← 首页
10JUL
AI前沿每日脉动
AI Frontier Pulse · 中英双语版 Bilingual Edition
2026.07.10 · 周五刊
19 位 Builder44 条推文1 期深度播客0 篇博客
GPT-5.6 SolChatGPT Work / CodexGemini 反馈闭环Schmidhuber 访谈
Richard Liu · 2026
Curated by Richard Liu
今日头条 · GPT-5.6 Sol01 / 15
“我们听到了企业对 AI 成本的担忧，5.6 Sol 是 dollars-per-task 的巨大进步，Terra 和 Luna 也是。”“5.6 Sol is a huge step forward for dollars-per-task, as are Terra and Luna.”
@sama X 原文 9,371 ❤ · 314 RT
GPT-5.6 Sol：OpenAI 把企业成本焦虑放到头版Dollars-Per-Task Becomes The Pitch
Sam Altman 明确把 GPT-5.6 Sol 的价值锚定在企业最敏感的指标：每个任务的美元成本。OpenAI 不再只讲“更聪明”，而是在讲“同样复杂工作能不能更便宜、更稳定地完成”。Terra、Luna 与 Sol 形成产品分层，说明模型发布正在从单一旗舰走向任务经济学。
Sam Altman framed GPT-5.6 Sol around the metric enterprises care about most: dollars per task. OpenAI is no longer only selling intelligence; it is selling cheaper and more reliable completion of complex work through a tiered Sol, Terra, and Luna family.
@samaX 原文9,371 ❤ · 314 RT · 545 💬
ChatGPT Work 与 Codex 24 小时两次重置限额Usage Limits As Launch Strategy
Thibault Sottiaux 宣布，为庆祝 GPT-5.6 Sol 发布，ChatGPT Work 与 Codex 的 rate limits 会在 24 小时内再次重置两次。OpenAI 不是只给 demo，而是主动制造“足够长的尝试窗口”，鼓励用户把模型推向更雄心勃勃的任务。
Thibault Sottiaux announced repeated rate-limit resets across ChatGPT Work and Codex so users have enough time to try ambitious tasks. The launch tactic is not just a demo; it creates a broad trial window for real work.
@thsottiauxX 原文6,244 ❤ · 471 RT · 840 💬
01 / 15
产品反馈 · OpenAI Work02 / 15
Codex 是新 Work 产品的核心，而不是被替代Codex Stays Core
Sam Altman 回应社区担忧：Codex 是 OpenAI 新 work product 的核心，正是它让产品好用，Codex 不会消失。这句话给出清晰方向：ChatGPT Work 不是把 coding agent 藏起来，而是把 Codex 能力扩展到更宽的白领工作流。
Sam Altman said Codex is the core of OpenAI’s new work product and is not going anywhere. ChatGPT Work is less about hiding the coding agent and more about expanding Codex-like agency into broader white-collar workflows.
@samaX 原文2,972 ❤ · 147 RT · 325 💬
Peter Yang：ChatGPT Work、Codex、Sol/Terra/Luna 的命名会让普通用户困惑Naming And UX Confusion
Peter Yang 给出高质量发布反馈：ChatGPT Work vs Codex 的分割容易让用户困惑，Sol/Terra/Luna 与 effort 设置也缺少清晰引导。他认为 OpenAI 最终应把体验统一成 ChatGPT / Codex，而不是让普通用户在模型、任务、聊天、插件之间来回判断。
Peter Yang praised the launch but pointed out confusing splits between ChatGPT Work and Codex, plus unclear choices across Sol, Terra, Luna, and effort settings. The product challenge is making agency legible to normal users.
@petergyangX 原文104 ❤ · 18 💬
02 / 15
企业智能 · Enterprise AI03 / 15
Box：Sol 在复杂企业文档任务上显著提升Complex Work Eval
Aaron Levie 透露 Box AI Complex Work eval 的结果：Sol 相比 GPT-5.5 在金融、医疗、公共部门、生命科学等数据密集型任务上显著提高。亮点不是通用聊天，而是模型能否从企业文档定义出发，避免早期错误假设，并把推理贯穿到真实决策数字。
Aaron Levie reported Box AI Complex Work eval gains for Sol over GPT-5.5 across finance, healthcare, public sector, and life sciences. The key is not chat polish, but document-grounded reasoning that avoids early wrong assumptions in real decisions.
