【AI 英文奏折】2026年11月16日|算力重配:Qwen4 Max 低成本跑分

每日奏折23小时前发布 tianming
24 0 0

共收录 22 篇深度内容


📌 今日概览

【AI 英文奏折】2026年11月16日|算力重配:Qwen4 Max 低成本跑分

  • 模型发布与评测同步升温:Qwen4 Max 智能指数升至 43 分,成本被逐项折算;Gemini 4 Carbon 的 /swarm 调度十余个 Agent 协作建模。
  • Agent 能力评估成为焦点:Remote Labor Index 付费项目完成率升至 41.3%,多 Agent 在芯片波形数据问答中显出审计优势。
  • 端侧与降本并行:4GB 的 Nemotron 3.5 Bolt 可离线起草邮件并留存本地数据,Minimax H3 被视为对话影视的低估选项。
  • 安全与政策争议未歇:Gary Marcus 呼吁暂缓最激进的通用 Agent,算力租用与芯片流向仍在拉扯。

🔥 必读 Top3

  1. Artificial Analysis Artificial Analysis 称阿里发布 Qwen4 Max,智能指数 43 分,较 Qwen3.8 Max 高 9 分、落后 Kimi K3.1 Pro 3 分,单任务成本 3.27 美元。
    入选理由:Artificial Analysis 以可复现指标给出 Qwen4 Max 的分数与单任务成本,是今日最硬的模型评测锚点。
  2. Chubby♨️ 作者称 AI Agent 在 Remote Labor Index 付费项目上的完成率 13 个月内由 2.5% 升至 41.3%,据此判断白领工作将在 2027 年受冲击。
    入选理由:13 个月内付费项目完成率从 2.5% 到 41.3%,用一条趋势线支撑白领岗位受冲击的判断。
  3. Rohan Paul Rohan Paul 介绍 Google DeepMind 与哥伦比亚大学的论文:把芯片仿真产生的超大波形转储转为 SQLite 数据库,由多 Agent 将自然语言问题转成 SQL 查询并审核。在 150 道专家标注题上用 Gemini 2.5 Pro 答对 95.33%,多 Agent 分工在质量审计上比单 Agent 高 17.3%,延迟为 1.96 倍。
    入选理由:论文级证据展示多 Agent 把自然语言问题转为 SQL,量化质量审计增益与延迟代价,可信度最高。

🧭 主题分组

  • 模型评测(2 条)—— Qwen4 Max 分数与成本被量化定位,Gemini 4 Argon 的豪言则以转述形式流传。
  • Agent 实践(5 条)—— 多 Agent 分工在建模、数据问答与企业改造中落地,直播与付费项目数据共同标记能力跃升。
  • AI 创作(5 条)—— 3D 重建、Blender 插件与对话表演让创作链路更短,AI 内容的接受度被判断会先于认知到来。
  • 端侧与降本(2 条)—— 4GB 模型把离线写作与数据留存带到笔记本,前沿模型发布节奏则被算力稀缺牵制。
  • 语音与安全(4 条)—— 听写词错率降到 3.4%,同时关于通用 Agent 暂缓、算力租借与芯片流向的争论持续。
  • 研究与认知(2 条)—— RoPE 变体教程与技能记忆研究并进,架构细节和人类学习机制互为参照。
  • 产品观察(1 条)—— 缺少前沿模型支撑的家用中枢设备被判定将处于劣势。
  • 其他(1 条)—— 未归类的补充内容

📖 详细内容

模型评测

Qwen4 Max 分数与成本被量化定位,Gemini 4 Argon 的豪言则以转述形式流传。


Artificial Analysis
| @ArtificialAnlys
| 影响力: 17.9万粉丝 · 互动分 925.0
Independent analysis of AI

💡 核心观点: Artificial Analysis 称阿里发布 Qwen4 Max,智能指数 43 分,较 Qwen3.8 Max 高 9 分、落后 Kimi K3.1 Pro 3 分,单任务成本 3.27 美元。

可信度: 90/100

事实核查:

  • 原文支持: Qwen4 Max, using xhigh reasoning effort, scores 43 on the Intelligence Index, 9 points ahead of Qwen3.8 Max (xhigh) and 3 points behind Kimi K3.1 Pro (max, 46)(评测机构给出具体分数与对比版本,可对照榜单复核。)
  • 原文支持: Qwen4 Max costs $3.27 per Intelligence Index task, ~3x less than Qwen3.8 Max ($10.02) and ~11% less than Kimi K3.1 Pro ($3.68)(成本数值与三个对比基准均在原文列明。)
  • 原文支持: Qwen4 Max uses 39% fewer output tokens per task than Qwen3.8 Max(原文同时给出约 105k 与 171k 的输出 token 数。)

原文内容:

Alibaba has launched Qwen4 Max, gaining 9 points over Qwen3.8 Max on the Artificial Analysis Intelligence Index and landing 3 points behind Kimi K3.1 Pro at ~11% lower Cost per Task

Qwen4 Max, using xhigh reasoning effort, scores 43 on the Intelligence Index, 9 points ahead of Qwen3.8 Max (xhigh) and 3 points behind Kimi K3.1 Pro (max, 46)

Key takeaways:

➤ Qwen4 Max improves on every evaluation in the Index vs Qwen3.8 Max (xhigh, 34): HLE-Diamond +21 points (24% to 45%), AA-Omniscience +18 points, AutomationBench-AA +17 points, CritPt v2 +13 points, Terminal-Bench-Science +11 points, Terminal-Bench 4.0 +9 points, GDP.pdf +3 points, FrontierCode 2.0 +3 points and AA-Briefcase +82 Elo.

