@hugues_va: JEPA has a huge blind spot no one talks about: it’s starving for stable, slow features. Add a SINGLE irrelevant patch, …
Summary
JEPA 有一个未被讨论的严重盲点：它严重依赖稳定、缓慢变化的特征。只要在每个 episode 中加入一个固定不变的无关 patch，规划就会崩溃；若让该 patch 不可预测，性能反而几乎恢复正常。
View Cached Full Text
Cached at: 10/01/26, 08:30 PM
JEPA has a huge blind spot no one talks about: it’s starving for stable, slow features.
Add a SINGLE irrelevant patch, keep it fixed within each episode, and planning falls apart. Make it unpredictable, and performance almost returns to normal. Wild!
More in a blog post soon. https://t.co/8nDNgtDiAF
Similar Articles
@realfxw: Lately, when running long tasks with Claude Code, do you often feel like the 5-hour quota is draining away like water? …
一位推主根据 Anthropic 工程师 Lydia Hallie 的官方解释，澄清了 Claude Code 的 auto-compact 机制：默认在接近 967K 上下文时才触发压缩，导致长任务持续携带巨量上下文而快速消耗 5 小时配额。文中建议通过 /autocompact 400k、autoCompactWindow、promptCacheTtl 等配置提前压缩并优化缓存。
@Vincent_AINotes: Anthropic officially demonstrates a more stable way to write prompts: don't mix background, materials, rules, examples, and output requirements into one paragraph. Instead, assign each type of information a clear position and use XML tags to delineate boundaries. The core idea is not to make prompts longer, but to minimize what the model needs to guess. I've compiled...
Anthropic officially demonstrates a method to structure prompts using XML tags, separating background, materials, rules, examples, and output requirements to reduce guesswork and improve stability.
@jakevin7: Current models all seem to have some serious issues—they've caught comment-mania syndrome.... This leads to big problems: lots of outdated information gets injected into the code as context, and this is done autonomously by agents, making it uncontrollable.
The author complains that current AI models generally have a tendency to over-comment, leading to outdated information being injected into code context, and this is done autonomously by agents, posing a risk of losing control.
@dongxi_nlp: Planning to dive deep into J-space this week. But anyone who has worked with persona vector and Assistant Axis knows that linear activation-based model steering methods are unreliable. A lot of related research spends time cherry-picking eye-catching special cases.
The author expresses skepticism about linear activation model steering methods like J-space, arguing that related research often relies on cherry-picked special cases and is therefore unreliable.
When working on agent workflows or autonomous agent loops, a major bottleneck affecting experience is: every small branch decision, tool routing, task completion check, or safety interception, if all handled by expensive reasoning LLMs, not only brings seconds of delay per call, but token costs also...
This article introduces the open-source list awesome-jev-projects, which decouples System 1 fast decision models from System 2 reasoning LLMs to optimize AI agent workflows, reduce latency and costs, and provide code-level references.