Animated SVG showing what the LLM actually receives: integer token IDs (one
layer deeper than llm-advanced.svg). Each of the 32 input tokens displays the
ID prominently with the token text in small italic underneath (e.g., 28133
"Does", 17554 " Chat", 162016 "GPT", 97481 " Claude", 29683 " Anth", 71571
"ropic"). Same 7-iteration autoregressive loop; generated tokens also shown
as IDs. Vocab size labeled V ≈ 200,000. Title formula: f: ℤᵏ → ℝⱽ;
next_id = argmax(f(ids)). ViewBox 1360×600 (wider than the other LLM SVGs).
Co-Authored-By: Claude <noreply@anthropic.com>
Animated SVG showing the same BPE-tokenized prompt from tokens.jpg (32 colored
subword tiles, e.g., "Anthropic" → "Anth"+"ropic", "Perplexity" →
"Per"+"plex"+"ity") feeding into the LLM and generating "Yes, they all use
BPE." token-by-token across 7 iterations. Combines tokenization and
autoregressive generation into one view.
Co-Authored-By: Claude <noreply@anthropic.com>
Screenshot of platform.openai.com/tokenizer showing the sentence "Does ChatGPT,
Claude, Anthropic, Llama, Mistral, Gemini, and Perplexity all use Byte-Pair
Encoding (BPE)?" tokenizing to 32 tokens / 105 characters. Visible tabs:
GPT-5.x & O1/3, GPT-4 & GPT-3.5 (legacy), GPT-3 (legacy) — illustrates that
different model generations use different tokenizers.
Co-Authored-By: Claude <noreply@anthropic.com>
Three-panel SVG (input context, LLM black box, predicted next token) with
7-iteration loop generating "The capital of Japan is Tokyo." from the prompt
"What is the capital of Japan?". Includes purple feedback loop showing each
predicted token appended back into the input.
Co-Authored-By: Claude <noreply@anthropic.com>
Visual aid showing how the model issues tool calls and the harness executes
them against the outside world. Used by the new Tool Calling slide.
Co-Authored-By: Claude <noreply@anthropic.com>
Vector replacement for the previous disrupt-logo.png raster, keeping crisp
rendering at any projection scale.
Co-Authored-By: Claude <noreply@anthropic.com>
Reorganize the screenshot under harness/ alongside other harness-limitation
assets. File contents unchanged (rename detected as 100% similarity).
Co-Authored-By: Claude <noreply@anthropic.com>
Used on learning-journey slide 4 to illustrate Claude guessing the wrong year
— a knowledge-cutoff failure that motivates injecting context (CLAUDE.md,
MCP) rather than relying on better prompting.
Co-Authored-By: Claude <noreply@anthropic.com>
Used on learning-journey slide 4 to illustrate Claude miscounting letters in
"strawberry" — a tokenization failure that motivates why tools (not prompts)
are the real fix.
Co-Authored-By: Claude <noreply@anthropic.com>
Shayan Rais