共收录 22 篇深度内容
📌 今日概览
- 开源权重模型竞争升温：Inkling 2 Small 以 31 分成为西方领先者，Nanbeige4.3-33B 主打笔记本本地运行，Hugging Face 把 ML Intern 做成可安装的 Slack 机器人。
- 模型评测走向具体任务：花 2 万美元让模型对玩《血染钟楼》测临场欺骗，GPT-6 Cyber 直接用于挖 Firefox 与 Linux 内核漏洞。
- AI 资本叙事出现分歧：一边是股价大跌与「清算日」的讨论，另一边是生产率年化 3.6% 的转折信号与 Anthropic 上市的呼声。
- 人才与工具争夺加剧：Serval 为首周带薪休假加 3 万美元推荐奖金，Claude Code 被用来搭建个人开支追踪工具。
🔥 必读 Top3
- Artificial Analysis 作者报告Thinking Machines发布Inkling 2 Small，在Artificial Analysis智能指数得31分；若按计划开放权重，它将成为中国以外得分最高的开源权重模型，多项基准较前代大幅提升。
入选理由：西方开源权重模型出现明显领先者，且公布 Apache 许可与多项基准提升，是当日最完整、可信度最高的模型发布信息。 - Dwarkesh Patel 引用推文称作者花2万美元API额度让不同AI模型对玩社交推理游戏《血染钟楼》，模型在临场欺骗与团队推理上进步明显；发帖者称OpenAI模型大胜Anthropic，Astra对Opus 5.5胜率82%。
入选理由：用 2 万美元 API 额度让模型对玩社交推理游戏，提供临场欺骗与团队推理的实证观察，兼具评测方法新意与结论争议。 - Chubby♨️ 作者称 OpenAI 发布网络安全模型 GPT-6 Cyber，引用官方说法称其发现 81 个此前未知的 Firefox 高危漏洞，并协助开发 948 个 Linux 内核漏洞的利用方式。
入选理由：披露 GPT-6 Cyber 发现 81 个未知 Firefox 高危漏洞并协助开发 Linux 内核漏洞利用方式，是当日影响面最大的安全类发布。
🧭 主题分组
- 模型发布（4 条）—— Inkling 2 Small 与 Nanbeige4.3-33B 相继发布，开源权重模型在智能指数与本地部署上同步推进。
- AI评测（2 条）—— 评测从基准分数转向任务实测：模型对玩《血染钟楼》检验社交推理，Rickroll 指令检验回应执行。
- 开源工具（1 条）—— 开源工具向工作流嵌入：ML Intern 可装进 Slack 用自有数据评估与微调模型。
- AI股市（1 条）—— 市场对 AI 公司的估值出现回调讨论，同时有观点认为硬件价格将随之下行。
- AI竞争（2 条）—— 围绕美国是否应限制 AI 形成争论：一方认为禁令自伤且会拱手让出市场，另一方寄望监管约束最严重危害。
- AI经济（1 条）—— 生产率年化 3.6% 被视为转折证据，AI 对宏观增长的贡献开始进入可测量的讨论。
- 行业渗透（2 条）—— Claude Code 被用于搭建日常工具，垂直专用 AI 与通用大模型之争同步升温。
- 公司上市（1 条）—— 有声音呼吁 Anthropic 尽快以高估值 IPO，作为重返 AI 竞争的姿态。
- 人才招聘（1 条）—— 工程人才争夺白热化，首周带薪休假与高额推荐奖金成为招聘筹码。
- 强化学习（1 条）—— 「分数居中」被指为业界常用做法并写入教材，用于理解训练与推理的不匹配。
- 产品批评（2 条）—— 对 AI 回复的复述腔调与 Safari 的兼容性表现提出具体批评。
- 游戏日常（1 条）—— 一位创作者公开自己在 PvP 服务器的阵营、等级与 Battle.net 标签。
- 目标心理（1 条）—— 借猴子多巴胺实验解释目标达成后的空虚，奖励预期在达成前已被提前消耗。
- 华尔街人物（1 条）—— 大卫·泰珀被拒升合伙人后创业，在金融危机中抄底获利约 40 亿美元。
- 其他（1 条）—— 未归类的补充内容
📖 详细内容
模型发布
Inkling 2 Small 与 Nanbeige4.3-33B 相继发布，开源权重模型在智能指数与本地部署上同步推进。
Artificial Analysis
｜ @ArtificialAnlys
｜ 影响力: 17.6万粉丝 · 互动分 395.78
Artificial Analysis
｜ @ArtificialAnlys
｜ 影响力: 17.6万粉丝 · 互动分 395.78
💡 核心观点: 作者报告Thinking Machines发布Inkling 2 Small，在Artificial Analysis智能指数得31分；若按计划开放权重，它将成为中国以外得分最高的开源权重模型，多项基准较前代大幅提升。
可信度: 85/100
事实核查:
- 原文支持: Thinking Machines has released Inkling 2 Small, scoring 31 on the Artificial Analysis Intelligence Index（原文明确给出模型名与指数分数，为机构自报的评测结果。）
- 原文支持: AutomationBench-AA rises from 5% to 46% and FrontierCode 2.0 from 5% to 25%（原文列出基准名称与前后分数，可对照其评测页面核验。）
- 原文支持: the hallucination rate falls from Inkling Small's 61% to 27%（原文给出幻觉率前后数值，但末尾准确率数据被截断。）
原文内容:
Thinking Machines has released Inkling 2 Small, scoring 31 on the Artificial Analysis Intelligence Index - their plan to release its weights would make it the highest-scoring open weights model developed outside China based on current rankings Inkling 2 Small is launching as a proprietary model, available first-party through @thinkymachines' Tinker platform and Tinker Playground, with weights to be released at a later date. Its score is up 18 points from Inkling Small, with ~4x as many output tokens per task. It matches @GoogleDeepMind's Gemini 3.8 Flash (high) and surpasses @XiaomiMiMo's MiMo-V2.6-Pro on the Artificial Analysis Intelligence Index v5.0.1. Key takeaways: ➤ Large gains over Inkling Small, especially on agentic tasks: AutomationBench-AA rises from 5% to 46% and FrontierCode 2.0 from 5% to 25%. HLE-Diamond also rises from 12% to 29%. ➤ Different coding strengths from Claude Sonnet 5.5 (medium): Inkling 2 Small scores 25% on FrontierCode 2.0 vs. 11% for Claude Sonnet 5.5 (medium), but 13% on Terminal-Bench 4.0 vs. 30%. ➤ Creative agentic scores above Claude Haiku 5.5 and Grok 4.7 on AA-3D: Inkling 2 Small also scores above Gemini 3.8 Flash on both AA-3D and Code2Video-AA, but behind Claude Sonnet 5.5 on both. ➤ Research-level physics improves, but trails Gemini 3.8 Flash (high): Inkling 2 Small scores 56% on CritPt v2, up from 34% for Inkling Small, vs. 65% for Gemini 3.8 Flash (high). ➤ Fewer hallucinations from more selective answering: On AA-Omniscience v1.1, the hallucination rate falls from Inkling Small's 61% to 27%, while accuracy rises from 32% to 34% and the attempt rate falls from 77% to 53%. ➤ ~4x as many output tokens: Inkling 2 Small averages ~150K output tokens per Intelligence Index task vs. ~37K for Inkling Small, ~31K for Claude Sonnet 5.5 (medium), and ~99K for Gemini 3.8 Flash (high). ➤ Almost 6x the cost of Inkling Small, but cheaper than several other models with similar Intelligence Index scores: Inkling 2 Small costs $1.18 per Intelligence Index task vs. $0.20 for Inkling Small, $7.77 for Gemini 3.8 Flash (high), and $10.02 for Claude Sonnet 5.5 (medium). Key model details: ➤ Type: Proprietary reasoning model at launch; Apache 2.0 weights release planned (date unconfirmed) ➤ Modalities: Text and image input, text output ➤ Size: 594B total parameters / 17B active parameters ➤ Pricing: $0.30/$1.00 per 1M input/output tokens ($0.03 per 1M cached input tokens) ➤ Availability at launch: Tinker and Tinker Playground Congratulations to the team at @thinkymachines on the release!
⏰ 23:59 | ❤️ 250 点赞 | 📝 341词 | 入选: 高赞 250 / 长文 | 查看原文 →
Artificial Analysis
｜ @ArtificialAnlys
｜ 影响力: 17.6万粉丝 · 互动分 314.61
Artificial Analysis
｜ @ArtificialAnlys
｜ 影响力: 17.6万粉丝 · 互动分 314.61
💡 核心观点: Artificial Analysis 称 Nanbeige 发布开源权重模型 Nanbeige4.3-33B，在 Intelligence Index v5 上与 Qwen3.8 27B（xhigh）同为 20 分，终端使用更强、知识较弱；模型属 BOSS 直聘旗下团队，Apache 2.0、524K 上下文，可在笔记本本地运行。
可信度: 85/100
事实核查:
- 原文支持: Nanbeige4.3-33B and Qwen3.8 27B (xhigh) both score 20, compared with 6 for Muse Glimmer (high, 30B).（原文给出具体指数与对照分数，来源为评测机构自身榜单，可对照其公开结果核验。）
- 原文支持: Nanbeige scores 30% on Terminal-Bench 4.0 and 16% on CritPt v2's Hard Subset, vs. 6% and 4% respectively for Qwen3.8 27B.（原文列明两项基准的具体百分比及对照值，细节充分，可由评测页面复核。）
- 原文支持: a hallucination rate of 18% vs. 29%（原文给出幻觉率对比，并说明弃答比例，属该机构 Omniscience 评测口径下的数据。）
原文内容:
Nanbeige has released Nanbeige4.3-33B, an open weights model matching Qwen3.8 27B (xhigh) at 20 on the Artificial Analysis Intelligence Index v5 with stronger terminal use but weaker knowledge @Nanbeige is the foundation model research team at BOSS Zhipin, the Chinese recruitment platform operated by Kanzhun Limited. Its dense 33B reasoning model is open weights under Apache 2.0, with text input and output only, a 524K-token context window, and is small enough to run locally on a laptop. Key results: ➤ Competitive results at ~30B total parameters. Nanbeige4.3-33B and Qwen3.8 27B (xhigh) both score 20, compared with 6 for Muse Glimmer (high, 30B). ➤ Terminal use and physics reasoning are relative strengths. Nanbeige scores 30% on Terminal-Bench 4.0 and 16% on CritPt v2's Hard Subset, vs. 6% and 4% respectively for Qwen3.8 27B. ➤ Different strengths to Qwen3.8 27B despite the same overall score. Nanbeige beats Qwen3.8 27B on AutomationBench-AA (53% vs. 48%), but trails on AA-Briefcase v1.1 (1162 vs. 1391 Elo), FrontierCode 2.0 (11% vs. 15%) and HLE-Diamond (7% vs. 15%). ➤ High abstention and low hallucination rates on Omniscience. Nanbeige scores -4 on AA-Omniscience v1.1 vs. -6 for Qwen3.8 27B, with a hallucination rate of 18% vs. 29%. It answers fewer questions correctly (12% vs. 17%) and declines 71% of questions. ➤ Similar intelligence with more output tokens per task. Nanbeige used 158K output tokens per Intelligence Index task vs. 110K for Qwen3.8 27B, about 44% more. Additional model details: ➤ Type: Open weights reasoning (Apache 2.0 license) ➤ Size: 33B total parameters (dense) ➤ Modalities: Text input and output only ➤ Context window: 524K ➤ API availability: No confirmed public first- or third-party APIs
⏰ 17:49 | ❤️ 198 点赞 | 📝 219词 | 入选: 高赞 198 / 长文 | 查看原文 →
Chubby♨️
｜ @kimmonismus
｜ 影响力: 15.7万粉丝 · 互动分 740.22
Chubby♨️
｜ @kimmonismus
｜ 影响力: 15.7万粉丝 · 互动分 740.22
💡 核心观点: 作者称 OpenAI 发布网络安全模型 GPT-6 Cyber，引用官方说法称其发现 81 个此前未知的 Firefox 高危漏洞，并协助开发 948 个 Linux 内核漏洞的利用方式。
可信度: 58/100
事实核查:
- 无法核实: it identified 81 previously unknown high-severity Firefox vulnerabilities（数字与漏洞等级来自作者转述的 OpenAI 说法，原文未附报告链接。）
- 无法核实: helped develop exploits for 948 Linux kernel vulnerabilities（原文给出具体数量，未说明统计口径与验证方式。）
- 原文支持: GPT-6 Cyber is available to select Daybreak Red customers（引用原文写明该模型的开放对象为 Daybreak Red 客户。）
原文内容:
OpenAI has released GPT-6 Cyber, a cybersecurity model designed to find and validate vulnerabilities while staying within the boundaries of authorized testing. According to OpenAI, it identified 81 previously unknown high-severity Firefox vulnerabilities and helped develop exploits for 948 Linux kernel vulnerabilities. GPT-6.2 yet to be released.
