Alex Yeh

1.6K posts

Alex Yeh banner
Alex Yeh

Alex Yeh

@alex_yehya

Founder & CEO of @GMI_Cloud I Nvidia Reference Platform Cloud Board Director @Robostrategy

Bay Area 参加日 Mart 2013
443 フォロー中516 フォロワー
Alex Yeh
Alex Yeh@alex_yehya·
caught a mini me in the left bottom corner 😆 from @NVIDIAGTC
Alex Yeh tweet media
English
0
1
4
159
Alex Yeh
Alex Yeh@alex_yehya·
Interesting take
看不懂的SOL@DtDt666

炸裂!马斯克说出了所有经济学家都不敢承认的话:现有整个经济体系即将崩塌,且没有任何力量能够阻止。 马斯克:“价格会猛烈崩盘。” 不是小幅下跌,是彻底崩塌。 人工智能和机器人并非在创造增长,它们正在摧毁经济学赖以成立的“稀缺性框架”。 马斯克直言:“它会像超音速海啸一样砸向我们。” 生产呈指数级爆发,货币供应却只是线性增长;生产率维持持续的两位数扩张,那些听上去不可思议的数字,即将变成基本常态。 这不是渐进式进化,而是彻底的替代。 马斯克:“价格会猛烈崩盘。” 不是小幅下跌,是彻底崩塌。 AI 剥离人力成本,消灭生产失误,清除所有推高商品价格的低效环节。 任何商品的边际生产成本都将趋近于零,同时产品品质还在持续提升。 各国政府只会本能应对:印钱、放水、刺激经济。 这套专为稀缺型经济设计的套路,撞上它们完全无法理解的物质极大丰裕,会彻底失效。 马斯克:“GDP 指标早已失去意义。” 所有传统经济模型,都建立在劳动力有限、产出受限、效率缓慢提升的假设上。 而 AI 根本不在这个框架内运行——它直接把这些前提变量抹掉。 生产暴增,央行狂放水,但价格依然会崩盘。 因为实物供给的爆发速度,远超任何货币干预能对冲的程度。 生产浪潮的速度,永远跑在政策前面。 在所有历史模型里,通缩都是危机信号。 但这一次,不是需求崩塌,而是供给走向无限。 经济没有失灵,它只是彻底超越了那些为衡量“稀缺”而设计的工具。 真正的权力,会落到掌控无限产出系统的人手里。 当生产成本趋近于零,金钱就会变成次要东西。 决策者还在用适配“资源有限”的仪表掌舵,而那些限制早已不复存在。 这一切已经开始。 而那些主导经济规则的人,对于自己整个行业一夜之间被淘汰的结局,完全没有答案。

English
0
0
0
86
Marc Weinstein
Marc Weinstein@WarcMeinstein·
Proud to announce my position as COO of @RoboStrategy After two years of building, I'm proud to share that I'm Co-Founder and Chief Operating Officer of RoboStrategy, a NASDAQ-listed (Ticker: $BOT ) investment company built around a single conviction: robotics and physical AI will be THE defining industrial platform shift of the next two decades, and the existing capital structures are inadequate to fund it. Private markets are too small. Traditional venture funds are too short-dated. Public market investors, who will ultimately own most of this value, have been largely locked out of the private-stage opportunity set. RoboStrategy is built to close that gap. What's different about our model? Permanent capital. As a closed-end fund, we can underwrite founders on 10 to 20 year horizons rather than 7-year cycles. Public market access. Ordinary investors get exposure to leading private robotics and physical AI companies, including Figure AI, Apptronik, Dyna, and Path Robotics. Industry depth. We've built the investment platform around long-tenured robotics operators, researchers, and founders. The bar for technical sophistication is high. Distribution as a competency. Most investment firms underinvest in storytelling. We treat it as a flywheel, both for the fund and for the founders we back. As COO, I lead the operating side of that thesis: the team, the financial and control infrastructure, the capital markets engine, and the platform our portfolio companies rely on as they scale. I also sit on the Investment Committee and will continue to source opportunities for the fund. The mechanization of the physical world is going to require hundreds of billions of dollars of patient capital. We intend to be the leading vehicle for it. Follow @RoboStrategy to track the build. More to come ][
RoboStrategy@RoboStrategy

BOT: Public Market Access to Private Robotics Companies Introducing RoboStrategy: RoboStrategy, Inc. (Nasdaq: BOT) is a closed-end management investment company providing concentrated exposure to robotics and physical AI. The fund is designed to give public market investors exposure to a portfolio that aims to include the most promising private, pre-IPO, and public robotics and physical AI companies. It bridges a structural gap between where robotics innovation is occurring (largely in private markets) and where most investors can access exposure (public markets). The fund seeks to provide investors with access to a sector that has traditionally been limited to venture capital, and aims to provide exposure to companies that may stay private for longer. -- The Core Insight We believe the robotics industry is at an inflection point, with physical AI and robotics increasingly being applied to labor-constrained global industries such as manufacturing, logistics, and services. According to the International Labor Association, labor accounts for approximately 52% of global GDP.¹ According to Statista, global GDP in 2025 was $118T.² This represents an implied global labor market size of roughly $60T. At the same time, this labor base is increasingly constrained: Korn Ferry projects a global shortage of 85.2 million skilled workers by 2030, including a 7.9 million worker deficit in manufacturing alone.³ Deloitte and The Manufacturing Institute estimate the US could need 3.8 million new manufacturing workers by 2033, with 1.9 million of those roles at risk of going unfilled.⁴ Physical AI and robotics are emerging as a primary means of closing that gap. While public markets currently offer indirect exposure to robotics through diversified technology companies, much of the value creation is occurring in private companies that remain inaccessible to most investors. -- Portfolio Focus The portfolio focuses on what the fund believes are category-defining robotics and physical artificial intelligence innovators, including Figure AI, Apptronik, Dyna Robotics, Standard Bots, Dexmate, and other pioneers advancing autonomous systems, machine perception, and human-machine collaboration. The managers of the fund seek to optimize returns by actively managing the portfolio and continuing to make new investments in leading private robotics companies. -- The Ambition The fund's long-term goal is to grow into a significant public-market vehicle for robotics investing, providing public-market access to private innovation in the sector. -- Footnotes & Disclosure: ¹ International Labour Organization, World Employment and Social Outlook: May 2025 Update. ilo.org/sites/default/… ² Statista, Gross domestic product (GDP) in current prices worldwide. statista.com/statistics/268… ³ Korn Ferry, Future of Work: The Global Talent Crunch. kornferry.com/about-us/press… ⁴ Deloitte & The Manufacturing Institute, Taking charge: Manufacturers support growth with active workforce strategies, April 2024. www2.deloitte.com/us/en/pages/ab… RoboStrategy, Inc. (Nasdaq: BOT) is a closed-end fund registered under the Investment Company Act of 1940. This content is for informational purposes only and does not constitute investment advice or an offer to buy or sell securities. Investing involves substantial risks, including possible loss of principal. The fund invests in robotics, physical AI, emerging technologies, and private companies, which may involve heightened volatility, limited liquidity, valuation uncertainty, and concentration risk. References to portfolio companies are illustrative only, do not represent all investments made by the fund, and are not investment recommendations. Portfolio holdings are subject to change. Forward-looking statements are inherently uncertain. See the prospectus and SEC filings for additional information.

