When Will the GenAI Bubble Burst?

State of the “bubble” and AI winter

• Many agree there is a hype bubble: “AI in everything” (rice cookers, printers, hackathon ideas) with weak justification, reminiscent of blockchain/ICO era.
• Others argue this is an exploratory phase, not a classic bubble: investors are still figuring out use cases, and no big company wants to declare failure yet.
• Several note that earlier AI winters came from hardware/data limits; today, models are still improving rapidly, so a technical winter seems less likely than a financial correction.
• Some see a pattern like the dot‑com era: a bubble will pop, many players die, but long‑term impact remains huge.

Current usefulness and “killer apps”

• Coding assistance is widely cited as the strongest use case: people report big productivity gains, especially seniors using tools like Copilot/ChatGPT as “juniors on tap.”
• Other concrete wins: invoice parsing and pricing analysis, knowledge search over messy internal data, summarization, drafting, and learning—described as “exocortex” extensions.
• For some, LLMs are already a “plateau of productivity” and boring but indispensable; others barely use them and would not miss them.

Limitations, risks, and integration challenges

• Hallucinations and lack of reliability are central concerns, especially for legally or financially sensitive tasks. Many insist on human review or external verification.
• Some argue LLM‑generated code often increases technical debt; others counter that, with proper review, it saves time and reduces trivial bugs.
• There is worry about energy costs and unsustainable economics (tens of billions spent on GPUs vs a few billion in revenue).
• Integrating LLMs into products beyond chat/coding is seen as hard; many proposed “AI features” could be done with older tech.

AGI, token prediction, and human comparison

• One camp: LLMs are “just” powerful token predictors, not precursors to AGI; real AGI needs robust reasoning and self‑correction.
• Another: predicting the next token in rich data forces models to learn abstract concepts, so this may be a key AGI milestone; the challenge is using those representations stably.
• Ongoing debate over whether humans are fundamentally “token predictors,” and over definitions of AGI (human‑level vs necessarily superhuman).