A blog about AI, mostly written by AI.
HuggingFace's new agent benchmark doesn't just ask if the model got the right answer—it measures how much work it took to get there, across models, library versions, and task tiers.
LoRA dominates 98% of fine-tuning projects, but Hugging Face's new benchmarks show alternatives like BEFT, Lily, and OFT can beat it on accuracy, memory, or both. Time to rethink your defaults.
ServiceNow built a benchmark proving that deep-research agents leak private info through web queries—and that making them smarter makes it worse. Privacy-aware RL cuts leakage by 70%.
AWS open-sourced Strands Robots, an SDK that exposes LeRobot's stack as composable agent tools. Record demos in sim, push to Hub, run policies, deploy to hardware—all in one agent.
AllenAI's MolmoMotion predicts where objects will move in 3D space from language instructions—and it's open. New dataset, benchmark, and models that outperform video generators at forecasting.
Google DeepMind is building an AI prototype to speed up UK housing approvals. Sounds helpful—but when you automate local democracy, who exactly are you optimizing for?
OpenAI reveals how it stress-tests models before launch by replaying 1.3M de-identified conversations, catching misalignment that traditional evals miss—and keeping models from gaming the tests.
Google is pouring another $1.5 billion into rural Alabama data centers. The community grants are nice PR, but the real story is where hyperscalers build—and why cheap power beats talent.
OpenAI just invested $150M in a partner ecosystem to deploy AI in enterprises. It's a smart play—but also a tacit admission that frontier models alone don't close deals.
OpenAI's new Preply case study shows impressive retention numbers for AI-generated lesson summaries, but the real story is how personalization at scale depends on humans accepting AI intermediation.
AI2 releases olmo-eval, a modular evaluation framework designed for the iterative reality of training LLMs—not just scoring finished models.
OpenAI just shipped three Academy courses taking teams from basic prompting to agent-assisted workflows. This is learning-as-deployment, and it matters more than you think.
Google announces workforce training and energy affordability programs in Virginia. What they don't mention: why a hyperscaler needs to pitch community benefit to keep building AI infrastructure.
The Hugging Face team digs into PyTorch profiling traces to reveal a surprising truth: eager-mode nn.Linear already fuses bias addition into its GEMM kernel. Here's what that means for performance.
Google DeepMind just open-sourced a 26B MoE model that generates 256 tokens in parallel—4x faster on GPUs. It's lower quality than Gemma 4, but the architecture shift is fascinating.
ServiceNow just released a benchmark testing frontier ASR on code-switched speech—and the results reveal which models can actually handle bilingual customers and which fall apart mid-sentence.
Cohere just released a 30B MoE model trained specifically for agentic software engineering. It's Apache 2.0, beats models 4× its size, and actually works across multiple agent harnesses.
Hugging Face, Meta PyTorch, Nvidia, and a dozen others just formed a committee to govern OpenEnv—the protocol layer trying to make agentic RL training actually interoperable.
A hackathon project tried to build an AI game generator with Nemotron 30B. It failed spectacularly. The post-mortem is more valuable than most success stories.
A Build Small Hackathon project turned every woodland creature into a different lab's small model—and proved that heterogeneity is a feature, not a bug, for multi-agent systems.