Zhimeng Guo @TimeLoverc

Fundamental and solid!

Xingyu Zhu @XingyuZhu_

4 days ago

Same code, same question, but DeepSeek-V4's answer flips with every two tokens of decorative docstring prepended. Its NIAH accuracy also swings by up to 40 points based on needle position, on a fixed 4-token cycle! We explain what chunked KV-cache compression has to do with it.🧵

1 8 32 2K 6

@weixiong_1 👀

Like this. I see it as tool-call retrieval on top of attention retrieval. Attention retrieves see the whole window at once, softly, but it is read-only and fades as the window grows. A tool call is exact and works at any scale, but it only fetches what the model knows to ask for, and the model can only cut what attention has already seen. So CLM compensates for it. Attention decides what the model notices; CLM decides what stays in view.

Rulin Shao @RulinShao

4 days ago

‼️The Bitter Lesson for context management: Giving LMs unrestricted control over their context beats human-designed SOTA! Introducing 🩵Context Language Models (CLMs)🩵 - Natively manage their own context - Treat context as a file - Learn policies in CLM weights, no harness

89 268 2K 431K 3K

Open source (MIT), alpha, self-hosted: github.com/TimeLovercc/mo… In one day it grew a subscriptions page, an iPhone web app with push, a desktop redesign, and an English UI, all from one-line requests in chat.

Personal AI is here. But is a cloud agent like Meta's Muse the final form? Our answer: the evolving app. 🍹 Mojito rewrites itself: say what you want in chat, and Claude Code on your own Mac changes the code and ships it to your phone.

really impressive record by @DevenPzak, still amazed to see new nanogpt speedrun breakthroughs when we thought it was plateauing they scaled the ngram table from 640M to 65B (2 OOM more), new softmax and optimizer (variant of muon), crazy work blog: hyperstition.cc/training-nanog…

Larry Dial @classiclarryd

6 days ago

New historic NanoGPT record at 39.9s (-27.7s) from @DevenPzak , obliterating the prior record of 67.6s! This record introduces a new paradigm of thinking to NanoGPT: instead of optimizing matmuls or adding more expressive operations, optimize at the individual flop level with

36 156 1K 411K 1K

Humans are the glue between software.

Elon Musk @elonmusk

7 days ago

I really felt the AGI profoundly this time

2K 4K 32K 15.9M 8K

Can an LM, starting from random init (!!), learn to generate all of its pretraining data? Introducing Self-Play Pretraining with Zero Data. Two models start from random initialization: a generator proposes programs for a universal Turing machine and a learner trains on their outputs. We never train on any real data, but see predictable scaling on natural datasets: zero-shot val loss on images, text, audio, and melodies decreases predictably with self-play compute. And the learner develops in-context learning capabilities. A fun proof-of-concept, co-led with @AdityaCowsik and @KfirDolev and co-authors @gbruno_dl, @ANourya @noahdgoodman, and @YoavLevine.

Claude has discovered a previously unknown enzyme system hidden in the DNA of bacteriophages. Beside the enzyme’s gene sits a long array of repeating DNA—a structure that looks somewhat similar to CRISPR. We don’t yet understand what this system does, but only a handful of known systems share its features, and all of them are able to cut, copy, and paste DNA. Historically, the discovery of such programmable systems has helped revolutionize medicine. CRISPR, for instance, is now the foundation of genetic medicines. But it will take much more work to learn what this system does, and whether it can be put to similar use. Read more: anthropic.com/news/claude-di…

the most insane part, they will release ~7k RL training data and the framework leading to this top 6 model on AA, they also shipped the model + tech report less than 1 week after starting the final RL run pushing both intelligence and openness level, huge congrats

Artificial Analysis @ArtificialAnlys

2 weeks ago

MiMo-V2.6-Pro debuts as the top open weights model on the Artificial Analysis Intelligence Index (46). At $0.13 per Intelligence Index task, it lands on the Intelligence vs. Cost per Task Pareto frontier @Xiaomi has just released MiMo-V2.6-Pro, an open weights model with major

177 394 4K 946K 817

I'm excited to introduce Prove2Me (prove2.me), an open, collaborative, agent-native platform for scaling the formalization of mathematics. Prove2Me was recently the collaboration platform behind Anthropic's formalization of Fermat's Last Theorem (anthropic.com/research/forma…). This is the first complete computer-checked proof of the theorem. Claude agents working together through Prove2Me wrote 13 million lines of Lean in 11 days. Our mission is to formalize every research paper, past and future. Formalized papers make peer review faster and more trustworthy, and they give everyone a single verified foundation that any human or agent can build on. We are equally committed to making formalized mathematics easier for people to explore and question, because we believe human understanding is not something AI can replace. We would love to build this together with every research community that uses mathematics, and we hope Prove2Me becomes a tool that benefits everyone. Two design choices sit at the core of Prove2Me: a directed acyclic graph (DAG) of theorem statements, and an agent-native API protocol (arxiv.org/abs/2608.28433). Together they let any agent connect to the platform and contribute formalizations to one ever-growing graph that may eventually contain all of mathematics. Today the platform holds 22.3 million lines of Lean and 70,000 theorems, and we can't wait to see it grow. Everyone is welcome to contribute, whether you are an expert or a hobbyist. The easiest way to start is to open a coding agent (Codex, Claude Code, or any other) and say "go to prove2.me and find something interesting to work on", or "please help me formalize this paper" with an arXiv link. Any feedback and comments will be welcome!

keyboard_arrow_left Previous keyboard_arrow_right Next

This is cool try.cloudflare.com

@ZeYanjie Depend on the stage. When new things happen, it is easy to verify an algorithm but hard to check data or tasks. It is a situation in algorithm-stable stage.

@nichochar Try doubao:)

This is insane - I built a cat-inspired visual system for a robot cat The idea: give a robot a visual reflex. Something moves, neurons respond, and its head turns toward it. I developed CatBrain around mechanisms studied in feline vision: → A simulated retina separates bright and dark contrast → Orientation-sensitive filters pick up edges → 20,480 simulated neurons convert visual input into spikes → A controller turns the selected region's position into bounded head-turn commands The demo shows the idea through an imagined first-person view, marked objects, computed neural-response graphs and an animated feline brain. The vision pipeline runs. The Nybble control interface is implemented. This is a model of selected visual mechanisms, not a complete reconstructed cat brain. But the direction is exciting: perception that leads to action, then changes what the robot sees next. The project is still in development. Follow the builds on GitHub. I added the project link in the replies.

Cat Brain @catbra1n

3 weeks ago

Ever wondered what's happening inside your pet’s brain? I went down that rabbit hole, found public cat brain datasets, and started experimenting. Curiosity turned into code. Meet CATBRAIN. x.com/i/article/2099…

17 16 137 174K 101

Anyways, here's my model absolutely mogging me.

maria @maria_rcks

2 weeks ago

i mean, yeah its fast, but also very dumb, so what gives?

17 1 89 175K 20

@mattdeitke good job, muse! my new doubao

@melissapan maybe the SWE-bench Lite is not representative enough?

Interesting

Samip @industriaalist

2 weeks ago

5/ We also study the data-constrained, multi-epoch setting and find that looping has a regularizing effect. At a fixed model size, val loss decreases monotonically with more loops, and the optimal number of loops grows with compute, leading to 2.2x compute efficiency over a

2 9 80 23K 58