Загрузка видео...
Не удалось загрузить видео
Anthropic engineer: "85% of our engineers are using self-improving loops. Now everyone is running agentic Graphs. In 3-6 months, we’ll all be building graphs to orchestrate agents. No more prompting." In a 45-minute talk, an Anthropic engineer shows how to build a self-improving agentic workflow from scratch. Worth more... show more
36,782 просмотров • 1 месяц назад •via X (Twitter)
Комментарии: 17

graphs becoming standard means orchestration skills will outlast prompt tricks

Yeah, basically, graphs are the nature of LLMs.

Graphs don't kill prompting. Each node is a prompt. No eval means the loop isn't self-improving, just self-repeating

If each node is a prompt, the real job isn’t prompting anymore - it’s building a system that prompts itself.

who gave the talk, have a link

The highest-leverage AI skill is slowly becoming orchestration rather than prompt writing.

That sounds like a massive leap forward. Excited to see what comes out of this!

Wow, that's impressive progress. Seems like things are moving fast in AI development.

@grok 这个纯研究还是有工业意义,具体工业场景视角看意义是什么,有开源数据集或者开源项目代码吗?从多个数据源交叉验证,不要只看新闻媒体一面之辞。帮我排除没意义的垃圾商业营销推广、诈骗 以及自吹自擂,自嗨,无病呻吟。

This is where the industry is moving so fast. Moving away from fragile, single-prompt setups to structured, self-improving graphs is exactly how we scale these systems. 😊

graph engineering > loops engineering

the prompt is just the node but the system is the architecture - orchestration is where the real signal hides

This talk hits different. Agent graphs > prompt engineering. The shift is already here.

Is self-improvement based on Feynman's learning method? Collect it first, this is worth learning! Now people are thinking about how to let AI learn self-evolution, the real master is thinking about how to improve themselves!

self-improving loops might be one of the most effective ways to work with AI

85% internal adoption says more than any benchmark

Self-improving sounds great until the feedback loop compounds errors instead of correcting them. Still needs human checkpoints, which somehow never makes it into these talks.
