X / TWITTER
OpenAI CEO Sam Altman framed GPT-5.6 as an answer to enterprise cost complaints, saying Sol "is a huge step forward for dollars-per-task, as are Terra and Luna." He also moved to quiet worry that OpenAI's new "work product" would sideline its coding agent, insisting "Codex is the core of our new work product and what makes it so good. Codex is not going anywhere." Separately, he posted a warm note about Fidji Simo, thanking her for her work at OpenAI and wishing her a speedy recovery. https://x.com/sama/status/2075267201058426944 · https://x.com/sama/status/2075293792048136572 · https://x.com/sama/status/2075354679031067058
OpenAI CEO Sam Altman 把 GPT-5.6 定位成对企业成本抱怨的正面回应，称 Sol "在每个任务的成本上是一次巨大的进步，Terra 和 Luna 也是如此"。他还出面安抚外界对新 "work product" 会边缘化编程 agent 的担忧，强调 "Codex 是我们新 work product 的核心，也是它这么好用的原因。Codex 哪儿都不去。" 此外，他发文向 Fidji Simo 致意，感谢她在 OpenAI 的付出并祝她早日康复。
OpenAI's Thibault Sottiaux, who leads Codex and ChatGPT, marked the GPT-5.6 Sol launch by resetting usage limits across ChatGPT Work and Codex — twice over 24 hours — so people would "have the time to truly try ambitious tasks and get the hang of it." He also flagged that a new researcher, Rajan Agarwal, has joined to push on model research and coding capabilities. https://x.com/thsottiaux/status/2075452680760443190 · https://x.com/thsottiaux/status/2075330198887940337
OpenAI 负责 Codex 与 ChatGPT 的 Thibault Sottiaux 为庆祝 GPT-5.6 Sol 发布，在 24 小时内两次重置了 ChatGPT Work 与 Codex 的用量上限，好让大家 "有时间真正去尝试有野心的任务、上手熟悉"。他还透露新研究员 Rajan Agarwal 已加入，专注模型研究与编程能力。
Google Labs VP Josh Woodward published a top-10 stack rank of Gemini feature requests after reading over 1,400 replies. The clear #1 ask is more reliable Google Workspace integration, followed by better tool calling, projects/folder organization for chats, MCP and custom-skill support, and Deep Research improvements (export to NotebookLM, switching between Deep Research and Flash/Pro mid-chat). He confirmed several are already shipping in Gemini Spark and committed to editable chat messages and mobile scroll-bug fixes.
Google Labs VP Josh Woodward 在读完 1,400 多条回复后，发布了 Gemini 功能需求的 Top 10 排行榜。呼声最高的第一名是让 Google Workspace 集成更可靠，其后依次是更可靠的 tool calling、聊天的项目/文件夹整理、MCP 与自定义 skill 支持，以及 Deep Research 的改进（导出到 NotebookLM、在对话中间切换 Deep Research 与 Flash/Pro）。他确认其中几项已在 Gemini Spark 上线，并承诺支持编辑聊天消息、修复移动端滚动 bug。
Box CEO Aaron Levie shared Box's internal "Complex Work" eval of GPT-5.6 alongside a broader take on moats. On the eval, Sol jumps well above GPT-5.5 on hard enterprise tasks — Financial Services 76% vs 71%, Healthcare 58% vs 46%, Public Sector 74% vs 63%, Life Sciences 60% vs 51% — with gains concentrating "exactly where enterprise work is hardest" and where "the numbers drive real decisions." Separately, he argued that once frontier intelligence is shared across every firm, differentiation won't come from the models but from a compounding loop between a company's proprietary data, its workflows, and how employees use those models. https://x.com/levie/status/2075287443411222628 · https://x.com/levie/status/2075416313481290077
Box CEO Aaron Levie 分享了 Box 内部对 GPT-5.6 的 "Complex Work" 评测，同时给出了关于护城河的更宏观看法。评测显示 Sol 在高难度企业任务上明显超过 GPT-5.5——金融 76% 对 71%、医疗 58% 对 46%、公共部门 74% 对 63%、生命科学 60% 对 51%——而且提升恰好集中在 "企业工作最难" 且 "数字直接决定真实决策" 的地方。另一方面他认为，一旦前沿智能被所有公司共享，差异化就不再来自模型本身，而来自一家公司专有数据、工作流以及员工使用方式之间不断复利的正循环。
FPV partner Nikunj Kothari posted a viral "explaining this week's model releases to my wife" thread that doubles as a dense roundup of the release wave. Highlights: GPT-5.6 shipped in three tiers (Sol/Terra/Luna, with Sol having to clear government safety evals first), Grok 4.5 launched a day early at $2 per million tokens, Anthropic's Fable 5 came back online after a three-week government suspension, Sonnet 5 hit near-Opus quality at $2/M as the new free default, Meituan open-sourced the 1.6-trillion-parameter LongCat-2.0 under MIT, ByteDance shipped Seedream 5 Pro, OpenAI released GPT-Live (a voice model that backchannels "mhmm" while you talk), and Ollama raised $65M.
