为什么造出大模型的研究者纷纷离开大厂创业
大模型研究者离开大厂创业,被文中解读为对持续学习、物理交互和前沿科学数据短板的回应。文中汇总的 45 家公司已确认融资合计 538 亿美元,其中 8 家未发布公开模型、接口或论文的公司拿走近 300 亿美元,约占 56%。文章提出,独家物理与生化实验数据、快速且低成本的自动验证机制,是创业团队抵御大厂竞争的两类筹码,并判断算力成本可能推动部分团队最终并入巨头。
大模型研究者离开大厂创业,被文中解读为对持续学习、物理交互和前沿科学数据短板的回应。文中汇总的 45 家公司已确认融资合计 538 亿美元,其中 8 家未发布公开模型、接口或论文的公司拿走近 300 亿美元,约占 56%。文章提出,独家物理与生化实验数据、快速且低成本的自动验证机制,是创业团队抵御大厂竞争的两类筹码,并判断算力成本可能推动部分团队最终并入巨头。
Josef Chen 参观宇树杭州工厂后,更看好西方制造机器人所需设备的机会,判断依据是中美人工成本差异与 AI 降低编程成本。宇树 2025 年生产近 32,000 台机器人,人工占制造成本的 8%;其新电机产线预计使一款机器人降本 1.7%,作者估算按美国工资可降本约 9%。他主张西方优先建设能快速适配尺寸变化的电机自动化产线,并披露自己投资了自动化制造及人形机器人相关企业。
The Decoder 梳理了 27 家研究驱动的独立 AI 实验室,将其长期技术押注分为持续学习、世界理解、独有数据、计算效率、研究自动化和模型应用缺口六类。
Genie 4 采用视频模型与三个 Gemini 模型协作的架构,文章据此分析世界模型端到端路线的局限、算力成本与工业应用方向。文中称,外围模型分别承担训练阶段的常识补充、运行阶段的任务引导和机器人棋步推理,单次交互虽延长至两分钟,仍无法保证全程一致性。文章区分视觉渲染、物理模拟与动作规划,主张由代码和三维网格管理空间规则、语言模型负责规划,世界模型负责合成供智能体训练的复杂环境。
前 DeepMind 研究员 Thore Graepel 创立 Metis Reasoning,提出用 AlphaGo 式搜索加显式知识账本替代把知识烘焙进模型权重,以减少幻觉并让结论可溯源。
工业智能将制造商的积累经验、运营实践与生产数据结合,让知识在决策现场可用并接入物理生产环节。Toyota Industries 打通碎片化喷漆车间数据后,分析周期从五天缩短到四小时以内,试点将相关喷漆缺陷减少约 25%。Rockwell Automation 新加坡工厂用维护 copilot 整合制造数据、设备手册与老员工经验,新员工掌握数百台设备排障的时间从九个月降至三个月。
Techable Ventures 创始人兼管理合伙人 Amulya Vishwanath 认为,物理 AI 目前仍处于“乐高时代”,而非外界期待的机器人版 ChatGPT 时刻。她指出,机器人手、执行器等基础软件与构件正被重新设计,通用与专用机器人同时在造,只有这些积木逐步拼合,创新才能规模化。她曾在英伟达做近七年产品营销,如今把连接创始人与顶尖人才视为基金护城河。
Waymo正在从区域出行服务走向更广泛的城市基础设施,无人驾驶的便利也伴随着乘客控制权、公共安全与责任归属的难题。公司宣称车辆比人类驾驶安全10倍,这一指标指造成严重伤害或更严重后果的碰撞减少92%,并未涵盖文中乘客经历的闯入施工路段、受袭后无法离开等情况。报道通过乘车体验和多方采访,记录了远程支持、无障碍出行及充电设施扰民等问题,区分自动驾驶的交通安全潜力与无人驾驶带来的私人空间体验。
Yao Shunyu said at Tencent Cloud’s AI industry application conference that Hunyuan 3 rebuilt pretraining and reinforcement-learning infrastructure, changed data and evaluation, and assigned its strongest post-training staff to improve Yuanbao first; he named coding agents, multimodality, and embodied AI as Tencent’s next focus areas.
Why it matters: Tencent's response to claims it lags in AI, plus how Hunyuan 3 rebuilt pretraining and RL and where post-training staff were reassigned.
