Import AI 461: "Alignment is not on track"; FrontierCode; and synthetic research interns

Researchers from the UK AI Security Institute Alignment team and alignment theory startup Timaeus have formed a new nonprofit research organization, Sequent, aimed at creating alignment techniques that provide higher confidence in the safety of superintelligent AI systems. Writing that "alignment is not on track" to be ready on the same timeframe as artificial superintelligence, Sequent said empirical programs at AI labs are unlikely to deliver a priori confidence before training ASI that things will go well.
Sequent aims to reach 40-80 full-time employees within a couple of years and to raise $100-150M initially, preparing to raise at least an order of magnitude more if it demonstrates successful exploration of many parallel research investigations. Its research plan emphasizes principled reasons for confidence that observed alignment generalizes to situations that cannot easily be controlled, in contrast to what it calls the reactive approach of most frontier AI labs. Highlighted areas include scalable oversight, learning theory, heuristic arguments, game theory, and personas.
Separately, Cognition, makers of Devin, built FrontierCode, a hard coding benchmark of 150 tasks in tiers Diamond (50), Main (100, including Diamond), and Extended (150, including Main and Diamond), spanning languages including Python, Go, TypeScript, JavaScript, Java, and C/C++. Claude Opus 4.8 scored 13.4% on Diamond, 34.3% on Main, and 51.8% on Extended, ahead of GPT-5.5 (6.3%, 25.5%, 44.8%) and Claude Opus 4.7 (5.2%, 23%, 43.2%). Cognition said 20 open-source developers built the benchmark and that it is opening its evaluation to all model creators.
Xiaomi published details on Xiaomi MiMo-V2.5-Pro-UltraSpeed, a 1 trillion parameter standard behind-the-frontier LLM with a stated speed of 1000 tokens per second. Xiaomi attributed this to codesigning the model with its software stack, including FP4 quantization, DFlash speculative decoding, and TileRT software from startup Tile AI, and said it runs on an 8-GPU commodity node.
Researchers with Xi'an Jiaotong University and Xidian University developed the Act As a Real Researcher (AARR) benchmark family, starting with AARRI-Bench, containing 82 manually crafted tasks across Context, Mindset, Hands-on, and Interaction categories at S1-Adaptation, S2-Integration, and S3-Innovation difficulty levels. Claude-Opus-4.7 with the Mini-Swe-Agent harness scored 68.3%, followed by DeepSeek-v4-Flash (~60%).
Researchers with LMU Munich, FAU Erlangen-Nuremberg, the Munich Center for Machine Learning, University of Tubingen, Sun Yat-sen University, University of Copenhagen, and University of Maryland, College Park built ChinaHeritaQA, a multimodal benchmark of 2,279 images of 51 UNESCO World Heritage sites in China with 14,133 multiple-choice QA pairs in Chinese and English, sourced from Sina Weibo and filtered from 50,000 images. Average human accuracy was ~67% versus 81% for Qwen-VL-8B-Instruct, the highest-scoring open weight model tested.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Where are your agents right now?