The Leader's Paradox: Microsoft CEO Against 'Tokenmaxxing'
Within Microsoft, one of the world's leaders in artificial intelligence, an internal paradox is brewing. The company's CEO, Satya Nadella, has publicly admitted that he himself suffers from an excessive obsession with neural networks and has urged the team to rethink their habits. According to him, uncontrolled use of AI not only reduces efficiency but also causes direct financial damage.
Nadella touched upon the topic of so-called 'tokenmaxxing' — a neologism describing the employees' drive to use the maximum number of tokens and queries to neural networks, even when there is no urgent need. Acknowledging the scale of the problem, the company's leader noted that this practice has become the norm within Microsoft.
"Very often. I engage in tokenmaxxing myself. It is addictive," Nadella honestly answered when asked how widespread such behavior is among employees. The admission by the head of one of the world's largest technology companies sounded like a warning signal: even professionals working with cutting-edge tools are not immune to algorithmic dependency.
The Economics of Attention and Token Costs
Nadella's call to use AI more wisely is dictated not only by productivity considerations but also by hard economics. Training and launching powerful models require colossal computing resources. Pointlessly running complex algorithms to solve simple tasks leads to unjustified budget expenditure.
The CEO emphasized that people should focus on end goals rather than the process of interacting with the model. When the novelty effect wears off, it is necessary to take a step back and ask oneself: "What exactly am I trying to create?". It is precisely this conscious approach that will allow separating real innovations from the empty waste of resources.
Automation of Choice: How Copilot Solves the Problem
The solution to the problem of excessive AI use lies not in banning technologies, but in their smarter integration into workflows. Nadella pointed out that the main task is to select the appropriate model for a specific task, rather than using "heavy artillery" to solve routine issues.
As an example of an optimal approach, he cited the Auto Mode in Copilot. This function automatically analyzes the user's request and selects the most suitable model for its execution. Such an approach eliminates the human factor and the emotional impulse to "play" with the neural network, ensuring a balance between result quality and the cost of obtaining it.