Lesson 3B: Capabilities & limitations | AI Fluency: Framework & Foundations Course

Main ideas & concepts (what generative AI/LLMs can do)

• Versatile language skills LLMs (e.g., Claude) can write and communicate effectively across many tasks, such as:

  • Drafting emails in a user’s “voice”
  • Condensing long reports into summaries
  • Translating between languages
  • Explaining complex topics across many domains (e.g., microbiology, marketing strategy)

• Multi-tasking without additional training The same model can switch between very different tasks (e.g., brainstorming party ideas to explaining quantum computing concepts) using only prompts/conversation.

• Context awareness within a conversation LLMs can maintain the “thread” of a conversation by remembering what was said earlier in that interaction. Example behavior described: if you mention a project deadline and refer back later, the AI typically understands.

• Extended capability via external tools/data Some modern LLMs can connect to external tools and information sources, such as:

  • Searching the web
  • Processing files
  • Using other applications This can expand what the AI can help with beyond its built-in knowledge.

---

Main limitations (what generative AI/LLMs can’t reliably do)

• Training-data boundary / knowledge cutoff LLMs only have innate knowledge up to their training time. If a model has a cutoff (e.g., November 2024), it may not have been trained on events after that. Implication: to learn recent developments, models may require tools like web search.

• No guarantee of factual correctness from training The training process may not “verify” every fact learned. As a result, models can reproduce inaccurate information from their training data.

• Hallucinations LLMs can produce responses that sound plausible but are incorrect. Mechanism described: unlike search engines that retrieve existing documents, LLMs generate text using statistical patterns, which can lead to confident errors.

• Context window limits LLMs can only process/retain a limited amount of information per interaction (the context window). If input exceeds the limit:

  • The AI won’t remember earlier parts outside the window
  • It typically behaves like first-in, first-out for dropped context Implication: large documents may not be fully handled, and long conversations may be partially forgotten.

• Non-deterministic outputs (variability) With the same prompt, an LLM may produce slightly different answers each time due to probabilistic choices in generating the “next token,” influenced by model settings. Some interfaces allow adjusting randomness via temperature: higher/lower temperature (as a concept) can affect variability/creativity. Implication: consistency and accuracy may require careful settings or verification.

• Limitations on complex reasoning (historical) LLMs have historically struggled with complex multi-step reasoning, especially math/logic problems. Newer “reasoning” or “extended thinking” models are improving step-by-step performance.

• Tool/data access limitations Even with tool access, models may still lack the specific data source or specialized tools required for certain tasks. Analogy used: a brilliant colleague without access to an internal company database. Implication: if the necessary data/tool isn’t available, the model can’t fully solve the task.

---

Methodologies / approaches mentioned for improving limitations

• Retrieval-Augmented Generation (RAG) Connects the model to external knowledge/data sources to reduce reliance on outdated or missing internal knowledge.

• Expanding tool use Improving how models use external tools/apps to perform tasks more effectively.

• Improving reasoning capabilities Using specialized “reasoning” or “extended thinking” approaches for better step-by-step problem solving.

---

Core lesson / practical guidance for “AI fluency”

• Learn the strengths and limitations to collaborate effectively Treat LLMs like a “new colleague”: understanding what they do well vs. where they fail improves outcomes.

• Use humans + AI together Best applications leverage complementary strengths:

  • Humans: critical thinking, judgment, creativity, ethical oversight
  • AI: speed, scale, pattern recognition, processing vast information Continued learning/experimentation helps users stay current with rapid changes.

• Hands-on practice in the course The course exercises involve interacting with Claude to build an intuitive sense of what AI can/can’t do and how to work with it.

---

Speakers / sources featured

• Claude (mentioned as the example LLM system used in the course exercises)
• The course instructor / course narrator (no individual name provided in the subtitles)