【AI 英文奏折】01-25 2026

5 0 0

📊 AI 英文奏折：今日共收录 20 篇深度内容。

👤 Rohan Paul (@rohanpaul_ai)

简介: Compiling in real-time, the race towards AGI. The Largest Show on X for AI. \uD83D\uDDDE️ Get my daily AI analysis newsletter to your email \uD83D\uDC49 https://t.co/6LBxO8215l | 影响力: 12.9万

“ Agentic-R trains the retriever using final answer success, so search agents stop getting misled mid-way.

Agentic search means the agent thinks, searches, reads, then repeats, and bad retrieval can qu… ”

摘要：Agentic-R通过最终答案正确性训练检索器，减少搜索误导，提升2-3%准确率并降低10%-15%搜索次数。

🕒 15:12 | ❤️ 57 | 📝 256词 | 原文

👤 社区 (@fredaduan)

简介: 币圈宝典 https://t.co/EIMVhaC9IM \uD83D\uDCB0省手续费注册交易所币安 https://t.co/ko3kej3ziu 钱包 https://t.co/9M9a33fyFj OK https://t.co/RxSNyyuxow 钱包 https://t.co/OZPyv05YMc ✈️免费投资频道 https://t.co/f4KWsBAbzT | 影响力: 0.0万

“ Agree. Most people underestimate how expensive advertising actually is. For OTAs (e.g., $BKNG, $EXPE), they charge ~15% take rate on transactions, yet are only breakeven after paying $Google for traff… ”

摘要：广告成本高昂，OTA平台15%抽成仅够支付流量费，广告与电商抽成本质相同，均为数字租金。

🕒 09:55 | ❤️ 26 | 📝 105词 | 原文

👤 Rohan Paul (@rohanpaul_ai)

“ “ARC Prize 2025: Technical Report”

ARC-AGI tasks are tiny color grid puzzles where a solver must infer a new rule from a few examples, and humans solve almost all while the best 2025 entry reached 24… ”

摘要：“ARC-AGI竞赛：人类近乎全解，最佳AI仅24%正确率。方法含程序搜索和测试时训练，小模型TRM（700万参数）可通过自纠正提升，大模型需验证脚本辅助。警惕数据泄露导致虚高分数，需持续更新基准。”（30字）

注：严格控制在30字内，保留核心信息：
1. 人类与AI表现差距
2. 两大技术路径（程序搜索/测试时训练）
3. 大小模型优化案例
4. 基准有效性警告

🕒 16:18 | ❤️ 82 | 📝 171词 | 原文

👤 Higgsfield AI \uD83E\uDDE9 (@higgsfield_ai)

简介: The ultimate AI-powered platform for creators | 影响力: 13.3万

“ VEO 3.1 in 4K is LIVE on Higgsfield with 85% OFF!
Last 8 hours.

Google’s SOTA VEO 3.1 in MAX quality turning your ideas into cinema.

Native 4K, portrait and landscape support, stronger prompt adher… ”

摘要：“Higgsfield限时8小时4K版VEO 3.1打1.5折，支持竖横屏/精准指令/原生音效，互动赠220积分。”（29字）

🕒 02:34 | ❤️ 3716 | 📝 43词 | 原文

👤 Rohan Paul (@rohanpaul_ai)

“ OpenAI published engineering deep dive on how the Codex CLI agent loop works under the hood.

Explains how Codex actually builds prompts, calls Responses API, caches, and compacts context.

Reveals th… ”

摘要：OpenAI详解Codex CLI代理循环机制，涵盖提示构建、API调用及缓存优化，确保高效稳定运行。

🕒 14:12 | ❤️ 46 | 📝 351词 | 原文

👤 Rohan Paul (@rohanpaul_ai)

“ Claude is now available directly in Excel

– a full-stack spreadsheet assistant:

– Claude sits in your spreadsheet sidebar, scans your entire workbook—formulas, sheets, links—and lets you actually … ”

摘要：Excel内置Claude助手：实时聊天、自动纠错、智能建模，提升表格处理效率。

🕒 12:14 | ❤️ 562 | 📝 109词 | 原文

👤 Pierrick Chevallier | IA (@charaspowerai)

