Sources
Andrew Ng Argues 'Tokenmaxxing' Is Finally Dying, and Warns the Labs Recommending It Have an Oil-Change Incentive
The Batch issue #365 (August 8) opens with Ng conceding that token consumption correlates with productive AI work but arguing the maximize-tokens trend overshoots: past a point, extra tokens hit organizational bottlenecks that more inference cannot dissolve. He explicitly notes frontier labs have a financial incentive to recommend heavier consumption, comparing it to manufacturers' oil-change intervals, and calls for independent guidance that optimizes for the practitioner rather than the vendor. For anyone running standing agent fleets, this is the first prominent pushback on scale-your-token-budget as a default strategy from a source with no inference to sell.
↳ Follow the thread