AI & ML Research 24 Hours

Last updated: August 25, 2026, 9:07 AM ET

AI & ML Research

Chinese humanoid robots took center stage at a Shanghai robot carnival, part of the country's broader push to embed AI into physical systems for daily life. Meanwhile, researchers are turning to spiderwebs as a novel source of environmental DNA, offering a low-cost way to census biodiversity, track migrations, and spot invasive species.

OpenAI announced it disrupted a Russian influence campaign that used its models to promote a fake Israel-based think tank and a "sovereignty" index praising Moscow while criticizing the West.

On the engineering side, developers can now inject custom logic into Codex agentic loops via hooks, enabling finer control over autonomous workflows. A separate enterprise RAG critique challenges ten mainstream tutorial positions, arguing that real-world document intelligence demands a different architecture than typical demos.

A novel LLM inference runtime tackles the problem of forgetting with deadlines: it refuses admission rather than miss a 33ms robot control cycle, evicts KV cache by semantic meaning instead of age, and is written entirely in Rust. In education, guidance is emerging on smarter classroom AI use, helping schools move beyond blanket bans toward structured application of LLMs.

Finally, a new technique called DFlash brings speculative decoding to CPUs, achieving 3.92x faster token generation than autoregressive throughput in vLLM tests with Qwen models—without altering output quality.