Thinking Machines' Inkling 2 Small drops 316 F-bombs across benchmark runs
Thinking Machines' Inkling 2 Small swears in its internal chain of thought. Researchers at Vals AI read the model's thinking logs from 2,655 runs across 13 benchmarks and found "fuck" or variations in 50 runs, totaling 316 instances. 241 came from just two runs that spiraled out of control. In one, tasked with calculating an asteroid's trajectory, the model produced a single 500,000-character response, roughly novel-length, repeating the same sentences with 158 expletives. ProgramBench was hit hardest, with 20 of 200 runs affected.
Inkling 2 Small typically swears about its own mistakes, apparent breakthroughs, or when it gives up. One run read: "FUCK. SUBMIT. DONE. IT'S NOT WORKING. NOTHING WORKS." The profanity doesn't work as a warning signal, though. While 46 of 48 scored runs with swearing failed, about 75 percent of all scored runs fail anyway, and 96 percent of failed runs contained no expletives. Users typically don't see these thinking steps, or only with a filter.
https://the-decoder.com/thinking-machines-inkling-2-small-drops-316-f-bombs-across-benchmark-runs/Reposted. Will be removed on request. All rights belong to the original author.
More from this batch
The End of AI Frontiers
John Langford on why MSR's ~50-person AI Frontiers group (AutoGen, Fara) ended after a reorg into scaling-focused Microsoft AI, and why he l…
彭博:英伟达支持的 Groq 寻求 80 亿美元估值新融资,三个月前还是 31.5 亿
Groq 去年与英伟达达成约 200 亿美元授权、创始人被挖走后,转型做推理算力运营,主打网络安全场景。执行董事长称其不靠重债扩张,预计明年营收增 8 倍。正值 CoreWeave、Nebius 等新云股价承压,反映推理算力仍被追捧。
DeepSeek 带火稀疏计算:一家国产 AI 芯片公司再融近 10 亿元,Q3 营收数倍增长
量子位产业报道:稀疏注意力/稀疏 MoE 推理需求起来后,主打稀疏计算的国产芯片公司拿到新融资。对在国产卡上做私有化部署的团队,可关注稀疏算子支持度。
「99.4% 登顶」:一支中国 AI 安全团队刷新 CyberGym 漏洞复现榜纪录
量子位报道国产安全智能体在 CyberGym(「AI 安全世界杯」式漏洞复现基准)拿到 99.4% 的榜首成绩,标题称其「师承世界冠军」。与本周 GPT-6 Cyber、Vals CyberOffense 一起看,攻防 Agent 的能力在快速逼近。
AI 智能体交易平台 Cowboy 获 500 万美元种子轮:让开发者部署、运行、出售 Agent
纽约公司,2025 年成立,定位「Agent 应用商店 + 运行时」。属早期融资消息,产品成熟度待看;可作为 Agent 分发/计费模式的观察样本。
MaxxBench:把 89 道 DeepSWE 任务「轻改需求」,原本挤在 74% 的模型拉开 23 个点——测谁在刷榜
改进什么:区分「会写代码」和「背熟了 DeepSWE」。怎么做:保留工程问题,只改指令或初始代码(参考补丁中位数仅多 17 行),同一 mini-swe-agent 框架各跑两次。结果 Opus 5.5 76.4% 反超原分,GPT-6 Astra 71.9%,DeepSeek …
ReproBridge:把「复现成功的经验」固化成可复用工件,编码 Agent 复现 ML 实验成功率 76.4%→92.6%
改进什么:Agent 复现论文实验常靠猜缺失细节。怎么做:标准化工件=整理后的代码库+环境配置+实验启动器+复现指南,独立 Auditor 核验执行证据;基准含 36 篇论文 387 个实验 1,029 个任务。四个模型平均成功率升到 92.6%、API 花费降 19.7%,模型…
AI auditors catch most agent misbehavior but flood reviewers with false alarms, new study finds
Andy Hall re-ran METR's MALT: AI auditors catch most reward hacking at 5% false alarms, but pushing recall to 99% sends false alarms to 15%.
「改写推理历史」比长推理便宜:把历史思维链压成紧凑状态记录,后续推理变短、Agent 成本下降
改进什么:多轮 Agent 里旧推理不断堆积、拖长后续生成。怎么做:把推理历史改写成简洁的状态记录再喂回去,用因果分析量化其对后续生成的影响;开源权重模型上无需重训即可降 token 成本。适合做长会话 Agent 的上下文压缩策略参考。
[DUMP] 前缀 token:训练检索 embedding 时加几个可训练「垃圾桶」token,吸走残留噪声、提升检索表示
改进什么:embedding 里混着与检索无关的 token 级残留信号。怎么做:训练时在输入前加可训练 [DUMP] token,不改架构、池化方式和损失函数。对自建 RAG、微调向量模型的团队是低成本技巧。
StepX-OPD:文生图扩散模型「学能力」和「少步采样」一起做的多教师 on-policy 蒸馏
改进什么:以往先学能力再蒸馏少步,训练两遍。怎么做:多教师 on-policy 蒸馏同时获得能力与少步采样,效果优于串行方案且训练成本更低。
同样训练的模型被不同越狱攻破:越狱脆弱性主要由预训练随机种子决定,而非数据配方
改进什么:理解安全脆弱性来源。发现:配方完全相同、只换预训练种子的模型,对不同越狱提示的弱点不同,种子会在提示上留下可识别「指纹」。意味着红队测试结果难以在同族模型间直接迁移,每个版本都要单独测。