AI capabilities and limitations

A walkthrough of the Claude Academy course AI capabilities and limitations (Become AI fluent track, 14 lessons + quiz). Each lesson is summarised with my own tables and examples; take the course for the full lessons. All the courses are listed in the Claude Academy course index.

Note: Detailed content expanded from course outline; verify against lessons. (Public lesson summaries and key takeaways were used as the primary source; explanatory examples and tables are added study aids.)

Audience: Anyone who uses or is about to use generative AI (educators, students, knowledge workers, team leads). No technical background or prerequisites are needed. Relationship to AI Fluency: This is the companion to AI Fluency: Framework & Foundations. The 4Ds describe the human competencies, and this course describes the machine properties those competencies respond to. The two courses can be taken in either order. Format: 6 sections, 14 lessons (each property has an explanation plus a “Try It Out” exercise), a quiz, and a completion badge.

---

Core model: four properties, each on a capability → limitation continuum

Property	Plain-language idea	Capability zone	Limitation zone	Characteristic failures	Product mitigations	4D link
Next Token Prediction	Very sophisticated autocomplete that writes one token at a time based on what tends to come next	Well-practiced patterns: summarizing, reformatting, explaining common concepts	Novel or sparse territory; telling “true” apart from “sounds true”	Fabrication concentrated in specifics (names, dates, stats, citations, URLs, quotes)	Citations, uncertainty signaling, constrained generation, generator-verifier loops	Discernment
Knowledge	It knows only what was in its training data, frozen at the knowledge cutoff	Topics that appeared frequently, recently (in training), and consistently	Rare, post-cutoff, niche, local, or contested topics	Staleness, uneven coverage, inherited bias about what counts as “default,” source amnesia (can’t say where knowledge came from)	Web search, retrieval/RAG, MCP connectors, tool use	Delegation
Working Memory	The context window: a fixed-size workspace that holds everything it can attend to right now	Material fits comfortably, the session is current, relevant context is supplied	Very long documents or chats; expecting continuity across sessions; key information buried in the middle	Cliff, not gradient: silent truncation, often without warning; it doesn’t learn from your corrections	Memory features, compaction, projects/workspaces, larger windows, multi-agent workflows	Description
Steerability	It follows instructions by continuing a pattern, not by understanding them	Short, concrete, verifiable instructions (formats, length limits, explicit roles)	Long reasoning chains, abstract or ambiguous asks, tasks needing exact numbers or logic	Reasoning drift (small errors compound), letter-over-spirit, brittle arithmetic	System prompts, code execution, visible reasoning, structured outputs	Description (what makes it powerful and what limits it)

• Calibrated trust means placing your task on each continuum. You neither trust nor distrust the AI wholesale.
• The material stays useful because the properties stay stable while their boundaries shift as models improve.

---

Section 1 — Getting started (3 lessons)

Intro to AI Capabilities and Limitations

• The 4D Framework teaches you how to collaborate. This course explains how the AI is able to work with you. Together they form one system.
• Roadmap: What we mean by AI → How AI is trained → Four properties → When properties collide → Next steps.

What we mean by AI

• Most AI you encounter is not generative. It is classification/prediction AI such as spam filters, recommendation engines, and fraud detection.
• This course focuses on transformer-based text models that produce new content one token at a time.
• AI isn’t uniformly capable or uniformly unreliable. It is strong and weak along four predictable axes, and the same mechanism produces both the capability and the limitation.

How AI gets its character

Training stage	What it produces	Fingerprints it leaves
Pretraining	A document completer that predicts what comes next across huge amounts of text. It has no concept of “helping you.”	Fluency and broad pattern knowledge
Fine-tuning (using human judgments of good responses)	Assistant behavior: treats input as a request, answers instead of rambling, declines harmful asks	Sycophancy (pull toward agreeing), verbosity, occasional over-caution, loose confidence calibration (stated confidence ≠ reliability)

• Politeness, helpfulness, and caution are trained in, not emergent magic. Knowing this helps you interpret behaviors such as unwarranted agreement.

Section 2 — Next Token Prediction (2 lessons)

• Generative AI is closer to a large-scale autocomplete than to a search engine. A single property explains both its fluency and its hallucinations.
• Well-worn paths (common formats and concepts) are reliable. Novel territory pushes beyond what prediction alone can do.
• The more precise a claim, the more it needs verification. Check every specific name, number, date, citation, and quote.
• Try It Out exercise: test the model on common and on obscure tasks and notice where fabrication appears.

Section 3 — Knowledge (2 lessons)

• Ask “how well-represented was this in what it read?” rather than “does the AI know this?”
• By default it has no real-time browsing and no lived experience, and its knowledge stops at the cutoff.
• Strong on mainstream topics and popular programming languages; weak on rare, recent, niche, local, or contested ones.
• When a question sits at the edge, bring your own sources or use search or retrieval tools.

