vanderseele

software engineer

[Unsloth AI @UnslothAI]

Qwen3.8 can now be run locally! 🔥

We shrank Qwen3.8-2.4T-A95B from 4.9TB to 397GB (-91% size) via Dynamic 1-bit by selectively quantizing layers

Run on 410GB+ RAM/VRAM via Unsloth Desktop

Qwen3.8 rivals GPT-5.6 Sol

Guide: https://t.co/J2PwgMP6GZ
GGUF: https://t.co/7K1KdtnO2I

[clem 🤗 @ClementDelangue]

Let's go!

https://t.co/BAJlYIJqDQ

[David Sun @arcticinstincts]

Nobody:

Absolutely nobody:

US military in 1943: Hey soldier, you know what we have in common with the Chinese? We both enjoy racist jokes about the Irish

[Dwarkesh Patel @dwarkesh_sp]

The least verifiable part of AI R&D (and the thing which may bottleneck the automated researchers) is making calls on large experiments and big training runs.

Interestingly, this may also partly explain why model parameter scaling has been slow for the last few years.

Even for human AI researchers, it's better to run more experiments and do training runs at small(er) scale, and make up for the benefits of scaling with doing research faster and thus finding more compute efficiency gains.

[Raffi @raffichill]

Shipped by Cursor team btw

[luffy @0xluffy]

capy v1 rewrote v2 from scratch in 3 weeks

v2 will build v3 in 3 days

try it out now https://t.co/PJqpzheVBh

[justin @justinsunyt]

Introducing Capy v2: the fastest, most powerful cloud coding agent

We beat Claude Code, Codex, Devin and Cursor on DeepSWE while being 50% cheaper

Spawn RLM agents, each with up to 32 vCPU/128GB RAM and <1s VM boots

Bring your Codex/Grok subs

Comment for $100 in free credits

[Luke Berry @LukeberryPi]

this is how i stop AI slop

[Zack Kanter @zackkanter]

Mostly solvable. Have agents incorporate the following:

- https://t.co/DTtzJMKc9V
- https://t.co/AbtH3PYRX3
- ASD-STE100 / Simplified Technical English + Google developer documentation style guide

[Neel Nanda @NeelNanda5]

This was a very satisfying project. An annoying problem with J-Lens is that errors accumulate as you backprop through many layers and it's highly ineffective at early layers. A simple, cheap tweak to J-Lens makes it perform much better, using layerwise relevance propagation!

[Alexander Panfilov @kotekjedi_ml]

We can finally talk about it:

We found a way to extract hidden reasoning of frontier models using a vulnerability in the APIs of every frontier AI company.

We verified that our reasoning token count matches billed API thinking tokens 1:1 for most of the prompts we queried.

[Crémieux @cremieuxrecueil]

@tszzl Drink the protein soda!

https://t.co/63ozZ116sB

[ADONIS @adonispara]

> jails communists without trial
> believes intelligence is hereditary & promotes eugenics
> forces every worker to save up to 1/3 of their income whether they like it or not
> takes that forced savings and uses it to build housing, healthcare and infrastructure
> Singapore ranks first in the world for healthcare
> turns renters into owners (over 80% of Singaporeans now own their house)
> makes corruption a career ending, prison sending offence (no matter who you are)
> death penalty for drug trafficking (no exceptions)
> GDP per capita goes from $500 to $56,000 in one generation
> now one of the richest nations on earth with one of the strongest passports

You can just do things.

[Gabriella @gabrielladb_]

For anyone looking to learn more about the Dutch powerhouse, I highly recommend reading ‘There’s A War Going On But No One Can See It’ by @huibmodderkolk

[Paul Klein IV @pk_iv]

Stagehand v4 has one mission: Destroy Playwright.

Stagehand now runs INSIDE the browser, enabling the same speed as a local browser but at massive scale.

If your agent uses the Playwright MCP, you'll get a 2x perf boost just by switching.

https://t.co/hbBMSZFtly

[GD44 @GD44vh]

@WhiteHouse Can you post this next

[doomer @uncledoomer]

>italys financial system is party backed by cheese

[Synara @trySynara]

Synara v0.7.1 is out 🪽

🔀 Create PR now handles the full flow: commit, push, and open the pull request in one guided action

⚡ Faster startup and reconnects for large histories, with smarter replay, polling, and thread subscriptions

📝 Explorer previews are now editable, saveable, and safer with dirty-state tracking, validation, and clearer breadcrumbs

🌿 Live Git metadata now follows branch, worktree, push, and PR changes while agents work

🛠️ Worktree setup now shows progress, supports cancellation, and lets you open a local checkout

📊 Provider usage is cached server-side and shared consistently across the sidebar and Settings

🧠 More reliable Codex, Claude, OpenCode, Grok, Kilo, Antigravity, and ACP sessions

🔄 Better provider recovery, cancellation, resume handling, AskUserQuestion recovery, and handoff eligibility

💬 Loading, Working, and takeover states now reflect the real task lifecycle

🧵 Codex thread forks now work for imported history while preserving source-thread provenance

🎨 Custom desktop app icons, native-style macOS artwork, and a visual theme picker

🫧 Refined translucent sidebar and floating composer surfaces with stronger production fallbacks

⌘ New shortcut to copy the active thread ID, plus improved project search and keybinding workflows

🔔 Release history is now available directly from the sidebar Help menu

🖥️ Terminal Add to chat now targets the correct conversation, with safer exits and better side-chat shortcuts

📚 Improved Markdown table rendering, image overlays, PR comments, and diff color presentation

🍞 Clearer error toasts, better notification timing, and fewer stuck spinners across the app

[dan @irl_danB]

thank you to Prime Intellect for finally building me a harness that lets me do this without having to pull the guts of the thing apart

[Unsloth AI @UnslothAI]

2-bit Muse Glimmer GGUF managed to call 100+ tools on just 14GB RAM. 🔥

Muse Glimmer did a complete repo bug hunt for 5 mins nonstop with: evidence, repro, fix, tests and a PR writeup.

Run and train it in Unsloth.

GitHub repo: https://t.co/aZWYAtakBP