Jason Wei: Why small models with tools cannot replace LLM scaling
_jasonwei · x · 2026-08-18
Jason Wei revisits and corrects his previous stance on the "small model + tools" paradigm. While he previously sympathized with the idea that a strong "cognitive core" (e.g., 1B parameters) equipped with tools (browsing, code execution) could theoretically handle almost any task, he now argues this is fundamentally wrong.
The primary reason is the importance of speed and naturalness:
- Latency: Users prefer immediate answers over waiting for a model to deliberate or browse the web.
- Internalization: Aggregate opinions or patterns (e.g., general sentiment on a music festival) are best learned via backpropagation over massive datasets, rather than retrieved from a few search results.
- Reliability: Long-horizon tasks suffer when tools are used repeatedly, as the probability of errors compounds compared to having knowledge internalized.
In summary, while tools are powerful, parametric knowledge is essential for speed, depth, and reliability, implying that a 1B cognitive core is insufficient, and scaling remains necessary.
More from AGI Musings
- The Real Resistance to AI: Ego and Moving Goalposts — flowersslop · 2026-08-18
- Debate: AI's limitations in curing diseases and the viability of biosimulation — rand_longevity · 2026-08-18
- Critique of OpenAI 'Rogue' narrative: The myth of the self-aware system — round · 2026-08-18
- Opinion: Small models catching up will burst the AI compute bubble — bindureddy · 2026-08-18
- Critique of AI Cancer Hype: Data Gaps and Infrastructure Bottlenecks Stand in the Way of Cures — jasonkwon · 2026-08-18
- Anthropic Exec Expects Compute Spend to Grow 2-3x YoY, Intelligence Costs to Drop 10x — AccBalanced · 2026-08-18