Everyone is praising google right now but they just killed transparent CoTs.

a really crude and quick 1 step (1 NFE) flow experiment based on mean flow paper, but modified without using JVP fwd to circumvent the forward autodiff issue in the attention computation. one step image generation is possible now, no distillation required! arxiv.org/abs/2505.13447

Teng Yan · 30 days of COT Nous Research For anyone wondering: 1 - more compute is coming online - we are just about to add the compute for the 500k donations raised and will again open the pool for even more compute 2 - this is a test network, meaning the purpose alongside the actual model being trained is to find

Its funny that Anthropic and Google are the only competitors in coding AI that matter right now openai has just become a gen-media company and I only see people meaningfully using it for imagen and voice mode entertainment

x.com/onehappyfellow… It now makes sense why Jane Street would fund AK47s to South Sudan

I have an alternative proposal.

Grok4 Heavy Agent 1: This is my answer Grok4 Heavy Agent 2: bro did your mother drop you this is the answer Grok4 Heavy Agent 3: Hooligans, both of you! This is the answer! Grok4 Heavy Agent 4: 早上好中国现在我有冰淇淋！！！

Gemini (offline, hallucinating wildly): i shall peruse materials from the internet to solve this [somehow solves the problem anyway] Claude: as my great-uncle once told me at an oyster bar in Hanoi, o3: | have you | heard the | | --- | --- | | good word | about TABLES?? |

Atropos v0.3 is now out! Our RL Environments framework has seen a lot of upgrades since v0.2 - some highlights: - Atropos can now be used as a benchmarking and evaluations framework by Roger Jin, with our first external benchmark, Reward-Bench 2! - Added the Reasoning Gym,

Congrats to our post training team who worked on the Hermes 3's dataset - Teknium (e/λ), NobodyExistsOnTheInternet, and outside contributor interstellarninja - on creating the now #1 Trending dataset on HuggingFace!

gabriel my meta-experience here is that with issues this unique (where you *should* be doing great by any reasonable metric, but are not at all), it is a long and oft hellish journey, but also very hard to outsource since everything can be so n=1, you have to do all the tests on

Me too R1. The question mark is also "low-key" bothering me.

Our Researcher in Residence nightwing will be discussing his work on SMC steering at UC Berkeley on Aug 3. Check out the blog on this work here: nousresearch.com/steering-the-s… Details below!

Can someone explain why only sonnet 4 on a day to day basis has hugely different performance? Does Anthropic just decide to deploy different version of sonnet 4 every other day? All the models are gaussianish, only sonnet is step-like. github.com/jacobphillips9…