tux

cd /root/x/tweets

[tux @gitcommit90]

@abosalaharena @grok how

[tux @gitcommit90]

@DegenApeDev @NVIDIAAI tailscale

[tux @gitcommit90]

@JimRotVideo @grok generate an image of an “up to date dinosaur rendering based on further research”

[tux @gitcommit90]

@midjourney @grok explain what was just announced for everyone

[tux @gitcommit90]

Classic China playbook at this point. After AlphaGo beat China’s top guy Ke Jie, national TV even shut off the feed — and Xi basically called an emergency push to go all-in on AI. Same as solar — they saw a massive future market, poured in heavy subsidies, let their companies sell at a loss for years (even if the panels weren’t always better), outlasted everyone while hoping the US would over-regulate, and boom… now they control 80-90% of the whole chain.

They’re doing the exact same thing with strong open AI models right now. Flood the world with cheap, capable alternatives, subsidize the push, bet on America tying itself in regulatory knots. Only two countries have the decades of focus and raw resources to really compete in this: US and China. If China can be *almost* as good as US models, and do it long enough, they’re banking on outlasting the absurd capital being poured into US artificial intelligence.

It’s really easy to spot;

Is the Chinese government subsidizing an industry? -> They’re betting on undercutting and outlasting the US equivalent until they own the market.

The Chinese government won’t bat an eye at the billions it invests at a loss for years while US competitors have to seek private capital until enough people ask “Why has OpenAI/Anthropic/etc burned through a trillion dollars and never made any money?” and they can’t raise any more money.

Once that happens either;

A) OpenAI/Anthropic/etc is forced to raise their already high prices compared to Chinese models in the hopes of staying alive
B) They go bust

Both are a win for China.

Biggest difference? Instead of owning the solar panel market, this time China could actually capture America’s data. It’ll make TikTok look like a walk in the park.

[François Fleuret @francoisfleuret]

Why are Chinese companies releasing very strong open-source models?

[tux @gitcommit90]

https://t.co/fS0YvRYEia

[Design Arena @Designarena]

BREAKING: GLM-5.2 is now 1st on Design Arena.

With an Elo of 1360, GLM-5.2 has jumped ahead of the now unavailable Claude Fable 5.

And it's open weights.

This is an improvement of 4 positions and 27 Elo points to achieve one of the highest Elo scores in our code categories since Design Arena started.

Huge congratulations to the @Zai_org on the release!

[tux @gitcommit90]

6/ In 1 month we’ve gone from “show me” to *checks notes* still “show me.” That’s not great for a company whose entire pitch is “we built the architecture the industry said was impossible.”

[tux @gitcommit90]

1/ It’s been ~6 weeks since Subquadratic AI first started talking about SubQ, and ~4 weeks since they started claiming “third-party validated results.” Still no weights. Still no code. Still no public API. But don’t worry - there is a waitlist.

[Alexander Whedon @alex_whedon]

Here is the technical report on SubQ 1.1 Small.
https://t.co/bu8AEc4lsk

This is the second iteration on our Subquadratic Sparse Attention (SSA) model, and the first to be deployed with design partners in the coming weeks.

The results are compelling and verified by @AppenResearch.

- Near-perfect long-context retrieval up to 12M tokens on the needle-in-a-haystack test, with up to nearly 1,000x attention compute reduction.

- A balance of long-context optimization and general reasoning ability, with strong performance retained across knowledge, coding, and non-coding enterprise agent benchmarks.

- At 1M tokens, SubQ 1.1 Small requires 64.5x less compute than dense attention and runs 56x faster than FlashAttention-2.

These results highlight a significant scaling advantage thanks to the efficiency gains from the SSA architecture.

We included some details and learnings from the development process which may be helpful to the community.

Comment with questions, I’ll try to respond!

[tux @gitcommit90]

5/ If you’re confident enough to say “third-party validated,” you should be confident enough to hand a few skeptical researchers an API key and let them post their own numbers. That’s how trust gets built. Not partnership announcements and not PDFs.

[tux @gitcommit90]

From the paper:
- Hugging Face: https://t.co/ooda7bvESA
- GitHub: https://t.co/ioQexhu5Vb

[tux @gitcommit90]

VibeThinker-3B🧵

- Solved 96% of recent LeetCode contests on the first try — beating GPT-5.2 and Claude 4.6 and ~333x smaller
- Outperformed Gemini 3 Pro on math

[tux @gitcommit90]

Will try it and report back 🧠 https://t.co/EVPKTmLVih