AI & LLM · 2026-11-03DeepMind: an AI agent compresses a drug-design iteration to about an hour
Google DeepMind introduced "Cheminformatics Skills" for Google Antigravity. On real lung-cancer drug data, the agent searches thousands of compounds, runs matched molecular pair analysis (MMPA) and proposes novel macrocycles with synthetic routes in about an hour — a loop that typically takes several days. DeepMind itself cautions that the designs are testable hypotheses that must still be synthesised and validated experimentally.
Source: Google DeepMind
Computer systems & chips · 2026-11-03Qualcomm validates first HBC silicon: 18x memory bandwidth over LPDDR
According to Futurum, Qualcomm has validated its High Bandwidth Compute (HBC) technology in the lab: a two-layer stacked-DRAM sample bonded to a compute die measured 18x the bandwidth of LPDDR and 5x the bandwidth per watt of HBM4. Qualcomm did not disclose the workload, model or precision, so this is a component-kernel result rather than end-to-end inference. The Dragonfly AI250, the first HBC platform, is due to sample to customers in the first half of 2027.
Source: Futurum Group
Computer systems & chips · 2026-11-03Nordic nRF71: Wi-Fi, Bluetooth LE, Thread and edge AI on one chip, up to 20x lower standby power
Nordic Semiconductor announced the nRF71 SoC series (nRF7120 with dual-band Wi-Fi 6, nRF7120E adding tri-band Wi-Fi 6E). The company says standby power is up to 20x lower than the current nRF70 series, and the integrated Axon NPU runs ML models 10–17x faster than the 256 MHz Cortex-M33 alone — enough for keyword spotting, anomaly detection and image classification on battery-powered devices. Figures are the manufacturer's; the chips are sampling.
Source: Nordic Semiconductor
Computer science & data · 2026-11-02KAIST: CluSTER cuts LLM fine-tuning time by up to 69.6% with nearly the same quality
Professor Jae-Gil Lee's team at KAIST proposes CluSTER, which groups and selects data with similar learning effect and spreads it evenly over multiple GPUs to reduce repeated learning. On ~7B and 13B models, three coding/medical datasets and eight benchmarks, models kept nearly the same performance with only about 31% of the data while training time dropped by up to 69.6% (10 hours to about 3). Experiments were mainly on 4 GPUs; presented at EMNLP 2026.
Source: KAIST