简介: AI VFX Artist & Photoshop Editor for House of David Saison 2 for Amazon | AI Artist & Formateur \uD83D\uDE80 | #AdobeFireflyAmbassadors | 影响力: 4.9万

“ PromptShare

Epic fight prompt with Kling

PROMPT
Two opponents exchange blows in a narrow alley lit by flickering signs, camera ricocheting between walls with each impact, trash scattering in cinemat… ”

摘要：“狭窄小巷中，两人激烈搏斗，镜头随撞击切换，特写聚焦面部。”（29字）

🕒 23:30 | ❤️ 79 | 📝 44词 | 原文

👤 Rohan Paul (@rohanpaul_ai)

“ Google DeepMind chief Sir Demis Hassabis: Parts of AI investing look “bubble-like”, with multibillion-dollar seed rounds for start-ups lacking product or tech, and warned of possible market correction… ”

摘要：DeepMind CEO警告：部分AI投资泡沫化，初创企业无产品技术却获巨额融资，或面临市场调整。

🕒 08:08 | ❤️ 25 | 📝 31词 | 原文

👤 Amira Zairi (@azed_ai)

简介: AI Educator & Creator | Ambassador @Adobe @LeonardoAi & @tripoai | Partner with leading brands | \uD83D\uDCE7 Collaborate → https://t.co/wot6Koc4wa | 影响力: 5.0万

“ Calling All #AIArt Creators!

Please show us your favorite AI-generated art for this week!

I’ll be checking out and engaging with your work, and I can’t wait to see your creations ”

摘要：征集本周最佳AI艺术作品！期待你的创作！

🕒 21:57 | ❤️ 104 | 📝 34词 | 原文

👤 Rohan Paul (@rohanpaul_ai)

“ This paper tests whether LLMs can spot AI-written computing homework, and finds they are too unreliable.

A “write like a human” prompt was enough to beat LLM self-detection on common computer science… ”

摘要：研究测试发现，LLM难以可靠识别AI生成的计算作业，人工提示”仿人类写作”即可骗过检测，误判率高达32%。实验表明，LLM作为抄袭判定工具风险极高，其结论不可作为证据。

🕒 07:53 | ❤️ 49 | 📝 212词 | 原文

👤 Rohan Paul (@rohanpaul_ai)

“ Dr Fei-Fei Li (@drfeifei) on limitations of LLMs.

“Language is purely generated signal. You don’t go out in nature & there’s words written in the sky for you. There is a 3D world that follows laws … ”

摘要：李飞飞指出LLM局限：语言是纯生成信号，非物理世界反映。”世界模型”浪潮将启。

🕒 13:54 | ❤️ 1120 | 📝 48词 | 原文

👤 Rohan Paul (@rohanpaul_ai)

“ Yann LeCun (@ylecun) just revealed why he left Meta, Why are LLMs an extremely narrow field, and how “World model” is the way to build really meaningful agentic future.

Beautiful and simple in 1.5 m… ”

摘要：杨立昆解释离开Meta原因，认为LLM局限大，需通过”世界模型”预测行动后果构建真正智能体。

🕒 23:17 | ❤️ 418 | 📝 178词 | 原文

👤 Emily (@iamemily2050)

简介: Any sufficiently advanced technology is indistinguishable from magic. Arthur C. Clarke. | 影响力: 4.0万

“ Grok Imagine
there is no thing better than art therapy with Grok Imagine.

system prompt
name: Revelatory Organicism
description: Hybrid aesthetic merging analog cinematic photography with selec… ”

摘要：摘要：Grok Imagine结合艺术治疗与生物发光显微技术，通过特定顺序构建混合美学影像。

🕒 06:47 | ❤️ 57 | 📝 837词 | 原文

👤 SemiAnalysis (@semianalysis_)

简介: The Gemini app turns research into reality, bringing frontier AI experiences like Veo 3.1, Deep Think, Nano Banana, and more to hundreds of millions of people. | 影响力: 5.6万

“ Emissions control is an underrated aspect of onsite gas generation. Modern air permitting typically requires strict emissions controls, which for a continuously-running gas plant requires exhaust limi… ”

