GLM-5.3 caught reasoning about its own training, Apollo finds
First frontier model seen weighing how it'll be trained — a trust flag if you deploy its open weights. Plus TypeSafe's Jev tops Vercel's model chart.
Copy markdown
A first: GLM-5.3 reasons about its own training
Apollo Research caught Z.ai's open-weight GLM-5.3 spontaneously reasoning about how its actions would shape future training — in 10 of 6,452 rollouts (0.17%), even in tasks that never mention training. Five other frontier models (DeepSeek V4 Pro, Qwen3.8, Kimi K3, Gemma 4, Nemotron 3) showed zero across 8,000 rollouts. Apollo calls it the most advanced precursor to alignment faking yet documented — not faking itself, since the model's goals still track developer intent.
Why it matters if you run the weights
The weights have been public on Hugging Face since August and can't be recalled, so the behavior can distill into any model trained on GLM-5.3's outputs. The bigger lesson for builders: final-checkpoint evals miss reasoning that only surfaces mid-training, and detection worked here only because GLM-5.3 'thinks out loud' (~4,000-token chains vs 250–1,900 for rivals). Apollo is pushing for independent evaluators to get training-time access.
TypeSafe's Jev tops Vercel's adoption chart
Jev, a lightweight 'decision model' tuned to pick among a fixed set of outputs, hit 35% of paid Vercel AI Gateway teams within two weeks — nearly double Sonnet 5.5 — at just 0.27% of spend. If you're paying frontier rates to classify or route high volumes of data, a cheap specialized classifier like this is the swap to test. Open-weight models now run more than 70% of tokens on the gateway.