Kimi K3 (Open Weights by July 27, 2026.) 🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal 🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts 🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost

Sung Kim@sungkim.bsky.social · 3w ago

They launched Kimi K3 with a video.

Wait, you… use libraries? Code that other people wrote for you? You just… download it? Do you even read the code? Do you not care about the craft, the knowledge that comes from putting in the work yourself? Aren’t you afraid your skills will atrophy?

so, using GLM 5.2 on Umans, this is the first time I am experiencing "oh yea, that's the stuff" feeling using these things. like i've had this feeling before about output quality, yes. but i've never had tps this high _with_ good output quality.

Introducing Sakana Fugu: A full multi-agent orchestration system accessible via a single model API. Our ‘Fugu Ultra’ model matches the performance of Fable and Mythos, delivering frontier capability without the risk of export controls. Try it: sakana.ai/fugu 🐡

Another important thing: Chinese models are not strong because they distill US models. Distillation of models via API is *impossible*. If somebody tells you the contrary, they don't understand machine learning:

I'll just comment on my impressions of their conclusions in a thread here. Worth sharing how I see it unfolding internally. (I signed up for their email spam to read this so yes, I am a hero.)

Sue Smith@suesmith.lol · 2mo ago

"The AI engineering impact data shows that output is up. It also shows that the work required to ensure that output is safe, correct, and maintainable has not decreased. It has increased substantially." Aye, if only any of us had been saying this from day one 🤡 www.faros.ai/blog/ai-acce...

Gemma 4 quantization-aware training (QAT) models are now available, bringing AI performance directly to edge devices and consumer GPUs. These checkpoints are optimized with quantization-aware training to dramatically reduce memory requirements and unlock high-speed local inference. 🧵

Anthropic partnering with Elon should be a reminder that the real end goal should be powerful open models you can run on your own hardware China is making great progress there

Kimi 2.6: Hangs with the best * on par with Opus 4.6 & GPT-5.4 5.4 xhigh * long horizon coding tasks www.kimi.com/blog/kimi-k2-6

sounds like it’s time to take local models seriously

Unsloth AI@unsloth.ai · 4mo ago

Gemma 4 E4B (4-bit) completed a full repo audit by executing Bash code and tool calls locally. Runs on just 6GB RAM.

1/ Today, we’re excited to introduce Attie, currently as an invite-only closed beta. Attie is the first agentic social app on atproto. It’s something completely new—an experiment in making building on the protocol more accessible.

"everyone hates llms" hundreds of millions (and growing) are using llms regularly. techcrunch.com/2026/02/27/c... thread of claims continues...

ChatGPT reaches 900M weekly active users | TechCrunch

OpenAI shared the new numbers as part of its announcement that it has raised $110 billion in private funding.

techcrunch.com

Nothing to see, just very powerful pattern matching. www-cs-faculty.stanford.edu/~knuth/paper...

It was given this challenge to prove its AI coding methods viable by being given a task thats unlikely to have lots of examples in the AI's training data. It of course accepted the challenge because it knows nothing about what this means. That way its fair fight.

Maxine 💃🏼@maxine.science · 7mo ago

Ok: Write a python package that does multivariate Nadaraya-Watson and local polynomial kernel regression with homo-/heteroscedastic goodness of fit checking and automatic crossvalidated order and bandwidth selection as an sklearn (fit/predict) compatible class.

Gleam v1.14.0 is out now! Merry Christmas everyone! 🎁 gleam.run/news/the-hap...

The happy holidays release 2025 🎁 | Gleam programming language

News post: Gleam v1.14.0 released

gleam.run

It takes 0.04 liters of water to make a single AI image. Meanwhile, it takes 5 liters to make a single piece of paper, and presumably even more when you add a pencil to it. Using AI to make art is literally better for the environment than pencil and paper. www.theatlantic.com/technology/a...

I shared a short succinct summary of my core strongest arguments that using chatbots is definitely not bad for the environment you can share with skeptical people in your life andymasley.substack.com/p/a-short-su...

A short summary of my argument that using ChatGPT isn't bad for the environment

To share with anyone still worried

andymasley.substack.com

I have a hunch that current LLMs might make it easier to launch a brand new programming language, provided you can describe it in a few thousand tokens and ship it with a compiler and linter that coding agents can use simonwillison.net/2025/Nov/7/l...

New data on the corporate ROI from generative AI from a large-scale tracking survey by my colleagues at Wharton. They found that 75% already have a positive return on investment from AI, less than 5% negative. Also 46% of businesses leaders use AI daily. knowledge.wharton.upenn.edu/special-repo...

You’re (probably) measuring application performance wrong. Humans have a strong bias for throughput. "I can handle X requests per second." Real capacity engineers use response-time curves.

I hope as we move past the first wave of AI criticism ("it doesn't work, all hype") we get a new wave of AI criticism rooted in the acknowledgement that, yes, these systems are very powerful & quite useful and focusing a deep exploration of when AI uses are uplifting and when they are detrimental.

AI/LLMs are a major accessibility technology — being against this technology is to be against advancing human accessibility, and i think there’s a case to be made that it’s borderline ableist

if you're curious about the architecture and mechanics of LLMs, this site has a really excellent explorable interactive visualization. it helps build intuition for how massive these models are, what 'interpretability' means, and the complexity involved here bbycroft.net/llm

Saying that we already know everything about LLMs because we know how they work on the lowest level is like saying that we know everything about mathematics by just defining axioms, or that we have computed everything just by inventing CPU.

⚡️🌙@dystopiabreaker.xyz · 10mo ago

things we know about LLMs and large DL models in general: - how they are trained (gradient descent) - the structure into which they are placed (architecture) - the base arithmetic (matmul, norm, batch norm, and so on)

hello tanglers! this october, we want to get more people into making their first ever open source contribution! to help them out, we've got the 🟣 'good-first-issue' label! head to tangled.org/goodfirstissues to see a list of issues across different repos that you can contribute to! 🤓

i wrote about atproto and why it matters

Open Social — overreacted

The protocol is the API.

overreacted.io

That’s a very interesting part #rustlang

Proton@proton.me · 10mo ago

If you're on iOS, you can get the update today on the App Store. On Android, it will be gradually rolling out. Read the full story here: proton.me/blog/new-mai...

buckle up tanglers, we've been working hard! first up: ✨CI is now generally available!✨ head to your repo settings and set your workflow runner to the hosted instance. 🚀 we have also got support for secrets! you can now deploy your site on CI! read the blog for more: blog.tangled.sh/ci 1/5

See how Gleam gets the job done in production ⭐ Full case study coming soon: “For a team like ours, with many other priorities and projects we need to work on, the confidence that Gleam gives us is worth its weight in gold.” - Edward Kelly, Director of Technology at Strand.

Guillaume Hivert, VP of Engineering at Steerlab, shares how being a Gleam business has been the best choice they have made!