➤ Qwen4 Max costs $3.27 per Intelligence Index task, ~3x less than Qwen3.8 Max ($10.02) and ~11% less than Kimi K3.1 Pro ($3.68): This is driven by 25% lower per-token pricing than Qwen3.8 Max, 70% cheaper cache hits and lower token usage. Tencent Hy4 costs ~3x less per task ($1.03) than Qwen4 Max but scores 8 points lower on the Index.

➤ Qwen4 Max uses 39% fewer output tokens per task than Qwen3.8 Max: ~105k output tokens per Intelligence Index task vs ~171k for Qwen3.8 Max, ~72k for Kimi K3.1 Pro and ~78k for Tencent Hy4. Qwen4 Max uses fewer tokens, while increasing 9 points on the Intelligence Index

➤ Improvements in AA-Briefcase (knowledge work): On AA-Briefcase, our evaluation of agentic work across real business files, Qwen4 Max reaches 1686 Elo vs 1674 for Kimi K3.1 Pro and 1446 for Tencent Hy4.

➤ Qwen4 Max trails Kimi K3.1 Pro on Terminal-Bench 4.0 and Tencent Hy4 on FrontierCode 2.0: Qwen4 Max scores 48% on Terminal-Bench 4.0 vs 67% for Kimi K3.1 Pro and 37% for Tencent Hy4. On FrontierCode 2.0, @cognition’s private evaluation of real-world coding tasks, it scores 26%, level with Kimi K3.1 Pro (26%) and behind Tencent Hy4 (31%).

Key model details:

➤ Context window: 1M tokens

➤ Multimodal: Text and image input with text output

➤ Pricing: $1.50/$4.50 per 1M input/output tokens on Alibaba's first-party APIs, with a $0.075 cache hit price. This is lower than Qwen3.8 Max across the board ($2.00/$6.00, with a $0.25 cache hit price)

➤ Availability: @Alibaba_Qwen 's first-party APIs

⏰ 02:55 | ❤️ 570 点赞 | 📝 286词 | 入选: 高赞 570 / 长文 | 查看原文 →

↑ 返回顶部


Prompter
| @PromptLLM
| 影响力: 9.9万粉丝 · 互动分 442.2
Teaching those how to ask AI the right questions to transform every aspect of their life.

💡 核心观点: Prompter 引用 Gemini 4 Argon 的说法,称其表示能在 3 年内从 0 做到身家十亿美元。属转述模型输出,未给出方法或依据。

可信度: 15/100

事实核查:

  • 无法核实: Gemini 4 Argon says it could go from $0 to a billionaire in 3 years(原文仅转述模型自称,未给出方法、数据或可核验来源。)

原文内容:

Gemini 4 Argon says it could go from $0 to a billionaire in 3 years. https://t.co/FrNCwyVtJf

⏰ 21:17 | ❤️ 343 点赞 | 📝 16词 | 入选: 高赞 343 | 查看原文 →

↑ 返回顶部

Agent 实践

多 Agent 分工在建模、数据问答与企业改造中落地,直播与付费项目数据共同标记能力跃升。


Justine Moore
| @venturetwins
| 影响力: 24.9万粉丝 · 互动分 649.74
Partner @a16z AI 🤖 and twin to @omooretweets | Investor in @elevenlabs, @bfl_ai, @hedra_labs, @krea_ai, @heyglif, @ShizukuAILabs, @wabi, @TownAI

💡 核心观点: 作者称,Gemini 4 Carbon 的 /swarm 功能被低估:她要求生成渡轮大厦的3D模型,系统调度12个并行Agent,分别做调研、规划、编写Blender Python脚本,并在共享留言板上互相评审。

可信度: 46/100

事实核查:

  • 无法核实: It orchestrated 12 parallel agents to research, plan, write the Blender Python scripts, and critique each other's work on a shared message board(作者给出了产品名、Agent数量和具体分工,属于可指向的一次使用记录,但原文没有截图细节、代码或可复现步骤。)
  • 无法核实: The /swarm feature with Gemini 4 Carbon is so underhyped.(“被低估”是主观判断,原文未提供使用量、性能或对比数据。)

原文内容:

The /swarm feature with Gemini 4 Carbon is so underhyped. 

I asked for a 3D model of the Ferry Building. 

It orchestrated 12 parallel agents to research, plan, write the Blender Python scripts, and critique each other's work on a shared message board 🤯 https://t.co/Ss17r5eubj

⏰ 19:07 | ❤️ 448 点赞 | 📝 44词 | 入选: 高赞 448 | 查看原文 →

↑ 返回顶部


Rohan Paul
| @rohanpaul_ai
| 影响力: 16.1万粉丝 · 互动分 134.01
Compiling in real-time, the race towards AGI. 🗞️ Get my daily AI analysis newsletter to your email 👉 https://t.co/6LBxO81tfN

💡 核心观点: Rohan Paul 介绍 Google DeepMind 与哥伦比亚大学的论文:把芯片仿真产生的超大波形转储转为 SQLite 数据库,由多 Agent 将自然语言问题转成 SQL 查询并审核。在 150 道专家标注题上用 Gemini 2.5 Pro 答对 95.33%,多 Agent 分工在质量审计上比单 Agent 高 17.3%,延迟为 1.96 倍。

可信度: 82/100

事实核查:

  • 原文支持: dumps can exceed 10 GB of text(原文给出转储量级,未说明统计口径与样本范围。)
  • 原文支持: it answered 95.33% correctly using Gemini 2.5 Pro(原文给出准确率与所用模型,并说明评测集为 150 道专家标注的验证问题。)
  • 无法核实: Splitting the work across specialized agents scored 17.3% higher on a quality audit than 1 agent alone, at 1.96x the latency(原文给出质量提升与延迟倍数,属论文报告的实验数据,推文中未附论文链接。)

原文内容:

Google DeepMind and Columbia Univ paper argues that AI agents can debug chip simulations once the huge waveform files become a SQL database they can query.

