Goldman Sachs: AI Is overhyped, expensive, and unreliable

Original Article ↗ Hacker News Discussion ↗

Goldman Sachs’ Motives and Credibility

  • Some argue financial firms never share valuable insights for free; public reports are seen as market signaling or propaganda.
  • Others counter that large institutions need accurate pricing and can be both self‑interested and broadly correct.
  • Several note the report’s headline language (“overhyped, wildly expensive, unreliable”) does not appear verbatim in the PDF.
  • There is debate over whether GS is expressing a genuine view, talking its book, or arriving late after already positioning.

AI vs Traditional Quant/Algo Trading

  • Commenters stress there’s no contradiction between GS using quant algorithms and criticizing current “AI.”
  • Algorithmic trading and ML/LLMs are framed as different categories; sophisticated models need not be “AI” in the buzzword sense.

Hype, Trajectory, and Historical Analogies

  • One camp: current generative AI is expensive, unreliable, and may never justify the massive capex; parallels drawn to crypto, web3, and self‑driving.
  • Another camp: the important factor is long‑term slope (2030–2040), comparing AI to early aviation or the internet in the 1990s, with transformative potential still ahead.
  • Skeptics respond that similar “it’s early days” narratives were used for bubbles that never delivered.

ROI, Investment Horizon, and AI Winter

  • Institutional investors are said to care about payoff within a few years, not distant decades; discounting makes far‑future gains less compelling.
  • Many foresee an “AI winter” in funding if near‑term returns disappoint, though AI as a deployed technology would persist.
  • Some argue current valuations resemble bubble assumptions (very high multiples, unrealistically perfect execution).

Current Usefulness and Limitations

  • Positive experiences: code autocomplete, text summarization, semantic search, translation, TTS/STT, classification, and productivity boosts for some users.
  • Negative experiences: hallucinations, generic answers, buggy code, weak search replacement, and “AI‑washed” products adding little value.
  • Distinction made between generative AI and long‑standing ML (recommendation, decision trees, moderation); the latter is already ubiquitous and impactful.

Societal and Business Effects

  • Concerns include job displacement as firms use AI for efficiency and potential over‑investment to justify layoffs.
  • Some see AI also enabling more powerful exploits, forcing simpler, more privacy‑preserving systems.
  • Pricing debates: some expect assistant tools to get much cheaper; others would pay high subscription fees for current capability.