Professional larper @Brave @Obscuravpn PhD in sleeping, prev a penguin
Pluto
Joined May 2023
- Tweets18K
- Following371
- Followers21.9K
- Likes127K
The amount of people saying that open-weight models are going to be terrible going forward due to American labs cracking down on distillation is insane to me.
Do you know how many ways there are to gather data for model training?
Opus 5.5 is an incredible model, but so is Ernie 5.5, GLM 5.3 Flash, Qwen 3.8 27B each of them have their time and place and it is really sad to me that people who claim to care about AI do not take the time to even test them and I mean genuinely test them not just create some slop test that evaluates how many M&Ms fit into a peanut butter jar if that jar was a car made in Mars with the gravity of the sun in the south east.
I mean using it day to day, getting it to work on stuff you genuinely care for.
I've been saying this for a very long time, you don't need SOTA models 95% of the time. At this point you don't need it 97% of the time and the smaller models are only going to get better and better.
It's also really sad to see that people have no interest in testing something because it looks bad on a benchmark, I think people are too lost in the sauce of what they hopefully can have vs what they do have currently.
Deepseek V4.1 Flash and GLM 5.3 Flash have genuinely been more useful to me than Opus and Astra ever will be, do I think Opus and Astra are bad models? No.
They just don't fit my use-case therefore aren't great for me.
There are so many research papers that labs put out especially the Chinese labs that are just a positive to the whole ecosystem, Deepseek R1 paper, how Deepseek found a way for KV Cache to be incredibly small and efficient and Xiaomi putting out the RL paper on how they trained the MiMo models in the open and how much it costed them.
Comparison is the thieve of joy, we only have so much joy left till it's all gone.
I want you guys to explore, you can do it at an incredibly speed now, it's great to get into coding, research whatever you want to do, you can be so incredibly.
I just wanted to ramble about how I feel about the current state of this platform when it comes to AI, and it wouldn't be a post by me if I didn't tell you to get a job because some people here have too much time and not enough going on.
Here are some great papers to read:
arxiv.org/pdf/2610.11959v1
arxiv.org/pdf/2501.12948
arxiv.org/pdf/2609.19969
arxiv.org/pdf/2609.26368
There are a bunch of them out there available for free that you can read ^^
1
12
349
Yes, I was impatient, I got the token plan, I'll tell you guys tomorrow how I feel about my Claude subscription.
Qwen4 is coming.
Deeper knowledge. Richer world understanding. Stronger coding and agentic capabilities.
Built for real-world tasks — from complex coding and knowledge work to front-end builds and professional software workflows.
Qwen4-Max is now in preview on Alibaba’s Token Plan. You can also get early access to the new model on QwenWork, Qoder, Qwen APP, and QwenStudio.
Put it to the test on a real problem. Show us what you build. 👀
Token Plan: qwencloud.com/pricing/token-…
I really hope the Qwen team resolves this because this is not fun to deal with.
xitter.twitterapi.xyz/gnukeith/status/212043…
This keeps coming up sometimes, I'm not sure if Qwen is just bad at following goals (entirely possible) or it thinks I'm actively trying to show it hentai (I am, but not right now).
19
1,638
Do you think it works to beg Framework to ship the Frameworks faster? Without being annoying?
2
34
1,596
Cheaper, better and more capable? I don't want to be that guy but what's happening at OpenAI?
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.
They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
8
101
3,330
MiMo 2.6 Flash is one of my favorite models to use because it's cyber capabilities is awesome but it does lack in comparison to DeepSeek V4.1 Flash and GLM 5.3 Flash but this architecture is AMAZING, I don't think you guys realize how good this architecture is.
MiMo-V3 is getting a new architecture. The core of it, HySparse2, is out today.
Less prefill, a smaller KV cache, better long-context retrieval—and we got all three at once.
Compared with MiMo-V2.6's Hybrid SWA architecture:
• 5.02× lower prefill FLOPs at 1M tokens
• 4.5× smaller KV cache at 1M tokens
• Better MRCRv2 and RULER-v2 scores, plus lower AgentPPL and LongPPL
Why build a new architecture?
Agentic inference is a very different workload. Each round, a short action can return a long observation that needs to be prefilled, while the context keeps growing. That puts prefill cost, KV-cache size, and retrieval accuracy on the critical path at the same time.
HySparse2 tackles all three with two levels of KV sharing:
• KV Bridging: Following YOCO, full-attention layers in the cross-decoder build their K/V from self-decoder hidden states.
• KV Reuse: Within each hybrid block, sparse layers reuse the preceding full-attention layer's KV cache and selection indices.
Two more changes: token-level selection replaces block-level selection, and a forced window of recent tokens replaces the separate SWA branch, so local and global tokens share one KV cache. Since all cross-decoder KV caches now come from the self-decoder, prefill can stop once the self-decoder finishes.
Paper: arxiv.org/pdf/2609.26368
2
46
1,766
I just got a 12U rack and 4 Framework DeskPi RackMates, I'll get more compute in January and then I think I've successfully achieved the 1TB RAM dream.
I'll be waiting for the Medusa Halo to replace some of the desktops later on and turn the other desktops into something entirely different.
I'm planning all of this to 2029 and obviously hoping that prices will go down.