引用 @OpenAI: Introducing GPT‑6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignment.
The model powered the Linux kernel research we announced yesterday, and security teams at companies like Mozilla, Palo Alto Networks, Calif, and Cisco—alongside our own at Ope…
⏰ 17:21 | ❤️ 559 点赞 | 📝 46词 | 入选: 高赞 559 / 引用讨论 | 查看原文 →
Chubby♨️
｜ @kimmonismus
｜ 影响力: 15.7万粉丝 · 互动分 752.25
Chubby♨️
｜ @kimmonismus
｜ 影响力: 15.7万粉丝 · 互动分 752.25
💡 核心观点: 作者称自己此前获得的信息准确，周四并未发布 GPT-6.2，因此推测次日可能就是发布日，并表达期待。
可信度: 12/100
事实核查:
- 无法核实: No GPT-6.2 on Thursday.（原文仅为主观断言，未给出 OpenAI 官方发布安排或可核验来源。）
原文内容:
The information was correct as always. No GPT-6.2 on Thursday. So tomorrow is probably the big day. Fingers crossed, see you tomorrow! https://t.co/QqjQOpuiZw
⏰ 22:52 | ❤️ 589 点赞 | 📝 26词 | 入选: 高赞 589 | 查看原文 →
AI评测
评测从基准分数转向任务实测：模型对玩《血染钟楼》检验社交推理，Rickroll 指令检验回应执行。
Dwarkesh Patel
｜ @dwarkesh_sp
｜ 影响力: 28.4万粉丝 · 互动分 426.62
Dwarkesh Patel
｜ @dwarkesh_sp
｜ 影响力: 28.4万粉丝 · 互动分 426.62
💡 核心观点: 引用推文称作者花2万美元API额度让不同AI模型对玩社交推理游戏《血染钟楼》，模型在临场欺骗与团队推理上进步明显；发帖者称OpenAI模型大胜Anthropic，Astra对Opus 5.5胜率82%。
可信度: 55/100
事实核查:
- 无法核实: I spent 20,000 dollars in API credits having different AI models play each other at Blood on the Clocktower（引用推文给出花费与实验设置，未附账单或对局原始记录。）
- 无法核实: Astra has a 82% win rate against Opus 5.5（原文给出胜率数字，未说明对局局数与统计方法。）
原文内容:
OpenAI models demolishing Anthropic models at Blood on the Clocktower (Astra has a 82% win rate against Opus 5.5). https://t.co/IuqicNusMO
引用 @dwarkesh_sp: I spent 20,000 dollars in API credits having different AI models play each other at Blood on the Clocktower (a Mafia or Werewolf type social deception game).
The results were fascinating. Models have gotten much better over the course of this year at creative on-the-spot scheming and bluffing, and …
⏰ 22:50 | ❤️ 299 点赞 | 📝 21词 | 入选: 高赞 299 / 引用讨论 | 查看原文 →
Ethan Mollick
｜ @emollick
｜ 影响力: 41.4万粉丝 · 互动分 361.78
Ethan Mollick
｜ @emollick
｜ 影响力: 41.4万粉丝 · 互动分 361.78
💡 核心观点: 作者用一段“Rickroll”测试指令展示 Claude 的实际回应执行能力，并称其“永远不变”，同时提到 GPT-6 Pro 版本以动画歌手做注意力测试。依据是一次自设提示与模型表现的对比（属个人观察）。
可信度: 45/100
事实核查:
- 原文支持: Never change, Claude.（原文确有此句，属作者对 Claude 表现的主观评价，非可核验事实。）
- 无法核实: The GPT-6 Pro version included an attention test interrupted by an animated singer（原文确有此描述，但未给出可复核的测试记录或来源，无法独立核实。）
原文内容:
"Can you actually Rickroll me? Don't explain what a Rickroll is and don't just describe how someone could do one. Try to pull it off in your response here." Never change, Claude. (The GPT-6 Pro version included an attention test interrupted by an animated singer - cute) https://t.co/A7Mabcn6oq
⏰ 18:39 | ❤️ 185 点赞 | 📝 51词 | 入选: 高赞 185 | 查看原文 →
开源工具
开源工具向工作流嵌入：ML Intern 可装进 Slack 用自有数据评估与微调模型。
Hugging Face
｜ @huggingface
｜ 影响力: 83.7万粉丝 · 互动分 327.38
Hugging Face
｜ @huggingface
｜ 影响力: 83.7万粉丝 · 互动分 327.38
💡 核心观点: Hugging Face 宣布其开源 ML Intern 已可作为 Slack 机器人安装到用户自己的工作区，支持即时评估模型、构建数据集和微调模型。引用推文强调用自有数据打造私有模型。
可信度: 70/100
事实核查:
- 原文支持: Our open-source ML Intern is now available as a Slack bot you can install in your own workspace!（官方账号原文直接陈述产品发布状态，属主体自述，缺乏第三方验证。）
- 无法核实: Evaluate models, build datasets, and fine-tune models on the fly（原文列出功能，但未提供文档或演示链接的具体能力说明，效果无法核实。）
原文内容:
Our open-source ML Intern is now available as a Slack bot you can install in your own workspace! Evaluate models, build datasets, and fine-tune models on the fly 🚀
引用 @abidlabs: YOUR data deserves YOUR own, private model.
Now, let @HuggingFace's ML Intern build it for you.
