Stoic Trooper

America, Freedom, Equality of Opportunity, Common Sense...let's use our brains to solve problems instead of exacerbating them like idgits

[Chubby♨️ @kimmonismus]

Qwen 3.8 Flash-Next official released: A 6B-active open model just beat Claude Opus 4.6 Max across 8 of 9 comparable benchmarks!

Qwen3.8-Flash-Next is a highly sparse MoE:

• 125B model parameters
• 51B additional n-gram embeddings
• Only 6B parameters active per token

It scores:
• 62.5 SWE-bench Pro
• 81.0 SWE-bench Multilingual
• 73.9 CoworkBench
• 55.7 JobBench
• 73.5 Toolathlon
• 81.3 IFBench
• 91.7 GPQA Diamond
• 91.9 LiveCodeBench

It also outperforms Qwen3.8-27B and DeepSeek-V4-Flash across most of the table.

Super cool release!!

[GBC @GBC_Press]

🚨 BREAKING:

US Rep. Thomas Massie reveals that Jewish billionaire Les Wexner was Epstein's first client and partner to be named in a child trafficking case.

He says: "Russian Jews working with Israel are trying to destroy America."

[Daryl McKenzie @theboysofdozier]

At this point, Thomas Massie is the most liked republican in the country. https://t.co/ZMFQgYKeCC

[Cruz @ViC305]

8GB RTX 4060 Laptop. ~16GB Ornith IQ3_S GGUF. 44.5 tok/s streaming locally.

FreeToken didn’t have the qwen35moe GGUF path needed for this model, so I implemented it and submitted PR #131 upstream. 👀

I also exposed the K/I GGUF quant types its CUDA kernels already supported through the Python loader.

This is the result.

Setup:

Ornith-1.5-35B-A3B
35B MoE / ~3B active
IQ3_S GGUF · ~16GB

RTX 4060 Laptop · 8GB VRAM
FreeToken · FlashInfer
MoE backend: offload

On the attached full-prompt run, I measured ~44.5 tok/s streaming on screen.

In the same runs, FreeToken’s server decode counter reported 46.7 → 50.1 tok/s.

VRAM stayed at 6,879 / 8,188 MiB, with 20GB host RAM and 98% GPU utilization during generation.

Load time: 65s.

My llama.cpp CPU reference on the same model was 11.1 tok/s.

The interesting part isn’t just “35B on an 8GB GPU.”

The code path itself was missing.

FreeToken’s CUDA side already supported more GGUF quant types than its Python loader exposed, including the K/I quants needed for this IQ3_S model.

And qwen35moe models like Ornith needed their own GGUF architecture/loading path.

I implemented both pieces in my fork and opened them upstream in PR #131.

Once loaded, FreeToken can use GPU + CPU + host RAM as an elastic MoE inference system instead of requiring the entire checkpoint to live in VRAM.

That fits Ornith well because only a subset of its experts are active per token.

Important caveat:

PR #131 is still open. This is not merged upstream yet.

I’m not calling this official Ornith support.

I’m showing a working implementation, the code is public, and the attached terminal video is the actual run.

PR #131:
https://t.co/uqZkjjaWAT

My fork:
https://t.co/Dcr4fruj4A

Next I’m going to test where this path holds up across more quants, longer context and real agent workloads.

[Candace Owens @RealCandaceO]

The Pentagon has been waging a PR war against me for investigating Charlie Kirk’s assasination.

Totally normal stuff.

[Yohei @yoheinakajima]

replaceable models + thin harness + modular company brain

[DAIR.AI @dair_ai]

Interesting new approach to enable memory in long-running agents.

Weighted Memory Tree organizes execution into tasks, subtasks, and actions, then gives every memory a retention score that moves. Event based updates raise it, selection based decay lowers it.

When a subtask finishes, its step by step detail collapses into a short summary and the full version stays retrievable. If a later step needs the details, the agent pulls them back.

Context trimming is usually permanent. Drop the wrong turn and the agent has no way to recover it. Folding gives you the same token savings with a way back.

On GAIA-Text with Qwen3-8B, Gemma 4 E4B, and Llama-3.1-8B, it beats linear memory by 9.97 points on average while using 32.8% fewer prompt tokens. Memory poisoning experiments show retention scoring limits how far unreliable information spreads.

Paper: https://t.co/2huRNYeOMz

Track more trending AI papers in our academy: https://t.co/LRnpZN7L4c

[Arpit Neewaliya @arpitneewaliya]

The most exhausting part of using AI for coding right now isn’t generating the code.

It’s the constant context switching:

- Explain the project again
- Paste the same files again
- Correct the same misunderstanding again
- Re-explain your coding style again

I want persistent project memory more than I want better models.

[Teknium 🪽 @Teknium]

Whoever claimed this is absolutely wrong. My Hermes runs 15+ subagents all the time whenever it makes sense to... We absolutely have no "1 agent in parallel limit"

I run 25~ sessions in parallel, each one using as much as 15-25 subagents per session.

We have no limits. This is just wrong lol

[The Startup Ideas Podcast (SIP) 🧃 @startupideaspod]

https://t.co/Baax4qVI0G

[Stoic Trooper @StoicTrooper]

@yama_boku Yes I concur

[Stoic Trooper @StoicTrooper]

@tunezbash @aka_govind Right on! “I’ve come to the right place!”😄

[Govind @aka_govind]

I want more builders on my timeline.

Not gurus.

Not people telling me how to make $100K/month.

People actually building stuff.

Where are you?

[The Disrespected Trucker @DisrespectedThe]

I'm not going to stop until I wake everyone up.

