What's really going on with AI, Expert weighs in | TheStandup

AI’s real-world impact is less about whether code generation is “good” and more about how organizations will operationalize it—through incentives, monitoring, and review pipelines that reshape engineering work for both junior and mid-to-senior staff. Dmitri, a long-time AI researcher, frames the central risk as career disruption: junior engineers may lose time to experimentation and re-skilling, while midcareer professionals face a sharper threat of becoming “useless” if their skills stop matching what teams can ship with AI.

The conversation then turns to why AI cost predictions—especially claims that token prices will fall 10x or 100x annually—should be treated cautiously. Dmitri separates the *content* of such claims from the *timing*, arguing that past technology promises (reusable rockets, full self-driving) often took years to materialize. He expects substantial efficiency gains are plausible, but he doubts that everyone can optimize today’s systems without being invalidated by next year’s architectural breakthroughs. A key practical uncertainty is whether current “token cost” is stable enough to optimize, given that hardware, electricity, infrastructure, and model architecture all evolve.

Casey adds a reality check from the infrastructure angle: if software alone could deliver dramatic cost drops, Google’s AI costs should already look far cheaper. The discussion highlights a common industry belief that Google has been quietly ahead on infrastructure—through long-running AI hardware work like TPUs and data advantages—while other players (including Nvidia-focused ecosystems) are more hybrid and may face longer hardware transition cycles. That leads into a speculative but concrete business thread: SpaceX-linked “orbital data centers.” The appeal isn’t just sci-fi coolness; it’s the prospect of free/cheaper power and easier cooling in space, plus a potential strategic moat if launch costs fall enough. The group debates the physics and business logic, landing on the idea that heat dissipation and launch economics—not marketing—will determine whether orbital compute becomes viable.

On the workplace question—whether AI will replace code review—Dmitri draws a boundary around “reliable AI”: tasks that can be executed hands-off with high trust, roughly up to a few thousand lines of relatively standard code. Beyond that, a review-heavy phase is likely because businesses will push AI into workflows faster than teams can validate outcomes. He describes how token monitoring can force behavior: linking token usage to PRs and KPIs can turn engineering into a compliance game, driving more PRs and more review overhead. Amazon’s “AI hero” style policy—requiring senior sign-off for junior/mid-level AI-generated changes—becomes an example of how organizations respond to risk and metrics, even if it strains senior capacity.

The group also questions whether token usage will remain the dominant KPI. Dmitri suggests the current push can last for years because adoption is powered by momentum and organizational incentives, not just technical merit. Even if AI quality doesn’t improve much, the workflow shift may still stick due to sunk costs and the sheer scale of investment. In that environment, the practical outcome may be less about “best practices” and more about which metrics companies choose to reward—until cost, uptime, and engineer burnout force a recalibration.