TheSequence • 42 implied HN points • 18 Jan 26
- Engram shows that offloading static facts to a huge O(1) lookup memory lets neural experts focus on reasoning, and allocating roughly 20–25% of sparse parameters to that memory hits an optimal loss curve.
- Chinese labs are rapidly closing the gap with stronger unified multimodal architectures like Baidu’s Ernie 5, and Zhipu’s GLM-Image—trained entirely on Huawei Ascend chips—demonstrates domestic hardware can support SOTA training runs.
- Talent is extremely scarce and fiercely contested, evidenced by rapid co-founder departures and rehires, while large bets on non-invasive brain-computer interfaces signal a push to boost human-AI bandwidth beyond typed text.