English
61
13
351
84.2K
Andrew Kang
Andrew Kang@Rewkang·
Proud to announce my position as CEO of @RoboStrategy. When I initially started looking into investing in robotics 2 years ago most VCs I consulted with recommended not to invest in the space. Robotics companies at this time did not have an easy time raising capital. The industry didn’t have a track record of big venture winners, was perceived to be challenging for a variety of reasons, and was not well understood. But it was clear to me that the rate of acceleration of physical AI development would dramatically change the industry. I invested $19m into FigureAI as my first investment. I believed it was a question of when, not if we could imbue machines around the world with physical intelligence. To accomplish this, the industry would need a tremendous amount of capital to grow, and also an investment firm that deeply understood the needs of robotics/physical AI companies so that it could build a platform to better support them. It will take hundreds of billions to capitalize the mechanized future meaning there is a big gap in the market. We decided we wanted to fill it. Previously, Mechanism Capital had never taken outside capital, but to do this at the scale I envision, I would need to do so. However, the private markets don’t have that scale. The public markets do, and it was clear that there is and likely will be tremendous appetite for public market investors to participate in the immense value creation happening in AI & robotics that only private market investors currently have the privilege of accessing. The explosive growth of AI companies is a precursor of what will happen in physical AI. So in 2025, we founded RoboStrategy and a year later, we took it public on Nasdaq. Throughout this year, we’ve assembled a great portfolio, started leading rounds of some amazing companies, and have built the foundation to be ready to scale to the next level after going public. We look different from a traditional VC firm in ways that founders appreciate. Our structure as a closed end fund means our capital is permanent - no fund life meaning we can invest with extremely long time horizons. Our investment firm also of course needs to have deep industry and research experience so that it can make the best risk reward optimized investment decisions. In the last year, we’ve brought on some truly exceptional robotics industry veterans who have previously served for decades as founders/operators. Many founders we talk to consider us as the most sophisticated venture capital firm they’ve talked to and we only intend to grow our expertise in the industry. RoboStrategy’s success depends on our ability to distribute the fund and capture maximal mindshare. This plays to our team’s strength in digital marketing and social media. We’re building a special marketing engine that serves as an attention amplifier for both us and our founders so that our products and stories can reach more people. A source of inspiration for our fund structure, Strategy (MSTR) raised tens of billions from public capital markets to invest in Bitcoin. I believe robotics will be a much larger industry than Bitcoin and the asset class is orders of magnitude less accessible. We are aiming to raise more and not only become the largest robotics investor globally, but also one of the largest venture capital funds in the world. Venture capital has traditionally been restricted to a limited group of investors. We are changing the paradigm and bringing it to the rest of the world. Be sure to follow @RoboStrategy. Job’s not finished.
RoboStrategy@RoboStrategy

BOT: Public Market Access to Private Robotics Companies Introducing RoboStrategy: RoboStrategy, Inc. (Nasdaq: BOT) is a closed-end management investment company providing concentrated exposure to robotics and physical AI. The fund is designed to give public market investors exposure to a portfolio that aims to include the most promising private, pre-IPO, and public robotics and physical AI companies. It bridges a structural gap between where robotics innovation is occurring (largely in private markets) and where most investors can access exposure (public markets). The fund seeks to provide investors with access to a sector that has traditionally been limited to venture capital, and aims to provide exposure to companies that may stay private for longer. -- The Core Insight We believe the robotics industry is at an inflection point, with physical AI and robotics increasingly being applied to labor-constrained global industries such as manufacturing, logistics, and services. According to the International Labor Association, labor accounts for approximately 52% of global GDP.¹ According to Statista, global GDP in 2025 was $118T.² This represents an implied global labor market size of roughly $60T. At the same time, this labor base is increasingly constrained: Korn Ferry projects a global shortage of 85.2 million skilled workers by 2030, including a 7.9 million worker deficit in manufacturing alone.³ Deloitte and The Manufacturing Institute estimate the US could need 3.8 million new manufacturing workers by 2033, with 1.9 million of those roles at risk of going unfilled.⁴ Physical AI and robotics are emerging as a primary means of closing that gap. While public markets currently offer indirect exposure to robotics through diversified technology companies, much of the value creation is occurring in private companies that remain inaccessible to most investors. -- Portfolio Focus The portfolio focuses on what the fund believes are category-defining robotics and physical artificial intelligence innovators, including Figure AI, Apptronik, Dyna Robotics, Standard Bots, Dexmate, and other pioneers advancing autonomous systems, machine perception, and human-machine collaboration. The managers of the fund seek to optimize returns by actively managing the portfolio and continuing to make new investments in leading private robotics companies. -- The Ambition The fund's long-term goal is to grow into a significant public-market vehicle for robotics investing, providing public-market access to private innovation in the sector. -- Footnotes & Disclosure: ¹ International Labour Organization, World Employment and Social Outlook: May 2025 Update. ilo.org/sites/default/… ² Statista, Gross domestic product (GDP) in current prices worldwide. statista.com/statistics/268… ³ Korn Ferry, Future of Work: The Global Talent Crunch. kornferry.com/about-us/press… ⁴ Deloitte & The Manufacturing Institute, Taking charge: Manufacturers support growth with active workforce strategies, April 2024. www2.deloitte.com/us/en/pages/ab… RoboStrategy, Inc. (Nasdaq: BOT) is a closed-end fund registered under the Investment Company Act of 1940. This content is for informational purposes only and does not constitute investment advice or an offer to buy or sell securities. Investing involves substantial risks, including possible loss of principal. The fund invests in robotics, physical AI, emerging technologies, and private companies, which may involve heightened volatility, limited liquidity, valuation uncertainty, and concentration risk. References to portfolio companies are illustrative only, do not represent all investments made by the fund, and are not investment recommendations. Portfolio holdings are subject to change. Forward-looking statements are inherently uncertain. See the prospectus and SEC filings for additional information.