Ruan Yifeng excerpts observations from U.S. analysts who visited 14 Chinese AI and robotics companies in early May: the article estimates U.S. AI compute at about 8 times China’s by the end of 2025, while Chinese firms’ intelligence output per unit of compute is estimated at 4-7 times naive scaling.
Why it matters: After visiting 14 Chinese AI and robotics companies, US analysts estimate US AI compute was about 8x China's by end-2025, with far fewer high-end GPUs available to Chinese firms.
Alex Imas and Phil Trammell argue that robots can be copied and scaled after AGI, while scarce human skills such as ballet performance remain fixed; the post does not disclose a quantitative model or timeline.
Why it matters: Two researchers argue that once AI can copy most products, human involvement becomes scarce as long as people care that humans did the work, with the fixed number of ballet dancers as evidence.
Yi-Ling Liu compares U.S. and Chinese AI anxiety through labor, companionship, and agency: over 70% of U.S. teenagers report using chatbots as companions, while China is projected to reach 200 million single-person households by 2030.
Why it matters: The author sees the same unease about AI among ordinary people in China and the US, and shows why the gains don't reach individuals.
Latent Space interviewed The Fourth Law founder Yaroslav Azhnyuk for a two-hour episode covering FPV drones, five levels of autonomy, eight dimensions of the autonomous battlefield, and China’s manufacturing advantage; the transcript claims Ukraine produced 4 million FPV drones last year and discusses a hypothetical Chinese capacity of 4 billion.
Why it matters: The Fourth Law founder Yaroslav Azhnyuk explains on a podcast how FPV drones became a battlefield staple, and lays out autonomy levels and cost structure.
Jiqizhixin translated a MoE Capital blog tracing two world-model lineages. The article says more than $10 billion entered the category over 18 months, and cites DreamDojo as using 44,711 hours of first-person video pretraining to reach r=0.995 correlation with real-world robot policy outcomes.
Why it matters: Maps two routes to world models, one from AI planning research, one from game and film 3D generation, with over $10B invested in 18 months.
Jensen Huang told Carnegie Mellon’s 2026 CS graduates that skilled trades have better prospects; Randstad says trade demand is growing three times faster than white-collar roles, with robotics technician jobs up 107%.
Why it matters: At Carnegie Mellon's computer science commencement, Jensen Huang said trades beat computer graduates, alongside data showing trade demand growing three times faster than white-collar and robotics technician roles up 107%.
Yann LeCun discussed LLM limitations on the Unsupervised Learning podcast, covering his 2027 forecast, AMI’s bet on world models, his reasons for leaving Meta, and major disagreements with Geoffrey Hinton and Yoshua Bengio over Turing Award-era views.
Why it matters: Yann LeCun said on a podcast that the large language model path has a ceiling, and explained why he is betting on world models after leaving Meta, giving a direct account of the split from a key figure.
Ted Xiao divides nearly a decade of robot learning into three eras: Google’s team trained RT-1 on 87,000 teleoperation trajectories, then adapted 5B to 55B VLMs into VLA policies for RT-2.
Why it matters: Ted Xiao walks through Google's robot learning path, from RT-1 trained on 87,000 teleoperation trajectories to RT-2 turning 5B-55B vision-language models into action policies.
The speaker presented a physical AGI roadmap with six named components: video world models, WAM, EgoScale, dexterity scaling laws, physical reinforcement learning, and DreamDojo; the snippet also mentions a 2016 OpenAI DGX-1 signing story with Jensen and Elon.
Why it matters: The speaker splits physical AGI into six modules and maps robot training onto the stages of large language model development, giving a shared framework for humanoid general intelligence.
Applied Intuition’s founders reviewed a 10-year physical AI path, with the company valued at $15B. The post cites 30+ products, 18 of the top 20 non-Chinese automakers as customers, and L4 driverless trucks in Japan. The key constraint is onboard deployment: millisecond latency, low power, small models, and safety validation.
Why it matters: Applied Intuition's CEO and CTO trace the path from a YC self-driving toolchain to a $15B physical AI company, with 30-plus products and 18 of the top 20 non-Chinese automakers as customers.
Honor's humanoid robot Lightning ran 50:26 at the 2026 Beijing Yizhuang half marathon, faster than the men's human world record of 57:20; the post also says Unitree H1 did a 1.9 km winding course in 4:13. The post cites nearly 200 embodied-AI financings and over RMB 30 billion in Q1 2026, plus Spirit AI's $455 million Pre-A on April 16. The real signal is capital shifting from robot hardware to model-centric 'brains.'
