The 'System 2' LLM: How Models Learn to Reason (o1, R1)

    Introduction: Two Systems of Thinking In cognitive science, Nobel laureate Daniel Kahneman described human thinking as two distinct systems: System 1: Fast, automatic, intuitive (e.g., recognizing faces, reading emotions) System 2: Slow, deliberate, analytical (e.g., solving math problems, planning) Traditional LLMs operate almost entirely in System 1 mode: they generate responses instantly, token by token, with no deliberate planning or self-reflection. Ask GPT-4 a question, and it starts answering immediately-no visible “thinking time.” ...

    February 10, 2025 · 11 min · Rafiul Alam

    The Zeigarnik Effect: Why Unfinished Tasks Haunt Us

    It’s 2 AM. I should be sleeping. Instead, I’m lying in bed thinking about that bug I almost fixed. I know exactly where the problem is. I know how to solve it. I just ran out of time. My brain won’t let it go. Or that blog post I started writing three days ago. I have the outline. I wrote the intro. But I haven’t finished it, and it’s nagging at me every time I sit down to work. ...

    June 24, 2024 · 13 min · Rafiul Alam