Description

Open weights caught the frontier and the model providers started picking sides on the same day: GLM-5.3 ties Fable 5 on the freshly recalibrated Terminal-Bench 4.0 while OpenAI moves to cut off SpaceX-owned Cursor, and underneath both, the hardware layer keeps attacking the memory wall — from Samsung computing inside DRAM banks to SpaceX putting an Nvidia rack in orbit.

1. OpenAI to cut off Cursor following SpaceX acquisition -
2. GLM-5.3 open weights land on Hugging Face -
3. Samsung's LPDDR5X processing-in-memory at Hot Chips 2026 -
4. htmx 4.0 released -
5. A rumour of a bug is now enough to produce a working exploit -
6. Booting a virtual iPhone on Apple's Virtualization.framework -
7. SpaceX to put an Nvidia Vera Rubin NVL72 rack in orbit -
8. AMD ships ROCm 10.0 -
9. Terminal-Bench 4.0 recalibrates the coding-agent leaderboard -
10. GUIs should be fully keyboard-driven -
11. US designates hosting collective Autistici/Inventati a global terrorist entity -
12. Turning LLM memory into program analysis with Datalog -
13. Google's Marvell deal pushes custom silicon past the TPU -
14. ROBBIN: injecting DNN backdoors via Rowhammer at inference time -
15. 181 tok/s aggregate on Qwen3.8-Flash-Next across 2x DGX Spark -
16. GSQ + RCO quantization ships Qwen3.8-27B at 2.5-3.0 bpw -
17. Ferroelectric tuning cuts optical-interconnect stalls in MoE LLM training -
18. Sapphire surface modification enables wafer-scale single-crystal 2D semiconductors -
19. Lean explained with TypeScript -
20. Crystal's execution contexts, now the default concurrency model -
21. Tesla's Cybercab takes the stage next week -
22. Microsoft's longest winning streak of the year as AI software fears fade -
23. Hiring slows again, and probably will not speed up soon -