LeMi

[LeMi @LeMiMind]

@h4x0r_dz its not like china was the only country could do that, if they didnt do it then other countries would do that, simple. Right now there is no point for other countries to make new ai models, its not worth it

[LeMi @LeMiMind]

@daniel_mac8 @rasbt @grok what do I do to implement this

[LeMi @LeMiMind]

@realproject516 @jun_song Not everyone is a software engineer or programmer

[LeMi @LeMiMind]

@Gusarich @grok what does this chart explains

[LeMi @LeMiMind]

You guys don’t understand what’s currently happening with AI — or what’s about to. Because of DeepSeek, models are going to get way more efficient and cost a lot less.

[LeMi @LeMiMind]

I love DeepSeek for dropping a new model so efficient it’s going to force proprietary labs to make theirs cheaper

[LeMi @LeMiMind]

@elonmusk You promised us that people would be on Mars by 2025; it's almost 2026, and we haven't even come close to that

[LeMi @LeMiMind]

@jturntdev i hope it's gonna be as efficient as Luna

[LeMi @LeMiMind]

@vikhyatk @grok is this true?

[LeMi @LeMiMind]

@theo I wonder where did you wasted it for and how did you even managed to get such a result

[LeMi @LeMiMind]

Source:

https://t.co/aNEIrells9

[LeMi @LeMiMind]

OpenAI’s next major model Astra just proved 10 new math results, including that nonsofic groups exist and a disproof of Connes’ Rigidity Conjecture.
They’re releasing the full Lean certificates + CoT walkthroughs.
AI is starting to do real frontier mathematics.

https://t.co/lva2ziftj1

[LeMi @LeMiMind]

@elonmusk are you planning to make your models opensource in the future?

[LeMi @LeMiMind]

Just wasted 7% of my weekly Plus usage using Luna + Fast mode trying to fix a bug that it didn’t accomplish, then switched to Sol High and it fixed right away using only 1% of usage.

[LeMi @LeMiMind]

@sama Excuse me, where is our promised 750t/s???

[LeMi @LeMiMind]

When are we getting it????

[LeMi @LeMiMind]

@OpenAIDevs @Weichaus when are we getting sol with 750t/s on cerebras?

[LeMi @LeMiMind]

Has anyone actually tested Luna for game development? I’d love to see how it compares to Opus 5, especially at such a lower price.

[LeMi @LeMiMind]

Holy 🔥 @UnslothAI just made 1-bit Kimi K3 (2.8T params) runnable locally.

Shrunk from 1.56TB → 594GB (-62%) while keeping ~78.9% accuracy.
Same prompt: interactive glass aquarium that cracks & bursts…

1-bit Kimi K3 GGUF matches GPT-5.6 + Claude Opus 5 quality, running on 4× B200s at 36 tok/s (or Mac Studio + 128GB).

Open-weight frontier models just got real.

https://t.co/hAXAl9aT0u

Guide + GGUF: https://t.co/SbdgtALgIS

[Unsloth AI @UnslothAI]

We compared 1-bit Kimi K3 to Claude Opus 5 and GPT 5.6.

We gave 4 models the same prompt:
Create a glass aquarium whose side panel develops a visible crack and then bursts.

1-bit Kimi K3 GGUF ran locally on 4x B200s at 36 tok/s.

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

[LeMi @LeMiMind]

Turns out GPT-5.6 Sol is actually SOTA on ARC-AGI-3.

Just two setting changes (retained reasoning + canonical compaction) tripled the score from 13.3% → 38.3% and cut tokens 6x.

Harness matters a lot.
https://t.co/tnrTJP8XHL