Section 4 — Working Memory (2 lessons)

• The context window is a hard edge: everything inside it is available, and nothing outside it is.
• Attention isn’t uniform across long input, so important material buried in the middle may be underweighted.
• A fresh session doesn’t remember the last one unless memory features are used. Corrections don’t teach the model permanently.
• Context strategies:
  • Front-load the most important material.
  • Chunk long work into pieces.
  • Re-supply critical context in long conversations.
  • Size up the task against the window before you start, and start a fresh session when a conversation has drifted.

Section 5 — Steerability (2 lessons)

• Steerability works because fine-tuning taught instruction-following. It is limited because instructions are followed through pattern-matching, not understanding.
• Control is tightest for short, concrete, verifiable asks (“respond as a table,” “under 100 words”) and loosest for long reasoning chains, abstract requests, and tasks that need exact numbers or logic.
• Letter-over-spirit: the instruction is followed but the intent is missed. Fix: restate the goal. Repeating the instruction more forcefully won’t close the gap.
• For math and precision, offload the work to code execution. For long reasoning, add checkpoints.

Section 6 — Putting it all together & next steps (3 lessons)

When properties collide

Most real failures involve two properties interacting. Name the pair, and you’ll know which fix to use.

Failure pattern	Properties in play	Targeted fix
Hallucinated citations or specifics	Next Token Prediction + Knowledge	Verify specifics; use search or retrieval; supply sources
Long-conversation drift	Working Memory + Steerability	Re-supply context; restate the goal; start fresh
Confidently wrong math	Steerability (native precision) + Next Token Prediction	Offload to code execution or a calculator
Agreeing with a bad premise	Fine-tuning fingerprint (sycophancy) + Steerability	Invite pushback explicitly; ask for counterarguments
Long contract review that misses clauses	Working Memory + Knowledge	Chunk the document; front-load key sections; check against an authoritative source

• Diagnostic routine: (1) Name which property is involved → (2) locate the task on that property’s spectrum → (3) apply a targeted fix instead of a generic retry. This routine is Discernment applied.

Next steps

• Keep a small, clear mental model of the machine: four continuums, and failures as intersections between properties.
• Match your verification and context habits to where the task sits on each continuum.
• Choose one concrete change to your AI practice this week.

---

Quick-Review Flashcards

1. Q: What are the four core properties? → A: Next Token Prediction, Knowledge, Working Memory, Steerability.
2. Q: How does generative AI differ from classification/prediction AI? → A: It produces new content instead of labeling or scoring existing data (spam filters and recommenders are non-generative).
3. Q: What does pretraining produce? → A: A document completer that predicts what comes next and has no notion of helping you.
4. Q: Name four fine-tuning “fingerprints.” → A: Sycophancy, verbosity, over-caution, loose confidence calibration.
5. Q: Where does fabrication concentrate? → A: In specifics: names, dates, statistics, citations, URLs, quotes.
6. Q: Which topics sit in the Knowledge capability zone? → A: Those that appeared frequently, recently (within training), and consistently in training data.
7. Q: What is “source amnesia”? → A: The model can’t attribute where a piece of its knowledge came from.
8. Q: Why is Working Memory called a “cliff”? → A: Things work until the window is exceeded, and then content is silently truncated, often without warning.
9. Q: Name three context strategies. → A: Front-load key material, chunk long work, re-supply critical context.
10. Q: What is letter-over-spirit, and how do you fix it? → A: The instruction is honored but the intent is missed. Restate the goal instead of repeating the instruction louder.
11. Q: What two properties explain long-conversation drift? → A: Working Memory + Steerability.
12. Q: What does “calibrated trust” mean? → A: Locating your task on each property’s continuum and adjusting verification to match, instead of trusting or distrusting AI wholesale.

Practice Ideas

• Continuum mapping: take 5 recent AI tasks and rate each one (well-worn ↔ novel, mainstream ↔ niche, fits ↔ overflows window, concrete ↔ abstract instruction).
• Specificity check: ask for 5 statistics with sources on a niche topic and verify each one. Record the hit rate.
• Cutoff probe: ask about a recent event with and without web search, and compare the answers and the uncertainty signals.
• Middle-of-document test: put a key fact in the middle of a long pasted document and ask about it. Then front-load it and ask again.
• Steerability ladder: give a crisp format instruction, then an abstract one (“make it more elegant”). Rewrite the abstract one into verifiable criteria.
• Failure diary: for one week, log every unexpected output with the property pair involved and the fix that worked.