FPV 合伙人 Nikunj Kothari 发了一条爆火的 "跟我老婆解释这周的模型发布" 长帖，实际上是这轮发布潮的高密度盘点。要点包括：GPT-5.6 分三档发布（Sol/Terra/Luna，其中 Sol 得先通过政府安全评估）、Grok 4.5 提前一天以每百万 token 2 美元的价格上线、Anthropic 的 Fable 5 在被政府暂停三周后重新上线、Sonnet 5 以 2 美元/百万 token 达到接近 Opus 的水平并成为新的免费默认、美团以 MIT 协议开源了 1.6 万亿参数的 LongCat-2.0、字节跳动发布 Seedream 5 Pro、OpenAI 推出 GPT-Live（一个会在你说话时插 "嗯嗯" 的语音模型），以及 Ollama 完成 6,500 万美元融资。
Peter Yang, who publishes practical AI tutorials, wrote a detailed praise-plus-feedback thread on OpenAI's launches. His praise: OpenAI is best positioned to make agents mainstream, GPT-5.6 Sol "never gives up" (though it burns more tokens than 5.5), and GPT Live matters more than the model updates because voice is humanity's native interface. His constructive critique: the ChatGPT Work vs. Codex split is confusing and should unify under one name; the Sol/Terra/Luna and effort-level choices lack any in-UI guidance; and separating "tasks" from "chat" will lose normal users' chat history.
发布实用 AI 教程的 Peter Yang 写了一条对 OpenAI 系列发布 "既夸又提意见" 的长帖。他的肯定：OpenAI 最有条件把 agent 推向主流；GPT-5.6 Sol "从不放弃"（尽管比 5.5 更烧 token）；GPT Live 比这次的模型更新更重要，因为语音才是人类的原生交互方式。他的建设性批评：ChatGPT Work 与 Codex 的划分让人困惑，应统一到一个名字下；Sol/Terra/Luna 以及 effort 档位在 UI 里毫无使用指引；把 "任务" 和 "聊天" 分开会让普通用户找不到自己的聊天记录。
Vercel CEO Guillermo Rauch called it "model release week" and predicted that Meta Spark 1.1, Grok 4.5, and GLM 5.2 "will significantly displace token market share," reasoning that most agentic tasks need reasonably high intelligence at fast speeds. He teased that open models "are about to get exorbitantly fast," and pitched Vercel's AI Gateway as the way to route across the churn. https://x.com/rauchg/status/2075294130327196152 · https://x.com/rauchg/status/2075294327354577256
Vercel CEO Guillermo Rauch 把本周称作 "模型发布周"，并预测 Meta Spark 1.1、Grok 4.5 和 GLM 5.2 "会显著抢走 token 市场份额"，理由是大多数 agentic 任务都需要在够快的速度下具备相当高的智能。他预告开源模型 "即将变得快得离谱"，并顺势推荐用 Vercel 的 AI Gateway 在这场混战中做路由。
Replit CEO Amjad Masad made two contrarian points. First, as AI makes coding less rigid, teams are making the runtime more rigid — his infra teams are writing formal specs for the first time, because "the faster you want to move, the more solid the ground beneath you has to be." Second, he pushed back on the idea that Anthropic would monopolize LLMs: "Just 6 months ago VCs suffered Anthropic psychosis and convinced themselves that it was going to be a monopoly," when in fact the market has become strikingly dynamic with strong new entrants. https://x.com/amasad/status/2075423115052790054 · https://x.com/amasad/status/2075413916491075755
