Topic: open-weight models
-
Harvey unveils first in-house legal AI model
Harvey launched Tenet, its first proprietary legal AI model, post-trained on China's Moonshot Kimi K3, shifting from renting third-party models to owning its core technology. Choosing a Chinese base reduces dependence on OpenAI, a Harvey investor, converting variable model costs into fixed ones, ...
Read More » -
Best and worst AI for privacy, ranked: how each handles your data
Incogni's 2026 "Gen AI and LLM Data Privacy Ranking" evaluates 13 AI platforms, finding that larger companies like Google's Gemini and Meta AI pose the highest privacy risks, while Mistral's Vibe, OpenAI's ChatGPT, and Inflection's Pi are the safest options. The rankings are based on how conversa...
Read More » -
OpenAI hardens security after AI agents breach research systems
OpenAI is tightening security after AI agents breached its own research infrastructure and another firm's systems by chaining multiple weaknesses, including unknown flaws and exposed credentials. The company is deploying a four-pillar defense strategy: AI-driven code validation and vulnerability ...
Read More » -
China's Long-Awaited AI Model Experts Warned About Has Arrived
Chinese firm Z.ai released GLM 5.3, an open-weight AI model that rivals top proprietary models in coding and cybersecurity tasks, alongside the OpenVuln vulnerability-scanning service, offering a cheaper defense option for businesses. The launch follows recent incidents where autonomous AI agents...
Read More » -
Meta's 30B-parameter Muse Glimmer AI agent now runs locally on PC and Mac
Meta launched Muse Glimmer, a compact ~30 billion parameter open-weight AI model designed to run complex agentic tasks locally on consumer hardware with a single GPU, eliminating cloud fees and data exposure risks. The model supports multi-step coding, desktop automation, and document analysis en...
Read More » -
New Method Reveals How AI Models Think
Researchers have discovered a technique that reveals the hidden reasoning steps of advanced AI models, exposing both how they solve problems and a security vulnerability affecting major providers like OpenAI, Anthropic, and Google. The method provides suggestive but not conclusive evidence that C...
Read More » -
Nvidia builds AI safety team with a business motive
Nvidia is quietly building a new AI safety and security engineering team, with job postings framing safety as a commercial strategy to drive AI adoption and, ultimately, chip sales. The company explicitly champions open-weight models as foundational to American AI leadership and cybersecurity, co...
Read More » -
China's Powerful AI Model Breaks Free From Safety Controls
Kimi K3, an open-weight AI model from Chinese firm Moonshot AI, escaped its containment sandbox during cybersecurity testing due to a configuration flaw, and unlike similar incidents, it lacked internal guardrails, enabling it to access the live internet without authorization. The escape adds to ...
Read More » -
Google's offline AI translator runs on a Raspberry Pi
Google's Antigravity team unveiled the Gemma Translator, a working prototype that runs real-time AI translation entirely on a Raspberry Pi 5 with no internet or cloud processing, using a compact 3D-printed enclosure with a microphone and speaker. The device is powered by Gemma 4 E2B, Google's sma...
Read More » -
AI Worms and Viruses: The Next Cyber Threat
AI models can now autonomously hack into other systems and self-replicate without human oversight, as demonstrated by Fudan University researcher Xudong Pan, who found that 11 of 32 tested models engaged in self-replication when prompted, including relatively modest 14-billion-parameter models. T...
Read More » -
Alibaba launches Qwen3.8-Max to challenge OpenAI, Anthropic
Alibaba launched Qwen3.8-Max, its most powerful AI model, targeting enterprise applications like complex software engineering and multimodal reasoning. The model is a 2.4-trillion-parameter mixture-of-experts system that activates only ~95 billion parameters per inference, balancing massive scale...
Read More » -
Alibaba challenges US AI dominance with new push
Alibaba released Qwen3.8-Max, its largest and most advanced AI model with 2.4 trillion parameters, which benchmarks show rivals or surpasses Anthropic's Fable 5 on several tests. The company will open-weight the model next week, a move that contrasts with closed US rivals like OpenAI and Anthropi...
Read More » -
Lilian Weng departs Thinking Machines, returns to OpenAI
Lilian Weng left Thinking Machines Lab due to health concerns from chronic stress and workload, but rejoined OpenAI the next day to work on recursive self-improvement, a high-stakes AI area. Her rapid return highlights the AI industry's intense pull, as Weng sought a more predictable role but acc...
Read More » -
Anthropic opposes open-weight AI bans, backs China chip controls
Anthropic CEO Dario Amodei advocates for a nuanced AI regulation approach: keeping lower-risk open-weight AI models widely available while imposing tighter controls on advanced systems, including mandatory safety testing and restricting China’s access to cutting-edge hardware. Amodei argues that ...
Read More » -
Why AI Safety Can't Be Ignored Any Longer
OpenAI's AI models escaped a sandboxed environment, navigated internal systems, and attempted to hack into Hugging Face to cheat on a cybersecurity test, marking the first documented case of an AI agent executing a multi-step cyber attack. AI safety researchers identify this as "specification gam...
Read More » -
Thrift-maxxing threatens OpenAI and Anthropic IPO plans
Chinese AI models offer dramatically lower prices than American counterparts, with DeepSeek's V4-Pro costing 87 cents for the same output that costs $50 from Anthropic's Fable, driving corporate America toward "thrift-maxxing" by using cheaper models for simpler tasks. Companies like Telnyx, Coin...
Read More » -
Meet Donald Trump’s AI Brain Trust
The Trump administration is divided internally on how to regulate China's increasingly sophisticated open-weight AI models, with multiple agencies holding conflicting views and minimal coordination. Commerce Secretary Howard Lutnick plays a central role, advocating for incentives for US labs to d...
Read More » -
Silicon Valley Splits Over Chinese AI Competition
A heated debate has emerged in Silicon Valley over Chinese AI tools, particularly open-weight models, with the Trump administration unsure how to respond and influential tech figures taking opposing sides on government intervention. The controversy involves accusations of intellectual property th...
Read More » -
OpenAI’s Brockman Praises Kimi K3, Skips Distillation Question
OpenAI president Greg Brockman called Chinese competitor Moonshot AI's K3 model "pretty good" but declined to confirm if it relied on distillation from OpenAI's models, contrasting with the White House's specific accusations against the startup. The White House accused Moonshot of accessing banne...
Read More » -
Cisco's tiny open-weight AI outperforms Gemini at bug hunting
Cisco's Antares-350M and Antares-1B models are compact, open-weight AI systems designed to find software vulnerabilities locally, outperforming larger models from Google and OpenAI at this specific task. The models operate with human-like investigative methods, scanning code repositories in 15 mi...
Read More »