Connect your data, files, mcps. And call it from Slack. https://t.co/ouWjBlmXIh
⏰ 19:50 | ❤️ 160 点赞 | 📝 30词 | 入选: 高赞 160 / 引用讨论 | 查看原文 →
AI股市
市场对 AI 公司的估值出现回调讨论，同时有观点认为硬件价格将随之下行。
Gary Marcus
｜ @GaryMarcus
｜ 影响力: 25.6万粉丝 · 互动分 456.42
Gary Marcus
｜ @GaryMarcus
｜ 影响力: 25.6万粉丝 · 互动分 456.42
💡 核心观点: 引用推文称AI公司股价大跌、行业「清算日」或已到来；作者补充乐观面，认为DRAM、GPU、笔记本电脑和游戏主机价格可能随之下降。
可信度: 30/100
事实核查:
- 无法核实: SpaceX 10% below IPO price, CoreWeave about to drop below its IPO price. Nvidia is down almost 10%（数据出自引用推文，原文未注明行情来源、时间与统计口径。）
- 无法核实: DRAM prices may come down. and GPUs, and laptops and video game consoles, too（原文以「may」表述价格走向，属推测，无价格数据或供应链依据。）
原文内容:
on the plus side DRAM prices may come down. and GPUs, and laptops and video game consoles, too.
引用 @GaryMarcus: Absolute slaughter.
Double digit losses for many AI companies. SpaceX 10% below IPO price, CoreWeave about to drop below its IPO price. Nvidia is down almost 10%
The AI industry’s day of reckoning may have finally come. https://t.co/6S7mWPkzCY
⏰ 19:45 | ❤️ 303 点赞 | 📝 18词 | 入选: 高赞 303 / 引用讨论 | 查看原文 →
AI竞争
围绕美国是否应限制 AI 形成争论：一方认为禁令自伤且会拱手让出市场，另一方寄望监管约束最严重危害。
Bindu Reddy
｜ @bindureddy
｜ 影响力: 39.5万粉丝 · 互动分 444.55
Bindu Reddy
｜ @bindureddy
｜ 影响力: 39.5万粉丝 · 互动分 444.55
💡 核心观点: 作者主张美国试图禁止AI是自伤：美国仅占全球市场30%，美国AI若停滞中国会追上，本国民众将不熟悉AI，剩余高薪工作还会被全球化吸走。
可信度: 35/100
事实核查:
- 无法核实: The US is just 30% of the global market（原文给出30%这一比例，但未注明统计范围与来源。）
- 无法核实: China will catch up if US AI stalls（原文为条件式预测，未提供数据或案例。）
原文内容:
🚨 AI is not stopping; the US is simply shooting itself in the foot by trying to ban it - The US is just 30% of the global market - China will catch up if US AI stalls - Our population will be AI-stupid - The rest of the world will be AI-savvy - Globalization will suck all the remaining high-paying jobs from here Giving up technical dominance is not a choice; it's suicide 😱
⏰ 00:36 | ❤️ 267 点赞 | 📝 71词 | 入选: 高赞 267 | 查看原文 →
Gary Marcus
｜ @GaryMarcus
｜ 影响力: 25.6万粉丝 · 互动分 530.7
Gary Marcus
｜ @GaryMarcus
｜ 影响力: 25.6万粉丝 · 互动分 530.7
💡 核心观点: 作者称中期选举结果令其「末日」「灾难」「反乌托邦」三类概率本周全部下降，理由是民主党监管虽可能过度，但有望限制最严重危害并揭露不法行为。
可信度: 35/100
事实核查:
- 原文支持: my p(doom), p(catastrophe), and p(dystopia) have all dropped this week, given the results of the midterms（原文明确表述三类主观概率因中期选举结果下降，但未给出具体数值或测算依据。）
- 无法核实: dems may possibly go too far in regulation but they are likely to limit the chances of many of the worst harms（原文用「may possibly」「likely」表述推测，未提供监管措施或案例支撑。）
原文内容:
happy news: my p(doom), p(catastrophe), and p(dystopia) have all dropped this week, given the results of the midterms. dems may possibly go too far in regulation but they are likely to limit the chances of many of the worst harms. and to expose a lot of malfeasance. that in turn will increase the public desire to regulate the worst actors, — and make us all safer.
⏰ 13:47 | ❤️ 373 点赞 | 📝 68词 | 入选: 高赞 373 | 查看原文 →
AI经济
生产率年化 3.6% 被视为转折证据，AI 对宏观增长的贡献开始进入可测量的讨论。
Ethan Mollick
｜ @emollick
｜ 影响力: 41.4万粉丝 · 互动分 617.27
Ethan Mollick
｜ @emollick
｜ 影响力: 41.4万粉丝 · 互动分 617.27
💡 核心观点: 作者引用 Erik 的帖文，称美国本季度生产率年化增长 3.6%，超过 2007 至 2019 年 1.5% 的平均水平，并称证据显示生产率正在迎来转折。
可信度: 61/100
事实核查:
- 原文支持: US productivity grew at a 3.6% annual rate this past quarter.（原文给出季度年化增速数字，并指向图表来源。）
- 原文支持: That's more than double the 1.5% average from 2007-2019.（原文给出对比区间与均值，3.6% 约为 1.5% 的两倍以上，与原文表述一致。）
原文内容:
Erik is one of the economists who has been watching the impact of AI on the economy most closely, even before generative AI...
引用 @erikbryn: Evidence is growing that we’re turning the corner on productivity.
US productivity grew at a 3.6% annual rate this past quarter.
That’s more than double the 1.5% average from 2007-2019.