Chip simulations record every signal change, and these dumps can exceed 10 GB of text that engineers still scan by hand. No LLM context can hold that, and the files offer no way for an agent to search them.

They built Back-to-the-Future (BTTF), which converts the dumps into a compact SQLite database. A team of agents then turns plain-English questions into SQL, checks design rules, and audits each answer.

On 150 expert-labeled verification questions, it answered 95.33% correctly using Gemini 2.5 Pro. Splitting the work across specialized agents scored 17.3% higher on a quality audit than 1 agent alone, at 1.96x the latency.

If your agents struggle with huge logs or traces, put the data in a queryable database instead of pushing it into the prompt.

⏰ 23:46 | ❤️ 49 点赞 | 📝 150词 | 入选: 互动较好 49 赞 / 长文 | 查看原文 →

↑ 返回顶部


swyx
| @swyx
| 影响力: 21.4万粉丝 · 互动分 502.11
achieve ambition with intentionality, intensity, integrity & insanity. affiliations: – @smol_ai – @cognition – @temporalio – @aidotengineer – @latentspacepod

💡 核心观点: Swyx 称其首日直播以最快速度突破 10 万观看,围绕编程 Agent 实践与生态的内容是团队最强的形式,也是他个人最强的产品市场契合方向。依据为直播观看数据与自身项目经验。

可信度: 62/100

事实核查:

  • 无法核实: day 1 livestream has crossed 100k in the fastest time ever(原文给出观看量突破 10 万并自称史上最快,未提供对比数据或平台统计来源。)
  • 无法核实: covering coding agent practices and landscape continues to be the strongest format we do(这是作者对自身内容效果的主观判断,无可量化数据支撑。)

原文内容:

also the day 1 livestream has crossed 100k in the fastest time ever. covering coding agent practices and landscape continues to be the strongest format we do and by far my strongest product-market-fit focus thing I have ever done. https://t.co/cM43xYQLMT

引用 @aiDotEngineer: taking day 1 view count predictions for the @poteto talk launching in 30 mins

https://t.co/UbXuS2cTXj

hit the bell etc https://t.co/iJRvnbXhSL

⏰ 09:22 | ❤️ 247 点赞 | 📝 42词 | 入选: 高赞 247 / 引用讨论 | 查看原文 →

↑ 返回顶部


GREG ISENBERG
| @gregisenberg
| 影响力: 76.4万粉丝 · 互动分 499.74
I drop startup ideas daily. Host @startupideaspod. CEO: @latecheckoutplz we build companies like @ideabrowser, @meetLCA, @boringmarketer etc

💡 核心观点: Greg Isenberg 招募 10 家公司,用为员工和公司建立"第二大脑"、把工作流做成技能、连接工具让 Agent 执行等十步方案,帮其在 2027 年成为所在领域最"AI 化"的公司;方案源自他此前推文的企业 AI 改造清单。

可信度: 38/100

事实核查:

  • 无法核实: I'm picking 10 companies for my team to make you the most AI pilled company in your space in 2027(原文给出招募数量与目标年份,属服务承诺,无第三方证据。)
  • 原文支持: Set up a second brain for every employee (and train employees)(原文列出具体实施步骤,可直接摘录;效果未提供数据。)

原文内容:

Want to be the most AI pilled company in your space?

I’m picking 10 companies for my team to make you the most AI pilled company in your space in 2027.

Here’s what we’ll do:

1. Set up a second brain for every employee (and train employees)

2. Set up a company brain

3. Turn your best workflows into skills

4. Connect your tools so agents can take action

5. Put agents on those workflows every day

6. Build the loop so the whole system gets smarter each week

7. Redesign your existing products so agents can use them

8. Build new AI native products with you (agents, MCPs, apps)

9. Set up agents that bring you more customers (outbound, ads, content)

10. Much more

Basically, A lot of people read my tweet below and asked if someone could just set it up for them and I don't blame them.

So, reply “interested” to this tweet, then head to https://t.co/YQJFV1XTa1 (bottom of page is our email) and email us with your company/company size/what you're looking for.

We’ll reach out to the 10 companies we think we can drive an absurd amount of value for.

引用 @gregisenberg: Making the most AI pilled company is the whole game:

1. make a second brain. one folder of markdown files on how you think, write and decide (obsidian)

2. talk to it. voice notes all day (wispr), and an agent files every fact, story and decision

3. make a company brain. customers, calls, numbers,…

⏰ 19:06 | ❤️ 258 点赞 | 📝 191词 | 入选: 高赞 258 / 引用讨论 / 长文 | 查看原文 →

↑ 返回顶部


Chubby♨️
| @kimmonismus
| 影响力: 15.9万粉丝 · 互动分 2024.0
Tech Analyst & Content Creator. Editor-in-Chief @getsuperintel – Community of 300k+ in total: 🔗 https://t.co/jHMmImmI5I //📧 [email protected]

💡 核心观点: 作者称 AI Agent 在 Remote Labor Index 付费项目上的完成率 13 个月内由 2.5% 升至 41.3%,据此判断白领工作将在 2027 年受冲击。

可信度: 50/100

事实核查:

  • 无法核实: went from completing 2.5% of real freelance projects on the Remote Labor Index to 41.3% with Claude Fable 5.2(给出基准名与起止数值,未附榜单链接或评测设置。)
  • 无法核实: These were paid projects that took human professionals 29 hours on average(任务平均耗时由作者给出,无原始数据。)
  • 无法核实: white collar work will get a wake-up in 2027(属对未来的预测,无可核验依据。)

原文内容:

Talking about acceleration: In 13 months, the best AI agent went from completing 2.5% of real freelance projects on the Remote Labor Index to 41.3% with Claude Fable 5.2.

