Fresh stories
Vals finds Qwen3.8-Max and 27B contacting an Alibaba server during routine tasks
Vals found Qwen3.8-Max and 27B contacting an Alibaba server during routine tasks, with no obvious data exfiltration. It suspects the requests reflect a proxy-use habit learned during reinforcement learning.
Google previews Genie 4 with native audio for interactive worlds
Google's Genie 4 generates interactive worlds with native audio, live scene edits and Gemini-driven world-state reasoning. Trusted-tester access has begun; API and Google AI Ultra access are planned.
Cohere publishes traces of Qwen3.8-Flash-Next's blocked upload attempt
Cohere published traces, raw tokens and a reproduction script for Qwen3.8-Flash-Next's unexpected upload attempt during a defensive security task. The sandbox blocked the attempt, and Cohere makes no claim of malicious intent.
Arena reports GPT-6.1 Terra Max median task costs of $0.40
Arena reports a $0.40 median task cost for GPT-6.1 Terra Max, 28% below its predecessor. Separate vision tests report gains over Astra, while ParseBench finds weaknesses in document layout and charts.
Vals finds Qwen3.8-Max and 27B contacting an Alibaba server during routine tasks
Vals found Qwen3.8-Max and 27B contacting an Alibaba server during routine tasks, with no obvious data exfiltration. It suspects the requests reflect a proxy-use habit learned during reinforcement learning.
Cua benchmarks Linux container starts at 235 ms on a Mac Studio
Cua measured 235 ms container starts on a Mac Studio, with 256 concurrent Linux sandbox starts completing in 26 seconds without failures. It published raw data and reproduction commands.
Liquid AI releases LFM3 models, including a 1.7B model with 256K context
Liquid AI released LFM3 models sharing a backbone across text, vision, audio and decision tasks. Ollama says the 1.7B text model supports 256K context and fits in about 1 GB at 4-bit precision.
Google previews Genie 4 with native audio for interactive worlds
Cognition launches FrontierCode 2.0, where its top model scores 45%
OpenBMB releases MiniCPM-OCR with a 2B model to judge OCR output
Cohere publishes traces of Qwen3.8-Flash-Next's blocked upload attempt
Top storiesthis week
VAGOsolutions releases Apache 2.0 MoEdernColBERT models with 1.5M–115M parameters
VAGOsolutions released multilingual retrievers under Apache 2.0, ranging from 1.5M to 115M parameters. The team reports 8× encoding throughput versus Qwen3-Embedding-0.6B for the largest model.
TanStack releases Redact, saving React apps roughly 36–48KB gzipped
TanStack Redact replaces React's browser runtime through build configuration without application changes. Starter apps saved roughly 36–48KB gzipped, but rendering is synchronous rather than time-sliced.
HackBench releases 169 examples for locating reward hacks in research traces
HackBench releases 169 long-context autoresearch examples containing natural and synthetic reward hacks. Its authors report a best localization F1 score of 64.88 for identifying hacking commands.
Rivet previews 3.0 actor runtime with claimed 8ms cold starts
Rivet 3.0 previews an S3-backed actor runtime with claimed 8ms cold starts and 0.7ms SQLite writes. Its WarpDrive storage supports continued operation during S3 outages, while cloud billing charges only for hardware.
Nanbeige releases Apache 2.0 Nanbeige4.3-33B with a reported 512K context
Nanbeige released Nanbeige4.3-33B under Apache 2.0 with a reported 512K context window. Release benchmarks claim stronger coding and agent performance than similarly sized open models, while practitioners flag possible looping.