@levieX 原文143 ❤ · 21 RT · 22 💬
当行业智能变得共享，公司如何重新差异化？Differentiation With Shared Intelligence
Levie 进一步提出未来十年的核心问题：如果 AI 学会了法律、金融、医疗等每个行业最好的数据，企业如何竞争？他的答案偏向复合循环：模型智能、公司自有数据、工作流连接、员工交互方式一起累积，使用得最好的公司仍会扩大优势。
Levie argues that if intelligence becomes broadly available inside every industry, differentiation will come from a reinforcing loop between models, proprietary data, workflow integration, and how employees interact with the system.
@levieX 原文144 ❤ · 20 RT · 19 💬
03 / 15
竞品闭环 · Gemini04 / 15
1,400+ Replies
Gemini 团队把用户痛点公开 stack rank：Workspace 集成、工具调用、项目组织、MCP/Skills、Deep Research、移动端 bug。Gemini turned public complaints into a ranked product backlog.
Josh Woodward：把 1,400+ 回复变成公开产品 backlogOpen Product Diagnosis
Josh Woodward 对前一天 Gemini 反馈帖做出完整回应，列出 Top 10 用户请求，并逐项说明状态。最重要的问题是 Workspace 集成可靠性、工具调用、项目/文件夹组织、MCP 与 Custom Skills 扩展、Deep Research 导出与模式切换。Google 正把模型竞争转化为产品体验修补速度的竞争。
Josh Woodward converted more than 1,400 replies into a public Top 10 backlog: Workspace reliability, tool calling, project organization, MCPs and custom skills, Deep Research flows, mobile bugs, and more. Model competition is becoming a product repair race.
@joshwoodwardX 原文1,058 ❤ · 65 RT · 153 💬
04 / 15
模型市场 · Market05 / 15
Amjad：LLM 市场正在快速摆脱单寡头叙事No Easy Monopoly
Amjad Masad 认为短短 6 个月内 LLM 市场已变得高度动态。Anthropic 会继续做出强模型，但其他公司和新进入者也会如此。模型层的竞争比许多投资者想象得更开放。
Amjad Masad argues the LLM market has become dynamic quickly. Anthropic will keep making great models, but so will others and new entrants. The model layer is not settling into a simple monopoly.
@amasadX 原文427 ❤ · 19 RT · 44 💬
Rauch：这是模型发布周，token 市占会被重排Model Release Week
Guillermo Rauch 认为 Meta Spark 1.1、Grok 4.5、GLM 5.2 会显著改变 token 市场份额。多数 agentic tasks 需要足够高智能和高速响应，模型网关与路由层因此变得战略性。
Guillermo Rauch says Meta Spark 1.1, Grok 4.5, and GLM 5.2 may reshape token market share. Agentic tasks need intelligence plus speed, making gateways and routing strategically important.
@rauchgX 原文409 ❤ · 8 RT · 44 💬
开放模型会变得异常快Open Models Get Fast
Rauch 的补充判断很短但重要：open models 即将变得“exorbitantly fast”。当开放模型速度提升，闭源旗舰的差异化压力会进一步转向工具链、上下文、数据与工作流。
Rauch adds that open models are about to get extremely fast. As open models improve on speed, closed frontier models will face more pressure to differentiate through tooling, context, data, and workflows.
@rauchgX 原文40 ❤ · 4 💬
Meta Muse Spark 1.1 在 OpenClaw 上表现亮眼Meta Model Signal
Garry Tan 表示 Meta Muse Spark 1.1 在 OpenClaw 上“真的很好”。这类来自真实产品/工具链的短反馈，往往比基准分更能说明模型是否进入开发者实际工作流。
Garry Tan says Meta Muse Spark 1.1 performed very well on OpenClaw. Short product-grounded feedback can matter more than benchmark charts when models enter real workflows.
