Models9月17日 19:19

Meta 推出最新版 Llama 模型

Keep reading

Recent stories with the same tag or section

Industry BuildersMulti-source00:07

Chinese contentMeta 借力咨询巨头向企业客户推广 AI 产品

Meta 卖企业 AI 卖不动，Amazon 和 Walmart 都怕数据被拿走，现在拉咨询公司当中间人来破局。对比 OpenAI、微软靠云渠道直销的路子，这条路径更迂回。做企业 AI 采购或者研究大厂商业化的人可以看看。

#Meta #Llama #企业AI #Amazon

Event

The Information@theinformationSource

Papers Builders15:56

Chinese content论文揭示微调引发人格泛化遵循跨模型共享的线性规律

微调一个领域会意外改变模型在其他领域的行为，这篇论文给出了可预测的规律：域内外人格率呈线性关系，Qwen 和 Llama 都成立。做微调的开发者能直接用上——只调后几层可以在不牺牲任务效果的情况下减少不安全行为迁移。

#微调 #emergent misalignment #Qwen #Llama

arXiv cs.LG@Olina Mukherjee et al. (4)Source

Industry US vs China13:24

Chinese contentMenlo Ventures 报告：中国开源模型占全球企业支出 89%，Kimi 居首

月之暗面 Kimi 拿下全球企业开源模型支出 61%，DeepSeek 和 GLM 分列二三，Meta Llama 才 7%，这组数据值得看看。

#Kimi #DeepSeek #GLM #Llama

IT之家Source

Papers BuildersSelected00:54

Stanford paper maps the in-context learning circuit in small LLMs

The practical implication is that few-shot prompting works by label lookup, not feature learning, so prompt design should repeat labels several times and use classes the model can already distinguish. Builders tuning small models for agent pipelines or classification tasks can use this to cut prompt size and failure cases. The finding also explains why tiny models fail on tasks needing unseen features.

#Qwen3 #Llama-3.2 #SmolLM3 #prompt engineering