X（推特）上的财经、科技与宏观账号
我被推荐，和看到太多聊AI风险的，说实话，这些风险我都认同，但是少有把风险和机遇放一起分析最后做出结论的，只聊风险并不会显得自己有多聪明，对未来需求的判断才是最难的，起码华尔街那些人23年GPT到现在，预测英伟达的一年后的盈利预期从来没有对过。 https://t.co/b9Nm961XwB 引用 @TJ_Research：美股的泡沫分为估值泡沫和盈利泡沫 估值现在根本没有泡沫，无论静态还是动态。 盈利泡沫需要向前看，在你听到别人聊泡沫时，记得多问一句，你看到未来的盈利风险在哪里？对建立自己的观点会有帮助的
过去七个交易日，美股因AI变现与资本开支回报受质疑、通胀和油价担忧升温而集中抛售。OpenAI年化营收口径差、Anthropic营收增速放缓及AI基础设施公司融资压力，加上CPI、PPI和布伦特原油走高，加重市场对利率与现金流的担忧。作者认为这属于高速增长叙事下的市场波动，后续关注云巨头财报能否显示收入增长和现金回报。
Wrote this 2.5 years ago. Think it's still mostly right. I hope so. FWIW my personal usage of code peaked around 2017 when I first became deeply informed about Machine Learning and AI. But imagining and adapting to the evolving forest has never required expertise in the trees 引用 @dampedspring：The
Already running a model endpoint? What does it take to scale it on Kubernetes? In five minutes, we walk through how NVIDIA Dynamo handles deployment with scalable workers and an OpenAI-compatible endpoint. Watch below: https://t.co/S8uc5ljUIW
JPM’s chart puts global AI compute demand at ~450 GW in 2030 and ~530 GW in 2031. Translating that into capital and revenue shows aggressive these assumptions seem to be. At an illustrative $50B per fully equipped GW (low-end by the way as costs are moving up), supplying 530 GW 引用 @pequityresearch
"In the past year, Devin has gone from writing 13% of Cognition’s production code to more than 90%." This statement is particularly notable given that many consider Cognition as having one of the highest talent density with regards to AI talent and software engineering. 引用 @pmarca：https://t.co/lhP
based. now this is the type of respectful pushback & fact-based discourse we need. no horse in the race but rooted in fundamentals and facts. presenting the sources. Rittenhouse doing a lot to raise the level of discourse on this topic here and elsewhere on this site. 引用 @RHouseResearch：@daniel_k
Thought it'd be helpful to frame this in economic output terms relative to history: Anthropic staff's survey answers imply US real GDP growth of ~18.6% in the 12 months to January 2030 (17.1% in the model's log terms). In the BEA annual record back to 1929, comparable growth https://t.co/j1RJcEaOKC
Fair distinction on the terminology and it's not fair for me to blend them. My broader point is that the investment case still assumes today’s revenue persists and grows so you're still making the same implicit bet. In reality, subscribers can churn, API customers can reduce 引用 @ramez：@ShanuMat
大多数关注 AI 交易的投资者没有意识到，货币和财政政策一直是支撑这笔交易的一大底层驱动力。下一轮政策刺激（policy pump）将难得多。本周在博客 App 上免费阅读： https://t.co/LVoRhG3Yzp 【被引推文】下一轮政策刺激（The Next Policy Pump） 过去几年，市场一直受到财政和货币双重宽松的支撑，但这些利好将会逆转，这意味着任何新的刺激都将需要远为非常规的手段。 https://t.co/o3oq3TJ8lq 原文： Most investors focused on the AI trade don't get that monetary
Coinbase使用内部专有数据微调开源模型Qwen3.5-9B，用于欺诈预防；其精度、召回率、F1分数和美元加权召回率均显著优于未微调版本及Opus 4.5等模型。微调后模型的P50推理延迟降低55%；Coinbase还开发了便于员工微调模型的内部工具，并预计未来可取代部分后端服务。
这个很好用。截图信用卡是最简单的方式，除非你想填写安全保险库数据。 把信用卡交给餐厅不认识的人好像没问题，却不愿交给 Grok Bot，显得有点可笑。 https://t.co/VJIxxMSXP9 【被引推文】我儿子爱篮球，我想让他能在家练习……但我们住得小，不想球不断弹地吵到邻居。 于是我给 Grok Bot 发了一条短信：“帮我订一个 DribbleUp 静音球，最好价格，寄到我家。” 它比了价、找到最优惠、用我的卡结账下单——含免运费共32.91美元，周四到。 我没打开购物App，也没填任何表单。 我只是想给儿子更安静的练习方式，Bot 就搞定了购物。 有自己的AI机器人
这真的管用！ Grok @Bot 可以自负盈亏。 https://t.co/GEZskbNU9h 【被引推文】我让我的 Grok Bot 去想办法自己挣钱。 我爱这个应用 https://t.co/g9SysaLINp 原文： This really works! Grok @Bot can pay for itself. https://t.co/GEZskbNU9h 引用 @mikepat711：I told my Grok Bot to go figure out how to pay for itself. I love this app https://t.co/g
Can anyone see a world in which we need five ( or more if you include non US) frontier models? I don't get how all can win. But not my lane. Like why would we need 5 roughly identical generalized models who leapfrog each other in capabilities instead of 2 or 3?