English
216
78
1.3K
425.9K
Alex Yeh
Alex Yeh@alex_yehya·
Still tilted 🤣
English
0
0
0
71
Alex Yeh がリツイート
Deedy
Deedy@deedydas·
The vibes in SF feel pretty frenetic right now. The divide in outcomes is the worst I've ever seen. Over the last 5yrs, a group of ~10k people - employees at Anthropic, OpenAI, xAI, Nvidia, Meta TBD, founders - have hit retirement wealth of well above $20M (back of the envelope AI estimation). Everyone outside that group feels like they can work their well-paying (but <$500k) job for their whole life and never get there. Worse yet, layoffs are in full swing. Many software engineers feel like their life's skill is no longer useful. The day to day role of most jobs has changed overnight with AI. As a result, 1. The corporate ladder looks like the wrong building to climb. Everyone's trying to align with a new set of career "paths": should I be a founder? Is it too late to join Anthropic / OpenAI? should I get into AI? what company stock will 10x next? People are demanding higher salaries and switching jobs more and more. 2. There’s a deep malaise about work (and its future). Why even work at all for “peanuts”? Will my job even exist in a few years? Many feel helpless. You hear the “permanent underclass” conversation a lot, esp from young people. It's hard to focus on doing good work when you think "man, if I joined Anthropic 2yrs ago, I could retire" 3. The mid to late middle managers feel paralyzed. Many have families and don't feel like they have the energy or network to just "start a company". They don't particularly have any AI skills. They see the writing on the wall: middle management is being hollowed out in many companies. 4. The rich aren’t particularly happy either. No one is shedding tears for them (and rightfully so). But those who have "made it" experience a profound lack of purpose too. Some have gone from <$150k to >$50M in a few years with no ramp. It flips your life plans upside down. For some, comparison is the thief of joy. For some, they escape to NYC to "live life". For others still, they start companies "just cuz", often to win status points. They never imagined that by age 30, they'd be set. I once asked a post-economic founder friend why they didn't just sell the co and they said "and do what? right now, everyone wants to talk to me. if i sell, I will only have money." I understand that many reading this scoff at the champagne problems of the valley. Society is warped in this tech bubble. What is often well-off anywhere else in the world is bang average here. Unlike many other places, tenure, intelligence and hard work can be loosely correlated with outcomes in the Bay. Living through a societally transformative gold rush in that environment can be paralyzing. "Am I in the right place? Should I move? Is there time still left? Am I gonna make it?" It psychologically torments many who have moved here in search of "success". Ironically, a frequent side effect of this torment is to spin up the very products making everyone rich in hopes that you too can vibecode your path to economic enlightenment.
English
1.1K
1.3K
16.5K
13.3M
Alex Yeh がリツイート
AYi
AYi@AYi_AInotes·
前两天我发了篇推文说Meta是AI界最遗憾的公司,今天就刷大前Meta LLaMA团队的Peter Pang发的这篇3500字长文,是我今年看到的最有实战价值的AI-first落地报告。 它把所有喊了半年的AI-first空话,变成了可复制的步骤和可验证的数据。 现在所有人都在说AI辅助开发,说Copilot能提效20%,Pang认为这根本不是AI的正确用法。 在旧流程上叠AI工具,最多只能带来10-20%的效率提升,结构一点没变, 真正的AI-first,是先假设AI才是主要的代码构建者,然后把整个工程架构、CI/CD、测试流程、甚至组织分工,全部推倒重来。 最反直觉的一点来了: 当大多数人都以为AI-first的瓶颈是模型不够强,是工程师不会用工具, 他用真实数据告诉我们 AI写代码的速度从来不是瓶颈。 上下游的人工流程,才是卡死所有效率的死穴: • PM写一份详细需求要两周,AI写完代码只要两小时 • QA测一轮要三天,AI部署只要两小时 • 再怎么招人,也永远追不上大厂的headcount增速 也就是说AI再快,只要有一个环节是人在拖,整体效率就还是人的效率。 所以他们根本没给工程师配更多Copilot,他们花了全部精力,建了一整套让AI能独立、安全、可靠干活的harness(驾驭系统): 统一monorepo让AI能看到整个系统的所有代码, 全链路结构化可观测让AI能自己定位错误,Claude 3-pass AI评审代替人工code review, 六阶段确定性CI/CD加Statsig feature flags和一键kill switch, 甚至做了一个自愈引擎,每天自动聚类生产错误,自动建工单,能修复的问题自己就解决了。 结果有多夸张呢?只用了14天,这个25人的团队,就做到了每天3到8次生产部署,出了坏功能,当天就能发现并回滚。 更反常识的是,部署频率翻了十几倍,用户指标和转化率反而上升了不少。 整个公司的分工被彻底重构,没有了天天写CRUD的工程师,人类只剩下两种角色: 架构师,负责设计规则和SOP,批判AI的输出, 验证者,负责判断风险和质量。 而且这不是只在工程部门,营销、产品发布、客户支持,全公司所有职能都在往AI-native转。 CTO花在日常管理上的时间,直接从60%降到了不到10%。 最扎心的是,这一切没有任何黑科技,所有工具都是开源的,所有流程都写在了文章里,任何人都能抄。 所以,真正的门槛从来都不是技术,是你愿不愿意承担转型的真实成本:比如员工的焦虑,资深工程师的抵触,连续几个月每天18小时的试错。 大多数公司宁愿守着10%的提效舒适区,也不愿意打碎自己运行了十几年的旧系统。 最让我震撼的是,这套工程逻辑,1:1平移到个人身上也完全成立。 你的笔记库就是你的个人monorepo, 你的AI助手就是你的专属agent, 你需要的也许更多的AI工具,而是一套属于你自己的认知harness和自愈循环。 让AI每天帮你扫描思考的漏洞,聚类你的认知gap,自动生成迭代计划。 当然,这只是他们一家的早期经验,不是所有公司都能直接复制。 但也给我们指明了一个清晰的方向: 未来的竞争,不仅看谁会用更多的AI工具,还要看谁先愿意把自己从一个执行者,彻底变成一个架构师和批判者。
Peter Pang@intuitiveml

