Alex Karp heads Palantir, an analytics technology company he co-founded in 2003 that now supplies AI and data platforms to militaries, intelligence agencies, and the world's largest corporations. He has worked with artificial intelligence for more than two decades, long before it became a buzzword.
He is the kind of man who speaks his mind bluntly. The Palantir CEO confirmed this once again at AIPCon 10, where, in a live interview on the TBPN platform, he discussed a trend that is known in the technology world but rarely talked about. Tokenmaxxing. The pursuit of the highest possible token consumption, rewarding employees for the volume of data processed, and corporate AI activity rankings. Karp compared this trend to pornography addiction. “Internally, we call it a stress reliever, simply a way to get it out of your system,” he said. “People just sit there all day; it's like pornography addiction.”
Tokens
A token is the basic building block of large language models. Models break text down into numerical units, with one token corresponding to roughly three-quarters of a word. Model providers charge based on the number of tokens consumed and the model used. Tokenmaxxing is a culture that emerged with the rise of AI agents. Companies began tracking token consumption on leaderboards, rewarding high usage, and treating the volume of data processed as a performance indicator. The more tokens, the better. But Karp disagrees.
Karp's position is not merely a personal opinion. It reflects the corporate philosophy of Palantir, which describes itself as a “slop-free zone.” The company's chief technology officer, Shyam Sankar, summed it up on an earnings call with analysts: “More tokens mean more slop. And the more commodity cognition you consume, the more you need a system that can prevent economic harm.” Both Sankar and Karp insist that cheaper AI and higher token consumption do not inherently mean greater value. Having access to a model is not enough. You need to know what you will use it for.
What AI Can and Cannot Do
Karp does not question that large language models are capable tools. They can write a report on Chinese GDP growth, compile an analysis, or answer a complex query. But then there are problems of a different kind.
“I want to understand the specialized way I extract oil and gas that is also legal, ethical, and lowers production costs. I want to change the supply chain in my industry, whether it is defense, manufacturing, or automotive,” Karp said. “These things require precise, ongoing processes.” And that is where a model alone is not enough. “They are enhanced by large language models. They are not replaced by them,” he added.
Karp also spoke openly about the atmosphere that prevailed in the technology world. He pointed out that doubts about AI did exist, but no one voiced them publicly for fear of looking foolish or behind the times. “Back then, when we first met, it was like: AI, maybe it's real. Then suddenly it was clear that it was real, but somehow it wasn't working. But we weren't allowed to say that publicly because we would look stupid,” he said.
Yet others are asking the same questions Karp raises. Uber Chief Operating Officer Andrew Macdonald recently admitted that the company sees no direct link between AI spending and measurable productivity improvements. And OpenAI's Sam Altman called this criticism “the fairest one AI faces right now.”
Sources: businessinsider.com and aol.com
