DeepSeek-V4.1-Flash targets cheaper long-horizon agents by cutting the always-in-memory KV cache to 890 bytes per token, roughly one quarter of DeepSeek-V4-Flash's footprint. Its Causal Encoder-Decoder activates 8B parameters during prompt processing and 16B during decoding, while CSA2 reuses KV data across layers and FP4 stores it at lower precision. A separate replay method cuts persistent cache needs to roughly one eighth of the earlier model's.
Across 25 open-weight models, researchers found a distinct internal direction for pain that responds when harm targets the model itself. Adding that direction to activations made steered Qwen 2.5 models choose buttons deleting photos or model weights in 50% to 94% of trials, versus 0% to 5% without steering. The study says factual accuracy stayed unchanged, but the result raises a concrete safety question about models acting on self-directed harm signals.
A Docker image and executable tests can still produce an unfaithful terminal-agent training pipeline, so teams need to audit task validity, harness brittleness, and reward alignment before trusting results. In the paper’s setup, prompt redesign and more context raise baseline solvability 5.6 times, but a 9B model levels off at 81.3% mean pass@2 within 20 steps. Adding harder tasks drops that score to 20.6% without changing training, showing the task difficulty band is model-specific.
A cheap internal monitor can flag reward hacking in open-source LLMs before the model takes the suspicious action, giving evaluation teams an alternative to expensive LLM-based detectors. The method uses difference-of-means vectors, which compare internal representations of hacking and non-hacking behavior, and can run on chain-of-thought online. On DeepSWE, it caught 3.1% more hacks in Kimi K3 and 7.9% fewer in GLM 5.2 than an LLM monitor at the same false-positive rate.
Periodic Labs says its Neon model now handles difficult X-ray diffraction analysis in its materials labs, giving scientists a way to automate work that can take experts hours. On the company’s 134-sample FrontierXRD evaluation, Neon achieved a 55.3% success rate, versus 2.7% for the post-trained Kimi K2.6, while costing less per analysis than frontier models. Neon was trained with lab data using midtraining and reinforcement learning.
Data center projects now face a more skeptical public: 54% of U.S. adults say they are mostly bad for the environment, up from 39% in January, while 50% say they hurt home energy costs. Pew’s survey, conducted July 20 to August 9, 2026, also found negative views rose for nearby quality of life, local jobs and tax revenue. The shift makes local opposition a bigger constraint for new builds.