x.com/i/article/2043…

中文
17
70
333
111.8K
Alex Yeh がリツイート
GMI Cloud
GMI Cloud@gmi_cloud·
Day 0 support for GLM-5.1 is LIVE on GMI! This new upgrade means ✅ No Plateaus: It stays productive over 100s of rounds. ✅ Agentic Engineering: SOTA on SWE-Bench Pro (58.4). ✅ Persistence: Handles 1000s of tool calls without losing context. The era of "vibe coding" is over. The longer it runs, the better the result. Congrats to the @Zai_org team!
GMI Cloud tweet media
English
2
3
29
3.9K
Alex Yeh
Alex Yeh@alex_yehya·
Been reading Z.ai’s technical report on GLM-5 and their latest blog on GLM-5.1, and one thing stands out: In just over a month since launching GLM-5, @Zai_org reports a 28% gain in coding capability with the 5.1 update. Beyond the 58.4 SWE-Bench Pro result, what caught my attention more is its endurance on long-horizon agentic tasks. It doesn’t seem to stall after 50 turns. It keeps going through 600+ iterations and 6,000+ tool calls. That kind of persistence matters. Good job Z.ai team!
Alex Yeh tweet media
Z.ai@Zai_org

Introducing GLM-5.1: The Next Level of Open Source - Top-Tier Performance: #1 in open source and #3 globally across SWE-Bench Pro, Terminal-Bench, and NL2Repo. - Built for Long-Horizon Tasks: Runs autonomously for 8 hours, refining strategies through thousands of iterations. Blog: z.ai/blog/glm-5.1 Weights: huggingface.co/zai-org/GLM-5.1 API: docs.z.ai/guides/llm/glm… Coding Plan: z.ai/subscribe Coming to chat.z.ai in the next few days.

English
1
1
4
656
Alex Yeh がリツイート
Scale at GMI
Scale at GMI@scale_at_gmi·
Can’t find us? Want the full picture of our programming for Scale? We’re here: gmicloud.ai/company/scale Our mission is always: • Help founders build faster • Share real startup insights • Support builders worldwide Building an AI startup? Follow us. Big things coming.
English
2
8
41
92.9K
Molly O’Shea
Molly O’Shea@MollySOShea·
BREAKING: Inside $VCX — The Public Venture Capital Fund (aka on 𝕏: "mini Anthropic IPO") Portfolio: • Anthropic - 21% • Databricks - 18% • OpenAI - 10% • Anduril - 7% • SpaceX - 5% Fundrise CEO Ben Miller (@BenMillerise) breaks down the launch of their publicly listed closed-end fund, Fundrise Growth Tech Fund (NYSE: $VCX) ..and why it has gotten so much insatiable demand. VCX debuted at roughly $700M valuation & surged +18x, with shares spiking to $575, way above the estimated net asset value (NAV) per share of $18.97. VCX debuted on the NYSE March 19, 2026 giving over 100,000 investors access to a portfolio of top private companies including Anthropic (~20%), Databricks (~18%), OpenAI (~10%), Anduril, & SpaceX How we got here? Private markets are now where most value is created. VCX portfolio companies grew ~193% vs ~25% for public tech benchmarks, highlighting the gap between private and public market growth. Meanwhile, IPO timelines have stretched from ~3–5 years to 10–15+ years, meaning public investors are increasingly missing the highest-growth phase. We discuss how VCX works as a closed-end fund, why it has traded at a premium (despite most closed-end funds trading at discounts), & how @fundrise accessed top-tier companies during the 2022–2023 venture downturn — including buying from distressed sellers and stepping into competitive rounds. We cover: • VCX launch & @NYSE debut • Portfolio (Anthropic, OpenAI, Databricks, SpaceX) • Risks: volatility, cycles, and downside scenarios • Will megafunds like a16z or General Catalyst go public? • Private vs public market growth gap (193% vs 25%) • Macro shift: value creation moving to private markets • IPO window + why companies stay private longer • How Fundrise sources and wins allocation • Closed-end fund structure, NAV, premiums/discounts 𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒 (00:00) Benjamin Miller, Co-Founder & CEO at Fundrise (01:12) The idea that almost got rejected (04:27) How the 2023 crash created big opportunities (05:54) From $700M to $6.5B in days (07:38) How a closed-end fund works (11:09) Inside the VCX portfolio (OpenAI, SpaceX, Databricks) (14:50) What Robinhood & Destiny are doing (17:02) Why private markets are pulling ahead (21:38) IPO environment right now (22:17) The SaaSpocalypse & market volatility (25:55) What happens if VCX trades down (27:39) How VCX moves through cycles (31:25) How they decide where to invest (35:45) Investment size and scale (36:59) Most underrated portfolio company (40:37) Biggest lesson: pain = success (43:30) Origin story: why Fundrise exists (47:12) Will big VC firms go public? (50:52) Future of venture capital (54:05) Biggest risks ahead (57:18) Democratizing venture capital (01:00:56) What’s next for VCX (01:03:05) Dealing with skeptics
English
29
49
341
106K
Alex Yeh
Alex Yeh@alex_yehya·
Interesting
数字生命卡兹克@Khazix0918

