AI Briefing#llm
The latest AI and developer news about #llm, with the original source and a short summary.
The latest AI and developer news about #llm, with the original source and a short summary.
Competing statements in context cause 9-110x higher loss than neutral filler, while length alone explains only 2-8% of variance.
The new PCIe card features 64 AI cores, 576 MB on-die SRAM, and 548 GB/s DRAM bandwidth for generative AI workloads.
Total voice agent costs combine the $0.08 platform rate with separate LLM and telephony charges, with plans ranging from a free 15-minute tier to custom Enterprise pricing.
Post-trained from GLM-5.3, the model achieves a 38.3% pass rate on FrontierScience Research, surpassing GPT-6 Astra's 33.3% at max thinking.
The feature searches the web before answering, raising accuracy on current-fact questions from 51% to 86% while keeping median latency under 400 ms.
GPT-6 Astra improved the MATH-500 score of Qwen3.5-0.8B to 48.1% by modifying its weights without training.
A viral 7.72 TB Hugging Face repository claiming to be a leaked Anthropic model was proven to be untrained data generated from SHA256-hashed tensor names.
Mistral released the weights for its 1T MoE model, Mistral Large 4, demonstrating token efficiency.
Sleep improved the AIME-24 score by 5.4%p through offline integration.
The shift to background, goal-oriented agents introduces unsolved problems in proactivity, memory, and alignment that standard coding agent patterns do not address.
The Qwen series has surpassed 3 billion cumulative global downloads, with over 460 open-source models released and more than 300,000 derivative models.
MIT CSAIL researchers demonstrate that GPT-6 Astra can read and write neural network weights, improving Qwen3.5-0.8B's MATH-500 score from 7.9% to 48.1% by editing 0.016% of its parameters to fix reasoning loops.
AnySpeechRecognizer, a Swift package that reimplements the Apple Speech API to enable the use of open models, and AnyDecisionModel for decision tasks have been released.
A study on Gemma 3 270M shows that training just 1,000 random weights (0.001% of the model) enables hidden skills like sorting and JSON output that prompts alone could not unlock.
In a single-file test of 100 WordPress 4.0 files, the best LLM left only 59 passing behavioral checks compared to 83 in the unchanged originals.
The 4GB model targets voice assistants and robotics on RTX, Jetson, and DGX Spark hardware.
The model uses a looped decoder to execute 56 layer passes from 32 distinct layers, supporting text and image inputs for agentic workflows.
The model outperforms GPT-6 Astra and Claude Sonnet 5.5 on plain prompts by using Persimmon as a simulated learner during training.
Amici curiae including the US Government, Authors Alliance, and independent creators argue that using copyrighted works for AI training is transformative and does not cause cognizable market harm.
Swift GLM-5.3-Flash achieves -61.4% thinking tokens on MMLU-PRO with -0.46% average accuracy loss across nine benchmarks.
The system achieved a 60% true-positive rate and engaged 589 engineers through a human-on-the-loop workflow.
The 27B parameter open-weights model is free for non-commercial use and outperforms real human writing in detector evasion tests.
The update improves agentic coding, defensive cybersecurity, and token efficiency without requiring changes to the existing model ID.
Liquid AI has released six 2B-scale LFM3 models supporting execution on edge devices and multimodal processing.