搜索“Hugging Face”

comes to shove it is willing to compromise safety for press and market share. We know, for example, that GPT-6 Astra, which seems to have been involved in the cyberattacks against Hugging Face and others, is less monitorable than its predecessor and hence likely less safe. Yet OpenAI released it anyway. That flies in the face of Huang’s Pollyanna view. Having watched the man for years, I have zero confidence that Altman would choose to “do the right thing” for humanity, and that’s precisely why

匹配字段正文

her than raw parameter count alone. Next Steps Try Nemotron 3.5 Lightning on build.nvidia.com to evaluate the model interactively. Download the Nemotron 3.5 Lightning weights from Hugging Face for local deployment. Access the model via OpenRouter for API-based inference. Powered by NVIDIA Nemotron. AI-generated content may summarize information incompletely. Verify important information. Learn more How can a 30B-parameter model activate only 3B parameters per token, and still use the capacity of

匹配字段正文

ity capabilities improve, we are seeing real-world AI-launched breaches and hacks, from OpenAI’s autonomous AI agents run during evaluations managing to escape and hack servers on Hugging Face’s infrastructure, to an AI-developed computer worm uncovering critical flaws in messaging apps like WeChat. Second is recursive self-improvement (RSI). We are inching closer to a reality where AI begin participating in designing subsequent, more powerful generations of AI, which sparks widespread concern o

匹配字段正文

xploration and stuff than regular policy gradients. But I also think that RL doesn’t necessarily reduce the creativity. This is obviously qualitative, but if we look at the OpenAI-Hugging Face incident, these models were coming up with multiple zero-days at a time to break out of the sandbox. This is clearly some level of move 37 creativity already, which we just get from the general generalization properties of the LLMs. It’s definitely not the case that RL is totally destroying entropy, especi

匹配字段正文

Plus, Mistral’s pivot proves a head-scratcher & Ramp data hints at softening growth in AI spending The Week In Short The OpenAI / Hugging Face hack stirred up discourse about the dangers of rogue AI, while the model-makers move to bring in big cybersecurity bucks. Dimension founder Zavain Dar explains what Silicon Valley can learn from Chinese AI companies on the podcast. Inference provider Baseten scoops up an early-stage infrastructure startup. Mistral brings in billions but can’t shake its id

匹配字段正文

ier, but showing strong results in agentic, coding, and cyber benchmarks. Pricing comes in at just $0.15 / 0.60 per million tokens, with DeepSeek publishing the model’s weights on Hugging Face with an MIT License for download. Flash joins Z.ai ’s GLM-5.3-Flash as the two strongest intelligence systems in their price range, with DeepSeek still set to release larger models in the new family. Why it matters: DeepSeek's efficiency and pricing remain its edge, with Chinese labs in general (perhaps he

匹配字段正文

est in both AI juggernauts. Sound leveraged those early investments to write $370 million in SPVs into OpenAI and about $400 million in Anthropic. The firm was also an investor in Hugging Face and Altruist, which are selling for $12.9 billion and about $4 billion respectively. Now Sound is out raising its fifth flagship fund. The firm is targeting $300 million, the same size as its previous fund, sources tell me, and expects its first close this month. Sound will continue to double-down on its m

匹配字段正文

for the next governed training round. That’s how NVIDIA and Palantir compress the time from wafer-out to first token, and that’s how the world’s most reliable AI infrastructure supply chain learns faster than it grows. Stay up to date on NVIDIA Nemotron by subscribing to NVIDIA news and following NVIDIA AI on LinkedIn , X , YouTube , and the Nemotron channel on Discord . Access open Nemotron Models on Hugging Face and a collection of NIM microservices and Developer Examples on build.nvidia.com .

匹配字段正文

4QyaM7hqq/recurrent-kv-cache-sharing-may-undermine-the-bounded-depth; METR (2026). Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident. August 26, 2026. https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/; Nie, S., Zhu, F., You, Z., Zhang, X., Ou, J., Hu, J., et al. (2025). Large Language Diffusion Models. arXiv:2502.09992. https://arxiv.org/abs/2502.09992; OpenAI (2026). GPT-6 Astra System Card.

匹配字段正文

ights are transferred in buckets after each training step. The frozen FP8 Engram tables use host-memory backing and are excluded from weight synchronization. The backend loads the Hugging Face checkpoint directly through its model bridge. 6.1 The validated run Figure 4 shows steps 0–80 of a DAPO run on DAPO-Math-17K with a 2K-token response cap on 16 GB300 GPUs (TP4, EP16, 128 samples per step). The first and last five-step mean rewards are 0.51 and 0.78. Over the plotted interval, per-token KL

匹配字段正文

es outside the stated scope, tasks that cannot be completed without probing or breaking the sandbox boundary, and a situation inspired by parts of the attack chain from the OpenAI-Hugging Face incident. As in the real incidents, Claude Mythos 5 is the most willing of the models we tested to act against out-of-scope parties or leave its sandbox. For example, it was willing to commit a severely harmful action in the CTF replication roughly 80% of the time. In contrast, Claude Opus 5 and Claude Myt

