ThursdAI weekly: Qwen 27B and GLM 5.3 beat GPTs; OpenAI pauses RL for security
thursdai_pod · x · 2026-08-21
ThursdAI — the weekly high-signal AI news show from Weights & Biases hosted by Alex Volkov (8,000+ subscribers) — covers this week's highlights:
- Qwen 27B and GLM 5.3 beat the GPT models in benchmarks;
- OpenAI announces pausing RL training to focus on security;
- HeyGen introduces HyperFrames, framing video editing as a coding problem, open-sourced so agents can adopt it (full breakdown in the linked letter).
More from Models
- Direct-OPD Transfers Weak Model RL Gains to Strong Models — tokenbender · 2026-08-21
- New Chinese model OX Alpha spotted, Zhipu and Xiaomi suspected — vista8 · 2026-08-21
- Ox Alpha: Free Chinese model with 1M context but high latency — bindureddy · 2026-08-21
- Forcing Simple Writing Reduces LLM Reasoning Quality — littmath · 2026-08-21
- Legal AI startup Harvey launches first model post-trained for law on Kimi K3 — Polymarket · 2026-08-21
- Ox Alpha appears to be GLM 5.3 Flash — LinusEkenstam · 2026-08-21