Palantir CEO Alex Karp isn’t buying into the “tokenmaxxing” hype.
Never one to mince words, Karp compared the urge to use AI in such an insatiable manner to watching pornography.
“Really, we call it the demastibatory, like get off masturbation thing internally,” Karp said, describing Palantir’s token tracking during a live interview on TBPN on the sidelines of Palantir’s AIP Con 10. “Sure, it’s like people are just sitting there all day kind of like a porn addiction.”
Karp’s comments echo what Palantir CTO Shyam Sankar told analysts on an earnings call last month about how the data analysis and technology firm calls itself “a no slop zone.” Sankar said companies need to understand that cheaper AI alone won’t create more value unless you have a system like Palantir’s AIP (Artificial Intelligence Platform) that can ground an AI model.
“More tokens means more slop,” Sankar said. “And the more commodity cognition you consume, the more you need a system that can prevent the economic harm, so you can harness the economic value.”
Tokens are the building blocks of large language models, which break down words into numerical units. A single token is roughly 3/4ths of a word. AI companies and AI model providers often charge based on the number of tokens consumed and the model that was used.
In the last few weeks, parts of Silicon Valley and the tech community have swung hard against “tokenmaxxing,” the culture that championed almost unfettered AI usage to match the rise and capability of AI agents.
Uber COO Andrew Macdonald highlighted those concerns when he said that the rideshare company was struggling to see the link between rising AI bills and meaningful returns, such as increased productivity. In reference to Macdonald’s comments, Karp said that until recently, perhaps it was not considered wise to publicly raise questions about AI.
“When we first met, it was like AI, maybe real,” Karp said, addressing when he first met the TBPN crew. “Then, I would say somehow until about two weeks ago there was like a holy fuck, this is real, but somehow it’s not working, but we’re not allowed to say it publicly because we’ll look stupid.”
Karp said people realize AI is real, but many of the issues surrounding the technology, including whether competitors may try to build something like Palantir’s ontology, come down to a matter of “taste.”
“All these things can be scaled in a very valuable but largely going to commodify way, but you can’t scale the taste of what is the business problem you want to have to solve and need to solve,” he said.
There are some problems that AI models can solve really well, Karp said. He gave an example of a prompt like, “I want to write a report on GDP growth in China.”
It’s the more complicated dilemmas where AI alone is not the answer, he said.
“‘I want to understand the specialized way I drill for oil and gas that’s both legal, ethical, and reduces the cost of production. I want to change the supply chain of my industry, whether that’s military or whether that’s building boxes or whether that’s cars.’ These things require actual, precise ongoing processes,” Karp said. “They are enhanced by large language models. They are not replaced by large language models.”