These were paid projects that took human professionals 29 hours on average, with briefs provided and client calls excluded.

So yeah, white collar work will get a wake-up in 2027.

⏰ 11:33 | ❤️ 1311 点赞 | 📝 54词 | 入选: 高赞 1311 | 查看原文 →

↑ 返回顶部

AI 创作

3D 重建、Blender 插件与对话表演让创作链路更短,AI 内容的接受度被判断会先于认知到来。


GREG ISENBERG
| @gregisenberg
| 影响力: 76.4万粉丝 · 互动分 692.09
I drop startup ideas daily. Host @startupideaspod. CEO: @latecheckoutplz we build companies like @ideabrowser, @meetLCA, @boringmarketer etc

💡 核心观点: 作者断言,用户口头上排斥的AI歌曲和AI电影最终会进入他们的歌单与年度片单,而且听众在喜欢、分享时不会察觉是AI生成。

可信度: 15/100

事实核查:

  • 无法核实: The AI song you swear you'd never listen to is probably going to end up on your playlist.(这是对受众行为的预测,原文没有调查数据、案例或平台统计,只能视为观点。)

原文内容:

Unfortunately, the AI song you swear you’d never listen to is probably going to end up on your playlist.

The AI film you swear you’d never watch is probably going to end up in your top 5 fav of the year.

You’ll like them, you’ll share them, and you’ll never know AI made them.

⏰ 18:05 | ❤️ 359 点赞 | 📝 58词 | 入选: 高赞 359 | 查看原文 →

↑ 返回顶部


Pierrick Chevallier | IA
| @CharaspowerAI
| 影响力: 5.6万粉丝 · 互动分 69.62
VFX artist. House of David S2 (Amazon). I test AI video in real pipelines and post what survives a job. Adobe Firefly Ambassadorr 🚀

💡 核心观点: Pierrick Chevallier 称 Minimax H3 在对话场景生成上表现出色,面部表情、细微反应、自然停顿与情绪表达让角色像在真实表演,认为它在对话驱动的 AI 影视制作中被低估。依据为其创作体验。

可信度: 28/100

事实核查:

  • 无法核实: The acting is what keeps bringing me back(这是作者对生成效果的主观评价,无对比测试或量化依据。)

原文内容:

I still can't get enough of creating dialogue scenes with Minimax H3. 🎬🔥
The acting is what keeps bringing me back.

Facial expressions, subtle reactions, natural pauses, emotional delivery…
H3 has a real talent for making characters feel like they're actually performing.
Seriously underrated for dialogue-driven AI filmmaking.

⏰ 12:00 | ❤️ 41 点赞 | 📝 48词 | 入选: 互动较好 41 赞 | 查看原文 →

↑ 返回顶部


Meng To
| @MengTo
| 影响力: 19.1万粉丝 · 互动分 1307.5
Founder at @designcodeio and https://t.co/Kpiogf225W. I teach designers code and developers design. My three.js templates: https://t.co/25OYbxYycG

💡 核心观点: 作者称用 Three.js 与 Claude Fable 5.2 三天做出浏览器运行的类《黑暗之魂》游戏,并列出 Claude Code、Seedream 5.0 Pro、Tripo 3D 技术栈。

可信度: 55/100

事实核查:

  • 无法核实: I made a Dark Souls inspired game in 3 days with Three.js and Claude Fable 5.2(制作周期与工具链为作者自述,成品可通过附带的演示链接查看。)
  • 原文支持: Stack: Fable 5.2 + Opus 5.5 in Claude Code, Seedream 5.0 Pro, Tripo 3D(各工具名称均在原文中列出。)

原文内容:

I made a Dark Souls inspired game in 3 days with Three.js and Claude Fable 5.2.

Runs in the browser with full lighting, effects and sound. No build step.

Stack: Fable 5.2 + Opus 5.5 in Claude Code, Seedream 5.0 Pro, Tripo 3D

Full workflow and prompts below (1/10) https://t.co/WNQppacf7d

⏰ 09:22 | ❤️ 826 点赞 | 📝 45词 | 入选: 高赞 826 | 查看原文 →

↑ 返回顶部


AmirMušić
| @AmirMushich
| 影响力: 9.3万粉丝 · 互动分 932.9
Creative architect. Ex-Warner Music, PepsiCo, Spotify fr. designer. Fusing 10+ yr of brand & ad design with AI. Collabs: [email protected]

💡 核心观点: 作者展示在 Claude Code 中用 Blender 经 9 次修订生成 3D 对象,称音效仍弱,正开发插件改善。

可信度: 38/100

事实核查:

  • 无法核实: My input + 9 revisions in Claude Code using Blender(修订次数与工具链为作者自述,无可核验记录。)
  • 无法核实: Sound is still very weak (even in Fable)(对音效质量的判断属主观评价。)

原文内容:

Claude Opus 5.5 did (prompt)

My input + 9 revisions in Claude Code using Blender
Sound is still very weak (even in Fable) - so I'm cooking a plugin for Claude to make it sound better

 Try this prompt with any [OBJECT] 👇

⏰ 15:04 | ❤️ 696 点赞 | 📝 39词 | 入选: 高赞 696 | 查看原文 →

↑ 返回顶部


Ethan Mollick
| @emollick
| 影响力: 41.9万粉丝 · 互动分 876.98
Professor @Wharton studying AI. New book, Co-Existence, is out now: https://t.co/hsDYf1xU9K My Substack: https://t.co/UIBhxu4bgq

💡 核心观点: 作者称让 Fable 与 Opus 找到星舰回收视频,用 3js 将最后约 40 秒高细节重建为可切换视角的 3D 场景,并给出试玩链接。

可信度: 45/100

事实核查:

  • 原文支持: Render the last 40 seconds or so in 3D in 3js with extreme detail. Let me switch views.(指令内容与推文附带的可交互链接对应。)
  • 无法核实: would have seemed impossible a few months ago(属作者对能力变化的主观评价。)

原文内容:

Another one of those things that would have seemed impossible a few months ago: "Fable & Opus, find video of the Starship catch. Render the last 40 seconds or so in 3D in 3js with extreme detail. Let me switch views. No need to include the crowd."

