5 level LLM with 50,000 H100 GPUs, a 10-person team, and six months of preparation
• They are recognized as an expert by >50k people

• 50% of their predictions about AI in the last 10 years have been correct

• They can teach a course on (all of) NLP, RL, CV, AI in Robotics, etc., with just one month of preparation
• They master frameworks like TensorFlow or PyTorch
• They can write efficient CUDA kernels
• They can write a Monte Carlo simulation from scratch
• They can describe gradient descent on the spot

In other words, they are a rather rare species. Good news is you NEVER need such a profile. 🎉

Ultimately, the only real AI experts that will ever exist are AI itself. The rest of us are merely trying to keep pace.