🪄Understand Emergent Capabilities in LLMs
See past the 'mystical jump' headlines: read the original emergence papers next to the 'mirage' rebuttal, watch the same task switch from jumpy to smooth when you swap the metric, and finish able to predict whether the capability you actually care about will scale in steps or in slopes.
Phase 1What Emergence Meant
Read the original emergence papers and the excitement
Emergence is a curve that looks like a cliff
6 minEmergence is a curve that looks like a cliff
If you can't extrapolate, you can't plan
6 minIf you can't extrapolate, you can't plan
Three tasks that defined the emergence canon
6 minThree tasks that defined the emergence canon
Scale isn't one number — it's at least three
6 minScale isn't one number — it's at least three
Phase 2Watch the Jump Disappear
Switch metrics and watch the jump disappear
A linear x-axis turns every curve into a wall
6 minA linear x-axis turns every curve into a wall
Exact-match accuracy is a step function
7 minExact-match accuracy is a step function
Replot 3-digit addition and watch the cliff melt
7 minReplot 3-digit addition and watch the cliff melt
Why composition turns slopes into cliffs
7 minWhy composition turns slopes into cliffs
Different runs, different thresholds
6 minDifferent runs, different thresholds
Phase 3Mirage, Real, or Both?
Reconcile the mirage rebuttal with what's still real
The rebuttal: emergence is a metric artifact
7 minThe rebuttal: emergence is a metric artifact
Both papers can be right at once
7 minBoth papers can be right at once
Some capabilities get worse with scale
7 minSome capabilities get worse with scale
The 2026 consensus is calmer than the 2022 hype
7 minThe 2026 consensus is calmer than the 2022 hype
Phase 4Predict Your Capability's Curve
Predict whether your capability will jump or slope
Predict the curve for one capability you actually care about
8 minPredict the curve for one capability you actually care about
Frequently asked questions
- What are emergent abilities in large language models?
- This is covered in the “Understand Emergent Capabilities in LLMs” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
- Are emergent abilities in LLMs real or just a measurement artifact?
- This is covered in the “Understand Emergent Capabilities in LLMs” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
- What did the 'Emergent Abilities Are a Mirage' paper actually show?
- This is covered in the “Understand Emergent Capabilities in LLMs” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
- How does the choice of metric affect whether a capability looks emergent?
- This is covered in the “Understand Emergent Capabilities in LLMs” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
- Can you predict in advance which LLM capabilities will be emergent?
- This is covered in the “Understand Emergent Capabilities in LLMs” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
Related paths
🐍Python Decorators Introduction
Build one mental model for Python decorators that covers closures, argument passing, functools.wraps, and stacking — then ship a working caching or logging decorator from scratch in under 30 lines.
🦀Rust Lifetimes Explained
Stop reading `'a` as line noise and start reading it as scope arithmetic — one failing snippet at a time — until you can thread lifetimes through a small parser or iterator adapter without fighting the borrow checker.
☸️Kubernetes Core Concepts
Stop drowning in 30+ resource types. Build the mental model one primitive at a time -- pods, deployments, services, ingress, config -- then deploy a real app with rolling updates and health checks.
📈Big O Intuition
Stop treating Big O as math you memorized for an interview — build the intuition to spot O(n²) disasters, pick the right data structure without thinking, and rewrite a slow function from O(n²) to O(n) in under five minutes.