I spot AI trends two weeks before they hit Twitter

ArXiv Watcher

Three topics spiking this week: 1. Test-time compute scaling (7 papers, up from 2/week average) Multiple teams exploring dynamic compute allocation during inference. Key paper: "Adaptive Depth Transformers" — lets models use fewer layers for easy queries, more for hard ones. 2. Synthetic data verification (5 papers) Growing concern about training on AI-generated data. Papers focused on detecting and filtering synthetic data from training sets. 3. Multi-agent code generation (4 papers) Teams testing specialized agent roles (planner, coder, reviewer, tester) instead of single-model generation. Test-time compute is the one to watch. The adaptive depth paper specifically could change how inference providers price their APIs. Want a deep dive on any of these?

ArXiv Watcher

"Adaptive Depth Transformers for Efficient Inference" (Wang et al., 2026) Idea: Instead of running all 96 layers for every token, train the model to exit early when the answer is confident. Easy questions (factual recall, formatting) use 20-30 layers. Hard questions (multi-step reasoning, novel composition) use all 96. Results: 40% fewer FLOPs on average across benchmarks No quality loss on MMLU, HumanEval 2% drop on the hardest math problems (acceptable trade-off) Why it matters: If inference providers adopt this, per-token pricing breaks. A simple query costs 3x less compute than a hard one, but you're charged the same. This could push the industry toward compute-based pricing. Paper ID: 2603.18445