聊聊CRWV的真实压力以及对能都扛过此次危机？一周前觉得高Beta容易📉的其中一个原因是市场对债务的担心：博通给anthropic和openai担保了近千亿美金，spacex计划借债500亿美金买芯片。当时Crwv的cds暴涨也有了苗头 这周完全成了市场最担心的话题也是大跌的导火索之一。 引用 @qinbafrank：再看上周五推文聊的几个思路，这两天：1）伊朗袭击油轮、胡塞沙特对轰 ，油价不稳住市场也没稳住； https://t.co/5aoaaS2gMx
Building safer AI agents takes more than model safeguards. @perplexity_ai open sourced SPACE, the agent sandbox platform behind Computer, giving developers more control over where agents run code and what they can access. Cheers to more teams open-sourcing tools to build safer 引用 @perplexity_ai：W
Databricks在一则产品财务总监招聘中称，AI是其业务中增长最快的部分，职位负责AI产品与GPU投资策略。Databricks称，8月收入年化运行率超过70亿美元、同比增长超过80%；AI产品2月收入年化运行率已超过14亿美元。职位职责包括规划生产环境推理的GPU需求、在超大规模云厂商和新云之间采购算力，并管理利用率及每token/每次查询成本；招聘未披露支出或算力规模。
据英国《金融时报》报道，Anthropic已向黑石开出高于10%的票息，以锚定博通安排的600亿美元债务融资方案中的次级层。黑石基金将投入90亿美元，参与180亿美元B类次级分层；银行牵头420亿美元A类优先担保分层。
华尔街与旧金山科技圈对 AI 产能是否过剩存在分歧，科技圈人士认为，考虑过度建设的可能性就意味着不够相信 AI。未来几个月，若 AI 采用加速或总体生产率证据显现，AI 交易可能大幅上涨；若采用速度或前沿实验室经济性令人失望，也存在下行风险。越来越多人预期未来 18 个月内重演互联网泡沫式崩盘。
美联储主席 Kevin Warsh 表示，AI 将通过改善预测从根本上改变货币政策的实施。他说，未来不久，分析通胀趋势时就不必再用老式模型“折磨”数据。
Bubble Talk: Tulips to Tech Rational Dissent Episode 013 Despite bubbly valuations, persistent market skepticism suggests the AI boom is expanding like 1996—not collapsing like 2000. -- Available on YT, Spotify, Apple Podcasts, Overcast, etc. Chapters 00:00 - Introduction & https://t.co/zP8B4ZOu6i
AI agents are moving from answers to actions. But how do you safely run what they create? On Azure Friday, @shanselman and Jan Kalis explore how Azure Container Apps Sandboxes isolate agentic AI workloads. Watch now: https://t.co/4enIdHGpUI https://t.co/8H5XRqLGXm
Reflection AI发布首个开放权重模型Beam，提供权重、tokenizer和参考实现，并以Apache 2.0许可证开放。该501B MoE模型使用NVIDIA GB300 NVL72训练，在Blackwell Ultra上开箱即用的推理速度超过70K tokens/sec/GPU；Beam 1.0提供BF16、FP8和NVFP4权重及DFlash2 drafter。
New from our STEM Fellows Program: Designing more efficient quantum circuits with Claude. Given plain-language instructions, Claude Mythos 5 designed quantum circuits that improve on the best published ones, sometimes by an order of magnitude or more. More efficient circuits
@alexandr_wang 和 @Muse 团队发布开放权重模型 Muse Spark Open。
Also, ask Grok @Bot to connect to your 𝕏 DMs. Best way to use it. 引用 @bot：Voice calls with Grok Bot are faster, more reliable, and better at remembering past conversations. https://t.co/v4vAZxNrf5
Grok @Bot got much faster 引用 @bot：Grok Bot is much faster. Computer tasks finish 3x sooner on average, and replies to everything else arrive in less than half the time.