@garrytanX 原文164 ❤ · 11 RT · 18 💬
05 / 15
Agent 工程 · Runtime06 / 15
AI 让编码更灵活，基础设施反而要更确定Rigid Runtime Under Flexible Coding
Amjad Masad 观察到一个反直觉变化：AI 让 coding less rigid，但 runtime 与 infra 需要更 rigid。团队开始写 formal specs、建设更确定和更韧性的系统。移动越快，脚下地基越要牢。
Amjad Masad notes a counterintuitive shift: AI makes coding less rigid while runtimes and infrastructure must become more deterministic. The faster teams move, the more solid the ground beneath them must be.
@amasadX 原文171 ❤ · 8 RT · 23 💬
Agentic coding 的核心技能：减少未知数Reduce Unknowns
Thariq 提醒：agentic coding 的核心技能之一是减少未知数。这句话和 Amjad 的基础设施观察互相呼应：AI 可以写更多代码，但人类的关键工作变成定义边界、收敛不确定性、给 agent 一个可验证的工作面。
Thariq says one core skill of agentic coding is reducing unknowns. AI can write more code, but humans increasingly define boundaries, reduce uncertainty, and create verifiable surfaces for agents.
@trq212X 原文16 ❤ · 4 💬
06 / 15
深度播客 · Schmidhuber07 / 15
Unsupervised Learning
“真正的 AI 不只是屏幕背后的 AI；没有像身体那样的硬件，就没有完整 AGI。”“True AI is not just the AI behind the screen.”
Jürgen Schmidhuber：AGI 不只是屏幕里的智能Embodiment And Artificial Scientists
Unsupervised Learning 访谈 Jürgen Schmidhuber，讨论今天模型缺什么、为什么他认为真正 AGI 需要硬件与机器人身体、人工科学家如何推动下一阶段进步，以及为什么他既看好 AI 技术、又对模型公司商业回报更谨慎。
Unsupervised Learning interviewed Jürgen Schmidhuber on what today’s models lack, why true AGI needs embodied hardware, how artificial scientists might drive progress, and why he is optimistic about AI technology but skeptical about model-company economics.
PodcastYouTube 完整播客2026-07-09
07 / 15
播客理念 · Key Ideas08 / 15
身体不是外设，而是智能边界Embodiment Matters
Schmidhuber 的观点提醒我们，纯屏幕智能虽然已能通过许多语言任务，但机器人硬件仍远弱于人体。真正进入物理世界，智能的瓶颈会从 token 推理转向感知、动作、反馈和能耗。
Schmidhuber’s point is that screen-based intelligence can pass many language tests, but robotic hardware remains far behind human bodies. In the physical world, bottlenecks move from token reasoning to perception, action, feedback, and energy.
Podcast source
人工科学家比聊天机器人更关键Artificial Scientists
访谈强调的另一条线是 artificial scientists：能提出假设、设计实验、验证和改进自身方法的系统。相比更会聊天的模型，这类系统更接近推动科学与工程进步的杠杆。
A second thread is artificial scientists: systems that propose hypotheses, design experiments, verify results, and improve methods. This may be a more important lever than models that merely chat better.
Podcast source
CapEx 热潮不等于公司护城河Technology vs. Company Returns
Schmidhuber 对 AI 技术乐观，但对模型公司更谨慎：算力投资可能过热，递归自我改进也未必自动成为公司护城河。技术进步与商业利润不会永远同向。
Schmidhuber is optimistic about AI technology but more cautious about model-company returns. CapEx may be overdone, and recursive self-improvement may not automatically become a corporate moat.
Podcast source
安全争论需要和能力边界一起更新Safety And Capability Boundaries
他对 AI safety 的担忧低于许多同行，这不代表安全问题消失，而是提醒讨论必须跟随能力、硬件、行动空间和经济激励的真实边界，而不是只在抽象恐惧中循环。
His lower level of AI safety concern does not make safety irrelevant. It suggests the debate should track actual boundaries in capability, hardware, action space, and incentives.
Podcast source
08 / 15
快讯速览 · Briefs09 / 15
Fable 回归成为编码模型焦点
Thariq 和 Alex Albert 都转发“More Fable”，显示 Fable 在开发者圈重新获得注意力。
Thariq and Alex Albert both amplified “More Fable,” signaling renewed developer attention around Fable.
