Zephyr

My Third Post Implications of the H20 It will have profound change on the dynamics of US-China AI race especially in the RL/inference age

I think these guys have solved P6 as well

Whale generally releases stuff on Monday

I'm long tech Cos Was short retail till Trump TACO

Bytedance, Whale, and Zhipu are top contenders Could be a government-backed lab as well

Which team??

I made a lot of money by buying Nvidia in 2018 & Broadcom in 2022, and shorting Tesla in December 2023

In 2016, I shared an elevator with the CFO of Wirecard in NY post a group meeting MS hosted with a room full of rabid short sellers post some high profile fraud allegations. We were short the stock - in size - because we thought they were probably a fraud and that they would at…

AI is very disruptive for policy research. Personally, my remaining alpha is almost entirely my: 1. network, 2. non-public info, 3. IRL persuasion ability, 4. platform. Less so my reasoning, writing ability, or mastery of public info, which are available in abundance to all.

First, it was Yookay, now Australia too Is this a roundabout way of targeting American tech companies now that Trump will block direct fines??

This could be very interesting

Why do FFNs use ReLU instead of more precise ones like Exp? "We propose the following hypothesis: A kernel with lower retrieval precision encourages a more polysemantic key–value memory: multiple unrelated facts can be stored under the same key space" Great and inspiring read!

TEMU HAS BEEN TOLD BY US COMPANIES AND SELLERS THAT IT CANNOT PROVIDE CHEAPER PRICES ON BRANDED PRODUCTS THAN THOSE OFFERED ON AMAZON - FT

These 4 will be the new stuff incorporated in V4/R2, along with more parameters & training data, and increased sparsity

ded rong NSA & SPCT GRM is far more impressive compared to Prover V2 What they haven't released this year is a new optimizer (I don't think they will use AdamW, especially when Muon exists), and a better RL algo GRPO has some issues and whale isn't as compute constrained anymore

Phase 1 of Physics of Language Models code release ✅our Part 3.1 + 4.1 = all you need to pretrain strong 8B base model in 42k GPU-hours ✅Canon layers = strong, scalable gains ✅Real open-source (data/train/weights) ✅Apache 2.0 license (commercial ok!) 🔗github.com/facebookresear…

Very cool

Huawei Ascend 910C chip analyze TSMC N7 Family Same chip with 910B Video on Youtube： youtu.be/0feji7gMrh8?si…

HUGE

.@basedjensen do u have this flag in ur house??

Ironic Schmidt was one of the chief architects of GPU and semiconductor equipment export controls