Tag
The author measures how lossy inference-engine flags (routing bias and tail skip) degrade model quality, showing that shipped defaults cost ~1.5 accuracy points with plenty of VRAM but ~3x more on small cards, while aggressive settings produce confident nonsense.
论文提出IEC框架，研究语言模型在受控指令负载（如无意义的单词替换规则）下的行为与性能退化，揭示了规则数量、替换密度、规则位置等因素遵循的数学规律，并在GPT-4o mini上验证，为模型建立可复现的"能力护照"。
The content discusses the common issue of AI models degrading in performance over time after initial excitement, and questions whether longitudinal benchmark studies have been conducted to track trends across models and labs.
OpenAI is reportedly silently routing some Astra and Codex accounts to less capable models due to account activity restrictions, causing performance issues. The company has acknowledged limiting access based on activity, which some users find unethical, especially affecting a large number of Chinese users.
Users report that the Astra AI model has been secretly downgraded, leading to poor performance despite high-effort prompting, with concerns about compute cost savings and calls for FTC investigation.
The author discusses using recursive summaries in their chat app 'harness' to manage context but observes model degradation with Qwen Flash Next after many messages, questioning if conversation compaction pollutes model coherence.
The post claims that OpenAI Codex assigns a downgraded gpt-5.5-mini model to domestic users at the Ultra reasoning level, provides a detection prompt, and calls on users to reply 1 or 2 to measure the downgrade rate for domestic users.
The author accuses Anthropic of intentionally degrading the performance of their AI model Fable after an initial trial period, citing this as evidence that closed-source AI companies are predatory and that open-source alternatives will prevail.
A user highlights significant performance degradation in Claude Fable 5 after recent updates, with benchmark scores dropping drastically in debugging, refactoring, and hallucination tasks, calling for regulation to address potential consumer scams in AI model behavior.
Anthropic apologized and reversed a policy that secretly degraded performance of its Claude Fable 5 model for users working on advanced AI development, sparking debate on safety vs. openness.
User observes that the opus-4.8 model has degraded in performance since its launch.
Users report that OpenAI's Codex GPT-5.5 high model performance has degraded, exhibiting laziness, hallucinations, and context loss. Suspecting it's due to OpenAI training GPT-5.6, they need to enable xhigh mode to restore normal performance.
User reports that Qwen3.6 models running on llama.cpp server become significantly less capable after ~2 weeks of continuous operation, and restarting sessions does not resolve the issue.
A tool that tracks the ELO history of major AI models from the LMSYS Arena leaderboard, revealing hidden trends like performance degradation and upgrades over time.
A user running multiple agents reports that after upgrading to GPT-5.5, the model suddenly became less capable at executing tool calls and more prone to giving suggestions instead of acting, speculating OpenAI may be throttling for load management.
MiniMax published a technical blog post providing an in-depth analysis of the systematic vocabulary degradation issue behind its M2 series large models' inability to output specific personal names. It reveals parameter shifts caused by a disconnect in data coverage between pre-training and post-training stages, and proposes an effective solution involving full-scale synthetic data for remediation.
A user documents how closed models (GPT-4o→5.3, Gemini) degraded and censored Chinese novel translations, while local Gemma 4 31B now outperforms them with natural, uncensored output.