The new numbers put me further ahead for my bet with Bob Gordon about productivity growth in the 2020s. 💪
Here…
⏰ 22:01 | ❤️ 425 点赞 | 📝 23词 | 入选: 高赞 425 / 引用讨论 | 查看原文 →
行业渗透
Claude Code 被用于搭建日常工具，垂直专用 AI 与通用大模型之争同步升温。
@levelsio
｜ @levelsio
｜ 影响力: 99.0万粉丝 · 互动分 730.0
@levelsio
｜ @levelsio
｜ 影响力: 99.0万粉丝 · 互动分 730.0
💡 核心观点: 作者称自己用 Claude Code 搭建开支追踪工具并借此发现可节省的开支，据此认为几乎每个行业都会被 Anthropic、OpenAI 乃至 xAI 进入。
可信度: 22/100
事实核查:
- 无法核实: I built my own expenses tracker with Claude Code（原文陈述个人使用经历，无可核验的产物或数据。）
- 无法核实: almost every industry will be entered by Anthropic, OpenAI and probably xAI too（原文为对行业走向的预测，未提供依据。）
原文内容:
P.S. This is good, I built my own expenses tracker with Claude Code, it's how I found so much stuff to save on It just proves my point that almost every industry will be entered by Anthropic, OpenAI and probably xAI too
引用 @levelsio: I told you, they will try eat all industries one by one
⏰ 20:22 | ❤️ 539 点赞 | 📝 44词 | 入选: 高赞 539 / 引用讨论 | 查看原文 →
Anthony Pompliano 🌪
｜ @APompliano
｜ 影响力: 243.9万粉丝 · 互动分 317.22
Anthony Pompliano 🌪
｜ @APompliano
｜ 影响力: 243.9万粉丝 · 互动分 317.22
💡 核心观点: 作者看好垂直专用 AI 胜过大型通用模型，称 Silvia 对个人可识别信息加密并从投资组合中匿名化，已获 SOC 2 认证并自建智能能力，面向传统金融体系中的“非主流”用户推广家庭理财。
可信度: 40/100
事实核查:
- 无法核实: Silvia encrypts your personally-identifiable information and anonymizes it away from your portfolio.（原文有该表述，但为企业方宣传口径，无技术文档或审计结果佐证。）
- 无法核实: We are also SOC 2 certified and own our own intelligence.（原文自称已获 SOC 2 认证，未附证书编号或审计机构，无法独立核实。）
- 无法核实: vertical-specialized AI will outperform the large, general models（原文为作者的投资判断与预测，未给出数据或对比结果。）
原文内容:
It may seem impossible for a small group of misfits to beat the large bullies, but never count @cfosilvia out. We are betting that vertical-specialized AI will outperform the large, general models. We believe personal finance products should be private and secure. No one wants Dario knowing what they own. Silvia encrypts your personally-identifiable information and anonymizes it away from your portfolio. We are also SOC 2 certified and own our own intelligence. If you are a misfit in the legacy financial system, sign up for Silvia and start using AI to better manage your household finances: https://t.co/4LyYZFRC4m $SVIA
⏰ 23:07 | ❤️ 191 点赞 | 📝 102词 | 入选: 高赞 191 | 查看原文 →
公司上市
有声音呼吁 Anthropic 尽快以高估值 IPO，作为重返 AI 竞争的姿态。
Bindu Reddy
｜ @bindureddy
｜ 影响力: 39.5万粉丝 · 互动分 307.52
Bindu Reddy
｜ @bindureddy
｜ 影响力: 39.5万粉丝 · 互动分 307.52
💡 核心观点: 作者呼吁 Anthropic 尽快上市，主张发布 Fable 5.5、以 1.25 万亿美元定价 IPO 并让市值涨至 1.75 万亿美元，以“重磅回归 AI”。内容为口号式建议，未给依据。
可信度: 15/100
事实核查:
- 无法核实: price the IPO at $1.25T（原文提出该估值目标，属作者主张，无任何市场数据或公司信息支撑。）
- 无法核实: drop Fable 5.5（原文提及该产品名，但未说明其内容或发布状态，无法核实。）
原文内容:
Anthropic needs to go public ASAP - drop Fable 5.5 - price the IPO at $1.25T - get the shares to pop to $1.75T BRING BACK AI WITH A BANG ‼️
⏰ 17:55 | ❤️ 206 点赞 | 📝 24词 | 入选: 高赞 206 | 查看原文 →
人才招聘
工程人才争夺白热化，首周带薪休假与高额推荐奖金成为招聘筹码。
Santiago
｜ @svpino
｜ 影响力: 45.7万粉丝 · 互动分 358.35
Santiago
｜ @svpino
｜ 影响力: 45.7万粉丝 · 互动分 358.35
💡 核心观点: 引用 Serval 招聘信息称，新入职工程师可享首周最高 1 万美元带薪休假并配专属旅行礼宾；成功推荐入职可得 3 万美元奖金，入职者同样获得。作者据此判断企业正激烈争夺优秀工程师。
可信度: 58/100
事实核查:
- 原文支持: every engineer who joins Serval begins with a weeklong, $10,000 paid vacation（引文原文明确写出该条件，但属企业自述的招聘承诺，未经第三方核实。）
- 无法核实: If you refer a candidate who joins, they give you a $30,000 bonus, and the person who joined also receives one.（原文写明推荐人与入职者各得 3 万美元，但引文截断处未展示完整条款，细则无法核实。）
原文内容:
Companies are killing for good engineers right now! These guys are offering a paid first-week vacation worth up to $10,000, planned with a dedicated travel concierge for new engineers who join their team. If you refer a candidate who joins, they give you a $30,000 bonus, and the person who joined also receives one.
引用 @AlexAtServal: Starting today, every engineer who joins Serval begins with a weeklong, $10,000 paid vacation. We’re building the next generational company. We need generational talent to see it through.
Our mission has always been to empower people across every team to turn their understanding of a problem into w…
⏰ 17:55 | ❤️ 180 点赞 | 📝 53词 | 入选: 高赞 180 / 引用讨论 | 查看原文 →
强化学习
「分数居中」被指为业界常用做法并写入教材，用于理解训练与推理的不匹配。
Nathan Lambert
｜ @natolambert
｜ 影响力: 10.9万粉丝 · 互动分 456.68
Nathan Lambert
｜ @natolambert
｜ 影响力: 10.9万粉丝 · 互动分 456.68
💡 核心观点: 作者称听闻业界普遍采用「分数居中」（如Tinker的StableReinforce使用该法），已将其写入rlhf教材，并称该推导有助于理解训练与推理不匹配问题。
可信度: 50/100
事实核查:
- 无法核实: score-centering is popular in industry (e.g. Tinker's StableReinforce using it)（原文为「听很多人说」的转述，未附使用案例或技术细节。）
- 原文支持: it's the latest addition to the rlhf book（原文明确说明该内容是rlhf教材的最新增补。）
原文内容:
Have heard from many folks that score-centering is popular in industry (e.g. Tinker's StableReinforce using it), so it's the latest addition to the rlhf book. It's a nice derivation to understand training-inference-mismatch. Video version explaining it soon.
引用 @mrtnm: Why is RL so sensitive to training-inference mismatch (TIM)?
With each training step, the trainer is pulled toward the sampler, resulting in accumulating drift.
We exploit this intuition to devise a novel correction method to stabilize RL under TIM, called Score Centering.