Why it matters: Honor's humanoid ran a half marathon in 50:26, faster than the human men's world record, while embodied AI drew nearly 200 funding deals worth over 30 billion yuan in Q1.
Companies and investors put $6.1 billion into humanoid robots in 2025, 4x 2024, and MIT Technology Review attributes the surge to a shift in how robots learn. The piece highlights two mechanisms: around 2015, simulation plus reward signals enabled millions of trial-and-error runs; after ChatGPT in 2022, robotics models took images, sensors, and joint states to predict dozens of motor commands per second. The key change is data-driven learning over hand-written rules; the provided text is truncated, so later examples are not fully disclosed.
Why it matters: Traces how humanoid robot learning moved from hand-written rules to simulation trial and error, against $6.1 billion of 2025 investment, four times the 2024 figure.
Micro1 hires thousands of contractors across 50+ countries to film chores at home with iPhones and sell that real-world data to humanoid robotics companies. The piece cites $15/hour pay for one worker, says robotics firms spend over $100 million a year on such data, and notes $6 billion+ went into humanoids in 2025. The real issue is data governance: workers know the footage trains robots, but the post shows they often do not know how it is stored, shared, or deleted.
Why it matters: Micro1 pays gig workers in more than 50 countries to film household chores on iPhones and sells the motion data to humanoid robot companies; a Nigerian medical student earns $15 an hour.
Waymo placed a $6.25 task on a delivery platform to send a rider 1 km away to close a robotaxi door, with another $5 after completion. The post frames this as software dispatching human labor, not a one-off gig, and argues platform workers are becoming a human API inside automated workflows. The point to watch is the AI-plus-labor loop; the post does not disclose Waymo's scale, frequency, or formal product design.
Why it matters: A robotaxi that can't close its own door pays a delivery rider $6.25 to do it, showing where automation still hands off to people.
The piece argues China has embedded AI into manufacturing, with 30,000+ smart factories, and over half of all industrial robots installed globally in 2024 going to Chinese plants. It cites shop-floor data: Zeekr's Ningbo plant uses 800+ robots, Xiaomi says its Beijing factory produces one car every 76 seconds, while only 18% of U.S. manufacturers report a formal AI strategy and two-thirds struggle to scale pilots. The real point is not frontier models but AI deployment in factory automation, scheduling, and inspection.
Why it matters: China has built over 30,000 smart factories, and more than half of 2024's new industrial robots went to Chinese plants, with Zeekr, Xiaomi and Midea showing AI on the line.
Jensen Huang said at Davos that global VC investment topped $100 billion in 2025, with most capital going to AI-native startups building the AI stack’s application and infrastructure layers. He described AI as a five-layer stack: energy, chips and computing infrastructure, cloud data centers, models, and applications, and cited a US nursing shortage of about 5 million where AI can handle charting and transcription. The key point for practitioners is that the bottleneck is not just models, but the full infrastructure and labor chain.
Why it matters: Jensen Huang split the AI stack into five layers at Davos, energy, chips, cloud data centers, models and applications, and mapped where over $100 billion in 2025 venture funding went.
At CES, Silicon Valley 101 discussed humanoid robot deployment and cited official figures: 21 of 38 humanoid exhibitors were Chinese companies. Guests noted Boston Dynamics plans Atlas deliveries in 2026 and 30,000 annual capacity by 2028, but argued scale claims do not prove product-market fit; in warehouses, wheeled bases plus arms often beat humanoids on ROI.
Why it matters: Fu Sheng talks to Silicon Valley 101 at CES, using on-site numbers like 21 of 38 exhibitors from China to weigh humanoid hype against warehouse reality.
CES 2026 drew 148,000+ attendees and 4,100+ exhibitors, with Chinese companies making up nearly a quarter and standing out in AI hardware and robotics. The post ties their optimism to manufacturing-led iteration speed, not one breakthrough; Lenovo Qira, Nvidia Vera Rubin, and AMD Helios show the race is shifting to cloud and hybrid AI.
Why it matters: Chinese exhibitors made up nearly a quarter of CES this year; on the floor the author saw home cleaning robots taking US share from Dyson and Shark, and most yard equipment also came from Shenzhen.