North America AI field notes: models accelerate, can the world keep up
An investor spent 15 days across Silicon Valley, Los Angeles, Seattle, Boston and New York and wrote up conversations with 101 people working on AI in North America. The notes describe researchers whose daily work is now “supervising and encouraging my 10,000 agents,” pretraining teams shrinking from 100–200 people to about 20, training data growing from 10T–20T tokens to 100T and model parameters moving from 1T–3T toward 10T–20T, with RL and executable, feedback-rich task environments seen as the next scaling axis. They also flag the delivery gap — one study of more than 500,000 developers found autonomous coding agents raised code commits 240% and projects 80% but actual releases only 30%, alongside “Tokenmaxxing” — and a power constraint in which a high-end GPU with networking and cooling draws roughly 1.5–2 kW, a million-GPU cluster approaches 1.5–2 GW, and grid interconnection can take five years or more.
It places frontline researchers’ day-to-day experience alongside hard constraints in compute, power and delivery, offering a rare cross-section for judging where the next stage of value may sit.
这趟行程密度很大,后劲很足,回到上海后满脑子还是15天里的各种画面: 一边,有Frontier Lab研究员波澜不惊地向我描述他的日常:“我的工作主要就是监督和鼓励我的1万个agent。”有数学博士对我说:“数学已经死了。” 在Mountain View的Red Rock Coffee里,连续几天见的多位“天才少年”都坚定表示“一定会创业”,也有不到30岁的一线研究员开玩笑地说“要退休了”。 另一边,则是热火朝天的科技新闻:OpenAI发布了Astra、公布Navier–Stokes千禧年难题解答、Meta推出了用户增速超过ChatGPT的AI个人助理Muse、Dario大声疾呼“pace the frontier”居然得到了Elon和死对头Sam的响应...... 新闻里的未来,和眼前研究员的日常,正在以一种奇妙的方式重叠。
同一轮浪潮里,人们正在经历不同的时间线
对于SevenX,我们始终相信,对任何技术和产业的理解,一定要站在最前沿、最一线,要和那些真正在其中挣扎过、绝望过、成功过的人去对话,才能完成一个关于未来的更全面、也更接近真实的拼图。于是9月初,带着新一轮的好奇和疑问,我在15天里横跨硅谷、洛杉矶、西雅图、波士顿和纽约,高密度地和101位新老朋友进行了不同颗粒度、但都非常有意思也极具启发的交流。
#OpenAI#Meta#Anthropic#Google#NVIDIA#SevenX Ventures#GPT-6 Astra#Meta Muse