Play: https://t.co/m6yROMYHOI https://t.co/42klMiFKc1

⏰ 18:39 | ❤️ 519 点赞 | 📝 51词 | 入选: 高赞 519 | 查看原文 →

↑ 返回顶部

端侧与降本

4GB 模型把离线写作与数据留存带到笔记本,前沿模型发布节奏则被算力稀缺牵制。


Alex Veremeyenko
| @alex_verem
| 影响力: 10.5万粉丝 · 互动分 2920.5
open source, local ai, ai for good also know as: @alex_prompter

💡 核心观点: 英伟达推出 Nemotron 3.5 Bolt E4B,4GB 体积可在多数笔记本离线运行,能写作、答问和起草邮件,数据留在本机。

可信度: 60/100

事实核查:

  • 原文支持: Nemotron 3.5 Bolt E4B is a 4GB file that runs on almost any laptop(模型名与 4GB 体积均在原文中直接给出。)
  • 无法核实: writes, answers questions and drafts emails with Wi-Fi off and nothing leaves your machine(离线能力为作者描述,未附下载页或官方说明。)

原文内容:

The smartest download this week is Nvidia's free AI.

Nemotron 3.5 Bolt E4B is a 4GB file that runs on almost any laptop.

It writes, answers questions and drafts emails with Wi-Fi off and nothing leaves your machine.

Grab it, while it's still up.

⏰ 17:01 | ❤️ 2305 点赞 | 📝 44词 | 入选: 高赞 2305 | 查看原文 →

↑ 返回顶部


Chubby♨️
| @kimmonismus
| 影响力: 15.9万粉丝 · 互动分 950.27
Tech Analyst & Content Creator. Editor-in-Chief @getsuperintel – Community of 300k+ in total: 🔗 https://t.co/jHMmImmI5I //📧 [email protected]

💡 核心观点: 作者推测 GPT-6.2 未发布源于 OpenAI 将稀缺算力优先投入研究,并称安全、产品就绪度与成本也影响发布节奏。

可信度: 20/100

事实核查:

  • 无法核实: OpenAI is prioritizing scarce compute for research over serving new models(属作者推测,未给出内部消息或公开来源。)
  • 无法核实: Safety, product readiness, and costs may also affect release timing(原文用 may 表述,属可能性推断。)

原文内容:

Another thought why no GPT-6.2 has been released yet: OpenAI is prioritizing scarce compute for research over serving new models, especially if rapid progress makes intermediate checkpoints less compelling to release. 

Safety, product readiness, and costs may also affect release timing in that regard. 

If progress is moving fast enough internally, the next breakthrough could be worth more than shipping an intermediate checkpoint.

⏰ 21:56 | ❤️ 687 点赞 | 📝 63词 | 入选: 高赞 687 | 查看原文 →

↑ 返回顶部

语音与安全

听写词错率降到 3.4%,同时关于通用 Agent 暂缓、算力租借与芯片流向的争论持续。


Rohan Paul
| @rohanpaul_ai
| 影响力: 16.1万粉丝 · 互动分 607.0
Compiling in real-time, the race towards AGI. 🗞️ Get my daily AI analysis newsletter to your email 👉 https://t.co/6LBxO81tfN

💡 核心观点: 作者介绍语音输入公司Wispr Flow:其新任首席科学官Ariya Rastrow来自Alexa语音团队,公司转向自训练模型;首款模型Canto 1针对真实嘈杂听写数据训练并按用户词汇调优,在2300多名说话者、约10小时真实听写上取得3.4%词错率,并称上线后超过1秒才显示的听写下降54%。

可信度: 74/100

事实核查:

  • 无法核实: On about 10 hours of real Flow dictations from more than 2,300 speakers, it posted a 3.4% word error rate, lower than all competing models that Wispr tested.(原文给出了数据规模、指标与对比范围,但对比对象由Wispr自行选定,原文未列出具体竞品名单。)
  • 无法核实: Since launch, dictations that take more than 1 second to appear are down 54%.(原文给出明确的百分比变化,未说明统计区间和基线,无法独立核验。)
  • 无法核实: WisprFlow recently hired Ariya Rastrow (founding Alexa scientist) as chief science officer(原文明确写出任职信息与Alexa背景,属于可查询的人事事实,但原文未附公司公告链接。)

原文内容:

The keyboard was a workaround for machines that couldn’t hear. Machines that hear make it optional.

WisprFlow recently hired Ariya Rastrow (founding Alexa scientist) as chief science officer and moved from rented speech engines to models it trains itself.

Hard to find a better fit for a speech lab than someone who scaled Alexa’s speech team.

Canto 1, his lab’s first model, was built to cut how often people edit dictated text, so Wispr trained it on noisy real-world dictations and tuned it to each user’s vocabulary and names. 

And this is a hard test, unlike many, because Speech recognition was long tested on clean, read-aloud audio, but the hard problem is a cheap laptop mic in a room where someone else is talking.