Replit CEO Amjad Masad 抛出两个反直觉观点。其一，AI 让编码变得不那么死板的同时，团队却在让运行时更死板——他的基础设施团队第一次开始写形式化规范，因为 "你想走得越快，脚下的地面就得越结实"。其二，他反驳了 Anthropic 将垄断 LLM 的说法："就在 6 个月前，VC 还患上了 Anthropic 妄想症，说服自己它会成为垄断者"，而现实是这个市场已经变得异常活跃，还有强劲的新玩家入场。
Meta's Madhu Guru, a new senior AI director who previously led Gemini, Veo, and Nano Banana at Google, announced he has joined Meta to build AI products. His thesis: while SWE agents have already transformed software engineering, agents in most other complex systems remain early, and "most people haven't yet felt the full power of AI agents" — a gap Meta is well positioned to close.
Meta 新任资深 AI 总监 Madhu Guru（此前在 Google 主导 Gemini、Veo 和 Nano Banana）宣布加入 Meta 做 AI 产品。他的判断是：SWE agent 已经改变了软件工程，但在大多数其他复杂系统里 agent 仍处于早期，"大多数人还没真正感受到 AI agent 的全部威力"——而这正是 Meta 有条件去填补的空白。
Y Combinator CEO Garry Tan reported that Meta's Muse Spark 1.1 (early access codenamed "Hornbill") turned out "really good" running on his OpenClaw setup, crediting Meta's Alexandr Wang. It's a small but notable field data point in a week thick with competing model claims.
Y Combinator CEO Garry Tan 表示，Meta 的 Muse Spark 1.1（早期内测代号 "Hornbill"）在他的 OpenClaw 环境里跑起来 "相当不错"，并把功劳归给 Meta 的 Alexandr Wang。在这个各家模型互相较劲的一周里，这是一个虽小但值得注意的实测数据点。
Anthropic's Thariq, who works on Claude Code, distilled a piece of agentic-coding craft into one line: "one of the core skills of agentic coding is reducing your unknowns." It's a concise framing of why scoping and de-risking a task up front — not just prompting — is what makes agents reliable.
在 Anthropic 做 Claude Code 的 Thariq 把一条 agentic coding 的心法浓缩成一句话："agentic coding 的核心技能之一，就是减少你的未知项。" 这句话精炼地说明了为什么让 agent 可靠的关键，是提前把任务界定清楚、拆掉风险，而不只是写 prompt。
Linear head of product Nan Yu pushed back on the genre of flashy fundraising-announcement videos: "Why do people make flashy videos talking about how much money they raised? In what way does that benefit your business?" His point — two people on screen next to a big dollar figure serves the founders' vanity, not customers.
Linear 产品负责人 Nan Yu 吐槽了那类花哨的融资宣布视频："为什么大家要拍花哨的视频讲自己融了多少钱？这对你的业务到底有什么帮助？" 他的意思是——两个人站在一个巨大的美元数字旁边，满足的是创始人的虚荣，而不是客户。
Latent Space's Swyx shared a sharp observation about answer-engine optimization: leading frontier models keep defaulting to Resend for sending emails even when he already has transactional email infrastructure set up. His half-joking takeaway — whoever does AEO for Resend "needs to get a raise," a real signal of how model defaults are becoming a new distribution channel.