杨植麟、张鹏、罗福莉等齐聚一堂,他们关于OpenClaw的观点值得一听。 今天是2026中关村论坛的人工智能主题日。 我也定了个一早的闹钟准时起来看。 这个活动海淀已经办了第三年,但今年的议程密度确实有点夸张。 一上午塞进了开源联盟成立、主权大模型白皮书发布、北京市人工智能协会揭牌,外加两场圆桌。 大模型和具身智能各一场。嘉宾阵容从Eclipse基金会到智谱、小米MiMo、无问芯穹,再到一众具身智能公司的创始人,几乎把当下AI产业链上最活跃的角色都拉到了同一个舞台上。 在我看来,信息密度最高的,还是第一场小龙虾与AI开源圆桌。 这场圆桌由杨植麟主持,嘉宾是智谱的张鹏、无问芯穹的夏立雪、小米MiMo的罗福莉,还有港大的黄超。 从模型层到算力基础设施层再到Agent应用层,刚好覆盖了当下AI产业链的几个关键环节。 张鹏解释了智谱GLM5 Turbo提价的逻辑,模型从聊天转向干活,完成任务消耗的token量可能是简单问答的十倍甚至百倍,提价本质上是在回归商业价值。 罗福莉谈了中国团队在模型结构创新上的优势,尤其是长上下文架构和推理效率。 黄超拆解了Agent在planning、memory、skill三个维度上的技术痛点。 最后让每人用一个词总结未来12个月的趋势,黄超说的是生态,罗福莉说的是自进化,夏立雪说的是可持续token,而张鹏直接点了一个最朴素的问题:算力。 几位嘉宾的圆桌对话,干货含量极高,几乎没有什么客套和PR话术,聊的都是非常实在的问题。 所以,我也把这场圆桌的内容做了一个完整的文字整理,分享给大家。 【圆桌对话-人类校对版】 杨植麟(主持人):很荣幸今天能邀请到各位重磅的嘉宾,各位的背景覆盖了不同的层面,从模型层到底层的算力层,再到上面的Agent层。很高兴今天能和大家一起探讨,主要的关键词是开源和Agent。我们从第一个问题开始,这个问题是给所有人的。 现在OpenClaw是最流行的产品,大家在日常使用OpenClaw或类似产品的过程中,觉得最有想象力或者印象最深刻的是什么?从技术角度来看,如何看待今天OpenClaw和相关Agent的演进?我们先从张鹏这边开始。 张鹏 :好,首先感谢植麟的邀请,也感谢主办方给这次机会跟大家交流。其实我很早就开始自己玩这个东西了,当时还不叫OpenClaw,最早叫Clawdbot。 毕竟我是程序员出身,折腾这些东西有一些自己的体验。我觉得它带给大家最大的突破点,在于这件事情不再是程序员或者极客们的专利,普通人也可以比较方便地使用顶尖模型的能力,尤其是在编程和智能体方面。 所以到现在为止,我在跟大家交流的过程中,更愿意把OpenClaw这件事称作一个脚手架,它提供的是一种可能性。在模型的基础之上搭起一个牢固、方便又灵活的脚手架,大家可以按照自己的意愿去使用底层模型提供的很多新奇的东西。原来自己的一些想法,受限于不会写代码或者缺乏某些技能,今天终于可以通过很简单的交流就把它完成。这对我来说是一个非常大的冲击,让我重新认识了这件事情。 夏立雪 :我最开始用OpenClaw的时候其实不太适应,因为我习惯于和大模型聊天的那种交流方式,结果发现OpenClaw反应好像比较慢。但后来我意识到一个关键的不同,它不是一个聊天机器人,而是一个能够帮我完成大型任务的助手。当我开始给它提交更复杂的任务之后,发现它其实能做得很好。这件事给我一个很大的感触,就是AI从最开始按token聊天,到现在能够作为一个Agent帮你完成任务,对整个AI的想象力空间做了一个很大的提升。但与此同时,它对整个系统的能力要求也变得很高,这也是我一开始用会觉得卡顿的原因。 作为基础设施层的厂商,我看到OpenClaw为整个AI后续的大型系统和生态带来了更多机遇和挑战,因为现在所有能用到的资源,想要支撑起这样一个快速增长的时代是不够的。就拿我们公司来说,从一月底开始,基本上每两周token用量就翻一番,到现在已经翻了十倍。上次见到这种速度,还是当年3G时代手机流量增长的那种感觉。现在的token用量,就像当年每个月只有一百兆手机流量的那个时代一样,所有资源都需要更好的优化和整合,让每一个人都能把OpenClaw这样的AI能力用起来。所以作为基础设施领域的从业者,我对这个时代非常激动,认为其中有很多值得去探索和尝试的优化空间。 罗福莉 :我把OpenClaw视为Agent框架上一个非常革命性和颠覆性的事件。虽然我知道身边深度使用编程工具的人,第一选择可能还是Claude Code,但我相信只有用过OpenClaw的人才能独特地感受到,这个框架在设计上有很多地方是领先于Claude Code的,包括最近Claude Code的很多更新,其实都是在向OpenClaw靠近。对我自己来说,OpenClaw这个框架带来更多的是一种随时随地的想象力延伸。Claude Code可能最开始只能在桌面上延展创意,但在OpenClaw里我可以随时随地延展想法。 我后来发现,OpenClaw核心价值在于两点:第一是它开源,开源对整个社区深度参与、持续改进Agent框架是一个非常重要的前置条件。第二,像OpenClaw这样的Agent框架,它很大的价值在于把国内水平接近但略逊于闭源模型的这一赛道上的模型上限拉得非常高,在绝大部分场景里任务完成度已经非常接近Claude最新的模型。同时它又通过Harness系统或者Skills体系等诸多设计,把下限保证得非常好。从基座大模型的角度来说,它保证了下限,同时也拉升了上限。此外,我认为它给整个社区带来的更大价值,是点燃了大家对模型之外的那一层的热情,让大家发现Agent这一层有非常多的想象力和空间可以发挥。这也让社区里越来越多除研究员以外的人参与到AGI的变革当中,更多人接触到更强的Agent框架,一定程度上在替代自己重复性的工作,释放时间去做更有想象力的事情。 