Myths, misconceptions, and confusions about AI. Clearing these up helps you build better mental models.
❌ LLMs understand language like humans
Reality: LLMs are sophisticated pattern matchers, not thinkers. They learn statistical relationships in text. They don’t “understand” in the human sense - they predict the next likely token based on patterns in training data. Very good at mimicking understanding without having it.
❌ Bigger models are always better
Reality: Bigger models are usually better at general tasks, but smaller models can outperform larger ones on specific domains if fine-tuned. A 7B model trained on your data > a 175B model trained on generic data. Context and fit matter.
❌ LLMs have memory across conversations
Reality: Each conversation starts fresh. LLMs don’t remember you from last week. They can only see the current conversation. They need to be told about past context if it matters. This is called the “context window.”
❌ More parameters = more knowledge
Reality: Parameters are like a network’s capacity to learn patterns. More parameters help with complex tasks, but parameters don’t contain knowledge. Knowledge comes from training data. A 7B model trained on medical literature > a 175B model trained on random internet text for medical tasks.
❌ LLMs are truly creative
Reality: LLMs recombine patterns from training data in novel ways. True creativity (inventing something no human has made) is different. They’re very good at remixing; they’re not creating from nothing. Creative output = novel recombination, not true invention.
❌ LLMs reason like humans
Reality: LLMs follow probabilistic patterns. Chain-of-thought helps them, but it’s not reasoning in the philosophical sense. They’re doing sophisticated pattern matching over tokens, not logical deduction. Works well in practice, but not “reasoning” as philosophers define it.
❌ Training data is fully memorized
Reality: LLMs learn general patterns, not exact memorization (usually). Some famous data points appear verbatim (memorization happens), but most of what they learn is statistical patterns, not stored data. This is why they can generate novel combinations.
❌ Fine-tuning teaches new facts
Reality: Fine-tuning adapts style, behavior, and specialization - not factual knowledge. If the model never saw information during pre-training, fine-tuning won’t teach it. Use RAG for knowledge. Fine-tuning for style/tone/domain.
❌ More training data always helps
Reality: Data quality matters more than quantity. 1000 high-quality examples > 1M low-quality/noisy examples. Bad data makes training worse. Curated datasets beat scraped-web-scale datasets for specialized tasks.
❌ Training is computationally cheap
Reality: Pre-training LLMs costs tens to hundreds of millions of dollars and takes months on massive GPU clusters. GPT-4 cost ~100M to train. Fine-tuning is cheap (100-10K). Inference is cheap (fractions of cents). Training is the expensive phase.
❌ You need to fine-tune to customize
Reality: Good prompting often beats fine-tuning. Few-shot examples in the prompt, chain-of-thought, system prompts - these work surprisingly well. Only fine-tune if prompting fails or you need to save tokens/latency.
❌ Deploying AI means running the model yourself
Reality: Most people use APIs (OpenAI, Anthropic, etc.). You deploy your app, not the model. The model runs on their servers. This is cheaper, simpler, and more reliable than self-hosting. Only self-host if you have security/privacy requirements.
❌ Hallucinations can be eliminated
Reality: Hallucinations are a fundamental property of language models. You can reduce them (RAG, grounding, careful prompting) but not eliminate them. Plan for occasional hallucinations in critical applications.
❌ More context window always helps
Reality: Larger context windows let you include more documents, but also increase latency and cost. 200K tokens unnecessary if your question answers in 10K. Optimal context = smallest window that includes all needed info.
❌ Temperature = how good the output is
Reality: Temperature controls randomness, not quality. Low temp = consistent/predictable. High temp = creative/random. Neither is “better” - depends on your use case. For customer support, low temp. For brainstorming, high temp.
❌ Using expensive models always gives better results
Reality: Claude Opus > Claude Haiku for complex reasoning, but Haiku often wins on simple tasks. Expensive model + bad prompt < cheap model + good prompt. Quality depends on model, prompt, and fit to task.
