Palantir CEO Alex Karp has a blunt tackle the AI trade’s hottest buzzword. Talking on the TBPN podcast on the sidelines of Palantir’s AIPCon 10 convention, Karp in contrast “tokenmaxxing”—the compulsive overuse of AI tokens—to a porn dependancy. His level was easy: lots of corporations are burning via tokens that appear like productiveness however ship nothing actual.Karp instructed the hosts that Palantir constructed an inside software to wean enterprises off the behavior. The identify is not match for air, however he described it as a “demasturbatory, get off masturbation” factor. “Individuals are simply sitting there all day like a porn dependancy,” he stated. He saved circling again to the comparability, even because the hosts tried to steer him towards the AI facet of the analogy.Tokens are the constructing blocks of enormous language fashions, breaking phrases into numerical items—a single token is roughly three-fourths of a phrase. AI suppliers often cost primarily based on what number of tokens you eat and which mannequin you decide. Over the previous few weeks, elements of Silicon Valley have turned laborious in opposition to tokenmaxxing, the tradition that cheered almost limitless AI use.
Tokenmaxxing backlash builds because the AI productiveness math stops including up
Firms like Meta and Amazon reportedly constructed inside scoreboards to trace token consumption, treating uncooked utilization as a proxy for productiveness. Most have since reduce. Uber COO Andrew Macdonald stated the rideshare agency was struggling to attach rising AI payments to any actual return. Karp stated that till about two weeks in the past, executives felt they could not query AI publicly with out trying silly—at the same time as many privately admitted it wasn’t fairly working.His CTO Shyam Sankar made a sharper model of the argument on a current earnings name, calling Palantir a “no slop zone.” Extra tokens, Sankar stated, means extra slop.
Why Karp thinks LLMs hit a wall on laborious, domain-specific issues
Karp did not dismiss AI outright. He known as it actual and stated it is helpful for slim duties—writing a report on China’s GDP progress, as an illustration. The difficulty begins with more durable, domain-specific issues, like determining a less expensive, authorized, and moral technique to drill for oil and gasoline. These want exact, ongoing processes that LLMs can improve however not exchange, he argued.In fact, a lot of this doubles as a gross sales pitch. Karp instructed prospects to spend a few days with frontier corporations like OpenAI or Anthropic first—and name him after they’re completed, as a result of purchasers often come again “clamoring.”