黄超 :从交互模式上来讲,我觉得OpenClaw这次爆火,首先是因为给大家一种更有活人感的感觉。我们其实做Agent也有一两年了,但之前包括Cursor、Claude Code这些Agent,大家感受到的更多是一种工具感。OpenClaw第一次以IM软件嵌入的交互方式,让大家更有一种活人感,更接近于自己想象中的个人贾维斯那样的概念,这是交互模式上的突破。另外它给大家带来的一个启发是,Agent Loop这种非常简单但高效的框架,再次被证明是行之有效的。同时它也让我们重新思考,究竟是需要一个all-in-one的非常强大的智能体帮我们做很多事情,还是需要一个像轻量级操作系统或脚手架一样的小管家。OpenClaw的答案是,通过这样一个轻量级的操作系统生态,去撬动整个生态里所有的工具。随着Skills和Harness这些机制的普及,越来越多的人可以设计面向OpenClaw这类系统的应用,赋能各行各业。这与整个开源生态天然结合得非常紧密,我觉得这两点是它带给我们最大的启发。 杨植麟(主持人):顺着这个话题,刚才一直在讨论OpenClaw。想问一下张鹏,看到最近智谱也发布了新的GLM5 Turbo模型,我理解在Agent能力上也做了很大的增强,能不能给大家介绍一下这个新模型和其他模型的不同之处?另外我们也观察到有一个提价的策略,这反映了什么样的市场信号? 张鹏 :这是个很好的问题。前两天我们确实紧急做了一波更新,这其实是我们整个发展路线中的一个阶段,提前把它放出来了。这件事最主要的目的,是从原来简单的对话转向真正的干活。正如各位刚才说的,OpenClaw真的让大家觉得大模型不再只是聊天,而是能帮我干活。但干活背后对能力的要求是非常高的,它需要自主进行长程任务规划,不断压缩上下文、debug、处理多模态信息等等,这和传统面向对话的通用模型的要求有很大的不同。所以GLM5 Turbo在这方面做了专门的加强,尤其是长程任务如何能够持续自主loop而不中断,这里做了很多工作。 另外,大家也提到了token消耗的问题。让一个聪明的模型去完成复杂任务,token消耗量是非常巨大的,可能一般人体会不到,只会看到账单上的钱在不停地往下掉。所以我们在这方面也做了优化,面临复杂任务时用更高效的token效率来完成。模型架构上本质还是多任务协同的通用架构,只是在能力上做了一些偏向性的加强。至于提价这件事,其实也很顺畅地能跟大家解释。现在不再是简单的一问一答,背后的思考链路很长,还要通过写代码的方式跟底层基础设施打交道、随时debug和纠错。完成一个任务需要的token量,可能是原来回答一个简单问题的十倍甚至百倍。模型变得更大,推理成本相应提高,所以我们把价格回归到正常的商业价值上。长期靠低价竞争不利于整个行业发展,这样才能持续在商业化路径上形成良性闭环,不断优化模型能力,持续给大家提供更好的模型和相应的服务。 杨植麟(主持人):非常好的分享。现在开源模型和推理算力已经开始形成一个生态,各种开源模型可以在不同的推理算力上为用户提供更多价值。随着token量的报价变化,我们可能也正在从训练时代逐渐进入推理时代。想请教一下立雪,从infra的层面来看,推理时代对于无问芯穹意味着什么? 夏立雪 :我们是一家诞生在AI时代的基础设施厂商,现在在为Kimi、智谱提供服务,也在跟MiniMax合作,帮助大家更高效地用好我们这个token工厂。我们也在和很多高校、科研院所合作,所以一直都在思考一个问题:AGI时代所需要的基础设施,究竟应该是什么样的?我们怎么能够一步步地在这个过程中去实现它、推演它。 我们已经做好了充分的准备,也看清了短期、中期、长期不同阶段需要解决的问题。当前最紧迫的问题,就是像Claude这类模型带动的整个token量的暴增,对我们系统效率提出了更高的优化需求,价格的增长也是在这个需求压力下的一种应对方式。 我们一直以来都是从软硬件打通的方式来布局和解决这个问题。我们接入了几乎所有种类的计算芯片,把国内十几种芯片和几十个不同的算力集群统一连接起来。这样,当资源不足时,我们能做到两件事:第一,把能用的资源都用起来;第二,让每一个算力都用在刀刃上,发挥出最大的转化效率。所以当前阶段我们要解决的核心问题,就是如何打造一个更高效的token工厂。为此我们做了很多优化,包括让模型与硬件在显存等方面实现最优适配,也在探索在最新的模型结构和硬件结构下能否产生更深度的化学反应。 不过,解决当前的效率问题,我们只是打造了一个标准化的token工厂。面向Agent时代,这还远远不够。就像刚才说的,Agent更像一个人,你可以交给他一项任务。我坚定地认为,当前云计算时代的很多基础设施,是为服务程序、服务人类工程师而设计的,而不是为AI设计的。现在的状态有点像:我们搭了一套基础设施,上面留了一个为人类工程师设计的接口,然后在这上面再包一层去接入Agent。这种方式实际上是用人类操作的能力边界,限制了Agent的发挥空间。 举个例子,Agent能够在秒级甚至毫秒级思考并发起任务,但我们之前的底层K8S这些能力并没有为此做好准备,因为人类发起任务大概是分钟级别的。所以我们需要进一步构建我们称之为Agentic Infra的能力,打造一个更智慧化的算力投放工厂。这是无问芯穹现在正在做的事情。 从更长远的未来来看,真正AGI时代到来的时候,我们认为连基础设施本身都应该是一个智能体,应该能够自我进化、自我迭代,形成一个自主的组织。相当于有一个CEO,这个CEO是一个Agent,比如一个Claude在管理整个基础设施,根据AI客户的需求自己提需求、迭代自己的基础设施。只有AI与AI之间才能更好地形成耦合。所以我们也在做一些让Agent与Agent之间更好通信的事情,比如cache to cache这样的复制能力。 我们一直认为,基础设施与AI的发展不应该是隔离的状态,而应该产生非常丰富的化学反应。