AI Capabilities and Limitations

What the course is

The companion course to ai-fluency-4d: where the 4D Framework teaches how to collaborate with AI, this course builds the mental model of the machine itself — why models are brilliant at some tasks and unreliable at others. The whole course hangs on four machine properties. Internalize these and most "weird" model behavior becomes predictable.

The four machine properties

1. Next-token prediction

Next-token prediction — the core generation mechanism: the model repeatedly predicts the most plausible next token given everything so far. It is not retrieving answers from a database; it is continuing text plausibly.
Consequences: fluency ≠ accuracy (plausible-sounding wrongness is the default failure mode), the model has no built-in "I don't know" reflex, and output quality is steered heavily by what precedes it — which is why prompting works at all (prompt-engineering-basics).

2. Knowledge

Training data — the model's knowledge comes from a fixed corpus assembled before training; it doesn't "look things up" unless given tools or documents.
Knowledge cutoff — the date after which the model has no information. Anything more recent must be supplied in the prompt or fetched via tools/search (rag-patterns).
Hallucination — confident generation of false or fabricated content. It's next-token prediction operating without grounding: when the training data is thin, the model interpolates something plausible. Mitigations: provide source material, ask for citations, permit "I don't know," and verify independently (Discernment in ai-fluency-4d).

3. Working memory

Context window — the model's working memory: the finite token budget holding the conversation, documents, and instructions currently "in view." Everything the model can use right now must fit here.
Consequences: long conversations degrade as early material competes with new material; oversized pastes crowd out instructions; and nothing persists between separate conversations unless deliberately carried over (Projects, standing context, or programmatic state — see context-management).
Practical habits: start fresh threads for new topics, summarize before continuing long work, and put the most important instructions where they won't get lost.

4. Steerability

Steerability — how strongly instructions, examples, and assigned roles shape output. The same model produces radically different work depending on system prompts, tone requests, and formatting constraints.
Character/persona — models like Claude also have trained-in dispositions (helpful, honest, harmless — see anthropic-safety) that interact with your steering; steering works with the grain of the character, not against hard safety limits.
Consequence: vague input → generic output. Steerability is the property that makes Description (in 4D terms) high-leverage.

When properties collide

The course's most exam-relevant section: real failures usually involve two properties interacting.

Collision	What happens	Fix
Knowledge × Next-token prediction	Hallucination — thin knowledge, fluent interpolation	Ground with sources; ask for uncertainty
Working memory × Knowledge	Long chat "forgets" your pasted document	Re-supply or summarize key material
Steerability × Knowledge	Model complies with a request it can't actually satisfy factually	Ask it to refuse/flag when unsure
Steerability × Working memory	Instructions buried in a huge context get diluted	Restate constraints near the end; keep context lean

Quick reference

Property	One-line definition	Chief limitation it explains
Next-token prediction	Generates by predicting plausible continuations	Fluent wrongness
Knowledge	Fixed training corpus with a cutoff	Stale info, hallucination
Working memory	Finite context window per conversation	Forgetting, dilution, no cross-chat memory
Steerability	Output shaped by instructions and character	Garbage-in prompting; also the biggest lever you own