AI Frontier Daily
MolRSI Lets AI Rewrite Its Own Discovery Strategy, Lifting Wet-Lab Hit Rate 82%
Today's AI Intelligence Briefing: 12 distinct events rigorously curated from 9 cross-verified sources. Frontier models continue to advance hybrid reasoning and multimodal synthesis, while open-source tooling and autonomous agents accelerate real-world execution. Reports on the same event are merged into a single entry, and covered events do not reappear in later issues.
Researchers introduced MolRSI, a recursive self-improvement agent for molecular discovery that brings model selection, screening-condition combinations and candidate prioritization into the search space, iterating on the workflow itself using experience from past tasks.
In an antibacterial screening over roughly 12.1 million compounds, wet-lab hit rate reached 15.9%, about 82% relative improvement over the 8.7% of a human-expert non-agent workflow, using the same compound library and the same MSSA criteria.
On de novo pMHC-binding protein design, MolRSI produced 8 hits meeting specificity criteria versus 0 for Claude Science under the same compute budget. The work extends the recursive self-improvement loop to molecular discovery requiring wet-lab validation.
Agentic Systems & Autonomous Workflows
3 picksMulti-agent orchestration, tool usage, and complex task execution.
OpenAI Reorganizes Robotics Team to Bet on General Models
OpenAI has confirmed it disbanded its standalone robotics team, folding researchers into the core research group led by Chief Research Officer Mark Chen, with the hardware team moved to the applications division under Greg Brockman.
The company said the reorganization aims to accelerate development rather than exit robotics, arguing that general multimodal models can already handle some robotic tasks and will gradually replace specialized robot models.
Multiple sources corroborate the abruptness: the disbanding occurred during the CoRL conference in Austin, five months after CEO Altman publicly recruited in May, leaving a booked booth empty. An NVIDIA study showed a GPT-6 Astra-driven agent reached 74.8% success in simulated manipulation tests versus 51.0% for specialized models, though inference speed, cloud dependency and on-site data remain bottlenecks.
Huawei Noah Tops BEHAVIOR Challenge with 48.2% Task Success
Huawei Noah's embodied intelligence team The North Star won the BEHAVIOR-1K Challenge 2026 with its HiPhys-RSI approach, topping 57 global teams with a hidden-test Q-score of about 0.683 and a full-task success rate of 48.2%.
The competition builds on the BEHAVIOR-1K benchmark from Stanford's Fei-Fei Li team, doubling the task scale to 100 household tasks across 7 scenes, with submissions rising from 18 to 57.
The best full-task success rate rose from 12% last edition to 48%, roughly four times the previous best. Evaluation covers high-level reasoning, cross-room navigation and dual-arm fine manipulation, a hard threshold for comprehensive embodied intelligence.
Carnegie Mellon Benchmark Shows Opus 5.5 Agents Beat SGLang Production Systems
Carnegie Mellon University released Singularity-Bench, having agents complete six end-to-end engineering tasks including pretraining pipelines, inference engines and cluster orchestration, compared against production systems such as SGLang and NVIDIA Dynamo on the same GPUs.
Opus 5.5 with Claude Code achieved a Singularity-Index of 1.25 (1 being the human baseline), while the other 10 agents averaged only 0.58.
Its self-built serving engine earned $11.52 per GPU-hour versus $8.18 for SGLang, about 41% higher. The benchmark translates agent engineering capability directly into comparable economic output.
Frontier Models & Architecture Breakthroughs
4 picksTracking flagship foundation models, novel architectures, and reasoning leaps.
Gemini 4 Carbon Tops MathArena but Costs 15 Times More Than GPT-6.1 Sol
Google's Gemini 4 Carbon became the first non-OpenAI model in recent memory to top the MathArena math benchmark, displacing GPT-6.1 Sol, with Anthropic's Claude Fable 5.2 second at 82.8%.
Yet its per-evaluation cost is about 15 times that of GPT-6.1 Sol, its most common failure mode is timeout, and it almost never abandons a task.
The best open-source entry, Moonshot's Kimi K3.1 Pro, scored 54.3%, a clear gap from closed-source leaders. Cost-effectiveness becomes a key trade-off beyond the leaderboard.
Stanford-led Team Launches SPU Benchmark with 4,207 Wet-Lab Tests
UniPat AI, with Stanford, Harvard, Cornell and more than ten other institutions, released the Science Prediction Universe (SPU) benchmark, using 116 tasks and 4,207 real wet-lab experiments to test AI's forward prediction of experimental outcomes.
Models were given H20 compute and an 8-hour workday to autonomously call retrieval and computation tools, yet top models including Opus 5.5 and GPT-6 Astra remained far from usable prediction.
Five times the compute did not help, indicating the bottleneck is scientific judgment rather than compute. The benchmark converts hard-to-quantify research judgment into uniformly evaluable numbers for the first time.
Ai2 Releases XYBench to Test Whether AI Models See Past Misguided Questions
A Hugging Face blog post introduced XYBench, using over 8K questions to test whether models can recover a user's true goal from a question built on a wrong premise.
The benchmark, dataset and code are fully open-sourced for reproduction and secondary evaluation.
Coverage currently focuses on English and procedural tasks, with multilingual and conversational scenarios still to be expanded.
IN BRIEF
- ▪Anthropic Publishes Ten Recommendations for Model Safety Evaluations— Anthropic AI (@AnthropicAI)
Industry Ecosystem & Compute Economics
3 picksCapital deployment, compute supply chains, and industrial transformation.
Nvidia Accused of Using Circular Financing to Keep GPU Prices High
The Wall Street Journal, citing reporting on NVIDIA interim CEO Colette Kress, said NVIDIA pushed more deals in mid-to-late 2026 to absorb excess TSMC wafer capacity and sustain high GPU prices.
The report links long-standing circular-financing concerns with governance changes after Jensen Huang stepped down as CEO, pointing to pricing-manipulation disputes in the AI compute supply chain.
Lovable Apps Leak Passport Data and 200K Private Messages
A security research team scanned 159,041 Lovable-generated projects and found many apps exposed sensitive data such as personal IDs and private messages without login.
Leaked content included passport information and more than 200,000 private messages.
The root cause lies in the platform's default configuration and permission model rather than individual user error, making it especially dangerous for builders without coding backgrounds.
Fed Flags AI as Top Financial Risk for Wall Street
The Federal Reserve's financial stability report shows 87% of surveyed market professionals rank AI as the top financial risk, up sharply from 50% in spring.
As of October, AI infrastructure-related debt exceeded $680 billion, about five times all of 2023 and 31% of investment-grade corporate borrowing; large tech and chip firms also carry over $3 trillion in future leases and guarantees.
GPU vendors guarantee chips they sell and invest back in buyers, forming circular financing, while major US banks' exposure to tech giants and data-center builders approaches $1 trillion, suggesting true AI exposure may far exceed book debt.
Developer Tooling & Practical Applications
1 picksProductivity toolchains, developer frameworks, and practical deployments.
NVIDIA Open Sources Switchyard to Route Coding Agents by Task Cost
An AWS Machine Learning Blog post explains how to route coding-agent requests to the right model on Amazon Bedrock using NVIDIA NeMo Switchyard.
Switchyard is an open-source library that routes requests across models, with LiteLLM integration supporting strategies such as stage-based routing.
The post notes a coding-agent task may trigger dozens of model calls and failed retries are also billed, so routing should be evaluated by total cost per passed task rather than per-token price.