What if A.I. doesn't get better than this?

Exponential vs S‑Curve Progress

• One camp argues current LLM gains are just the steep part of an S‑curve; aviation and spaceflight are cited as examples where early rapid progress plateaued.
• Others counter that we can’t know where on the curve we are; 10–75‑year extrapolations are seen as speculative and historically unreliable.
• Several note that all real‑world “exponentials” saturate, but that doesn’t tell us what the ceiling is for LLMs.

Labor, Politics, and Social Risk

• Some fear mass automation without safety nets (like UBI) will push people into “scraps” work and possible unrest or revolution.
• Others think current LLMs are more augmentation tools than replacements; transformative change may be slow, like the Internet’s diffusion.
• There’s concern AI will amplify elite control and surveillance, intensifying long‑standing labor–management conflicts.

AI vs LLMs and Mislabeling

• Many dislike the article’s conflation of “AI” with LLMs, arguing AI also includes search, planning, classic ML, etc.
• Others say the linguistic battle is effectively lost: in popular usage “AI = LLM chatbot products,” and media just reflects that.
• Some see this conflation as part of a “grift” that exaggerates intelligence to justify huge investment.

Capabilities, Orchestration, and Integration

• Several argue models are already powerful; the real frontier is orchestration: multi‑step workflows, agents, tool use, and system integration.
• Even without model improvements, better protocols, sensors, and cross‑system interfaces could yield major practical impact (and also dystopian scenarios).
• Others are skeptical: if it were truly “powerful,” use cases would be more obvious and self‑justifying.

Economics, Business Models, and Competition

• Massive AI capex versus modest current revenues leads to predictions of a shakeout or collapse for firms betting solely on frontier models.
• Open and cheap competitors (e.g., DeepSeek) are viewed as limiting price power and moat formation.
• One view: big players are in a user‑acquisition phase, aiming to monetize later via ads embedded in AI outputs; whether ad budgets can support this at scale is contested.
• Debate over inference costs: some claim serving millions with low latency and high uptime is expensive; others argue that, at scale and with hardware amortized, tokens can be cheaper than human labor.

Have We Hit a Plateau?

• Some participants perceive slowed progress: model differences feel incremental, coding help seems logarithmic, and products regress on certain tasks.
• Others point to recent competition performance (IMO/IOI medals) and new “reasoning” models as evidence that frontier capability is still rising.
• There’s disagreement over whether 2023 expectations for GPT‑5–style breakthroughs (sometimes framed as near‑AGI) have been met.

Data, Training Paradigms, and Cognition

• One view: language is a weak, high‑level substrate for cognition, yielding broad but shallow, brittle models; future systems should learn from lower‑level or real‑world data at huge scale.
• Others stress that not all AI is linguistic; “performing” systems (vision, control, optimization) may keep advancing even if pure LLMs stall.
• Some discussion touches on human thought: subconscious decisions preceding language, suggesting internal “conceptual” processing distinct from verbalization.

Reliability, Hallucinations, and Trust

• A reported test found GPT‑5 hallucinating most scientific citations (fabricated titles, authors, or mismatched journals), reinforcing claims that LLMs can’t be trusted as fact‑grounded systems.
• Some argue true trustworthiness requires integrated citation/claim‑checking pipelines outside the base model; products that solve this could be decisive.
• Others note that certain tools (e.g., web‑powered “deep research”) already partially address this, but are still imperfect.

Local Models, Infrastructure, and Medium‑Term Outlook

• Several expect local or on‑device models to erode centralized providers’ margins once quality is “good enough,” especially for coding and niche tasks.
• Infrastructure, integrations, and UX are seen as lagging far behind model capability; building robust AI‑aware systems is viewed as at least half the challenge.
• Some foresee an AI hype “trough of disillusionment,” with many VC‑funded players burning out, while big incumbents survive thanks to diversified profits.