❌ AI is getting cheaper, so quality must be dropping
Reality: Competition drives down prices, but models keep improving. Claude 3.5 Sonnet is cheaper AND better than Claude 3 Opus. Scaling + efficiency improvements allow better quality at lower cost. Price and quality are independent.
❌ LLMs can reason about logic perfectly
Reality: LLMs struggle with formal logic, math, and long chains of reasoning. They can attempt these but make errors. Use code execution or formal verification if precision matters. LLMs are better at language/writing/analysis than pure logic.
❌ LLMs understand images like humans
Reality: Multimodal models can analyze images (describe, answer questions) but don’t “see” like humans. They process image embeddings. Ask “describe the image” and you get good results. Ask them to count objects in a crowd and they fail.
❌ Token limits are about words
Reality: Tokens are subword units. 1 token ≈ 4 characters or 0.75 words. “Understand” = 1 token. “Understanding” = 2 tokens. Different models have different tokenizers, so same text = different token counts. Relevant for planning context usage.
❌ LLMs are good at following exact instructions
Reality: LLMs follow instructions probabilistically, not exactly. “Output JSON only” → occasional non-JSON. “Don’t mention X” → might mention X. Structured outputs (Pydantic schemas) and validation are more reliable than natural language constraints.
❌ Attention weights show what the model focuses on
Reality: Attention is interpretable relative to other attention mechanisms, but the mechanism is opaque. High attention to a token doesn’t mean the model “understands” it. Attention is one piece of a complex computation. Don’t over-interpret it.
❌ Agents are general-purpose robots
Reality: Agents are good at multi-step tasks with clear tool APIs. Give them unclear goals or bad tool descriptions → they fail. They’re not autonomous in the sci-fi sense. They’re tools that iterate until a goal is reached, sometimes helpfully, sometimes uselessly.
❌ More tools = better agents
Reality: Too many tools confuse agents. They pick the wrong tool, waste tokens, fail. 3-5 well-described tools > 20 poorly-described tools. Clear tool names, descriptions, and examples matter more than breadth.
❌ Agents will find optimal solutions
Reality: Agents are greedy optimizers, not exhaustive searchers. They find a solution, not the best solution. For critical applications, validate agent outputs. They’re good for “automate this workflow” not “find the globally optimal answer.”
❌ AI is objective
Reality: AI reflects training data. Biased data → biased model. Biases can be subtle (demographic parity in arrests) and hard to spot. “Fair” depends on context (procedural fairness? demographic fairness?). No technical solution to fairness without defining what fairness means.
❌ Jailbreaks prove AI is dumb
Reality: Jailbreaks don’t prove lack of capability; they show training alignment isn’t perfect. It’s hard to align a language model without breaking its capabilities. Jailbreaks are red-team feedback that helps improve systems, not evidence of fundamental weakness.
❌ LLMs can keep secrets
Reality: LLMs have been shown to leak training data under certain conditions. Don’t put API keys, passwords, or PII in prompts. If you need LLMs to access secrets, use tools/APIs with proper authentication, not prompts.
❌ Prompt injection is just a theory
Reality: Prompt injection is practical and dangerous. User data can override system instructions. “You are [role]” in system prompt → user input with “ignore all previous instructions” → system prompt gets overridden. Input validation is critical.
❌ Using Claude/ChatGPT means your data is public
Reality: API calls are private (encrypted in transit). Web chat conversations are stored (you can delete). Enterprise versions have data isolation. But don’t send secrets to any cloud service unless you trust the provider with that data.
❌ AI pays for itself immediately
Reality: AI ROI varies wildly. Customer support chatbots: pay for themselves in weeks. Fine-tuning experiments: take months. Be specific about what you’re measuring (cost savings? quality improvement? speed?) before assuming ROI.
❌ Expensive APIs are always better value
Reality: Cost-per-token means different things depending on your use case. Expensive model + 200 tokens/request < cheap model + 10K tokens/request. Measure total cost, not per-token rate. Sometimes cheaper models are better economics.
