Uber’s COO says it’s getting harder to justify money spent on tokenmaxxing
Uber’s COO has signaled that the company’s aggressive “tokenmaxxing” push—encouraging engineers to maximize AI token usage—no longer looks financially justifiable, after data showed soaring spend without a proportional increase in useful product features. Commenters argue this was a predictable case of Goodhart’s law, likening token-based performance metrics to past missteps like judging developers by lines of code and warning that employees quickly learn to game such measures. Many expect a shift toward more disciplined AI use, cheaper or local models, and evaluation based on actual outcomes rather than raw token consumption, with some predicting that over time only a small, highly effective subset of engineers will be given large AI budgets.
What “Tokenmaxxing” Actually Is
- Companies incentivize or require engineers to maximize AI token usage, sometimes with leaderboards and perceived ties to performance reviews.
- Reported Uber numbers (via linked coverage): average AI spend of ~$150–$250 per engineer/month, with “power users” at $500–$2,000.
- Some devs say they now feed models pointless tasks just to avoid looking like “low users.”
Costs, Subsidies, and Economics
- Many argue current token prices are heavily subsidized by VC or model vendors; once subsidies end, heavy agentic use may rival or exceed junior salaries.
- Others counter that inference is highly optimizable, open-weight models are improving fast, and labor has a hard cost floor, so AI should eventually be cheaper than humans for many tasks.
- Skepticism that major model providers are truly profitable; comparisons to prior bubbles and expectations that a recession will purge weak AI offerings.
Goodhart’s Law and Management Critique
- Strong consensus that “tokens used” is a bad productivity metric, likened to lines-of-code or hours-at-desk.
- Example behaviors: agents chatting with each other, redundant tool calls, and feature work driven by metric-chasing rather than user value.
- Some see executives as herd-followers reacting to investor/board FOMO; others defend “burn money now” as an intentional exploration phase to find high-value workflows.
Productivity, Quality, and Engineering Culture
- Reported gains range from “not helpful” to ~20% boost, with many warning that massive claimed multipliers are hype.
- Concerns that junior engineers using AI for everything don’t build real understanding; rise of fragile, “vibe-coded” PoCs that become production-critical.
- Teams struggle to measure real ROI: many tokens can produce many low-impact features; fewer tokens can yield one high-impact change.
Local Models, Security, and Architecture
- Debate over centralized data centers vs. GPUs per dev; some believe local/open models (e.g., DeepSeek, Mistral) plus internal hosting will undercut cloud APIs.
- Security worries about sending sensitive data to third-party APIs, especially foreign ones; local hosting seen as a mitigation.
Uber-Specific and Broader Strategy Questions
- Some question why a 16-year-old “taxi + food delivery” firm still needs thousands of engineers; others point to enormous regulatory, tax, payment, and locale-specific complexity.
- Many expect a shift from “use as much AI as possible” to “use the right model, minimally, for clear business value.”