🔥 Search Hot Tweets
Search and analyze hot tweets from KOL accounts list (list: https://x.com/i/lists/1961235697677017443) within 6 hours. Use SoPilot plugin to quickly comment and occupy the comment section.
发现一个合法节约 AI 订阅费的方法，以 Claude 为例： 1. 先选最低的那一档，比如 $20 2. 用完了升级下一档，比如 $100 3. 以此类推，最终 $200 通常等你到最后要升级到 $200 的时候可能这个月都要结束了，甚至都不需要。 唯一的问题是操作比较麻烦 这事也许可以找个 AI Agent 比如 Claude Cowork 自动帮我做，快月底了降级订阅，用完了升级
哇咔咔咔！今天第一次做了手抓饭！铸铁塔吉锅一锅出！肥肥的羊排特别给力！超满足，吃了大半锅，睡前再玩会儿壶铃减减肥 https://t.co/jFSWMGFwCS
Revealing the answer: Eigent = Eigen + Agent Eigen means “own” in German and is also related to the concept of eigenvalues. Product vision: a fully local AI agent that is open at every level, from the model and agent framework to the full-stack application. So it truly becomes your “own” agent. Research vision: identifying the principal components, the most important “eigenvectors” of the covariance matrix, behind the scaling laws of agents. Personal reason: my 5-year-old cat is named Eigen 🐱
写了一个 Skill 调MCP 把DAN KOE这篇神文转成6分钟解读视频。 仔细读有些启发，融合了很多人和经典书中的想法，但原创性跟纳瓦尔比差些。 虽然纳瓦尔也集合了前人的智慧，但第一次读“杠杆”、“幸福”等概念的解读，耳目一新，不知道大家的感受如何。 https://t.co/rzWSh5iEUt
我现在写代码尽量用 Codex，其他任务用 Claude Code。 Codex CLI 虽然不好用，但是模型好，写代码稳，并且 Token 量大。 Claude Code 好用，模型写代码不如 Codex，但是通用任务执行的更好。 这里其实 OpenAI Codex 团队是要反思一下后续走向的，空费了这么好的模型，整天做些华而不实的更新。
除了 Conductor，又出现两新的 parallel agent（多 agent 并行） 客户端软件。 - https://t.co/Ci3iMxVxhR - https://t.co/iC388PSFgd - https://t.co/yAiPAUE7Xj https://t.co/qVdQM2fpEl
Guohao Li 🐫
One (open) secret why (terminal) coding agent RL training is easier than other domains is the structure in rewards. You can do many things with it. Here is how we design the reward function. We followed GRPO training setup. The binary 0/1 reward based on task success or failure is too sparse for calculating advantage (the model doesn't learn from its near-misses), so we turn to calculate the unit test pass ratio for each rollout as reward. An additional bonus reward 1 was assigned to fully succeeded rollout to encourage completing the tasks. There are even more you can do with the reward shaping. For example, you can tune the weights of the unit tests based on their structural dependencies which could be considered as a graph. By doing that you can encourage the model to learn the most critical unit test to pass first and then succeed in the other tests. Will update more on this once we get some GPUs to finish the remaining experiments
Est. 100 views for your reply