@trq212 原文6,264 ❤
Meta 人才继续流入 AI 产品
Madhu Guru 加入 Meta 构建 AI products，认为 SWE agents 已改变软件工程，但多数复杂系统里的 agents 仍很早期。
Madhu Guru joined Meta to build AI products, arguing that agents outside software engineering are still early.
@realmadhuguru 原文465 ❤
AEO 开始影响模型默认推荐
Swyx 发现前沿模型反复推荐 Resend，即使已有交易邮件设施，AEO 可能成为新分发入口。
Swyx notes frontier models repeatedly suggest Resend, hinting that AEO may become a new distribution channel.
@swyx 原文97 ❤
AI 发布视频的形式泡沫被吐槽
Nan Yu 质疑“融资金额+对屏幕聊天”的 flashy videos 对业务有什么帮助。
Nan Yu questions flashy fundraise videos with big dollar graphics and screen-talking formats.
@thenanyu 原文79 ❤
模型发布周已经变成社区段子
Nikunj 用长段子总结 GPT-5.6、Grok 4.5、Fable、Sonnet、LongCat 等密集发布，说明市场节奏已超出正常人类跟踪能力。
Nikunj’s comic summary of the release week shows how model velocity has exceeded normal human tracking capacity.
@nikunj 原文94 ❤
空间创业与硬科技仍在 SPC 视野内
Aditya Agarwal 访谈印度宇航员 Gagan，讨论 ISS、微重力、新空间经济和创业机会。
Aditya Agarwal interviewed astronaut Gagan on the ISS, microgravity, the new space economy, and entrepreneurship.
@adityaag 原文31 ❤
09 / 15
数据洞察 · Data10 / 15
今日数据概览Today Stats
19 位活跃 Builder
44 条推文收录
1 期深度播客
0 篇新博客
9,371 最高赞：@sama GPT-5.6 Sol 成本论述
840 最高回复：ChatGPT Work / Codex 限额重置
19 builders, 44 tweets, 1 podcast, 0 blog posts. Top engagement came from Sam Altman framing GPT-5.6 Sol around enterprise dollars per task and Thibault Sottiaux resetting ChatGPT Work/Codex limits.
follow-buildersfeed generated 07:27Z
5 条关键洞察5 Key Takeaways
成本成为旗舰卖点：GPT-5.6 Sol 直接围绕 dollars-per-task 讲企业价值。
Cost per task is now a first-class launch message.
发布即试用窗口：限额重置是增长策略，也是产品信心测试。
Rate-limit resets turn launch attention into real trials.
企业 eval 走向真实文档：Box 的复杂工作测试比通用 benchmark 更贴近采购问题。
Enterprise evals are moving toward real document workflows.
模型市场多极化：Meta、xAI、GLM、开放模型都在争夺 agent workload。
Model competition is becoming multipolar.
Agent 越灵活，系统越要确定：规范、runtime、验证面成为新的工程核心。
Flexible agents need deterministic infrastructure.
10 / 15
本周之声 1
我们听到了企业对 AI 成本的担忧，5.6 Sol 是 dollars-per-task 的巨大进步。
“We have heard enterprises on their concerns about AI costs, and 5.6 Sol is a huge step forward for dollars-per-task.”
@sama · 原文链接
11 / 15
本周之声 2
感谢前 12 小时 1,400+ 条回复。我昨晚读完了所有回复。
“Thanks to the 1,400+ replies in the first 12 hours! I read all of them last night.”
@joshwoodward · 原文链接
12 / 15
本周之声 3
你移动得越快，脚下的地基就越要牢固。
“The faster you want to move, the more solid the ground beneath you has to be.”
@amasad · 原文链接
13 / 15
AI前沿每日脉动
AI Frontier Pulse · 2026.07.10
GPT-5.6 Sol 把“成本/任务”推到舞台中央；Gemini 把用户反馈变成公开路线图；企业 AI 的竞争开始落到真实文档、真实工作流和真实基础设施上。
GPT-5.6 Sol puts cost per task at center stage; Gemini turns feedback into public roadmap; enterprise AI competition moves into real documents, workflows, and infrastructure.
Richard Liu · AI前沿每日脉动 · 2026
14 / 15