“Emergent” abilities in LLMs actually develop gradually and predictably – study

Meaning of “emergent”

• Strong disagreement over what “emergent” means.
• In ML, one influential paper defines emergent abilities as behaviors absent in small models but appearing suddenly in larger ones, not predictable by extrapolating smaller models.
• Others argue this is a misuse of a broader scientific/philosophical term where emergence is about higher‑level properties from simpler parts, not necessarily sudden jumps.
• Several comments highlight polysemy: words legitimately mean different things in different domains, but that also creates confusion and misplaced arguments.

Measurement and metrics

• Core claim of the discussed study: many supposed “emergent abilities” disappear when using smoother metrics (e.g., partial credit, edit distance) instead of binary pass/fail.
• Participants note that any thresholded metric (e.g., “all test items correct”) will naturally produce apparent phase changes.
• Some agree that continuous metrics are better for understanding gradual improvement and extrapolating progress.
• Others warn that partial credit can be a poor proxy when tasks demand exactness (e.g., arithmetic in real applications).

Arithmetic as a test of reasoning

• One camp calls arithmetic a poor or “pointless” benchmark, since good LLMs often succeed by writing and running code rather than doing math internally.
• Another camp defends arithmetic as a canonical structured task and proxy for learning formal rules; if models truly internalized arithmetic, it would indicate strong general reasoning.
• Examples are given where LLMs fail at basic reversals, ordering, and time reasoning, unless they rely on external tools or very careful prompting—seen as evidence of limited algorithmic capability and lack of internal state.

Capabilities, limits, and architecture

• Debate over whether surprising LLM skills are genuinely “unpredictable” or just reflect our poor understanding of training dynamics.
• Some argue that scaling inevitably yields more capabilities; others stress that architecture, data, and optimization also matter, so “emergence” claims are premature.
• LLMs are characterized by some as “plausibility engines.” There’s disagreement on whether that inherently caps their intelligence or parallels human reliance on plausibility, mitigated by tools like the scientific method.

Forecasting progress and hype

• Many see value in the study: smoother metrics suggest capabilities grow more continuously, enabling better forecasting of when a model will cross a usability threshold.
• Some worry the earlier “sudden emergence” narrative fed hype and could contribute to an AI bubble; this work is viewed as a corrective.
• Others note that even under revised metrics, a few tasks still look emergent, and caution against over-interpreting a single paper.