▼Frontier Model Efficiency Gains

Score history

Daily conviction score, 0 to 100. Higher means the thesis is more strongly corroborated.

Now 94 · -2 since Aug 27 · ranged 94 to 96

Showing the last few days. Unlock full score history.

Why the conviction moved

• Aug 27
  Strengthened +7

  Alibaba's Qwen3.8-Flash-Next previews the Qwen4 architecture with only 6 billion active parameters, pairing a 125-billion-parameter mixture-of-experts core with 51 billion parameters of N-gram lookup memory, and was trained for a ninth of its predecessor's cost. A 9x training-cost reduction achieved by moving capability into retrieval-style memory rather than active compute further decouples capability from raw scale spend.

• Aug 26
  Weakened

  Meta's internal documents, reported by The Information, say the October Watermelon model required ten times the compute of its predecessor to reach claimed GPT-5.5 parity on Meta's own benchmarks. A 10x training-compute step to reach the current frontier is the scale-bound path the thesis expects to be loosening, not evidence of capability per unit compute improving.

Showing the last 2 days. Unlock the full record.

Source trail

• Supporting · August 27, 2026

  Alibaba's Qwen Team Previews Its Qwen4 Architecture With a 6-Billion-Active-Parameter Open Model

  Alibaba's Qwen3.8-Flash-Next previews the Qwen4 architecture with only 6 billion active parameters, pairing a 125-billion-parameter mixture-of-experts core with 51 billion parameters of N-gram lookup memory, and was trained for a ninth of its predecessor's cost. A 9x training-cost reduction achieved by moving capability into retrieval-style memory rather than active compute further decouples capability from raw scale spend.

  AI ML Big Data (Telegram)

• Contradicting · August 26, 2026

  Meta Prepares a Consumer AI Agent Called Hatch and a New Model It Says Needed Ten Times the Compute

  Meta's internal documents, reported by The Information, say the October Watermelon model required ten times the compute of its predecessor to reach claimed GPT-5.5 parity on Meta's own benchmarks. A 10x training-compute step to reach the current frontier is the scale-bound path the thesis expects to be loosening, not evidence of capability per unit compute improving.

  AI Post (Telegram)