[Mark Ruffalo @MarkRuffalo]

The accusation that I am antisemitic is appalling and fundamentally dishonest. Criticizing the actions of the Israeli prime minister, a military technology contract, or the executives who supply it is not the same as criticizing Jewish people. This critical and necessary dialogue is then dishonestly framed as being anti-Israel. To be clear, my views come from my own political convictions and should never be interpreted as hostility toward Jewish people, for whom I have deep love and respect.  Everything I know about acting, activism, and humanism has been profoundly shaped by the Jewish friends, colleagues, and loved ones who have been integral and family throughout every point of my life.

This merger has real consequences for real people, and for the entire country. Scrutinizing the Ellisons, including Oracle’s business built on data, surveillance technology and government contracts, and the serious threat to editorial freedom and the loss of a livelihood for thousands of families, is fair and necessary. The $111 billion deal would hand one family control over CNN, HBO and Warner Bros., backed in part by foreign money whose influence on editorial decisions has never been fully explained to the public. 

Lawmakers on both sides of the aisle have called for serious national security review, and regulators still haven’t given the public a real answer. Until they do, the merger shouldn’t move forward. Now is the time to dive boldly into all these issues, not step back or concede.

[Stoic Trooper @StoicTrooper]

@theo @henrypl_dev Would you be willing to share the setup with instructions for AI on a GitHub repo?

[Mario Nawfal @MarioNawfal]

This guy built over 1,000 homes for squirrels and fitted some of them with cameras so he could see how they live

Welcome to the cutest reality TV show out there

Writer: Ian

[Stoic Trooper @StoicTrooper]

@chamath “To stop AI deepfakes” just like we need policed internet access “to save the children” lol

[Max For AI @MaxForAI]

爆了，Pi 居然在他们新的开发笔记里把其他家的Agent都锐评了一下🤣

@pidotdev 官方表示，DSH、Claude Code、OpenCode 这一类方案，它关注的已经不是谁多几个 Tool、谁的 Demo 更炫，而是一个更麻烦的问题：

如果一个 Agent 连续跑 50 个小时，它还知道自己到底干过什么吗？

他们先点了 DeepSeek Harness。

Pi 直接从 DSH 的 Context 管理开刀。
DSH 有一个 tool result pruner，工具输出太长以后，会保留前面一部分和后面一部分，中间直接裁掉。

Pi 那边给的评价很不客气：permanently-lossy pruner。

因为删掉的内容以后找不回来了。
比如一个 Agent 跑测试吐出 3 万字符日志，真正的错误刚好在第 12000 个字符，压缩完以后这段已经消失。后面模型再聪明，也没法从空气里重新读出来。

所以 Pi 做的版本是 prune + spill。

Context 里依然只放一小段，完整 Tool Result 同时落到磁盘，Context 留文件路径和 offset。模型需要的时候自己 grep、sed、read 回去。
他们拿 GLM-5.3 和 DeepSeek V4 Flash 跑了 19 个真实 Session，Context 占用降低 26%～35%，uncached prefill 降低 72%～88%，信息还能恢复。
PR 里甚至直接写strictly better than dsh’s permanently-lossy pruner😂。

然后又点了 Claude Code。

Pi 表示有一个挺有意思的发现：
新一代 Claude 在第三方 Harness 上，有时候反而更容易把 Tool 调错。
最典型的情况是，Pi 给模型的 edit tool schema 明明写得很清楚，模型却会自己生成 requireUnique、matchCase、oldText2 这种 Pi 根本没有的参数。
一个很合理的猜测是，Claude 的后训练越来越适配 Claude Code 自己的 Harness 和 Tool Schema。
模型已经学会了 Claude Code 里某个工具应该长什么样。
然后你把它放进 Pi，它脑子里还是那一套。
所以现在模型能力越强，模型和 Harness 的绑定可能也越深。
以后评价一个 Coding Model，单独看模型可能越来越没意义。
Claude + Claude Code、DeepSeek + DSH、GPT + Codex，很可能逐渐变成一个整体。

讲到这里 Pi 表示他们最近重写了 Harness，开始把一次 Agent Run 当成接近数据库事务的东西来设计。

Tool Call 执行之前，先写 intent。

执行结束，再写结果。

进程崩溃以后，系统得知道这个 Tool 到底执行过没有。

Session Tree、运行状态、Lane、Operation Log 分开保存。

一个 Agent 可以有多个 Lane 并行跑，每个 Lane 都知道自己的当前位置、队列和未完成操作。

Context 允许压缩，原始执行历史继续保留。

甚至连进程刚好死在 Tool 执行到一半这种恶心情况，他们都专门设计了 recovery。

你再回头看现在很多 Agent Framework，会发现大家以前真的挺勇的：

一个 while loop，塞一个模型，塞几个 Tool，Context 快满了就总结一下，然后祈祷它一直别挂。

跑十分钟当然没什么。

Agent 开始跑几小时、几天以后，问题全出来了。

这些已经越来越像操作系统和数据库问题了。

有意思的是Pi 甚至对 Extension 的态度也在变。

现在大家都喜欢说「一切皆插件」，DSH 的 Cordis 就走得很远。

Pi 这边最近却越来越强调边界：

哪些是 Conversation，哪些是 Runtime State，哪些是 UI，哪些是 Extension，哪些东西能持久化，哪些东西只能观察。

因为插件能力一旦无限扩大，Agent 又长期运行，状态会变得非常难推理。

OpenCode、Claude Code、DSH、Pi 现在看着都是 Coding Agent，底层其实已经开始走不同路线了。

Harness 这东西，开始越来越像 AI 时代的操作系统了🤔

[BRICS News @BRICSinfo]

JUST IN: 🇹🇷🇮🇱 Turkey issues arrest warrant for Israeli Prime Minister Netanyahu.