Model scaling promises that larger models, more data, and more compute will resolve quality and reasoning failures. Capability has risen, but hallucination and the need for external verification have persisted across generations. section 1.2
Use scaling as a capability strategy, not as a substitute for verification:
- Increasing model size, data volume, and related resources can produce predictable performance gains. section 1.2
- Benchmark improvements do not eliminate hallucination, because probable next-token predictions can diverge from truth. section 1.2
- Treat successive releases as iterative refinement and preserve external checks instead of assuming a larger model knows when it is wrong. section 1.2
Capability without reliability
Across the progression from the first GPT systems to current frontier models, reasoning, competitive-programming, and agentic benchmark scores improve while the underlying hallucination failure mode remains. section 1.2
