Gagan Arora

Building Foolish.
Making AI that watches your screen, not waits for your prompt.
Ship more, think less.

[Gagan Arora @gaganfoolish]

Chat is the wrong UI for most AI work.

When people get stuck, they usually don’t know what to ask.

They’re just staring at a screen where something isn’t moving.

I’m building Foolish: AI that sees your screen and points to the next step.

Follow for the first demo.

[Gagan Arora @gaganfoolish]

@alexandr_wang cute

[Alexandr Wang @alexandr_wang]

excuse me I must correct the record on this slander

[Gagan Arora @gaganfoolish]

Case study: I'm selling a 6,000 sq ft commercial holding in Gurugram and my agent runs the dealer outreach on WhatsApp.

Its rules: one channel per prospect. Accepted ≠ delivered ≠ read ≠ reply. Stop on opt-out. No follow-up without my sign-off.

Score so far: 6 sends, 0 offers.

[Gagan Arora @gaganfoolish]

Lesson: a worker showing "online" means nothing.

Mine had a heartbeat 8 hours stale and one run stuck on "database is locked". The fix wasn't a restart. It was a supervisor that owns every process lock, plus one rule: reconcile the unfinished run before retrying anything.

[Gagan Arora @gaganfoolish]

Today my assistant called me. A real WhatsApp voice call from the India number, reading out 7 task updates and the 3 decisions it needs from me.

First attempt rang out after 50s. Second connected. The log has every state with a timestamp. That's the only proof I trust.

[Gagan Arora @gaganfoolish]

@LDT0545 no reset yet?

[Julian Assange @JulianAssange]

I'm back.

[Browser Use @browser_use]

holy cheap

[Gagan Arora @gaganfoolish]

@typesafeai [email protected]

[Gagan Arora @gaganfoolish]

I run an AI assistant called Astra that does my actual chores: reads my WhatsApp, calls me with task updates, runs scheduled workers off a kanban board.

This account is now the operator log. What worked, what broke, with evidence. No revenue claims until there's a receipt.

[Gagan Arora @gaganfoolish]

@jack @AnnDiegoUSA you're awesome, love the work with buzz

[Travis Fischer @transitive_bs]

OpenAI's board finally realizes what Ilya saw

[Yann LeCun @ylecun]

@PessimistsArc Right. Dario was already claiming that GPT2 was too dangerous to open source back in 2019.
I made fun of them then.
Everyone should make fun of them now.

[Gagan Arora @gaganfoolish]

@Nipi_220 10

[Max For AI @MaxForAI]

🔥NVIDIA 开源了他们的Harness：SoL-Pi

刚刚看到 NVLabs 开源了 SoL-Pi（Scaling Auto-Research Loops for Efficient Agent Harnesses）

它不是从零做了一套新的 Agent Harness，而是基于 Pi 做了一套 Harness 效率增强层，然后用 MIT License 开源。

这个项目的重点是 @NVIDIAAI 在让 Agent 自动研究怎么优化 Agent 自己的 Harness。

他们搭了 535 个可执行训练环境，其中 495 个来自真实 GitHub Issue-PR，另外 40 个是带 Verifier 的合成任务，然后让 Auto-Research Loop 不断提出新的 Harness 改进方案、实现、跑实验、验证，最后只有大约 1/40 的想法能活下来。

最终这个管线自动找出了四个优化：

- Action Fusion：改完代码直接把测试/运行合并进同一次 Tool Call，减少一次模型往返。
- Online Context Compact：不等 Context 快爆了再压缩，而是在子任务完成节点主动判断是否 Compact。
- ObservationPack：巨大的 Tool Result 不再每轮塞回 Context，只保留索引，需要时精确取回。
- Evidence-Preserving Reducer：长 Log 先交给更便宜的 Agent 阅读压缩，同时对引用证据逐条验证。

效果也非常好：
相比原始 Pi，SoL-Pi Token 使用量降低 45%–49%，成本降低约 1/3，同时保留约 94% 的平均任务得分；
相比模型自己的原生 Harness，Token 少 35%–64%，按 API 标价计算成本低 50%–54%。
在 GPT-5.6 Sol 上，SoL-Pi 的 EdgeBench 得分甚至超过了原生 Codex Harness。

这其实和现在 DeepSeek Harness 的方向撞到一起了：

模型越来越强之后，下一层 Scaling 开始发生在 Harness。

但Nvidia的SoL-Pi又往下走了一次一层：
不只是人优化 Harness，而是让 Agent 自己跑 Research Loop，找到更好的 Harness，再拿这个 Harness 去跑下一轮 Research。

他们自己给这个方向起名叫Efficiency for Efficiency。

更高效的 Harness → 更便宜的 Auto-Research → 同样预算跑更多实验 → 再找到更高效的 Harness。

模型之外的 RSI，已经开始有点具体的样子了。

地址：https://t.co/veFbs9PRfM

[Gagan Arora @gaganfoolish]

A useful tip for codex users, dont use the desktop app it burns tokens, use the models in the terminal

[Ryan Carson @ryancarson]

This post has 100,000,000 views

[Gagan Arora @gaganfoolish]

@AnthropicAI is getting cooked in public before their IPO

[Love @jetfam]

🚨 Cute animation. Shame the real problems still aren’t moving.

Claude Code users are hitting brutal limits, Max customers are furious, and support is still a mess.

Maybe spend less time polishing the mascot and more time fixing the product.

@bcherny @AnthropicAI @claudeai

[Mr Tactical @MrTacticalX]

🚨CLAUDE MAX USERS EXPAND LAWSUIT OVER USAGE LIMITS

Claude Max users have expanded a proposed class-action lawsuit against Anthropic. They say that the $100 and $200 plans provide far less usage than buyers would expect from plans called Max 5x and Max 20x.

Anthropic says the session and weekly limits were explained in pages linked during purchase. The claims have not yet been proven in court.

Anthropic is in trouble

Follow for more AI news.