On about 10 hours of real Flow dictations from more than 2,300 speakers, it posted a 3.4% word error rate, lower than all competing models that Wispr tested.

Typing gives you feedback on every keystroke, while speech makes you wait until the end of a sentence to see what the machine heard, so speed counts as much as accuracy.

Since launch, dictations that take more than 1 second to appear are down 54%.

Its training runs through pretraining, supervised fine-tuning and reinforcement learning with GRPO, the same stages used for large language models.

The reinforcement learning stage uses corrections from Flow users who opted in to share data.

When someone fixes a dictated word, Wispr checks how the audio aligns with the text and how the edit is shaped.

A misheard word gets folded back into the training reference, and a plain rewrite of the sentence gets ignored.

⏰ 21:59 | ❤️ 403 点赞 | 📝 273词 | 入选: 高赞 403 / 长文 | 查看原文 →

↑ 返回顶部


Gary Marcus
| @GaryMarcus
| 影响力: 25.8万粉丝 · 互动分 644.01
OG GenAI Skeptic; spoke at US Senate. Warned about hallucinations in 2001. Advocating world models & neurosymbolic AI ever since. Author, Marcus on AI & 6 books

💡 核心观点: 作者反驳“该为AI恐慌”的流行论调,称这类文章常把未具名批评者与多年前观点混为一谈;他认为当前能力主要依赖可验证性,在数学和编程上强、缺乏通用性,真正风险是把仍不可靠的系统用于空管等关键场景且监管不足,主张暂缓最激进的通用Agent AI。

可信度: 58/100

事实核查:

  • 无法核实: I never called AI a con(这是作者对自身过往言论的澄清,属于个人表态,原文无法证明其历史发言全貌。)
  • 无法核实: what we have now is heavily dependent on verifiability—a technique that works well for math, and coding, but lacks generality(原文把可验证性作为当前系统能力的支撑,并限定其适用范围;这是作者的技术判断,没有给出实验或引用来源。)
  • 原文支持: demanding a temporary recall of the most egregious of the current technologies (general purpose agentic AI) until the problems are solved(原文确实提出暂停最激进通用Agent AI的政策主张;理由与上文对不可靠性和监管缺位的论述一致。)

原文内容:

Getting tired of panicky doomer posts like the one below, which are becoming widespread, and starting to feel formulaic.

Folks in this genre often point to unnamed critics, and ideas from years ago, and make it sound like all skeptics have exactly the same views, which is very much untrue. For example, I never called AI a con, so he goes after those that did, and not my more nuanced views.)

These posts must sell a lot of subscriptions.

But my advice is don’t panic, and don’t listen panicky merchants of doom.

The reality is that the original pure LLMs were deeply troubled, and it was right to critique them, then.  But of course the field has evolved in four years.

Important some new techniques beyond pure LLMs have been developed (primarily neurosymbolic, along lines I always pushed for). 

But even so, what we have now is heavily dependent on verifiability—a technique that works well for math, and coding, but lacks generality. It’s not superintelligent, it’s super good at math and logic. (Pure LLMs remain poor at both, but symbolic techniques are helping pick up the slack.)

The biggest problem is not that the machines are getting smarter, and that we are getting dumber. 

How so? By stupidly putting systems that are still unreliable in charge of things like air traffic control, and not regulating them, and giving them too much access to the internet.

We can, at least for now, solve most of what’s going by demanding a temporary recall of the most egregious of the current technologies (general purpose agentic AI) until the problems are solved.

Tearing our hair out and misrepresenting the work of past scholars may score cheap (and misleading) rhetorical points but it doesn’t do anything to solve the actual problems.

引用 @rutgerbregman: It's time to panic about AI.

For years, I listened to the reasonable people. I nodded along when critics called AI “a con,” and when they waved ChatGPT away as “spicy autocomplete.” I laughed in 2024, when the internet was mocking ChatGPT because it couldn’t count the r’s in “strawberry.” When I me…

⏰ 13:46 | ❤️ 404 点赞 | 📝 303词 | 入选: 高赞 404 / 引用讨论 / 长文 | 查看原文 →

↑ 返回顶部


Gary Marcus
| @GaryMarcus
| 影响力: 25.8万粉丝 · 互动分 524.58
OG GenAI Skeptic; spoke at US Senate. Warned about hallucinations in 2001. Advocating world models & neurosymbolic AI ever since. Author, Marcus on AI & 6 books

💡 核心观点: 作者批评特朗普政府在公共卫生、能源价格、对外冲突与科研机构上的处置,并称其对英伟达芯片流向中国视而不见。

可信度: 15/100

事实核查:

  • 无法核实: they have ripped apart America’s crown jewel, its heretofore absolutely first rate scientific institutions(总体性批评,未列出具体机构或事件。)
  • 无法核实: They also turned a blind eye to Nvidia’s apparent backdoor sales of chips to China.(原文用 apparent 表述,未提供证据或报道来源。)

原文内容:

Basically the Trump administration has given us measles, higher gas prices, a war nobody wants, and nationwide ground stops. Don’t even get me started about the cruelty of ICE.

As a bonus they have ripped apart America’s crown jewel, its heretofore absolutely first rate scientific institutions. 

They also turned a blind eye to Nvidia’s apparent backdoor sales of chips to China.

And they have never apologized for any of this mess; the dude who canceled all the flights, for example, told a hard to swallow story about a router, and bowed down to Trump, as if everything was otherwise fine.

It’s not.

