← Extractions

Next Token! — Chris Hasiński on LLM falsehoods

Krzysztof 'Chris' Hasiński's wroclove.rb 2025 talk debunks common LLM misconceptions — they don't really chat, think, or use tools; they only generate tokens. He walks through tokenization, stop tokens, reasoning models, tool calling, embeddings, RAG, MCP servers, and structured output, surveying the Ruby AI ecosystem and noting that everything is outdated within weeks.

Model

claude-opus-4-7

Ingestion

cfe5be57

Input tokens

305,629

fresh: 225,633
cached: 68,568
cache write: 11,428

Output tokens

14,492

Duration

228.1s

Roundtrips

Tool calls

Cost

$0.00

Nodes/edges extracted

23 / 52

Read set (nodes/edges)

115 / 17

Nodes (23)

update Next Token! talk

attrs {"type" => "talk"} → {"date" => "2025-03-14", "type" => "talk"}

description Talk at wroclove.rb 2025. → Krzysztof 'Chris' Hasiński's wroclove.rb 2025 single-speaker talk, reusing slides he has had to re-update for every c...

short_description Talk at wroclove.rb 2025. → wroclove.rb 2025 talk debunking falsehoods about LLMs via tokens and stop tokens.

update Krzysztof Hasiński person

description Ruby developer who goes by 'Chris' because his Polish name is hard for English speakers to pronounce. Filled a missin... → Ruby developer who goes by 'Chris' because his Polish name is hard for English speakers to pronounce. A regular speak...

short_description Ruby developer; wroclove.rb speaker on developer feedback loops and databases. → Ruby developer; wroclove.rb speaker on developer feedback loops, databases, and LLMs.

create LLM Tokens concept

kind (empty) → concept

name (empty) → LLM Tokens

slug (empty) → llm-tokens

attrs (empty) → {"category" => "architecture"}

description (empty) → Underlying output unit of an LLM. Text is broken into tokens that are really numbers; OpenAI's tokenizer visualizes h...

short_description (empty) → Numeric units LLMs emit one-at-a-time when generating text.

create Stop Tokens concept

kind (empty) → concept

name (empty) → Stop Tokens

slug (empty) → stop-tokens

attrs (empty) → {"category" => "architecture"}

description (empty) → Special tokens a program running an LLM watches for to halt generation. Chat is implemented entirely via stop tokens:...

short_description (empty) → Tokens that tell the LLM runner to halt generation.

create llamafile tool

kind (empty) → tool

name (empty) → llamafile

slug (empty) → llamafile

attrs (empty) → {"license" => "open-source", "category" => "tool"}

description (empty) → Project that packages llama.cpp and a model into a single portable binary that runs on macOS, Windows and Linux witho...

short_description (empty) → Single-binary distribution of llama.cpp that runs models across macOS, Windows, and Linux.

create llama.cpp tool

kind (empty) → tool

name (empty) → llama.cpp

slug (empty) → llama-cpp

attrs (empty) → {"license" => "open-source", "category" => "tool"}

description (empty) → Open-source C/C++ engine for running LLMs locally. Underlies llamafile. Exposes parameters like minimum token counts ...

short_description (empty) → C/C++ inference engine for running LLMs locally.

create OpenAI Tokenizer tool

kind (empty) → tool

name (empty) → OpenAI Tokenizer

slug (empty) → openai-tokenizer

attrs (empty) → {"category" => "service"}

description (empty) → Web tool from OpenAI for calculating how many tokens a piece of text will burn. Visualizes how words map to tokens — ...

short_description (empty) → OpenAI web tool for visualizing and counting tokens in text.

create Reinforcement Learning from Human Feedback concept

kind (empty) → concept

name (empty) → Reinforcement Learning from Human Feedback

slug (empty) → reinforcement-learning-from-human-feedback

attrs (empty) → {"category" => "methodology"}

description (empty) → Training step layered on top of base LLM training where models are rewarded for outputs trainers actually preferred. ...

short_description (empty) → Training step where models are tuned to produce outputs humans prefer.

create Reasoning Models concept

kind (empty) → concept

name (empty) → Reasoning Models

slug (empty) → reasoning-models

attrs (empty) → {"category" => "architecture"}

description (empty) → LLMs like OpenAI o1/o3 and DeepSeek that appear to 'think' via a thinking/reasoning box in the UI. Hasiński's debunki...

short_description (empty) → LLMs that emit hidden chain-of-thought tokens before the user-visible answer.

create LLM Tool Calling concept

kind (empty) → concept

name (empty) → LLM Tool Calling

slug (empty) → llm-tool-calling

attrs (empty) → {"category" => "pattern"}

description (empty) → Agentic tool use as implemented today: the LLM does not actually call anything. The system prompt defines a 'tool' ro...

short_description (empty) → Pattern where LLMs 'call tools' by emitting structured invocations caught via stop tokens.

create MCP Server concept

kind (empty) → concept

name (empty) → MCP Server

slug (empty) → mcp-server

attrs (empty) → {"category" => "architecture"}

description (empty) → Model Context Protocol server: a meta-tool that exposes a list of tools the LLM can discover and invoke, so only rele...

short_description (empty) → Meta-tool protocol letting an LLM discover and call other tools on demand.

create Ruby LLM tool

kind (empty) → tool

name (empty) → Ruby LLM

slug (empty) → ruby-llm

attrs (empty) → {"license" => "open-source", "category" => "library", "language" => "Ruby"}

description (empty) → Ruby gem providing a cleaner, more readable wrapper over LLM APIs than early LangChain-style scanning for magic marke...

