GitHub's Rubicon Rewrites Git Storage to Handle 7,200 AI Agent Pushes per Second
GitHub's new Rubicon architecture decouples Git storage from compute, hitting 7,200 pushes per second on a single repo under agent-scale load.
20 stories
GitHub's new Rubicon architecture decouples Git storage from compute, hitting 7,200 pushes per second on a single repo under agent-scale load.
A telecom-specialized 31B model post-trained from Gemma 4 on 440B tokens of standards data lands on Hugging Face with QLoRA efficiency.
Nous Research post-trained Qwen3.8-27B to strip state-aligned behavior, cutting its propaganda rate from 22% to 1% with no capability loss.
Microsoft Foundry adds Kling, LTX, and MiniMax video models through a fal partnership, giving builders first-party access to five generative video options.
A month of software tuning pushed DeepSeek V4.1 Flash to 1.5x faster per-agent speeds on GB300, while Vera Rubin numbers land 5x throughput at high speeds.
A new benchmark shows open-weight models autonomously compromising simulated enterprise networks at roughly one-tenth the cost of restricted frontier systems.
NVIDIA's Lyra 2.0 turns a single photograph into a navigable 3D scene by generating a walkthrough video and lifting it into Gaussian splats.
OpenAI, Anthropic, Google DeepMind, Microsoft, and Amazon publish a shared scoring system for cyber jailbreaks, with risk bands from Negligible to Critical.
OpenAI and biotech firm Ligo released a 100-task benchmark that tests whether AI agents can actually interpret messy wet-lab evidence.
A new study shows that mixing benign examples into poisoned fine-tuning data can install backdoors that neither dataset would create alone.
Meta is distilling Muse Wonder's Active Memory and Active Focus into the models powering Muse and Meta AI, cutting forgetting by 2.4x.
Google DeepMind paired LLM agents with the Lean theorem prover to design new quantum algorithms, with machine-checked proofs and up to 290x cost reductions.
A new ICML paper formalizes when RLHF beats DPO, when DPO wins, and when online DPO quietly outperforms both.
Artificial Analysis launches VIALS-AA, a leaderboard scoring frontier models on 161 scientific image reasoning tasks spanning gels, flow cytometry, and phylogenies.
Qwen3-ASR-1.7B now ships as GGUF quantizations for transcribe.cpp, bringing a state-of-the-art multilingual speech recognizer to local CPU inference.
Meta FAIR's e-means replaces Lloyd's 68-year-old k-means routine with an EMA mini-batch optimizer, hitting lower loss in roughly one-tenth the epochs.
Ai2 releases an open simulated tabletop with 85K demonstrations, trained policies, and tooling to trace robot learning failures to their causes.
A hobbyist cloned Wispr Flow into a free, local Mac app by fine-tuning LiquidAI's LFM3-1.7B with DeepSeek-generated targets.
A new benchmark shows a tiny supervised model with 5-7 labeled failures beats complex unsupervised anomaly detectors by up to 70% on industrial time series.
Sandflow's sand3d turns 3D modeling into short Python programs that LLMs can write in a few hundred tokens, with datasets of real objects.