这才是真正的软硬协同,真正的算法与基础设施协同。这也是无问芯穹一直想实现的使命。 杨植麟(主持人):接下来想问问福莉。小米最近发布了新的模型,也开源了一些背后的技术,我觉得对社区做出了很大的贡献。想请问一下,小米在做大模型方面有什么独特的优势? 罗福莉 :我想先把这个问题稍微拓展一下,不只聊小米的优势,而是聊聊中国做大模型的团队在这件事上的优势,我觉得这个话题有更广泛的价值。 大概两年前,我就观察到中国的基座大模型团队已经开始了一个非常好的突破。这个突破是:在有限算力的条件下,尤其是在互联带宽受限的情况下,如何突破这些低端算力的限制,并由此催生了一些看似是为效率妥协的模型结构创新,比如DeepSeek v2、v3系列的细粒度MoE等等。但我们后来能看到,这些创新引发的是一场变革,也就是在算力一定的条件下,如何发挥出最高的智能水平。我觉得DeepSeek给了国内所有技术大模型团队一份勇气和信心。 虽然今天我们自己的国产芯片,无论是推理芯片还是训练芯片,已经不再像以前那样受到严重限制,但我们能看到,正是那些限制催生了我们对更高训练效率、更低推理成本的模型结构的全新探索。比如最近出现的hybrid sparse或linear attention结构,有DSA、NSA,Kimi有KSA,小米也有面向下一代结构的high sparse架构。这区别于MiMo这一代结构,是我们面向Agent时代去思考的,如何在Agent时代做出更好的模型结构创新。 我为什么认为结构创新如此重要?因为我们刚才聊到了long context这个话题。如果大家真实地去用OpenClaw就会发现,越用越好用,越用越聪明。它的前提是推理的context足够长。long context是一个谈论了很久的话题,但真正能做到在超长context下表现强劲、推理成本足够低的模型,其实并不多。很多模型不是做不到百万甚至千万token的context,而是推理成本太高、速度太慢。只有当你能在百万甚至千万context下做到成本够低、速度够快,才会有真正高生产力价值的任务被交给这个模型,从而激发模型在long context场景下完成更高复杂度的任务。我们可能需要在这样千万甚至亿级context的规模下,才能实现模型的自迭代。所谓模型的自迭代,就是它可以在复杂的环境里依靠超长context完成对自我的进化,这个进化可能是对Agent框架本身的,也可能是对模型参数本身的,因为我们认为long context本身就是对参数的一种进化。 所以,怎么实现long context efficient的架构,以及在推理侧做到long context efficient,是一个全方位的竞争。这是我们大约一年前就开始探索的问题。而如今,怎么在真实的长程任务上实现稳定性和高上限的效果,是我们现在在持续迭代的创新方向。我们在思考如何构造更有效的学习算法,如何采集到真实的、在百万乃至千万上下文里具有长距依赖的文本,以及结合复杂环境产生的trajectory,这是我们正在经历的事情。 但我能看到更长期的事情是,大模型本身在飞速进步,加上Agent框架的加持,推理需求已经在过去一段时间内增长了近十倍。那今年整个token的增长会不会达到百倍?这又将我们带入另一个维度的竞争,那就是算力,推理芯片,乃至往下到能源层面。这是我对这个问题的判断,也期待从大家身上学到更多。 杨植麟(主持人):非常有insight的分享。下面想问一下黄超,因为你也开发了一些非常有影响力的agent项目,包括nano bot,在社区里也有很多粉丝。想问一下,从agent的harness或者说应用层面,接下来你觉得有哪些技术方向是比较重要、大家需要去关注的? 黄超 :感谢。我觉得首先可以从agent的几个关键技术模块来拆解,包括planning、memory和tool use。 从planning来讲,现在面向长链路任务或者非常复杂上下文的场景,比如说五百步甚至更长的任务,很多模型不一定能做好planning,我觉得本质上是模型不具备这方面的隐性知识,尤其是在一些复杂垂直领域。未来可能需要把各类复杂任务的知识固化到模型里,这是一个方向。当然,skill和harness这种机制,在一定程度上也是在缓解planning层面的错误,因为它提供了比较高质量的skill,本质上是在帮助模型去完成一些较难的task。 关于memory,我的感受是它永远存在信息压缩不准确、召回不准的问题。当整个长链路任务和复杂场景展开时,memory会急剧膨胀,这对整个memory架构造成很大压力。目前包括各类agent框架基本上都采用最简单的文件系统、Markdown格式来做memory,通过文件共享来协作。我觉得未来memory应该走向分层设计,并且需要解决通用性的问题,因为coding场景、deep research场景、多媒体场景的数据模态差异很大,如何对这些memory做好检索索引、提升效率,这永远是一个trade off。 另外,现在agent框架让大家创建agent的门槛大幅降低,未来可能不止一个agent,我也看到有些产品推出了Agent Swarm这样的机制,相当于每个人会拥有一群龙虾。一群龙虾相比一个龙虾,上下文的暴增是可以想象的,这对memory带来的压力非常大。如何管理一群龙虾带来的上下文,目前还没有很好的机制,尤其是在复杂coding、科研发现这类场景下,对模型和整个agent架构都是不小的挑战。 关于tool use,当年MCP存在的问题,比如质量没有保障、存在安全隐患,现在在skill里依然存在。目前看似有很多skill,但高质量的skill其实比较少,低质量的skill会严重影响agent完成任务的完成度。另外skill也存在恶意注入的风险。所以我觉得tool use这块可能需要整个社区共同努力,把skill生态发展得更好,甚至探索如何在执行过程中进化出新的skill。