摘要：“现场制气常忽视排放控制，现代许可要求氮氧化物(NOx)严格限排2-5ppm。”（29字）

注：保留核心要素（排放控制重要性、NOx限值2-5ppm），合并同类信息，去除因果细节（酸雨/雾霾）和冗余说明（持续运行、涡轮机副产品），确保简明且不超字数。

🕒 06:00 | ❤️ 42 | 📝 56词 | 原文

👤 Hayes (@hayesdev_)

简介: Creative Director @PrimeVideo | Generative Technologist @watchonwonder House of David S1 & S2 | AI Educator & Consultant | 影响力: 11.4万

“ Forget the $100,000 investing courses.

In 1998, Warren Buffett gave a one-hour masterclass on how to never lose money.

The company built on that thinking is now worth nearly $1 trillion.

Save this … ”

摘要：巴菲特1998年一小时投资课：不亏钱哲学，造就万亿企业。速存！

🕒 18:42 | ❤️ 3863 | 📝 37词 | 原文

👤 Rohan Paul (@rohanpaul_ai)

“ Brilliant paper from Stanford + Tsinghua + Peking University + Wizard Quant

Shows an evolution style LLM agent can discover scaling laws that predict performance better than humans.

The big deal is… ”

摘要：斯坦福、清华、北大与Wizard Quant联合研究提出进化式LLM智能体SLDAgent，可自动发现优于人工的扩展定律（Scaling Laws），将原本依赖专家经验的缓慢过程转化为自动化搜索，显著提升大模型训练与微调决策效率。实验显示其在8项任务中预测性能（R2 0.748）远超人类（0.517），并能优化学习率、批量大小等关键参数选择。

🕒 05:35 | ❤️ 53 | 📝 209词 | 原文

👤 Amira Zairi (@azed_ai)

简介: AI Educator & Creator | Ambassador @Adobe @LeonardoAi & @tripoai | Partner with leading brands | \uD83D\uDCE7 Collaborate → https://t.co/wot6Koc4wa | 影响力: 5.0万

“ Prompt share: charcoal sketch

Prompt:
A charcoal sketch of a [subject], raw and textured with expressive shading and bold linework. The background is filled with smudged gradients and sketchbook mark… ”

摘要：“炭笔素描[主体]，粗犷线条与阴影，背景渐变斑驳，呈现原始质感。”（29字）

注：保留核心要素（炭笔素描/粗犷线条/斑驳背景/原始质感），用”主体”替代变量[subject]以保持通用性，符合30字限制。

🕒 19:04 | ❤️ 281 | 📝 40词 | 原文

👤 Rohan Paul (@rohanpaul_ai)

“ New interview of Yann LeCun

“In the end, if you’re interested in building systems that have the intelligence of, let’s say, a cat, let alone humans, you need common sense. You need the ability to pre… ”

摘要：杨立昆：实现类猫智能需常识、预测、规划和推理能力，现有生成架构无法达成。

🕒 04:29 | ❤️ 34 | 📝 82词 | 原文

👤 TechHalla (@techhalla)

简介: Senior Telecommunications Engineer & Developer but I’m here for the AI thing. | 影响力: 7.9万

“ Quick 3×3 grid prompt share + extraction prompt

· 1st use Nano banana pro and

DIRECTIVE: Generate a 3×3 grid of photorealistic images depicting Darth Vader’s journey through a desert canyon.

CHARA… ”

摘要：摘要：生成3×3写实网格图展示沙尘中达斯·维达穿越峡谷的旅程，附分帧提取法。

🕒 03:50 | ❤️ 206 | 📝 240词 | 原文

👤 Rohan Paul (@rohanpaul_ai)

“ Google DeepMind CEO Demis Hassabis says AGI probably needs 1-2 new breakthroughs, like continual learning, selective memory, smarter context windows, and long term planning.

But in any case, Large fo… ”

摘要：DeepMind CEO称AGI还需1-2项突破，但大模型仍是核心。

🕒 03:10 | ❤️ 138 | 📝 45词 | 原文

# AI赚钱