KKY

Data scientist | Kaggle Global Rank 35th | 分享多模态、LLMs、生成视觉模型的实践和思考

[KKY @evilpsycho42]

Codex + gpt 5.5 任劳任怨，头脑灵光，手脚不利索
Claude Code + opus 4.8 很机敏，但是爱摸鱼
Claude Code + DeepSeek-v4-pro 蠢得眼神清澈，但不撒谎...

ds乱删除后，codex还是把屁股擦干净了。当你codex没额度换其它model + harness的时候，慎重

[KKY @evilpsycho42]

Codex 昨晚goal跑没额度了，而且怀疑酱汁或者5.5被路由到了5.4-mini, repo残留了很多东西。 不得已接claude code + deepseek-v4-pro 来收拾下残局

结果，NM直接删除了昨晚Codex还没提交的几个有效solver文件，然后它开始尝试逆向pyc回复，开始无限tool call error.

我QNM，从5.4/4.6开始我都是用的yolo模式，今天被恶心到了。还好只是几个solver files, 还可以尝试从codex session files里恢复

ds是酱汁了吗？我走的官网API

[KKY @evilpsycho42]

Agent on Kaggle

我不确定和前排agent only teams的差距是harness还是 token数量，所以准备充值试试。

万一就是token数量呢？

目前效率是 7% codex prox20周限额 = 50分

[KKY @evilpsycho42]

I am sorry but this implementation is not dynamic workflows.

The dynamic, emergent nature of later phases (where phase N's shape is determined by phase N-1's output) is a defining characteristic of Claude Code's workflow model, and it's one of the places where pi-dynamic-workflows currently takes a simpler, more static approach.

[KKY @evilpsycho42]

@antirez m5ultral would be the killer

[KKY @evilpsycho42]

Hermes goal的实现，很好，我讨厌随便注册tools.

- No additional agent tools
- Goal state judged by another model
- Simplified mental model

[KKY @evilpsycho42]

用codex做kaggle目前感觉几个问题:
1. subagent + goal 一起使用并没有起到明显的作用，上下文和调度效率都比较低
2. 主agent带宽瓶颈明显，和codex harness以及 tps都有关系
3. 400子任务优化过程中，实验记录数据/no-go-note/...逐渐爆炸，哪怕加入了裁剪机制

打算：
1. 重构repo, 对子任务上下文空间进行深度隔离
2. 以pi agent为基座，深度融合/goal /autoresearch /kanban 功能，正在研究harnes agent对应实现，需要做一些针对性优化
3. 对于io层agent采用tps更快的模型...
4. 尝试在worker层级做多种agent harness + model组合拓宽思路

目标：
提升单位时间和单位token对应的score提升效率

[KKY @evilpsycho42]

@Teknium OpenSource must win, again

[KKY @evilpsycho42]

gemini 3.5 flash 傻得让我想清仓GOOG

[KKY @evilpsycho42]

背景是个人量化场景数据收集任务
对比：
codex + gpt-5.4-mini low
pi + deepseek-v4-flash think on

结论对比如图

纯用apoli肯定ds价格是5.4-mini的1/20，但是如果考虑订阅的话，gpt-5.4-mini似乎性价比还更好，200刀订阅一个月可以跑8000刀以上，1/40 api价格。

[KKY @evilpsycho42]

Manage skills across multiple agents and devices
- personal skills
- npm skills managed thirdparty skills

https://t.co/KmDXSDhLiA

[KKY @evilpsycho42]

@zcbenz amazing. so all CUDA ecosystems will OPEN the door to Mac?

[KKY @evilpsycho42]

昨天我优化了一下autoresearch skill，让main agent尽量利用subagent 进行并行研究，采用类似于beam search的思路维持多个研究路线。

今早起来我看到lb提升稳定但是幅度很低，我以为进入瓶颈了，结果分析发现是之前某个限制导致agent只提交本地最好的144个tasks结果...

去除限制后可以的最新提交比昨天提升了接近1000分，还是很不错了，目前应该还远远没有饱和，接近饱和后我会开放其访问kaggle discussion和notebook的权限

[KKY @evilpsycho42]

/ralph loop => A script maintains the loop and force the inner agent to solve a single task per iteration.

/autoresearch => The vanilla autoresearch requires the agent maintain the loop and optimize a slice of code per iteration. Some models are struggling of maintaining the loop after 5~10 iterations.

/goal => lightweight /ralph + /autoresearch

Pi Agent is easy to implemente all of them through its extensions system.

@badlogicgames

[Mario Zechner @badlogicgames]

looked into /goal in both cofex and claude codr and all i'm seeing are inferior versions of autoresearch. what am i missing?

available from the @ShopifyDevs folks as a pi extension.

https://t.co/ml6PfaS6UM

[KKY @evilpsycho42]

@badlogicgames @ShopifyDevs basically yes, /goal an inferior version of auto research. but it solves how to force the agent iterates in the loop by sending continuous prompt.

which is more lightweight for user than writing the loop script. some models can't maintain the loop by itself.

[KKY @evilpsycho42]

ds-v4-flash is struggling with invoking a customized `send_message` tool. I defined 10 eval examples and splitted them into `train` and `test` and ask the agent optimize tool description, guidelines to make it work in `train`.

After 10 iterations, the model still can't pass the train cases... so i ask the agent to include the tool_name into optmization scope ....

Does the agent overfit on specific harness too much? or i missed some keypoints?

[KKY @evilpsycho42]

Codex is leading the trend : )

@thsottiaux what's the next ^ ^

[Daniel San @dani_avila7]

Claude Code 2.1.139 added /goal

You set a completion condition and Claude keeps working across turns until it's met

Works in interactive, -p, and Remote Control 👏

[KKY @evilpsycho42]

@yihui_indie 后台agent监控调用频率定期自己清理归档，global统一放.agent/skills

项目级别的不做处理，除非escalate到global

[KKY @evilpsycho42]

Understand Pi Agent system prompt in a single image

[KKY @evilpsycho42]

/goal + autoresearch

The agent has been working continuously for 24 hours and keeps climbing the leaderboard. The goal is set to reach 5,000 points on the leaderboard.

I don’t know how long this will take. let's watch.

BTW, I've restricted its network access.