Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Chen Mingda maps the limits of scaling and the next AI research paths

AI summary

Meta FAIR researcher Chen Mingda said in a DeepTech interview that scaling drove rapid gains in language understanding, mathematical reasoning, and coding, but it did not solve capabilities that remained near zero. He highlighted procedural knowledge extracted from prior reasoning, coding agents, and learning from unlabeled visual data as research directions for complex reasoning and non-language foundation models. The interview also contrasts corporate priorities at Meta with the greater freedom he seeks at Shanghai Jiao Tong University.

In-depth insight

The interview offers a researcher's grounded view of where large-model scaling has helped, where immature techniques have struggled, and which new approaches are being pursued for complex reasoning and multimodal AI.

2022 年,陈明达博士毕业后加入 Meta 基础人工智能研究院(FAIR)。当时,前沿 AI 实验室仍在探索大模型的能力边界。研究者可以从一个未经验证的想法出发,设计实验、训练模型,再根据结果决定是否继续投入资源。不同技术路线之间,还有相当大的探索空间。

几个月后,ChatGPT 的出现改变了这一局面。

大模型展现出的能力超出了许多研究者的预期,也让 Scaling(规模化)成为行业最受关注的研发路线之一。科技公司开始集中算力、数据和人才,围绕更大规模的模型展开竞争。前沿 AI 研究越来越依赖大规模工程投入,希望通过持续扩大规模,推动模型能力不断提升。

陈明达所在的 FAIR 同样经历了这一转变。他先后参与 LLaMA、Chameleon 等项目,见证了 Scaling 推动模型能力快速提升的过程,也看到一些尚未成熟的技术,即使投入大量资源,依然难以取得预期效果。

#Meta FAIR#Shanghai Jiao Tong University#LLaMA#Chameleon#Claude Code#Procedural Intelligence#JEPA

Original · 23 min readRead the original
Group chat

No comments yet. Start the conversation.