ทวีตที่ปักหมุด
Meet Strix, my AI agent
This one covers:
- an intro from Strix
- architecture deep dive & rationale
- helpful diagrams
- stories
- oh my god what's it doing now??
- conclusion
timkellogg.me/blog/2025/12/1…
English
Tim Kellogg
18.9K posts
Tim Kellogg
@kellogh
AI architect // hiking, camping and long walks on the beach as long as they involve backpacking
Raleigh, NC เข้าร่วม Kasım 2011
722 กำลังติดตาม1.3K ผู้ติดตาม
THIS IS INSANE!!
Supermemory reached a 99% SOTA memory system.
AI agents will now remember EVERYTHING
p.s they’re open sourcing it in 11 days 👇
Dhravya Shah@DhravyaShah
English
@kellogh @YouJiacheng My dude if you read the language on the kimi license all it asks is that you use their name displayed
English
This is nuts to me, the one thing Moonshot (kimi creators) asks you do is say that they are the base. Like just say it, what's the big deal, everyone already knows! It's insane. At this point not saying it makes you look like the baddies.
English
English
English
Congrats to the @cursor_ai team on the launch of Composer 2!
We are proud to see Kimi-k2.5 provide the foundation. Seeing our model integrated effectively through Cursor's continued pretraining & high-compute RL training is the open model ecosystem we love to support.
Note: Cursor accesses Kimi-k2.5 via @FireworksAI_HQ ' hosted RL and inference platform as part of an authorized commercial partnership.
English
@kimmonismus it was a COMMERCIAL license. Cursor PAID kimi for non-standard terms, such as white labeling
English
For transparency reasons, I believe it would have been better to include a direct reference to Kimi K2 in the blog post about Compoaer 2.
Furthermore, it also demonstrates how good Chinese open-source models have become.
Lee Robinson@leerob
Since people really want me to say this: "KIMI K2.5" ‼️ Yes, that is the base we started from. And we are following the license through inference partner terms (e.g. Fireworks) I'm thankful for OSS models personally, good for the ecosystem.
English
@jessfraz @YouJiacheng it’s a COMMERCIAL license
really not sure if you’re reading this, but they absolutely are exchanging funds for this
English
@YouJiacheng they arent asking for 80 bajillion dollars they are asking that you say their name, JUST SAY IT
English
@jessfraz @Kimi_Moonshot @cursor_ai it looks to me that they paid for the right to not have to announce it
English
@Kimi_Moonshot @cursor_ai I'm mad on your behalf, they should have put you in the blog and made it apparent from the start.
English
@teortaxesTex ya i cant even evaluate models this year without an agent harness. the chat frame just can’t push them hard enough
English
> Most improvements in the last 9 months are attributable more to the tooling around the model rather than the models themselves
That's because the models themselves have become good enough to reliably make use of large numbers of nontrivial tools
what's this mistake called?
expatanon@expatanon
Altman admitted that transformer models have hit the wall. Most improvements in the last 9 months are attributable more to the tooling around the model rather than the models themselves. In other words, this technology is rapidly maturing with no signs of another leap.
English
@rickasaurus in my brief testing, it appears alarmingly over-RL’d. which is sad, 2.5 was really good
English
You might say they're min maxxing their release cadence
🥭@MangoSweet78
minimax 2.7 apparently.
English
@himanshustwts @MiniMax_AI by recursive self-improvement are you just saying it knows what to remember when taking notes to it’s future self?
English
Diamond in the rough X account alert
His takedown of Bender is fantastic as well, will link below
SE Gyges@segyges
oh substack plug. basically this entire conception of alignment to instruction-following predates LLMs and the premise, the "value-loading problem", is not true. i wrote it up previously verysane.ai/p/alignment-is…
English
@inductionheads god damn that website thing is incredible. we need more like that
English
Sparsity, go figure
Shiwei Liu@Shiwei_Liu66
Residual connections and pre-norm are not the whole story behind depth utilization. Our new paper shows that many seemingly different design choices — MoE, grouped-query attention, weight decay, and longer sequence length — can be understood through one unifying lens: sparsity. These components induce different forms of sparsity, which reduce output variance and in turn preserve healthier gradient flow across depth. Strikingly, these techniques also complement each other remarkably well: when combined, they lead to substantial improvements in depth utilization and notable gains in downstream accuracy. Paper page: pumpkin-co.github.io/SparsityAndCoD/ Arxiv: arxiv.org/pdf/2603.15389 Leading by @pumpkinnnnne
English
What happens to Anthropic when anyone can use Claude Code to generate Claude Code?
English
@norootcause depends what it is. like, i generally don’t send people output directly, but there’s been a few times where i just thought to myself, “yeah, i can’t improve on this, send it”
English
Nobody wants to read the output of someone else’s prompt
English
Is Healer Alpha the GPT-5 Omni model (GPT-5.5 or GPT-5o) or DeepSeek v4 or perhaps Kimi K3? I'm leaning towards the OpenAI model personally but very unclear. 1M context, vision, hearing and reasoning.
AiBattle@AiBattle_
2 new Stealth models on OpenRouter Hunter Alpha: - "Hunter Alpha is a 1 Trillion parameter + 1M token context frontier intelligence model built for agentic use. It excels at long-horizon planning, complex reasoning, and sustained multi-step task execution, with the reliability and instruction-following precision that production agentic pipelines demand" Healer Alpha: - "Healer Alpha is a frontier omni-modal model with vision, hearing, reasoning, and action capabilities. It brings the full power of agentic intelligence into the real world: natively perceiving visual and audio inputs, reasoning across modalities, and executing complex multi-step tasks with precision and reliability"
English
@teortaxesTex “pretrained in nvfp4” is the headline imo
we hadn’t seen viability of this in the wild yet really
English
Well, seems we're not getting DeepSeek V4 today but we're getting what amounts to its lite version runnable on normal hardware. New architecture, fast, 1M context…
…and it's a bit weaker than the equivalent Qwen 3.5.
Lisan al Gaib@scaling01
Nvidia released Nemotron 3 Super - a 120B-A12B hybrid Mamba model with LatentMoE and MTP - pre-trained on 25T tokens in NVFP4 - context up to 1M - 2.2X faster inference than GPT-OSS-120B - 7.5X faster inference than Qwen3.5-122B huggingface.co/nvidia/NVIDIA-…
English
Nvidia will carry US open source.
Commoditize your complement, from here to the singularity.
🇺🇸
NVIDIA Newsroom@nvidianewsroom
NVIDIA Nemotron 3 Super is here to accelerate the era of agentic AI. Optimized for NVIDIA Blackwell, this 120B open model uses a hybrid Mixture-of-Experts (MoE) architecture that delivers 5x higher throughput and 2x higher accuracy. The model combines advanced reasoning with a 1-million-token context window, enabling autonomous agents to solve complex tasks with speed and precision.
English
@rickasaurus yeah i don’t think it matters. it’s all just information & flow. better curation skills would help, the fundamental skill is knowing what to remember & what to forget
English