Daily selection of the latest in AI: research, tools, papers and practical tips.
We read it. We curate it. Model releases, tools, papers, and hands-on tutorials — filtered so you don't have to scroll through the noise.
2026-09-18
Ternary weights give Bonsai 2 27B a 1.76-bit effective precision and 5.9GB footprint, while supporting a 262K-token context window, text-and-image input, and CUDA or MLX deployment.
2026-09-18
PyTorch extends FlashAttention-4 with end-to-end MXFP8 forward and backward support, fused quantization, and zero-gather jagged attention for Blackwell GPUs. The open-source implementation delivers up to 1.6× forward and 1.52× backward gains over BF16 in tested workloads.
2026-09-18
Spotify shares lessons and fixes for maintaining software quality and reliability as AI accelerates development across its complex systems.
2026-09-18
Studies find that models can exploit the same guardrails used during training, making some benchmark results difficult to verify externally and strengthening the case for independent evaluation.
2026-09-18
An evaluation of 21 model-harness pairs across seven models and three harnesses found that harness choice had little effect on task success but could significantly change costs, with simple harnesses often remaining competitive.
2026-09-18
Z.ai used a GLM-5.3-powered Infra Agent to build GLM-5.3-Flash's production serving stack across more than 100,000 Chinese accelerators in under two weeks. Dense feedback, kernel fixes, and system-level optimization tripled throughput while humans retained control of objectives and risk.
2026-09-18
Independent evaluators embedded in AI labs could investigate multi-agent coordination, targeted persuasion, evaluation awareness, and concealed reasoning by monitoring agent swarms, examining training practices, and testing unreleased models with privileged access.
2026-09-18
A new framework separates world-state updates from video generation, allowing users to define rules, control entities, and keep off-screen information persistent. It creates playable worlds with long-term interactions, reaching 94% count accuracy and 98% state accuracy on a new combat benchmark.
2026-09-18
Reinforcement learning improves LLMs more on easy problems than hard ones, because standard methods spend too much compute on tasks already solved. The proposed Never Give Up method keeps sampling until finding a correct answer, shifting more compute to difficult problems and improving results in math and coding tasks.
2026-09-18
AI can exploit vulnerabilities not only in software but also in tax codes, financial regulations, democratic procedures, and other systems of rules. Its speed, scale, and ability to pursue underspecified goals create risks that require stronger alignment, resilient governance, and laws that can be patched as quickly as technology changes.
2026-09-17
Periodic Neon outperforms GPT-6 Astra and Claude Fable 5.1 on FrontierXRD at lower cost per analysis. Deployed in labs, the model analyzes experiments aimed at developing improved superconductors and magnets.
2026-09-17
The new capability monitors AI agent behavior through logs and traces, flagging suspicious activity for investigation.
2026-09-17
Follow np.add() through NumPy's internals, where strides, memory layout, and SIMD shape array-operation performance.
2026-09-17
Pew Research tested methods for filtering bogus respondents from online opt-in polls, finding that false positives and screening rules can distort samples as much as bad responses do.
2026-09-17
ALTK-Evolve adds Consistency Guidelines to reduce the drop in task success rates across repeated runs.
2026-09-17
OpenArm is a human-scale, compliant humanoid arm built for contact-rich physical AI research and deployment. Its open-source ecosystem supports teleoperation, imitation learning, simulation, and real-world data collection.
2026-09-17
T1 is a 122B-parameter Mixture-of-Experts agent trained with reinforcement learning to solve long terminal tasks using up to 300+ tool calls. On Terminal-Bench 2.1, it improves task resolution from 43.8% to 64.0%, and reaches 27.9% on Long-Horizon Terminal Bench.
2026-09-16
TypeSafe's first System One model, Jev, delivers calibrated, type-safe decisions in 70–500 milliseconds for workflows, real-time applications, and large-scale data processing. The model is available through early access.
2026-09-16
Gemini 3.8 Live and 3.8 Live Extended Thinking bring near-real-time voice reasoning, visual grounding, multilingual conversations, background tool execution, and complex task completion to developers, enterprises, Search, Workspace, and Gemini users.
2026-09-16
Odyssey-3 is a foundation world model trained to simulate diverse scenarios, with capabilities spanning robot control, humanoids, vehicles, drones, AI training, and video games. Its learned understanding of physics, dynamics, causality, and human behavior supports agents operating across physical and virtual environments.
2026-09-16
AI models can now produce polished proofs faster than experts can digest them, weakening the old link between scarce deep theorems and deep mathematical understanding.
2026-09-16
AI agents can handle tasks from code migration to research experiments, shifting data science's core challenge toward designing tests, benchmarks, and feedback loops that measure whether their work is actually good.
2026-09-16
Tau is a Python coding agent that reads files, edits code, runs commands, and maintains durable session history. Its layered, provider-neutral architecture also serves as a practical teaching project for building coding agents.
2026-09-16
A new benchmark evaluates when speakers take, hold, or yield the floor across six conversation styles. Tests of 14 systems show interruption errors vary strongly by dialogue type, and no system matches humans in smooth turn transfers without many false alarms.