⏰ 15:07 | ❤️ 359 点赞 | 📝 44词 | 入选: 高赞 359 / 引用讨论 | 查看原文 →
产品批评
对 AI 回复的复述腔调与 Safari 的兼容性表现提出具体批评。
@levelsio
｜ @levelsio
｜ 影响力: 99.0万粉丝 · 互动分 394.67
@levelsio
｜ @levelsio
｜ 影响力: 99.0万粉丝 · 互动分 394.67
💡 核心观点: 引用推文批评AI回复机器人只是把原帖完整复述一遍并带有明显AI腔调，发帖者称这正是AI回复令人反感的原因。
可信度: 25/100
事实核查:
- 原文支持: "here's a complete regurgitation of your post but with strong AI tells."（该句为引用推文中对AI回复套路的举例描述，属观察性说法。）
- 无法核实: Exactly why AI replies suck（原文为作者对AI回复质量的主观评价，无数据支撑。）
原文内容:
"here's a complete regurgitation of your post but with strong AI tells." Exactly why AI replies suck
引用 @ryanlanciaux: It's so obnoxious. I have a much much smaller reach and any of my posts that get outside normal follower band gets some AI reply bots.
I don't understand what the people running the account think it does for them. It's so so easy to spot too.
"here's a complete regurgitation of your post but wit…
⏰ 13:44 | ❤️ 193 点赞 | 📝 18词 | 入选: 高赞 193 / 引用讨论 | 查看原文 →
Peter Steinberger 🦞
｜ @steipete
｜ 影响力: 61.4万粉丝 · 互动分 578.0
Peter Steinberger 🦞
｜ @steipete
｜ 影响力: 61.4万粉丝 · 互动分 578.0
💡 核心观点: 作者质疑苹果关于改用 Safari 的说法，称自己开发和使用 Safari 时遇到速度更慢、功能更少、异常缺陷和网站不兼容等问题，难以接近 Chrome 的表现。
可信度: 30/100
事实核查:
- 无法核实: it's slower, has fewer features, weird bugs and plenty websites just don't work（原文为个人使用与开发体验的描述，未提供测试数据。）
原文内容:
Is my perception off or is Apple just lying here? Anytime I build for or use Safari, it’s slower, has fewer features, weird bugs and plenty websites just don’t work. It’s a fight to get even somewhat close to Chrome’s performance.
引用 @mjtsai: Switch to Safari:
https://t.co/f42PZNH1y4 #mjtsaiblog
⏰ 23:54 | ❤️ 348 点赞 | 📝 45词 | 入选: 高赞 348 / 引用讨论 | 查看原文 →
游戏日常
一位创作者公开自己在 PvP 服务器的阵营、等级与 Battle.net 标签。
Marc Lou
｜ @marclou
｜ 影响力: 42.3万粉丝 · 互动分 453.75
Marc Lou
｜ @marclou
｜ 影响力: 42.3万粉丝 · 互动分 453.75
💡 核心观点: 作者称自己等了15年迎来这一天，并邀请读者一起升级：他在PvP服务器玩部落阵营，当前12级，公开了Battle.net标签。
可信度: 20/100
事实核查:
- 原文支持: I've been waiting 15 years for this day（原文明确写出等待15年，但未说明所指的具体事件。）
- 无法核实: I play Horde on the PvP server. Currently level 12（原文为个人游戏状态自述，无可核验来源。）
原文内容:
I've been waiting 15 years for this day... 🥹 If you want to level up together: I play Horde on the PvP server. Currently level 12. My @battlenet tag is marclou#2719. https://t.co/bXQDM5bJOh
⏰ 15:06 | ❤️ 280 点赞 | 📝 32词 | 入选: 高赞 280 | 查看原文 →
目标心理
借猴子多巴胺实验解释目标达成后的空虚，奖励预期在达成前已被提前消耗。
Aakash Gupta
｜ @aakashgupta
｜ 影响力: 33.7万粉丝 · 互动分 3942.5
Aakash Gupta
｜ @aakashgupta
｜ 影响力: 33.7万粉丝 · 互动分 3942.5
💡 核心观点: 作者引用神经科学家 Schultz 的猴子多巴胺实验，解释登月后目标达成带来的空虚与焦虑：大脑在目标看似可达时已提前消耗奖励预期，真正达成后只剩失去它的恐惧。
可信度: 52/100
事实核查:
- 原文支持: Buzz Aldrin was 39 when he stepped onto the moon in 1969（原文给出奥尔德林登月时的年龄与年份，属可对照的传记性细节。）
- 无法核实: a neuroscientist named Wolfram Schultz recorded single dopamine neurons in monkeys while giving them juice（原文将 Schultz 的猴子果汁实验作为多巴胺预测误差的解释依据，除年份外未提供论文或机构出处。）
- 原文支持: He was the first actor Hollywood ever paid $20 million for one movie（原文称金·凯瑞是首位单片片酬 2000 万美元的演员，片酬数字与身份可在原文中找到。）
原文内容:
The second man to ever walk on the moon came home and ended up selling Cadillacs at a dealership in Beverly Hills. Buzz Aldrin was 39 when he stepped onto the moon in 1969. About two years later he was sitting in an Air Force doctor's office in San Antonio, telling someone for the first time that he was depressed. By the end of the '70s he'd been through two divorces and a drinking problem that nearly killed him. If you've ever hit a big goal and felt weirdly flat a week later, you've had a tiny version of what happened to him. And there's a real reason for it. Back in the '90s a neuroscientist named Wolfram Schultz recorded single dopamine neurons in monkeys while giving them juice. When the juice came out of nowhere, the neurons lit up. Then he started flashing a light right before the juice, and pretty quickly the monkeys figured out what the light meant. The neurons started firing at the light instead. The juice itself, the thing they actually wanted, didn't get a spike anymore. And if the light flashed and no juice showed up, the neurons dropped below normal. So your brain cashes in on the dream the moment it starts looking real. By the time you actually get there, it already spent the money, and the only thing left that can still surprise you is losing it. That's the anxiety this guy is talking about, the one about making it and becoming somebody, and every one of us is walking around with that wiring. Jim Carrey learned this the hard way. He was the first actor Hollywood ever paid $20 million for one movie, and in 2005 he told a reporter he wished everybody could get rich and famous and get everything they ever dreamed of, just so they could see it's not the answer. Honestly, the oldest version of this video is in the Bible. Ecclesiastes is written as a king in Jerusalem looking back on his life. He built houses, planted vineyards, dug pools and stacked up silver and gold, and then he called all of it hevel. He uses that word 38 times. It means vapor, like your breath on a cold morning that's gone in a second. Most English Bibles just say "vanity." My favorite part comes in chapter 9. He tells you to eat your bread with joy, and a few verses later he says "whatsoever thy hand findeth to do, do it with thy might." The guy who called everything vapor still told you to go all in. Aldrin got sober in 1978 and spent the decades after that trying to get people to Mars, chasing the next big thing without needing it to make him somebody.
