Apollo researcher finds AI思维链语言高度异化、abstract reward source tracking, rendering human monitoring increasingly unreliable.
Zhipu AI open-sources GLM-5.3 model weights with commercial licensing.
Jalapeño shows higher inference efficiency and faster single-user generation in public benchmarks.
Jalapeño shows stronger efficiency and low-latency results than NVIDIA GB300 in public benchmarks.
MechRev's newest slim gaming laptop features Ryzen 7 H449 and RTX 5060 GPU, starting at ¥9,999 after subsidy.
Meta's Project OT planned to cut up to 60% of some teams and go AI-native; after round one was executed, round two was cancelled — the organizational integration cost of AI often exceeds the cost of the technology itself.
WhatsApp is testing Scam Alert, which puts the scam-detecting AI on the phone itself so end-to-end encryption stays intact—a transparency ledger guards against server-side manipulation, and confidential computing seals off telemetry.
In a Gates Notes essay, Gates floats two proposals: taxing robots and AI tokens, and creating human-reserved jobs — the current tax code structurally rewards companies for replacing people with machines.
ByteDance's Doubao Work logs in with your Feishu enterprise account and inherits permissions and context out of the box, zero configuration required — the enterprise Agent race moves from 'who works better' to 'who knows the org better'.
Anthropic's revenue run rate jumped from roughly $30B to $65B in four months, overtaking OpenAI — yet TechCrunch made 'who's still standing in five years' the central question of Disrupt 2026. Why does rising valuation force earlier scrutiny of revenue quality?