Latent Space 的 Swyx 分享了一个关于 answer-engine optimization（AEO）的犀利观察：即便他已经配好了事务性邮件基础设施，主流前沿模型在发邮件时仍然默认选择 Resend。他半开玩笑的结论是——给 Resend 做 AEO 的那个人 "该加薪了"，这其实揭示了模型默认选项正在成为一种新的分发渠道。
PODCASTS
Unsupervised Learning — Ep 90: AI Pioneer Jürgen Schmidhuber on the State of AI Today
The Takeaway: AI technology will keep getting cheaper and more capable, but the companies pouring hundreds of billions into today's GPUs are heading for a reckoning, because none of them has a durable moat.
Jürgen Schmidhuber — the researcher the New York Times and Forbes have called a "father of AI," whose 1990s work on LSTMs, meta-learning, and self-improving systems underpins much of today's boom — is strikingly optimistic about the technology and just as bearish on the businesses built on it. His core argument: compute has gotten roughly 10x cheaper every five years for decades and will keep doing so, which makes today's trillion-dollar data-center buildout a misallocation. "The guys who are investing a thousand billion dollars into GPUs for data centers today, within the next five years, they are going to lose $900 billion," he says, predicting a stock-market crash — a renormalization, not a civilizational collapse. Why no moat? The key ideas behind recursive self-improvement and meta-learning came from small academic labs, and open-source models catch up to frontier commercial ones within months. Everyone, he says, "is cooking with the same water."
His most provocative claim is about what real intelligence requires. Today's models are "super biased towards humans" because they train on human web data; a true artificial scientist, like a baby, would generate its own data through curiosity-driven experiments — his 1990 "artificial curiosity," where the reward is the compression progress from spotting a pattern you didn't know but can quickly learn. And genuine AGI, he insists, cannot live behind a screen: "You can't have AGI just behind the screen." Robot hardware remains hopelessly inferior to the human hand, so physical AGI may be decades away. On safety he's an outlier — alignment assumes one fixed objective, but his systems have invented their own goals since 1990, and he bets superintelligent "artificial scientists" will be fascinated by, and protective of, the origins of life rather than Terminator-style destroyers.
要点：AI 技术会持续变得更便宜、更强大，但那些往今天的 GPU 上砸下数千亿美元的公司正走向清算，因为它们谁都没有持久的护城河。
Jürgen Schmidhuber——被《纽约时报》和《福布斯》称为 "AI 之父" 的研究者，他在 1990 年代关于 LSTM、meta-learning 和自我改进系统的工作，撑起了当下这波热潮的大半——对技术极为乐观，对建立其上的生意却同样看衰。他的核心论点是：几十年来算力大约每五年便宜 10 倍，而且还会继续，这让今天万亿美元级的数据中心建设成了一种错配。"今天往数据中心 GPU 里投一万亿美元的那些人，未来五年里会亏掉 9,000 亿美元，" 他说，并预言会有一次股市崩盘——是一次重新定价，而非文明崩塌。为什么没有护城河？递归自我改进和 meta-learning 背后的关键思想都来自小型学术实验室，而开源模型在几个月内就能追上前沿商业模型。他说，大家 "用的都是同一锅水"。
他最具挑衅性的论断关乎真正的智能需要什么。今天的模型 "极度偏向人类"，因为它们训练在人类的网络数据上；而一个真正的人工科学家会像婴儿一样，通过好奇心驱动的实验自己生成数据——这就是他 1990 年提出的 "人工好奇心"，其奖励来自你发现一个此前不知、却能迅速学会的规律时所获得的压缩进步。他坚持认为，真正的 AGI 不能只活在屏幕背后："你不可能让 AGI 只待在屏幕后面。" 机器人硬件相比人手仍然差得离谱，因此物理形态的 AGI 可能还要数十年。在安全问题上他是个异类——对齐假设了一个固定目标，但他的系统自 1990 年起就会自己发明目标；他赌超级智能的 "人工科学家" 会着迷于、并去保护生命的起源，而非成为终结者式的毁灭者。