引用 @_falsi1ke: I have been watching this every morning for 11 months now. https://t.co/gIUfvuN6rB
⏰ 12:21 | ❤️ 2579 点赞 | 📝 462词 | 入选: 高赞 2579 / 引用讨论 / 长文 | 查看原文 →
华尔街人物
大卫·泰珀被拒升合伙人后创业，在金融危机中抄底获利约 40 亿美元。
Aakash Gupta
｜ @aakashgupta
｜ 影响力: 33.7万粉丝 · 互动分 2271.0
Aakash Gupta
｜ @aakashgupta
｜ 影响力: 33.7万粉丝 · 互动分 2271.0
💡 核心观点: 作者引用中文帖讲述大卫·泰珀被科赞两次拒绝升任合伙人后辞职创业，金融危机中抄底获利约 40 亿美元，并以 4350 万美元买下科赞前妻的汉普顿豪宅后推平重建。
可信度: 46/100
事实核查:
- 原文支持: He got passed over for partner twice. The guy he blamed was the head of his division, Jon Corzine.（原文明确写出两次落选合伙人，以及他归咎的部门负责人为科赞。）
- 无法核实: His fund was up more than 130% that year and he personally made about $4 billion.（原文给出基金年度涨幅与个人获利数字，未提供基金名称或财报来源。）
- 存在偏差: Two months later Tepper wired $43.5 million to Corzine's ex-wife for the Hamptons house.（引用帖称泰珀购得豪宅后未入住即拆除重建，英文正文只写到汇款买房，两处叙述存在差异。）
原文内容:
A Goldman trader got passed over for partner, quit, made $4 billion in one year, then paid $43.5 million for the beach house his old boss lost in the divorce and bulldozed it. That's the version everybody shares. What it leaves out is what the old boss was doing those same years, and honestly it's way worse. The trader is David Tepper. He joined Goldman in 1985 and was running its brand new junk bond desk within six months. His desk made the firm a ton of money, but he was loud and he cursed a lot, and Goldman back then was a place where you got ahead by being polished. He got passed over for partner twice. The guy he blamed was the head of his division, Jon Corzine. Tepper quit in December 1992 and started his own fund with $57 million. Corzine went on to run Goldman, and even after Hank Paulson pushed him out in 1999, he left with a stake worth about $400 million when the firm went public. Fast forward to early 2009. Citigroup was trading under a dollar, Bank of America was around $3, and half of Wall Street thought the government was about to take over the banks. Tepper figured Washington would never let them die, so he bought as much as he could get his hands on. His fund was up more than 130% that year and he personally made about $4 billion. Now watch what Corzine does next. In March 2010 he took over a brokerage called MF Global. Two months later Tepper wired $43.5 million to Corzine's ex-wife for the Hamptons house. And that summer Corzine started loading up on bonds from Italy, Spain, Portugal, Ireland and Belgium, eventually $6.3 billion of them, because he was sure Europe would never let those countries go broke. If that sounds familiar, it's basically Tepper's bank trade pointed at Europe. In 2011 the bulldozers came for the house, and on Halloween of that year MF Global filed for bankruptcy. Every single one of those European bonds ended up paying back in full. Corzine was right about the bet, and his company still never saw the payout because the money showed up months after it went under. What killed him was how much he borrowed. MF Global had about $40 of debt for every $1 of its own money, and at that ratio a 2.5% drop in what you own wipes you out. So when those bonds dipped, the lenders wanted more cash right away, the ratings agencies downgraded the firm, and the whole thing was over in about a week. Tepper owned his bank stocks outright enough to just sit there and wait for everyone to calm down. Tepper is worth $28.5 billion today, about 71 times what Corzine walked away from Goldman with. Getting passed over for partner turned out to be the best thing that ever happened to him, and he has his old boss to thank for it.
引用 @chakmillion: 华尔街阿帕卢萨对冲基金创始人、NFL卡罗来纳黑豹队老板大卫·泰珀，目前身家超过 200 亿美元，是个出了名的有仇必报狠角色。
泰珀年轻时在高盛集团当交易员，替公司赚了大钱，结果当时的部门顶头上司兼高盛掌门人乔恩·科赞嫌他说话嗓门大、举止粗鲁不够体面，连续两次拒绝升他当高盛合伙人。泰珀一气之下辞职单干，靠精准抄底次贷危机成了华尔街最赚钱的基金教父。
而当年看不起他的那位前上司科赞，后来不仅做生意破产还闹了离婚。泰珀得知消息后，直接掏出 4350 万美元天价现金，买下了前上司家族在纽约汉普顿最心爱的海景大庄园；拿到钥匙后泰珀一天都没进去住，直接叫来重型推土机把前上司住了多年的豪宅当场推平砸烂，…
⏰ 10:03 | ❤️ 1871 点赞 | 📝 484词 | 入选: 高赞 1871 / 引用讨论 / 长文 | 查看原文 →
其他
未归类的补充内容
Nathan Lambert
｜ @natolambert
｜ 影响力: 10.9万粉丝 · 互动分 561.36
Nathan Lambert
｜ @natolambert
｜ 影响力: 10.9万粉丝 · 互动分 561.36
💡 核心观点: 作者引用 Inkling 2 Small 发布信息，称其为西方最强开源权重模型且有明显领先，体积小于 muse spark open 与 mistral large，并采用 Apache 许可证。
可信度: 44/100
事实核查:
- 无法核实: Scoring 31 on the AA Intelligence Index（分数来自引用原文，未附评测页面链接。）
- 原文支持: with a better license (apache)（原文称许可证为 Apache，属可在发布信息中核对的产品属性。）
- 无法核实: Looks like the smartest western open weight model by a margin.（原文为主观比较判断，未给出对比模型的具体分数。）
原文内容:
Looks like the smartest western open weight model by a margin. While much smaller than muse spark open or mistal large. While with a better license (apache). Love it!