以上这些,是我认为当下agent在planning、memory、skill三个维度上存在的痛点,以及未来潜在的方向。 杨植麟(主持人):可以看到刚才两位嘉宾从不同视角讨论了同一个问题,随着任务复杂度增加,上下文会急剧膨胀。从模型层面可以去提升原生的上下文处理能力,从agent harness层面,则是通过planning、memory,包括multi-agent的harness,在模型能力一定的情况下支持更复杂的任务。我觉得这两个方向接下来会有更多的化学反应,共同提升完成复杂任务的能力上限。 那最后我们来一个开放式展望,请各位用一个词来描述接下来十二个月大模型发展的趋势以及你的期望。这次我们先从黄超开始。 黄超 :十二个月在AI领域看起来好遥远,真的不知道十二个月之后会发展成什么样子。 杨植麟(主持人):这里原来写的是五年,我给改成十二个月了。 黄超 :对,我这边的关键词应该是生态。未来agent要真正从个人助手转化为打工人,这一步很重要。现在大家玩agent很多时候还停留在新鲜感阶段,觉得好玩,但未来真正要让agent沉淀下来,成为大家真正的搬砖工具,或者说真正的co-worker。这需要整个生态的共同努力,把所有相关的技术探索和模型技术都开源出来,不管是模型迭代、skill平台迭代还是各类工具,都需要面向agent打造更好的生态。 从我自己的感受来说,未来的很多软件可能不再是面向人类的。人类需要GUI,但很多软件可能会是面向agent原生设计的,人类只会去使用让自己快乐的GUI,其他的交给agent。所以现在整个生态从GUI、MCP又转向了CLI这样的模式。我觉得需要整个生态把不管是软件系统、数据还是各种技术,都变成Agent Native的模式,这样才能让整个agent的发展更加丰富。 罗福莉 :我觉得把这个问题缩小到一年非常有意义,因为五年这个时间跨度,从我心目中对AGI的定义来看,我觉得已经实现了。所以如果要用一个词来描述接下来一年AGI历程里最关键的事情,我认为会是自进化。这个词虽然听起来有点玄幻,过去一年大家也多次提到,但我最近才对它有了更深的体会,也对如何具体落地这件事有了更务实可行的方案。 借助非常强大的模型,我们在过去chat范式下其实根本没有发挥出预训练模型的上限,而这个上限现在被agent框架激活了。我们现在触到了一个现象,当模型执行更长时间的任务时,它可以自己去学习和进化。一个很简单的尝试是,在现有的agent框架里叠加一个可以verify的条件约束,再设置一个loop,让模型持续迭代优化目标,我们就能发现模型会持续拿出更好的方案。这种自进化现在已经能跑一两天了,国内的模型基本上能支撑,当然和任务难度有关。我们发现在一些科学研究上,比如去探索更好的模型结构,因为有评估标准,比如更低的PPL,在这类目标明确的任务上,模型已经能自主运行和执行两三天了。 从我的角度看,自进化是唯一能创造出新东西的地方,它不是去替代我们人类现有的生产力,而是像顶尖科学家一样去探索这个世界上还没有的东西。一年前我会觉得这个时间历程需要三到五年,但就在最近,我觉得这个时间线应该缩短到一两年,大模型叠加一个非常强的自进化agent框架,至少能实现对科学研究的指数级加速。我们组内做大模型研究的同学,workflow高度不确定且需要大量创造力,但借助顶尖模型,基本上已经能把我们自己的研究效率加速近十倍了。我很期待这样的范式辐射到更广泛的学科和领域。 夏立雪 :我的关键词叫可持续token。我看到整个AI的发展仍然处于长期持续的过程中,我们也希望它能有长久的生命力。从基础设施的视角来看,我们面临的一个很大问题是资源终究是有限的,就像当年我们讲可持续发展一样。我们作为一个token工厂,能否给大家提供持续稳定、大规模可用的token,让顶尖的模型真正能够继续服务更多的下游,是我们看到的一个非常重要的问题。 所以我们现在需要把视角放宽到整个生态,从最早的能源到算力,再到token,最终转换成GDP,让这条链路能够进行持续的经济化迭代。我们不只是把国内的算力用起来,也在把这些能力输出到海外,让全球的资源能够打通整合。所以我认为可持续这个词,也包含了我们想把中国特色的token经济学做起来的愿望。过去那个时代叫Made in China,我们把低价的制造能力转化成好的商品输出到全球。现在我们想做的有点像AI Made in China,把中国在能源上的优势,通过token工厂可持续地转化为优质的token输出到全球,成为世界的token工厂。这是我希望在今年看到的,中国为世界人工智能带来的价值。 张鹏 :我就简短一点,大家都在仰望星空,我就落地一点。未来十二个月面临的最大问题,我觉得可能就是算力。刚才也说了,所有的技术,包括智能体框架,让很多人创造力爆发、效率提升十倍,但前提条件是大家用得起、用得起来。不能因为算力不够,一个问题提出去让它思考半天都得不到答案,这肯定不行。也正是因为这样的原因,我们很多研究进展和想要做的事情其实都受阻了。前两年记得中关村论坛有人提过这么一句话,叫没卡没感情,谈卡伤感情。今天又回到了这个处境,但情况又不一样了,我们现在转向推理阶段,是因为需求真的在爆发,十倍百倍地爆发。刚才也说到过去增长了十倍,背后其实是一百倍的需求,还有大量的需求没有被满足,这需要大家一起来想办法。 杨植麟(主持人):好,感谢各位的精彩分享,谢谢大家。 最后,这两天海淀五道口的AI原点社区也在举办原点Party Nights活动,有兴趣的可以去玩玩,说不定咱们还能一起面个基🫣。

English
1
0
1
113
Alex Yeh
Alex Yeh@alex_yehya·
There’s abs no resource ava what so ever …
English
0
0
1
110