AI Adventures: A Programmer’s Journey

Why "Out-of-Distribution" Misses the Point About AI Creativity

A common worry: LLMs can only remix their training data, so they can’t produce genuinely new knowledge.

Oct 15 • 

Zbigniew Łukasiak

Reinventing Daydreaming Machines

A Search‑Space–First Experiment

Oct 13 • 

Zbigniew Łukasiak

July 2025

Dreaming Machines

Why Gwern's AI Proposal Needs More Than Random Connections

Jul 19 • 

Zbigniew Łukasiak

2

Behind the Scenes

Why I'm Publishing My AI Writing Process

Jul 17 • 

Zbigniew Łukasiak

December 2024

What is a response?

The Missing Piece in LLM Compatibility Layers: Comprehensive Documentation

Dec 16, 2024 • 

Zbigniew Łukasiak

Two roles of tool schemas

Today I learned

Dec 12, 2024 • 

Zbigniew Łukasiak

October 2024

Semantic leakage - continued

How do we decide if an LLM answer is wrong?

Oct 27, 2024 • 

Zbigniew Łukasiak

Semantic leakage - quick notes

The canonical example from Does Liking Yellow Imply Driving a School Bus?:

Oct 25, 2024 • 

Zbigniew Łukasiak

2

LLMs and reasoning

Multi-level Pattern Matching and why better sampling seems to be a low hanging fruit right now

Oct 19, 2024 • 

Zbigniew Łukasiak

1

o1, inference time, Turing Machines and stuff

Why Turing Completeness in LLMs is a red herring

Oct 16, 2024 • 

Zbigniew Łukasiak

1

August 2024

Separating Data from Instructions

Introducing Prompete - a LiteLLM wrapper for prompting with templates

Aug 24, 2024 • 

Zbigniew Łukasiak

1

June 2024