Tokenmaxxing is dead, long live tokenmaxxing

Original Article ↗ Hacker News Discussion ↗

What “tokenmaxxing” referred to

  • Using AI token spend as a visible metric and incentive, sometimes tied to performance reviews or informal leaderboards.
  • In practice, often meant “use AI as much as possible” without clear success criteria or guardrails.

Why companies did it (competing explanations)

  • Charitable view: a blunt but intentional way to force widespread AI experimentation in large orgs where bottom‑up adoption was too slow.
  • Less charitable view: classic hype/FOMO, copying competitors and analyst slides, with little understanding of AI or ROI.
  • Some point to consultants and vendors aggressively selling AI as transformational (huge profit boosts, cost cuts), driving execs to pre‑buy tokens and then push usage.

Extent and consequences

  • Some say only a minority of companies did strict tokenmaxxing; others claim multiple big firms burned billions per quarter on tokens.
  • Reported effects include: wasted money, pointless loops that churn tokens, and even layoffs or pressure framed around “underperforming in token spend.”
  • Others report modest budgets and careful cost‑control, with tokenmax anxiety spreading socially even where no leaderboard existed.

Employee experience & culture

  • Many engineers resent AI mandates, seeing them as faddish micromanagement or humiliation rituals that ignore their expertise.
  • Some feel it erodes their status and autonomy, shifting power toward PMs and executives; others consciously opt out of AI and plan to leave the industry.
  • A subset report genuine enthusiasm, saying heavy AI use greatly increases their output and makes them internal “AI leaders.”

Productivity and “compounding correctness”

  • One prominent claim: newer agents and multi‑agent systems show “compounding correctness,” where more tokens generally mean better results.
  • Many commenters strongly dispute this, likening it to worshiping lines of code or “more sawdust = more furniture,” and demanding real evidence.
  • Practical experiences vary widely: some hit limits even on expensive plans and claim huge gains; others rarely use tokens and see little benefit.

Alternatives & metrics

  • Suggested better approaches:
    • Assign specific people/teams to experiment and report back.
    • Measure business outcomes, quality, and bugs rather than raw tokens.
    • Let people “freely experiment” but don’t grade them on token burn.
  • Tokenmaxxing is widely criticized as a textbook Goodhart’s‑law failure: once the metric is targeted, people optimize tokens, not value.

Broader pattern

  • Many frame tokenmaxxing as another hype wave (after blockchain, metaverse, big data, cloud) driven by financial markets and managerial herd behavior.
  • Several highlight how often large companies burn cash on dubious initiatives, challenging the idea that corporate capitalism is reliably efficient.