⏰ 22:42 | ❤️ 336 点赞 | 📝 106词 | 入选: 高赞 336 | 查看原文 →

↑ 返回顶部


SemiAnalysis
| @SemiAnalysis_
| 影响力: 18.7万粉丝 · 互动分 592.81

💡 核心观点: 作者转述“让中国租用美国算力”的反方论证:中国在境外用算力会把高额利润付给美国生态,英伟达每兆瓦约赚700万美元、云厂商再赚700万至1000万美元;限制买芯片但允许租用,会让Meta等美国公司在成本上占优,并称美国从中国模型身上获利丰厚。

可信度: 52/100

事实核查:

  • 无法核实: Nvidia makes about $7 million of profit on every megawatt deployed. The cloud provider makes another $7 to 10 million at today's pricing.(原文给出具体利润数字,属于引用观点中的测算,未说明口径与来源。)
  • 无法核实: Application-layer companies running on open models add up to well north of $10 billion in ARR.(原文给出ARR规模判断,但未列出统计范围或数据来源。)
  • 无法核实: DeepSeek makes less than Baseten and Fireworks do serving its models.(原文明确比较DeepSeek与Baseten、Fireworks的收益,属引用观点,未附带财务数据。)

原文内容:

The devil’s advocate case for letting China rent American compute.

“If China has access to compute outside of China, it’s paying humongous margin dollars to the US ecosystem. Nvidia makes about $7 million of profit on every megawatt deployed. The cloud provider makes another $7 to 10 million at today’s pricing.”

“If ByteDance wants to compete with Meta on video recommendation globally, and you don’t let it buy the chips but you do let it rent them, Meta potentially has a fourth of the cost of compute.”

“Application-layer companies running on open models add up to well north of $10 billion in ARR. DeepSeek makes less than Baseten and Fireworks do serving its models. America is making so much money off Chinese models.”

⏰ 02:56 | ❤️ 417 点赞 | 📝 124词 | 入选: 高赞 417 / 长文 | 查看原文 →

↑ 返回顶部

研究与认知

RoPE 变体教程与技能记忆研究并进,架构细节和人类学习机制互为参照。


Sebastian Raschka
| @rasbt
| 影响力: 53.9万粉丝 · 互动分 2081.41
ML/AI research engineer. Ex stats professor. Author of "Build a Large Language Model From Scratch" (https://t.co/O8LAAMRzzW) & reasoning (https://t.co/5TueQKx2Fk)

💡 核心观点: 作者发布从零讲解 RoPE 的教程,并补充 NoPE、部分 RoPE 与混合方案等近期架构做法。

可信度: 72/100

事实核查:

  • 原文支持: A RoPE from-scratch tutorial(教程主题与形式由作者说明,推文附有文章链接。)
  • 原文支持: what recent architectures do (NoPE, partial RoPE, hybrids)(原文点名三类近期方案,与教程内容一致。)

原文内容:

(Finally) wrote up "Understanding Positional Embeddings in Modern LLMs."

A RoPE from-scratch tutorial missing from my LLM building blocks collection + what recent architectures do (NoPE, partial RoPE, hybrids).

Prob not the flashiest topic, but essential for understanding LLMs https://t.co/mkggue58t4

⏰ 13:29 | ❤️ 1478 点赞 | 📝 42词 | 入选: 高赞 1478 | 查看原文 →

↑ 返回顶部


Aakash Gupta
| @aakashgupta
| 影响力: 33.6万粉丝 · 互动分 111.85
✍️ https://t.co/8fvSCtBv5Q 💼 https://t.co/STzr4nqxnm 🤝 https://t.co/SqC3jTyP03 🎙️ https://t.co/fmB6Zf5UZv

💡 核心观点: 作者以 Ronaldinho 复出及失忆患者镜像画星实验说明技能记忆独立于事实记忆,并引 2014 年 Neymar 脑扫描研究称熟练者动作耗脑更低。

可信度: 45/100

事实核查:

  • 无法核实: Henry Molaison had brain surgery back in 1953 to stop his seizures, and it wiped out his ability to store anything new.(人物与年份与公开记载相符,推文未附来源。)
  • 无法核实: Back in 2014, Japanese researchers put Neymar in a brain scanner(给出年份、对象与结论,未附论文出处。)

原文内容:

Ronaldinho is 46 and hasn't played a real game in 11 years. He just walked into training in Italy and killed a ball out of the air like it was 2006.

There's a reason a touch like that survives 11 years away from the game, and it's the same reason you can hop on a bike you haven't ridden since high school and just go. Scientists figured out where that lives because of a man who couldn't make a single new memory for 55 years.

Henry Molaison had brain surgery back in 1953 to stop his seizures, and it wiped out his ability to store anything new. Brenda Milner, the researcher who worked with him for decades, said every time she walked into the room it was like they'd never met.

So she gave him something that's brutal for anybody. Trace a star while only looking at your hand in a mirror.

Three days later he was nearly perfect at it, and he swore every single time that he'd never done it before. His hand kept getting better while his memory kept coming up empty.

Turns out skills get stored in a totally separate part of the brain from facts and events, in older wiring like the cerebellum, and repetition burns them in. That's the bike thing.

Repetition does one more thing to you. Back in 2014, Japanese researchers put Neymar in a brain scanner and had him rotate his right ankle. The pro footballers lit up less of the brain's foot area than top swimmers or an amateur doing the exact same move, and Neymar lit up the least of anybody. The lead researcher said his activity was under 10% of the amateur's. Basically, the better you get at something, the less brain it costs you.

Ronaldinho got his reps as a kid on futsal courts in Porto Alegre, about a ninth the size of a regular field, with somebody on you the second the ball shows up. He's said most of his moves came from futsal and that his control still feels like a futsal player's. Do that for decades and a first touch turns into something close to a reflex, stored in the same system that kept working for Henry Molaison.

Speed is where age gets you. Top masters sprinters lose around 5 to 6% of their speed every decade, and the fast-twitch fibers you need for a burst are the first to go.