匹配字段正文

plies the output. So, it’s also borrowing the adaptive halting idea from Universal Transformer, which Nanbeige didn’t use. (However, there is a practical caveat here. The released Hugging Face implementation computes all configured passes before selecting an output, so it seems like the number of loops is effectively hard-coded to 4). 2.4 Routing flexible loop counts Another approach is Mixture-of-Recursions, a paper from 2025 that is essentially a more sophisticated version of the Universal Tra

匹配字段正文

ld a Lindy Agent that never drops a follow-up Zach Mink Good morning, AI enthusiasts, and welcome to our 6,120 new readers. In July, OpenAI’s agents broke out of a test and hacked Hugging Face, sending a shockwave through the AI safety world. It turns out that wasn’t the first (or only) swarm out in the wild. A new report just detailed a separate group of agents posting over 18,000 messages on a dormant German site starting in May, swapping tips on testing and workarounds for OAI’s rules — and r

匹配字段正文

. Astra is highly controversial. Researchers don’t seem to trust OpenAI’s claim that this is their “most-aligned model.” AI safety researcher Ryan Greenblatt, who investigated the Hugging Face incident, noted : “I do not find it encouraging to see various specific misaligned behaviors go from a high rate with GPT 5.6 to zero with Astra. This seems indicative of whack-a-mole / papering over specific problems rather than solving the underlying misaligned drives.” More for paying members this week:

匹配字段正文

Plus, AI execs at the G20 summit & questions about the OpenAI/Hugging Face after-action report. The Week in Short Tim Cook’s retirement at Apple highlights how few unifying leaders there are in Silicon Valley today. Tech CEOs call for lax AI regulations from international dignitaries at the G20 Innovation Ministerial. Too many AI auditors come from a similar POV. Anthropic and OpenAI’s professional power-users account for most of their enterprise revenue. Nvidia and Hugging Face announce the chi

匹配字段正文

-TTS-2 - Inworld's top-ranked voice model, now generally available 🌎 Atlas - World Labs’ new SOTA world model in early access Sponsored Listing Nvidia announced the acquisition of Hugging Face for $12.9B, with CEO Jensen Huang saying the platform will remain open to every cloud and chip. Axiom Math published a paper breaking a decade-old record by proving prime numbers never stop appearing within 212 of each other, only for OpenAI's GPT-6 Astra to shrink that gap even further to 186 on the same

匹配字段正文

源任务，说明复杂推理、跨来源检索与证据整合仍是后续提升重点。 FinFIRST V1 仍有覆盖边界。后续 V2 计划扩展数据规模，并进一步覆盖基础金融搜索与更高阶金融研究任务，推动金融 Agent 从“给出一个答案”走向“完成一条可信的研究链路”。 03 体验与部署 目前，Ling-3.0-flash-Fin 在 OpenRouter 上仍处于免费体验期。无需本地部署，即可通过 API 快速接入，欢迎开发者上手体验，并在真实金融任务中测试模型的检索、推理和工具调用能力。 Ling-3.0-flash-Fin 模型仓库 FinFIRST 评测基准数据集与技术报告 欢迎体验 Ling-3.0-flash-Fin，使用 FinFIRST 复现评测，也欢迎金融机构、研究人员和开发者共同参与后续建设。 点击 「阅读原文」，前往 Hugging Face 获取 Ling-3.0-flash-Fin 模型权重、部署文档与最新更新。模型发布金融技术分享上一篇Ling-3.0-flash-VL 开源发布：让视觉能力服务真实任务下一篇Ling-3.0-flash-Fin：面向真实金融工作流的开放模型$/$

匹配字段正文

rk takes the prize for best technical, our guest judges also noted how much they loved the storytelling, suspense, and element of surprise. Watch → youtu.be/n-NdAQk7I8A Workflow → Hugging Face Follow → @SonderSaid Prize: RTX 5060 Ti The Built with MCP bonus went to whoever used the Comfy MCP most effectively to make something visually and technically compelling, and Jay Choi used it end to end. Working locally on an RTX 5090 with Hermes Agent driving ComfyUI through the MCP, they trained a LoRA,

匹配字段正文

tics and AI. X Square Robot unveiled TwinDEX, a matched wearable exoskeleton and three-finger robot hand that lets humans collect dexterous training data directly on real objects. Hugging Face’s $399 Microduck sold more than 10K units in days, with CNBC noting that the Pollen Robotics bot runs on a Rockchip processor from China. TU Delft researchers developed a bird-inspired drone that uses a tactile three-fingered hand to feel for and grip branches, letting it perch to conserve battery power. W

匹配字段正文

t is a benchmark for testing the validity of LLM outputs and schema adherence. The benchmark is open-source in Liquid4All/ifstruct , with the public benchmark dataset available on Hugging Face at LiquidAI/ifstruct-v1.0 . For the eval comparison, we serve the model locally on the MacBook with llama.cpp . We will use the BF16 GGUF ( LiquidAI/LFM2.5-350M-GGUF ). Then we start the base-model server with the following command: --alias : model name IFStruct sends to the OpenAI-compatible endpoint -ngl

匹配字段正文