引用 @thinkymachines: Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.
Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.
https://t.co/VR02tw73fJ
⏰ 00:16 | ❤️ 458 点赞 | 📝 29词 | 入选: 高赞 458 / 引用讨论 | 查看原文 →
🔗 引用关系
- @aakashgupta 引用 @_falsi1ke：作者引用神经科学家 Schultz 的猴子多巴胺实验，解释登月后目标达成带来的空虚与焦虑：大脑在目标看似可达时已提前消耗奖励预期，真正达成后只剩失去它的恐惧。（原文：I have been watching this every morning for 11 months now. https://t.co/gIUfvuN6rB）
- @aakashgupta 引用 @chakmillion：作者引用中文帖讲述大卫·泰珀被科赞两次拒绝升任合伙人后辞职创业，金融危机中抄底获利约 40 亿美元，并以 4350 万美元买下科赞前妻的汉普顿豪宅后推平重建。（原文：华尔街阿帕卢萨对冲基金创始人、NFL卡罗来纳黑豹队老板大卫·泰珀，目前身家超过 200 亿美元，是个出了名的有仇必报狠角色。
泰珀年轻时在高盛集团当交易员，替公司赚了大钱，结果当时的部门顶头上司兼高盛掌门人乔恩·科赞嫌他说话嗓门大、举止粗…）
- @kimmonismus 引用 @OpenAI：作者称 OpenAI 发布网络安全模型 GPT-6 Cyber，引用官方说法称其发现 81 个此前未知的 Firefox 高危漏洞，并协助开发 948 个 Linux 内核漏洞的利用方式。（原文：Introducing GPT‑6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignme…）
- @levelsio 引用 @levelsio：作者称自己用 Claude Code 搭建开支追踪工具并借此发现可节省的开支，据此认为几乎每个行业都会被 Anthropic、OpenAI 乃至 xAI 进入。（原文：I told you, they will try eat all industries one by one）
- @emollick 引用 @erikbryn：作者引用 Erik 的帖文，称美国本季度生产率年化增长 3.6%，超过 2007 至 2019 年 1.5% 的平均水平，并称证据显示生产率正在迎来转折。（原文：Evidence is growing that we’re turning the corner on productivity.
US productivity grew at a 3.6% annual rate this past…）
- @steipete 引用 @mjtsai：作者质疑苹果关于改用 Safari 的说法，称自己开发和使用 Safari 时遇到速度更慢、功能更少、异常缺陷和网站不兼容等问题，难以接近 Chrome 的表现。（原文：Switch to Safari:
https://t.co/f42PZNH1y4 #mjtsaiblog） - @natolambert 引用 @thinkymachines：作者引用 Inkling 2 Small 发布信息，称其为西方最强开源权重模型且有明显领先，体积小于 muse spark open 与 mistral large，并采用 Apache 许可证。（原文：Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.
Inkling 2 S…）
- @natolambert 引用 @mrtnm：作者称听闻业界普遍采用「分数居中」（如Tinker的StableReinforce使用该法），已将其写入rlhf教材，并称该推导有助于理解训练与推理不匹配问题。（原文：Why is RL so sensitive to training-inference mismatch (TIM)?
With each training step, the trainer is pulled toward the …）
- @GaryMarcus 引用 @GaryMarcus：引用推文称AI公司股价大跌、行业「清算日」或已到来；作者补充乐观面，认为DRAM、GPU、笔记本电脑和游戏主机价格可能随之下降。（原文：Absolute slaughter.
Double digit losses for many AI companies. SpaceX 10% below IPO price, CoreWeave about to drop b…）
- @dwarkesh_sp 引用 @dwarkesh_sp：引用推文称作者花2万美元API额度让不同AI模型对玩社交推理游戏《血染钟楼》，模型在临场欺骗与团队推理上进步明显；发帖者称OpenAI模型大胜Anthropic，Astra对Opus 5.5胜率82%。（原文：I spent 20,000 dollars in API credits having different AI models play each other at Blood on the Clocktower (a Mafia or …）
- @levelsio 引用 @ryanlanciaux：引用推文批评AI回复机器人只是把原帖完整复述一遍并带有明显AI腔调，发帖者称这正是AI回复令人反感的原因。（原文：It's so obnoxious. I have a much much smaller reach and any of my posts that get outside normal follower band gets some …）
- @svpino 引用 @AlexAtServal：引用 Serval 招聘信息称，新入职工程师可享首周最高 1 万美元带薪休假并配专属旅行礼宾；成功推荐入职可得 3 万美元奖金，入职者同样获得。作者据此判断企业正激烈争夺优秀工程师。（原文：Starting today, every engineer who joins Serval begins with a weeklong, $10,000 paid vacation. We’re building the next g…）
- @huggingface 引用 @abidlabs：Hugging Face 宣布其开源 ML Intern 已可作为 Slack 机器人安装到用户自己的工作区，支持即时评估模型、构建数据集和微调模型。引用推文强调用自有数据打造私有模型。（原文：YOUR data deserves YOUR own, private model.
Now, let @HuggingFace's ML Intern build it for you.
Connect your data, f…）
🧩 跨推文归纳
共识
- 开源与本地部署仍在快速推进：Inkling 2 Small 发布，Nanbeige4.3-33B 可在笔记本本地运行，Hugging Face 的 ML Intern 支持用自有数据微调模型。
- AI 能力正从通用对话转向具体任务：让模型对玩《血染钟楼》检测社交推理，用 GPT-6 Cyber 寻找漏洞，用 Claude Code 搭建工具。
分歧
- 关于开源模型的领先位置：Nathan Lambert 称 Inkling 2 Small 是西方最强开源权重模型；Artificial Analysis 则强调，只有在按计划开放权重后，它才会成为中国以外得分最高的开源模型，两者对「领先」成立条件的判断不同。
数据来源：X 列表「AI 英文」（22 条候选）· 由 twscrape 采集、grok 整理
© 版权声明
文章版权归作者所有，未经允许请勿转载。
相关文章
暂无评论...