X 트렌드 요약: Grok 4.5·Hermes Agent·NVIDIA 옵티마이저 등 최신 흐름

Elon Musk

@elonmusk

Grok 4.5 is a solid workhorse

Elon Musk

@elonmusk

Grok Build /deep-research

Tibo

@thsottiaux

Let ChatGPT *work* for you. How many time have you wanted to negotiate your internet bill, get rid of all those spam emails you’re subscribed to, or find the perfect deal for something you wanted to do or buy. It’s quite literally just one prompt away, all from the comfort of your phone. It does at least 20 things for me every single day and I’m still surprised.

Greg Brockman

@gdb

ChatGPT is increasingly becoming your personal AGI

Teknium

@Teknium

Hermes Agent now scales to a near infinite number of MCP and plugin tools with high accuracy. You save a ton of money on input tokens and lose almost no accuracy in our internal testing. We've implemented progressive disclosure for tools, so they behave much more like skills. All MCP tools now sit behind a single search-and-execute tool. When your MCP tools use little context, their names and descriptions are shown to the model directly. If that would use more than 5% of your available context, the agent is instead told it has tools from X MCP behind tool search, so it always knows they're available.

Oh, lava lava baby! i did it! Hermes and i built a fully functional Lava Chicken shop in Minecraft. 1,865 verified blocks 563 note events One button. One hot serving. Built with Hermes Agent Local. Bounded. Verified. #Minecraft #AI #AIAgents #HermesAgent @NousResearch @Teknium

New research from NVIDIA. Does AdamW have a scale ceiling? This work claims yes, and shows where it sits. At batch sizes up to 100M tokens for next-token prediction, SOAP and Muon maintain training stability and quality while AdamW degrades. Higher-order optimizers have promised faster convergence for a while. The standing objection has been computational cost and numerical stability at scale. The team identifies instabilities in SOAP at large batch sizes and eliminates the loss spikes with per-step QR orthogonalization and improved preconditioning strategies. They also measure Muon's orthogonalization quality empirically rather than assuming it. On multi-billion-parameter models trained over trillions of tokens, both optimizers consistently beat AdamW. A layer-wise distributed optimizer compatible with Megatron-LM balances memory and hides communication without approximating the optimizer math, so the convergence benefit survives the systems layer. Paper: https:// arxiv.org/abs/2607.20548 Learn to build effective AI agents in our academy: https:// academy.dair.ai

Cohere has released open-source models Transcribe, Command A+, and North Mini Code so far this year, all available under Apache 2.0. With more to come... Own your AI.

A $5 USB stick just Run a fully local uncensored AI model 0 refusal just plug and play, - Uncensord supported model Mistral 7B, Qwen 2.5 7B Instruct , Phi-3.5, NemoMix etc - Runs 100% offline - Zero traces on the host machine - Works on Windows, Mac & Linux - Comes with uncensored models - No install. No cloud. No data left behind, Everything models, chats, settings stays on the stick. Just plug it in and talk to your own AI.

- http:// github.com/techjarves/Por table-AI-USB …

Gemini turn 6 hours of manual data work into 2 minutes. Gemini inside Colab just did exactly that planned, analyzed, and visualized a 100k-row dataset without me writing a single line of code. Completely hands-off. Worth a look if you spend any time in notebooks.