aiweekly

Gradient Filter Finds Safety-Eroding Training Data LLM Judges Miss, Cuts Attack Rate Up to 22%

Fine-tuning on benign-looking data can quietly erode safety guardrails in ways standard content filters cannot detect.
aiweekly

Tsinghua and Berkeley Uni-Robot-Agent Harness Matches Fine-Tuned pi 0.5 on 19 Real-World Tabletop Tasks

Tsinghua University and UC Berkeley introduce a modular uni-robot-agent harness built around open models. Paired with Qwen3.8-Flash-Next, it scores ...
aiweekly

Google Launches Dreambeans in India in 13 Local Languages

Google has launched Dreambeans in India, bringing its proactive AI story app to 13 local languages including Hindi, Punjabi, Tamil, Telugu, and ...
aiweekly

Tsinghua, Berkeley Harness Lets Qwen3.8 Edge Fine-Tuned π₀.₅

Qwen3.8-Flash-Next, slotted into a new open-weight agent harness, scored 31.39 on the RoboDojo simulation benchmark. GPT-6 Astra–RoboProbe's ...
aiweekly

Larridin Raises $30M Series A to Audit Enterprise AI Spending

Want only the AI that hits your stack? Build an agent that tracks your companies and topics, and skips everything else. Larridin, a startup co-founded by two former Comscore executives, has raised $30 ...
aiweekly

NorthCUA Pro 397B Tops Qwen3.8-Max, Kimi-K3 on OSWorld-Verified

Want only the AI that hits your stack? Build an agent that tracks your companies and topics, and skips everything else. Open-source has now claimed computer-use-agent parity at a size (397B) that ...
aiweekly

SJTU, Ant Group Survey Charts Three Paths to Omnimodal Embeddings

Want only the AI that hits your stack? Build an agent that tracks your companies and topics, and skips everything else. Embedding research across text, code, image and video, audio, graph, table and ...
aiweekly

OSWAM cuts driving world-model inference to one step at 40ms

OSWAM, a new world action model for autonomous driving and embodied control, generates driving actions in a single diffusion step and runs with ...
aiweekly

The Information: Huawei Ascend 960DT Will Rely Heavily on Samsung and SK Hynix HBM Routed Through South Korea

We scan all of AI. You track a slice of it. Build an agent that watches only the companies, topics and people you choose. It reads everything our scanners pull and sends you an edition made of just ...
aiweekly

Wave Forcing Hits 134.8 End-to-End FPS Streaming Video on Eight H200s, 7x Over Single-GPU Baseline

We scan all of AI. You track a slice of it. Build an agent that watches only the companies, topics and people you choose. It reads everything our scanners pull and sends you an edition made of just ...
aiweekly

NorthCUA Open 397B Computer-Use Model Beats Trillion-Param Closed Rivals on OSWorld-Verified

We scan all of AI. You track a slice of it. Build an agent that watches only the companies, topics and people you choose. It reads everything our scanners pull and sends you an edition made of just ...
aiweekly

One Researcher, 1/12 the Training Data: Metis-1.6 Logs Ranks 7th on 40 Benchmarks

We scan all of AI. You track a slice of it. Build an agent that watches only the companies, topics and people you choose. It reads everything our scanners pull and sends you an edition made of just ...