Debate: Was Tool Calling an Unpredictable Leap, or Predictable Since 2022?
OwariDa · x · 2026-11-16
In a thread sparked by Gary Marcus and ardanmichel1 arguing over whether "Claude Code is just a giant LLM," OwariDa pushes back: calling tool use an unpredictable leap that turned LLMs into the neurosymbolic AI Marcus allegedly envisioned is silly. He cites his own 2022 reply to Eliezer Yudkowsky — scaling LLMs alone won't yield AGI, but training models to emit special sequences for interacting with external systems, receiving output, and continuously training might — and notes he built an agent harness working with both GPT-3 and the first GPT-3.5-based ChatGPT days after that tweet, showing the "LLM + tools" idea was anticipated long ago.
More from AGI Musings
- Gary Marcus challenges dev: show symbols in your agent harness or you're confused — GaryMarcus · 2026-11-16
- Dean Ball: top AI models mode-collapse into 2020-era governance discourse — deanwball · 2026-11-16
- JOBhakdi pushes back on 'Musk just got lucky with money' view of innovation — JOBhakdi · 2026-11-16
- The AI doom market sells psychological certainty, not probability — YogeshMalik · 2026-11-16
- Runway CEO: AI's history is an ideology of 'everyone gets to make' — c_valenzuelab · 2026-11-16
- The Autonomy Gradient: a framework for how much to delegate to AI — DrKavner · 2026-11-16