CHAPTERS
- 0:00 – 0:32
What’s really happening when an AI “thinks” before replying
Jane reframes the “thinking” pause: the model isn’t browsing the internet or pulling from a database in real time. Instead, it generates a reply by predicting what should come next based on what it learned during training.
- •AI responses are generated, not fetched from a live database
- •The model produces text incrementally, token by token (appearing word-by-word)
- •Prediction is the core mechanism behind modern language models
- •Sets up misconceptions (search engine, copying, reading the internet) to correct them
- 0:32 – 1:02
From keyboard autocomplete to large-scale prediction
The video uses phone autocorrect as a familiar example of prediction, then explains why LLMs are fundamentally more capable. Unlike simple next-word suggestions that only look at a few prior words, an AI model draws on broad learned patterns.
- •Autocomplete is also prediction, but very shallow
- •Simple predictors use only a tiny local context (last few words)
- •LLMs leverage patterns learned from massive training data
- •The depth of context and learned structure enables coherent long-form output
- 1:02 – 1:33
Training: billions of next-word guesses and gradual adjustment
Jane describes pretraining as repeated cycles of guessing the next word, checking the error, and adjusting the model. This iterative process over huge datasets builds the model’s ability to generate fluent, relevant text.
- •Pretraining loop: see text → predict next word → measure error → update
- •Repeated at enormous scale (billions of rounds)
- •Training data spans vast amounts of text and other information
- •Generative ability emerges from accumulating many small adjustments
- 1:33 – 2:04
Fine-tuning: shaping helpfulness and reducing harm
After pretraining, a fine-tuning stage rates full answers to steer the model toward more useful and safer behavior. Ratings can come from people or from written guidelines used to evaluate outputs.
- •Fine-tuning happens after base training
- •Full responses are evaluated (human or guideline-based)
- •Model is nudged toward helpful, less misleading, less harmful outputs
- •Both pretraining and fine-tuning contribute to overall capabilities
- 2:04 – 2:34
Training cutoff and when the model may need tools like web search
The model’s knowledge is bounded by a training cutoff date; after that, it won’t reliably know new facts without tools. Even if search exists, users shouldn’t assume it was used unless explicitly requested and cited.
- •Knowledge is reliable only up to the training cutoff date
- •For recent info, the model may need browsing/search tools
- •Don’t assume web search occurred automatically
- •When asked to search, the model can provide sources for verification
- 2:34 – 3:05
How the model ‘thinks’: global context behind next-token prediction
Although output is produced one token at a time, good prediction requires modeling the direction of the sentence and the goal of the answer. The model conditions on everything available in the conversation and instructions to decide each next piece of text.
- •Token-by-token output doesn’t mean token-by-token reasoning
- •Strong prediction requires planning-like context sensitivity
- •The model uses prior messages, prompts, system instructions, and uploads
- •This conditioning process repeats for every next token until completion
- 3:05 – 3:35
Why this matters: generation explains creativity and confident errors
Understanding the model as a prediction system helps explain both its strengths and its failure modes. It can produce novel text because it’s generating patterns, and it can hallucinate because a plausible-sounding continuation isn’t always true.
- •Generation (not retrieval) enables novel, never-before-seen responses
- •Confidence and polish are stylistic artifacts of generation
- •Hallucinations occur when “good-looking” answers diverge from reality
- •Human judgment remains essential
- 3:35 – 4:05
Habit 1: Provide context and define what ‘good’ looks like
Jane’s first practical tip is to give the model the situational context it needs to predict a useful completion. Details about goals, audience, constraints, and success criteria improve results.
- •Share who you are and what you’re working on
- •Specify goals, constraints, and desired format/quality bar
- •Context becomes part of the pattern the model continues
- •Better inputs produce more aligned outputs
- 4:05 – 4:35
Habits 2–4: Watch the cutoff, request variations, and verify outputs
Jane finishes with three more habits: account for outdated knowledge, ask for multiple options since there’s no single stored answer, and double-check results—especially when stakes are high. The takeaway is to treat outputs as drafts that require scrutiny.
- •Remember the cutoff for anything time-sensitive (news, prices, recent events)
- •Ask for options or different tones to explore better drafts
- •Don’t equate confidence with correctness
- •Verify facts proportional to the consequences of being wrong
- 4:35 – 4:47
Wrap-up: pattern completion and experimenting across models
The closing ties it together: every prompt is the start of a pattern the model completes using training. Since different models complete patterns differently, experimentation helps users find what works best, and the channel points to further research updates.
- •Every message seeds a pattern the model completes
- •Completions are shaped by learned training patterns
- •Different models may produce different completions for the same prompt
- •Further learning resources are available via Anthropic’s research/blog