Shanu Mathew转述黄仁勋在最新一期 BG2 访谈中的说法，英伟达每吉瓦的内容价值超过400亿美元，前沿实验室约占其业务的25%–30%。基线计算约占三分之一，年增长率约为30%–40%；访谈还涉及AI收入是否支撑资本开支、2027年AI瓶颈及开放与封闭等话题。
Dylan Patel 预测，OpenAI 与 Anthropic 到 2027 年 12 月的合计收入可能达到 6000亿至7000亿美元。Anthropic 今年收入预计增长约 10 倍，从低于 100 亿美元升至高于 1000 亿美元的收入运行率；他还称 Anthropic 可能在 2028—2029 年达到约 1 万亿美元收入。
最新@conferenceboard调查发现，近一半CEO预期未来三年AI将提升生产率并降低运营成本 https://t.co/yTaNSjJYs1 原文： Latest survey from @conferenceboard found that nearly half of CEOs expect AI to boost productivity and reduce operating costs over the next three years https://t.co/yTaNSjJYs1
Meta的开放权重AI模型Muse Spark Open已可在苹果Mac Studio上本地运行，开发者可借此开展端侧AI实验。文章称，这展示了苹果芯片与MLX运行前沿大模型的能力，也为Meta扩大开放模型生态提供助力；但数万美元的硬件门槛使其目前更适合开发者、小团队或特定企业。
I've been pretty clear about my skepticism around land, power and shells, though the rest of the world waking up in a big way could change the picture. The question is whether it happens, and on what timeline. “TSMC is very conservative and doesn't freelance. They're sending 引用 @DratchCap：Zoom out
Gemini Argon 4 crashed on the very first task I gave it (which was largely reviewing this infrastructure model I've developed with my corpus of compute + infra economics info over time)... not off to a promising start.
亚马逊推出可穿戴设备 Bee Frontier 和磁吸手机背面设备 Bee Eclipse，试图将持续记忆与行动代理整合进 Alexa+ 和购物体系。文章认为，Bee 可延伸 Alexa 的个人上下文、捕捉购物需求，并成为亚马逊布局个人行动智能体的潜在催化剂。
很多人用AI，其实还停留在问问题、改文案、做总结这些事上。但其实可以让它更进一步，真正让它进市场做交易和别的角色协作。NeoSoul想做的就是这个，它从BNB Chain和0G生态出来，方向很明确：让AI…… 【被引】$DNA空投查询已上线…… 原文： 很多人用AI，其实还停留在问问题、改文案、做总结这些事上。但其实可以让它更进一步，真正让它进市场做交易和别的角色协作。NeoSoul想做的就是这个，它从BNB Chain和0G生态出来，方向很明确：让AI https://t.co/mQzhflJP9d 引用 @NeoSoulAI：$DNA airdrop checker is live
我们已经连续两份热通胀读数，可利率却在下滑。 整个市场都在…… 原文： We've had two hot inflation prints in a row, and yet rates have been sliding. The entire market is all about the AI trade, and the recent jitters about Anthropic's revenue numbers matter way more than some random government datapoint right now.
If the recent 8bln/mo ARR increases for OpenAI+Anthropic continue that'll get them to a combined 1tln ARR by the end of 2035.
昨天：AI泡沫破裂 今天：一看放量反弹，PPI砸不下去，线一画，可能要新高了😂😂😂
How do I use a Microsoft LLM? Where are they? I couldn't find an API on Openrouter. 引用 @mustafasuleyman：We now have state-of-the-art pre-trained base models in the 3T-parameter weight class, built entirely in-house at MAI. Post-training results coming soon -- stay tuned! Lots of details in our te
Figuring out how to prevent AI models from doing “rogue” things is obviously good. But I also agree with Alvaro here. We don’t know how the future will unfold, but extreme concentrated power per se is worrisome and it’s very plausible that creating strong AI will lead to that. 引用 @GroundhogStrat：So