Unverified50% confidenceFactExact time
Scaling Law主要适用于预训练损失（交叉熵）这一代理指标，而非所有下游任务的实际性能
1
Sources
50%
Confidence
Long-term
Relevance
7/15/2026
First Seen
Sources
追新AI模型值得吗？大厂工程师的灵魂拷问与理性答案
Related Claims
UnverifiedApplication-layer jailbreak defenses must simultaneously rely on robustness improvements at the training stage and dynamic monitoring at the inference stage, as reinforcement along a single dimension yields limited results.70% similarUnverified对于特定的垂直任务，模型不必越大越好；训练数据质量和音乐表示方式往往比参数规模更能决定最终效果68% similarUnverifiedThe Chinchilla law states that under a fixed compute budget, appropriately reducing model size and increasing training data volume often yields better performance.67% similarUnverified分配给预训练的算力越多，下游任务的pass@1奖励就越高，预训练损失越低模型在后续RL中获得的收益越大67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/517804API
curl https://kongchang.com/api/v1/knowledge/claims/517804MCP
get_claim(id=517804)