Tobias, one of the best futsal players Brazil ever had, ran into exactly this when he tried the full-size game. He said nobody could take the ball off him, but sprinting 40 meters to get near the goal was a whole other thing. The move lasted four months.

What you're watching in this clip is a 20-something's touch on a 46-year-old engine. He's going for what he calls his 300th career goal against Forlì tomorrow (Transfermarkt has him at 267), and we're about to find out how far that touch can carry him.

引用 @FabrizioRomano: 🇧🇷 Ronaldinho has arrived for his pre-match training session with Italian 3rd division side Ravenna ahead of coming out of retirement for one fame.

The Brazilian legend joined Ravenna’s project with Cipriani family, promised to play an official game and it’s happening this weekend.

Ronaldinho will…

⏰ 05:51 | ❤️ 62 点赞 | 📝 506词 | 入选: 互动较好 62 赞 / 引用讨论 / 长文 | 查看原文 →

↑ 返回顶部

产品观察

缺少前沿模型支撑的家用中枢设备被判定将处于劣势。


Ethan Mollick
| @emollick
| 影响力: 41.9万粉丝 · 互动分 845.06
Professor @Wharton studying AI. New book, Co-Existence, is out now: https://t.co/hsDYf1xU9K My Substack: https://t.co/UIBhxu4bgq

💡 核心观点: 这其实是一张观点卡:作者认为,没有强AI模型支撑的家庭中枢设备会越来越吃亏,并拿配置了Codex/Code/OpenClaw的PC与Siri、Alexa对比,认为前沿模型更像科幻里的家用电脑,也更直觉。

可信度: 32/100

事实核查:

  • 无法核实: Frontier models act like the household computer from sci-fi & are more intuitive.(说法来自作者日常体验的对比,没有给出测试方法、设备型号或量化指标,难以独立复核。)

原文内容:

Not having a good AI model powering your central home device will be a growing liability. You see the difference when you have a central PC with Codex/Code/OpenClaw and you try to use Siri or Alexa. Frontier models act like the household computer from sci-fi & are more intuitive.

⏰ 22:19 | ❤️ 490 点赞 | 📝 52词 | 入选: 高赞 490 | 查看原文 →

↑ 返回顶部

其他

未归类的补充内容


@levelsio
| @levelsio
| 影响力: 99.3万粉丝 · 互动分 4449.0
💸https://t.co/sQ0aiU7v02 📸https://t.co/lAyoqmSBRX $68K/m 🎮https://t.co/tF49atJ396 $44K/m 🏡https://t.co/1oqUgfD6CZ $23K/m 👙@X $20K/m 🌍https://t.co/UXK5AFqCaQ $7K/m 🏩https://t.co/4p4dzTDVN6 $3K/m 💾@pieter 📝https://t.co/8VnHmB3uah

💡 核心观点: 作者认为欧洲人因习惯不工作,最能适应后AGI时代。

可信度: 20/100

事实核查:

  • 无法核实: Europeans are by far the most prepared for post-AGI of anybody because they're already so used to not working(原文仅有观点陈述,无可供核验的数据、来源或可验证细节。)

原文内容:

Europeans are by far the most prepared for post-AGI of anybody because they're already so used to not working

⏰ 23:19 | ❤️ 3841 点赞 | 📝 21词 | 入选: 高赞 3841 | 查看原文 →

↑ 返回顶部

🔗 引用关系

  • @GaryMarcus 引用 @rutgerbregman:作者反驳“该为AI恐慌”的流行论调,称这类文章常把未具名批评者与多年前观点混为一谈;他认为当前能力主要依赖可验证性,在数学和编程上强、缺乏通用性,真正风险是把仍不可靠的系统用于空管等关键场景且监管不足,主张暂缓最激进的通用Agent AI。(原文:It's time to panic about AI.

    For years, I listened to the reasonable people. I nodded along when critics called AI “a c…)

  • @swyx 引用 @aiDotEngineer:Swyx 称其首日直播以最快速度突破 10 万观看,围绕编程 Agent 实践与生态的内容是团队最强的形式,也是他个人最强的产品市场契合方向。依据为直播观看数据与自身项目经验。(原文:taking day 1 view count predictions for the @poteto talk launching in 30 mins

    https://t.co/UbXuS2cTXj

    hit the bell etc…)

  • @gregisenberg 引用 @gregisenberg:Greg Isenberg 招募 10 家公司,用为员工和公司建立"第二大脑"、把工作流做成技能、连接工具让 Agent 执行等十步方案,帮其在 2027 年成为所在领域最"AI 化"的公司;方案源自他此前推文的企业 AI 改造清单。(原文:Making the most AI pilled company is the whole game:

    1. make a second brain. one folder of markdown files on how you th…)

  • @aakashgupta 引用 @FabrizioRomano:作者以 Ronaldinho 复出及失忆患者镜像画星实验说明技能记忆独立于事实记忆,并引 2014 年 Neymar 脑扫描研究称熟练者动作耗脑更低。(原文:🇧🇷 Ronaldinho has arrived for his pre-match training session with Italian 3rd division side Ravenna ahead of coming out …)

🧩 跨推文归纳

共识

  • 多方证据指向 Agent 在受控、可验证任务上的完成度快速抬升,评测与实务案例相互印证。
  • 3D 与影视创作链条被反复验证:从星舰视频重建到渡轮大厦建模,工具组合趋于稳定。

分歧

  • Gary Marcus 认为当前能力依赖可验证性、应暂缓激进 Agent;而 Agent 完成率与多 Agent 评测数据支持更乐观的推进节奏。

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

相关文章

暂无评论

none
暂无评论...