mark

[mark @SpXMerlin1D]

@teortaxesTex compute-constrained is easy to tackled, relative to a unscalable architecture. I hope it is merely compute-constrained problem and everything goes well with moonshot.

[mark @SpXMerlin1D]

@teortaxesTex I'm saying the KIMI K3, and source is non official, so don't worry about it too much

[mark @SpXMerlin1D]

@teortaxesTex DeepSeek is definitely willing to adopt other's work when they can prove it's usable and scalable. And when it comes to AttnRes, I'm worrying Moonshot might encountered some severe problem when trying to train KDA+AttnRes. somebody said moonshot failed training many times.

[mark @SpXMerlin1D]

@teortaxesTex Jianlin Su's word should be remembered by every researcher. You shouldn't give up one of the right path,even one of the most efficient path, in practise, just to show that you're different.
(btw, for paper-writing, you still need something new, or nobody will accept this paper.)

[mark @SpXMerlin1D]

@teortaxesTex maybe using seed 2.1

[mark @SpXMerlin1D]

@teortaxesTex I know GLM-5.2 isn't an all-rounder, but V4Pro GA should at least having GLM-5.2/GPT-5.4 level performance when it comes to coding, or it's a total failure. An 1.6T model failed to beat ~700B model.

[mark @SpXMerlin1D]

@eliebakouch wondering why 31B Dense pre-training with significant less token than E2B or even 26B-A4B? because PLE need more training?

[mark @SpXMerlin1D]

@teortaxesTex but now "european company", obviously I'm saying MistralAI and maybe Blackforest Lab is commiting suicide by strict censorship, yet bad performance. I won't expect russia will surpass China or US, since AI is still compute-hungry and power-hungry, but at least nobody will ignore.

[mark @SpXMerlin1D]

@teortaxesTex at least it's a SOTA model in russia, and also not that kind of tiny model (below 100B), and deliver a reasonable performance, so not bad for 🇷🇺 huh?

[mark @SpXMerlin1D]

@teortaxesTex I hope so, since it seems like V4-Flash is better in both price and performance than Pro when it comes to agentic job like coding in Claude Code. since it's the first time that DeepSeek announce the approx release date of their new model, I believe Wenfeng is cooking something.

[mark @SpXMerlin1D]

@Xudong07452910 https://t.co/Lli7o8prl3

[mark @SpXMerlin1D]

@QuantumTransf 我看到第一眼在想KPRO是什么新算法

[mark @SpXMerlin1D]

@gurtej__gill_ wtf are u talking about? just released from prison and haven't followed up?

[mark @SpXMerlin1D]

@0xSweep @Abyss_ix it's copied from here https://t.co/2iuE6VOHl0

[Razed Esports @RazedEsport]

According to some studies, Generation Z prefers watching a NiKo match over having sex.

[mark @SpXMerlin1D]

@localoptimiser @teortaxesTex quite a classic solution, but AI need HIGH RND 4K IOPS

[mark @SpXMerlin1D]

@saraichiso @robmooreEsports DANK1NG is former TYLOO CSGO player

[mark @SpXMerlin1D]

@wightmanr ununu-Net

[mark @SpXMerlin1D]

@teortaxesTex reasonable for this result, GLM-5.2 from my personal experience, is specifically designed for agent, like GLM-5-Turbo, or at least not designed for one-shot job. because 5.2's one-shot performance is worse than 5.1