sda4

資訊工程 → AI & Code
Exploring intelligent systems
mounted on /home/sda4/
📧 [email protected]

[0xWulf @hexawulf]

I was still using Fable 5 yesterday. today the model simply says "not available".

If model access can simply be switched off by jurisdiction, we as global users no longer share the same
AI productivity layer.

The US and cleared institutions get the frontier, everyone else gets weak tea. A two-class AI world is not only coming, it has already appeared today in the model selector.

For Europe, the lesson is brutal but familiar: over-regulation with zero sovereign compute, models and cloud infrastructure leaves you dependent on the US or China's export-control switch.

[Peter Steinberger 🦞 @steipete]

Here’s your monthly reminder that you shouldn’t be prompting coding agents anymore.

You should be designing loops that prompt your agents.

[0xWulf @hexawulf]

Qwen3.7-Max is a real milestone for Chinese AI.

On Code Arena, Alibaba's latest model is now being ranked in the same top coding tier as Claude and ahead of several frontier models from OpenAI, Google, Zhipu and Moonshot.

What matters is the method: blind human comparisons of real code outputs, not just synthetic benchmark flexing. You can see the live Code Arena coding leaderboard here: https://t.co/bpokNehKL0

LLM Stats coding leaderboard tracks performance across practical tasks like React apps, games, data viz, 3D scenes, SVGs and animation. That’s much closer to how developers actually use these models.

The takeaway: the US still leads the top of the AI coding stack, especially via Anthropic. But the gap is no longer uncontested. Qwen, DeepSeek, GLM and Kimi show that Chinese labs are now competing across the frontier — not as copycats, but as serious model builders.

For developers, that’s good news: more competition, more API options, and faster pressure on price/performance.

[Garry Tan @garrytan]

Thin Harness / Fat Skills is actually a powerful mindset shift that helps you reduce your bugs in agentic flows

[Peter Steinberger 🦞 @steipete]

Me and codex were busy.

🔊 https://t.co/kAbQGMTQIQ — Sonos
🗃️ https://t.co/okyk5oZOSZ — WhatsApp
🪶 https://t.co/IOOLpksihC — X archive
🧰 https://t.co/8pYSuKt0Ea — GitHub archive
🛰️ https://t.co/MErsuc1FO7 — Discord archive
🎧 https://t.co/E4FCEXEixU — Spotify
💬 https://t.co/cN5M5iRiQs — iMessage
🧳 https://t.co/f6WRrsXVaj — MCP to CLI
🗣️ https://t.co/aYKEugKwUC — ElevenLabs voice
🧿 https://t.co/q6DP88loU9 — second opinion

Upgrading the 🦞 OpenClaw army.

[Prasenjit @Star_Knight12]

CLAUDE OPUS 4.7 USING 500K TOKENS TO RENAME A VARIABLE

[Mitchell Hashimoto @mitchellh]

Ghostty is leaving GitHub. I'm GitHub user 1299, joined Feb 2008. I've visited GitHub almost every single day for over 18 years. It's never been a question for me where I'd put my projects: always GitHub. I'm super sad to say this, but its time to go. https://t.co/DQDemHdytV

[Peter Steinberger 🦞 @steipete]

I smell a leak.

[Kyle Chan @kyleichan]

Must-listen interview by @Changxche with ex-ByteDance AI researcher:
- Benchmaxxing
- Distillation on US models
- Poor data quality and infra
- Compute constraints

"I don’t even agree with the assumption that Chinese models are catching up — I believe we’re still far behind. I guess the gap is getting larger, very sadly." https://t.co/OC2o0NV4eI

[Chubby♨️ @kimmonismus]

Yeah, GPT image 2 is *that* good.

It’s just so freaking accurate.

Image: a 20 person horde raid is fighting Sam Altman in 2004 world of Warcraft style. One shortted.

[Crystal @crystalsssup]

Kimi K2.6, let's goooooooo🥳

Open-source SOTA on HLE w/ tools (54.0), SWE-Bench Pro (58.6), SWE-bench Multilingual (76.7), BrowseComp (83.2), Toolathlon (50.0), Charxiv w/ python(86.7), Math Vision w/ python (93.2)

[Researcher_王十三 @ResearchWang]

https://t.co/KQqwgy2XhL

[Yutong @gnotuy]

We open sourced Kimi K2.6. The next frontier in test-time compute isn't bigger models. It's better organizations of intelligence.

The hardest things were never built by one person. They require coordination. Different skills, different contexts, different minds arguing until something better emerges.

[Teknium 🪽 @Teknium]

Hermes Agent just hit 100,000 stars on GitHub!!!

Thank you everyone!!

[Greg Brockman @gdb]

codex is becoming a full agentic IDE

[rari @0xwhrrari]

OpenAI just mass-fired their robotics team. One of the engineers DM'd me 20 minutes later.

I didn't know him. He found me through a Polymarket thread.

His first message:

"I have 30 days of severance and nothing to lose. Let me tell you what we actually use internally. It's not GPT"

I thought he was trolling.

"Every serious team at OpenAI prototypes on Claude Code. Not ChatGPT. Not the API. Claude Code connected to a repo. That's the actual workflow"

I asked why.

"Because Claude reads the codebase. GPT reads the prompt. There's a difference. One guesses. The other one understands the full context and builds on top of it"

He sent me one link.

https://t.co/X1p0WUtEqy

86 million Polymarket trades. Every wallet. Every entry. Every exit.

Open source. Free.

"Point Claude Code at this. Say - find every wallet with 70%+ win rate and 100+ trades. Watch what happens"

I did it that night. Claude pulled 47 wallets in 4 minutes.

Average profit: $214K. Hold time: 7 hours.

91% exit BEFORE resolution. They never wait for the outcome.

"Now look at how they exit"

Top wallets capture 86% of the move and cut at 12%.

Everyone else captures 58% and holds losers to 41%.

Same entries. Completely different results.

He sent another link.

https://t.co/VvcsrL29AO

"Three commands. Your bot sees 500+ markets. No key needed. Read-only. Claude scores them in 20 minutes"

I asked why he's telling me all this.

"Because I just got fired for saying we should open-source more. So here I am open-sourcing everything I know"

Then he sent me an article where someone built the full bot from these repos in a weekend -> https://t.co/VvXrNhc5d4

Three exit triggers:

Target 85% of move. Volume spike x3 - smart money out. 24h silence - thesis dead.

I copied the stack. Claude Code $20. VPS $5. $25/month.

No team. No office. No GPT subscription.

17 days. 191 trades. 73% win rate. $850 seed.

+$9,400.

I sent him my results. He replied:

"This is exactly what I built as a side project at OpenAI. They made me delete it"

I asked if I could post this.

"Post it. What are they gonna do. Fire me again?"

[Sigrid Jin 🌈🙏 @realsigridjin]

yesterday I read Moshe's thread (@Mosescreates) about his Hermes setup

he mentioned six agent profiles, one shared self-hosted memory store, Claude Code hooked in so coding sessions write into the same memory the agents read from

I spent days building a stripped-down version. Two profiles instead of six, one machine instead of two, but the core architecture holds

The interesting piece is the shared memory layer that lets Claude Code sessions feed directly into agent context

[Garry Tan @garrytan]

OpenClaw is the most significant innovation since ChatGPT and Claude Code

It’s the third shockwave that will be felt by all