Things You Never Dared to Ask About LLMs

Explore the hidden corners of Large Language Model (LLM) technology in this 46-minute Devoxx conference talk by Guillaume Laforge. Delve into the enigmatic aspects of LLMs that often leave developers puzzled, addressing questions you've never dared to ask. Learn why LLMs generate tokens instead of words and where these tokens originate from. Understand the distinction between "foundation"/"pre-trained" models and "instruction-tuned" ones. Discover how hyperparameters like temperature, top-p, and top-k actually influence token selection. Grasp the meaning behind various number encodings in quantization (fp32, bfloat16, int8) that make models smaller. Find out if LLMs truly understand the Base64 language despite their translation capabilities. Examine LLMs' imperfections, including hallucinations (or confabulations), the reversal curse that causes models to ignore facts from different viewpoints, and how context influences responses even at low temperatures. Buckle up to dispel the magic of LLMs and confront those questions you've been hesitant to ask!