A trillion dollars (potentially) wasted on gen-AI

Scaling vs. limitations of LLMs

• One side invokes the “bitter lesson”: bigger models + more data + more compute have repeatedly delivered breakthroughs since the 2010s; perceptrons only took off once scaled.
• Others push back that “scaling is all that matters” is historically wrong and technically shallow: many AI waves (expert systems, early MT, speech) hit walls where missing cognitive abilities couldn’t be fixed by more compute.
• Critics reference diminishing returns, No Free Lunch, narrow specialist models, and current issues with transfer, hierarchy, causality, and world models as evidence that LLMs are not the final paradigm.

Was the trillion dollars “wasted”?

• Supporters of the boom argue investment isn’t wasted just because AGI isn’t reached; LLMs already power useful services, much like cloud/SaaS or earlier infrastructure bubbles.
• Others stress opportunity cost: trillions on datacenters and GPUs vs. funding many researchers or other societal needs; magnitude matters for “waste.”
• Some see a classic bubble: speculative promises of AGI and “every job automated,” VC incentives to pump a story, and systemic risk if pensions and broad capital pools are exposed.
• Counterpoint: tech booms have always overbuilt and misallocated in the short term but left valuable infrastructure and knowledge.

Definitions and status of AGI

• There’s no consensus on “AGI”: views range from “we’re already past it” (LLMs outperform average humans on many cognitive tasks) to “we’re nowhere close” (hallucinations, lack of agency, no stable self-knowledge).
• Proposed criteria include: lack of hallucinations, stable reasoning, synthesis, recursive self-improvement, and the ability to operate autonomously in the real world (power plants, fabs, etc.).
• Many see current systems as “human-in-the-loop AGI” or on a continuum: superhuman in some domains, subhuman in others, with “jagged” capabilities.

Real‑world utility and risks

• Multiple practitioners report 2–3× productivity gains, especially in coding, refactoring, debugging, documentation, and research, plus lower mental fatigue.
• Others emphasize unreliability, hallucinations, and the need for expert oversight; inexperienced users may be misled, making LLMs “dangerous conveniences.”
• There’s frustration at both overhyping (“AGI 2027”, total job automation) and at critics who ignore clear practical value.

Economic, hardware, and bubble dynamics

• Some welcome the rich burning capital on GPUs, noting spillovers: better tools, open-source models, and possibly cheaper compute later.
• Others worry about environmental impact, RAM and power prices, short GPU lifecycles, and eventual e‑waste if this stack is abandoned.
• Comparisons are made to dot‑com fiber overbuild (later useful) vs. housing/crypto bubbles (primarily wealth transfer and waste).

Views on the article and research trajectory

• Several commenters see the article as overstating “waste” and rehashing long‑running “deep learning is hitting a wall” claims that already missed major progress.
• Others say its technical critiques (data hunger, weak generalization, opacity, lack of causality) still largely hold, and LLMs alone won’t deliver AGI or promised returns.
• Broad agreement that future progress likely needs new architectures and hybrid approaches, but disagreement on whether today’s scaling push was a necessary step or a trillion‑dollar detour.