short_description (empty) → Ruby gem wrapping LLM APIs with readable code; growing very fast.

create neighbor tool

kind (empty) → tool

name (empty) → neighbor

slug (empty) → neighbor

attrs (empty) → {"license" => "open-source", "category" => "library", "language" => "Ruby"}

description (empty) → Ruby gem by Andrew Kane that adds vector/embedding search to Active Record over Postgres (pgvector) and SQLite, enabl...

short_description (empty) → Ruby gem by Andrew Kane adding vector search to Postgres and SQLite via Active Record.

create Baron tool

kind (empty) → tool

name (empty) → Baron

slug (empty) → baron

attrs (empty) → {"category" => "library", "language" => "Ruby"}

description (empty) → Ruby gem for splitting (chunking) content into meaningful parts so it embeds better for RAG pipelines. Hasiński menti...

short_description (empty) → Ruby gem for chunking content into meaningful pieces for embedding.

create SigLIP tool

kind (empty) → tool

name (empty) → SigLIP

slug (empty) → siglip

attrs (empty) → {"category" => "service"}

description (empty) → Multimodal embedding model from Google that Hasiński's workshop at wroclove.rb 2025 used. Enables embedding of images...

short_description (empty) → Google multimodal embedding model used during wroclove.rb 2025 workshop.

create Hybrid Search concept

kind (empty) → concept

name (empty) → Hybrid Search

slug (empty) → hybrid-search

attrs (empty) → {"category" => "pattern"}

description (empty) → Common retrieval strategy for LLM-powered applications: mix classic word-based search with vector-embedding search (a...

short_description (empty) → Combining classic keyword lookup with vector/embedding search for retrieval.

create Structured LLM Output concept

kind (empty) → concept

name (empty) → Structured LLM Output

slug (empty) → structured-llm-output

attrs (empty) → {"category" => "pattern"}

description (empty) → Technique for getting parseable data out of an LLM despite it only emitting text. Old approach (still visible in Lang...

short_description (empty) → Forcing LLMs to emit valid JSON/YAML via schema-prompting or server-side enforcement.

create Falsehoods Programmers Believe About Time resource

kind (empty) → resource

name (empty) → Falsehoods Programmers Believe About Time

slug (empty) → falsehoods-programmers-believe-about-time

attrs (empty) → {"type" => "article"}

description (empty) → Widely-cited article listing misconceptions programmers carry about time — e.g. 'a month always has 30 or 31 days' (n...

short_description (empty) → Popular article listing misconceptions programmers hold about time.

create Falsehoods Programmers Believe About LLMs concept

kind (empty) → concept

name (empty) → Falsehoods Programmers Believe About LLMs

slug (empty) → falsehoods-programmers-believe-about-llms

attrs (empty) → {"category" => "practice"}

description (empty) → Hasiński's contribution to the falsehoods genre: a list of misconceptions programmers hold about LLMs. Headlined fals...

short_description (empty) → List of common misconceptions programmers hold about how LLMs work.

create LLMs Are Just Token Generators takeaway

kind (empty) → takeaway

name (empty) → LLMs Are Just Token Generators

slug (empty) → llms-are-just-token-generators

attrs (empty) → {"type" => "insight"}

description (empty) → Central takeaway of 'Next Token!': LLMs can't chat, think, or use tools — they are a magical next-token generator. Ch...

short_description (empty) → Everything beyond 'predict next token' is scaffolding written around the LLM.

create MCP Servers Have No Security Model takeaway

kind (empty) → takeaway

name (empty) → MCP Servers Have No Security Model

slug (empty) → mcp-servers-have-no-security-model

attrs (empty) → {"type" => "warning"}

description (empty) → Warning from Hasiński at wroclove.rb 2025: MCP servers are a nice way to expose many tools to an LLM with efficient c...

short_description (empty) → Using an MCP server effectively hands control of your LLM to whoever runs it.

create LangChain Is Dead takeaway

kind (empty) → takeaway

name (empty) → LangChain Is Dead

slug (empty) → langchain-is-dead

attrs (empty) → {"type" => "insight"}

description (empty) → Hasiński's 2025 update: LangChain used to serve as the 'ORM for your LLM' when token-stream scanning and custom forma...

short_description (empty) → LangChain's abstractions are obsolete now that LLM servers provide unified APIs.

create Fault-tolerant token-by-token parsing for structured output question

kind (empty) → question

name (empty) → Fault-tolerant token-by-token parsing for structured output

slug (empty) → fault-tolerant-token-by-token-parsing-for-structured-output

attrs (empty) → {"answer_summary" => "Yes — llama.cpp already does something like this for structured output, and proprietary provide...

description (empty) → Audience question: since LLMs generate one token at a time with full context visibility, could you hook a fault-toler...

short_description (empty) → Can a parser validate each generated token live and retry invalid ones?

Edges (52)

update Krzysztof Hasiński → authored → Next Token!

context (empty) → Hasiński delivered the 'Next Token!' talk at wroclove.rb 2025.

update Next Token! → presented_at → wroclove.rb 2025

context (empty) → Delivered on 2025-03-14 at wroclove.rb 2025.

create Krzysztof Hasiński → attended → wroclove.rb 2025

context (empty) → Speaker at wroclove.rb 2025.

relation (empty) → attended

source_node_id (empty) → 5d1051ea-ed4d-4fd6-a87f-5976ec8aa8f8

target_node_id (empty) → 4a9688cf-482e-48ac-a0b5-de88900bdb2b

create Next Token! → about → Large Language Models

context (empty) → Core subject — debunking how LLMs actually work.

relation (empty) → about