오픈 웨이트 논쟁과 로컬 에이전트 가속화 동향

OpenAI just signed "Open Weights and American AI Leadership" letter. it sad Anthropic and Google still haven't signed. The same coalition that NVIDIA, xAI, Microsoft, and Meta started yesterday now has OpenAI on board. Anthropic and Google are now the only major frontier labs still holding out. For anyone running models locally, fine-tuning on consumer GPUs, or shipping sovereign AI this is meaningful.

Harrison Chase

@hwchase17

1달 전

Every company that wants to build a business around AI will need to own their intelligence As @JensenHuang @satyanadella and others wrote yesterday, a huge part of that is open models But it’s more than that as well! This my take on what it means to own your intelligence

Your next LLM didn’t need a GPU at all it was just silicon, No VRAM drama. Taalas hardwired an entire LLM into silicon, AI agents that can actually think at light speed locally Taalas HC1 just showed a PCIe ASIC board running Llama 3.1 8B at with - 17,000 tokens per second. - On a PCIe card - No weights loading. They’re already planning medium-sized models running a 27B-class model at 10k+ tokens/sec on your desktop with almost zero power draw. Hyper-fast local agents without renting cloud GPUs suddenly feel a lot closer.

Agent Zero beat Codex by 2.6× on token efficiency while building a better Qbert game. Same local model (Qwen 27B). Same task. Agent Zero : 239K tokens Codex : 627K tokens The difference? Agent Zero only sends what’s new. Codex was refilling the entire conversation history every turn bloating context and causing hallucinations. Way less tokens. https:// x.com/atomic_chat_hq /status/2080382294833926278/video/1 …

a hybrid-quantized Laguna-S-2.1 at 2.75 bpw (50GB). Used REAP-style expert saliency to keep the important experts in NVFP4 while pushing the tails to calibrated EXL3 Trellis-2. The technique is hybrid expert quantization (REAP approach): - Hot/salient experts stay in higher precision - Everything else gets aggressively quantized - No expert pruning all 256 still available In cyber Frontier-level agentic coding model trained with cybersecurity in the calibration mix just got compressed hardand it’s still competent at agentic terminal work. - 117.6B to 50GB with hybrid quantization. - all 256 experts still routable, - and it’s already putting up 6 passes on Terminal-Bench 2.1. This is the kind of compression work that actually moves the needle for local agent builders. Thanks @0xSero for dropped this

Md Ismail Šojal

@0x0SojalSec

1달 전

- be Elliot Arledge - 14.5 hours. - One bedroom. - One whiteboard. - building the infrastructure that makes better GPU kernels possible. This is what building production-grade Kernel RL infrastructure actually looks like. - Elliot just pushed his kernel RL engine much further with D-Spark higher throughput, better concurrency handling, and fewer compilation bottlenecks. - At the same time he’s stress-testing Poolside’s new Laguna 118B model & running kernel benchmarks (both multi-GPU and the fast mini version). The future of fast local models and agentic systems will be built by people willing to do this kind of focused, unglamorous work for hours. Timelapses like this are underrated motivation.

A Fully local Open-source model cleared the X-ray. Should every medical scan get a second AI reader looked at the actual pixels and caught a clinic name the first one missed. 27 identifiers handled, 0 survived. In this demo: - OpenMed’s 1.4B PII filter handles text & multilingual identifiers - Qwen3-VL 4B acts as the second reader on pixels - Both re-scanned. 27 identifiers removed, 0 survived. https:// x.com/MaziyarPanahi/ status/2079899270925144085/video/1 …

Crazy, you know? Your Smart TV is secretly fingerprinting everything on your screen every 15 seconds even when you use it as a monitor for your laptop or PS5. Peer-reviewed research confirmed it from UC Davis, UCL & UC3M proves it. LG TVs Samsung capture every 15 seconds, They send it to their servers Even via HDMI. Smart TVs use Automatic Content Recognition (ACR) basically Shazam for video. It captures frames from whatever is on screen, turns them into fingerprints (hashes), and sends them to the manufacturer’s servers to identify exactly what you’re watching. LG sends data every 15 seconds. Samsung sends data every minute. A moment that surprised researchers: ACR doesn’t only track content from Netflix, YouTube, or the TV’s own apps. It tracks whatever is displayed on screen including when you connect a laptop, gaming console, or cable box via HDMI. Direct finding from the paper: ACR network traffic exists when watching linear TV and when using smart TV as an external display using HDMI. Lol Your TV is no longer just a screen. - Why do they do this? Simple answer Money bro Some TV companies now make more revenue from ads and data than from selling the hardware itself. - Vizio’s ad & data revenue hit $598 million in 2023. - LG’s ad business made nearly $700 million in 2024. They’re not just selling you a TV they’re selling your attention and viewing habits. Main thing is How to turn it off (do this today): - for Samsung : Settings to All Settings then General & Privacy then Terms & Privacy Then Uncheck Viewing Information Services LG - for LG :Settings to General then System to Additional Settings then Turn OFF Live Plus And Turn off Viewing Information Did this others or ask ChatGPT or Grok etc Important: Software updates can reset these settings. Re-check every few months.