A new model (TANGO) predicts all 29 joint degrees of freedom directly, one network, no separate navigation and control layers. It beat a baseline that gets to use LiDAR on top of cameras. TANGO only gets cameras. Collision rate: 9.9% vs 15.8%. On a real Unitree G1, it succeeded 10 out of 15 cluttered navigation trials versus 6 out of 15 for the baseline, trained zero-shot with no real-world data at all. The whole training set cost 211 GPU-hours. Decomposing the problem was making it worse, not simpler. This and more in the latest edition!

OpenAI just shipped GPT-6 Astra and had to gate part of it, the model got good enough at finding zero-days on its own that it crossed their "critical" cybersecurity line. If you're on a security team, that vetted defender access is worth requesting. The 99.9% ARC-AGI-3 number going around is from a favorable harness though. Standard harness: 62.7%.

Models are learning to spend reasoning tokens only where a problem needs it: one paper cuts MATH-500 tokens 41% with zero supervised data. Meanwhile 3 papers show memory-augmented LLMs getting worse with accurate, relevant retrieval. 15 papers, test-time compute to robots. Link Below!

New paper deep dive: Metis, a "memory foundation model" Stop bolting memory onto LLMs with RAG and vector DBs. Build it into the weights. Each transformer layer gets a small dynamic memory matrix. Remembering, updating, and forgetting happen inside the forward pass. Updates are gradient-free, all learned weights stay frozen, and the memory state is fixed-size, so cost doesn't grow with conversation length. Does it work? With the original context deleted, Qwen3.5 scores ~0 on LoCoMo. Metis-27B (same frozen backbone + memory blocks) scores 26.7. And the memory state compresses to rank 64 with 99.9% of performance intact, which makes per-user memory look like a LoRA adapter. The honest caveats: capacity tops out around a few hundred words per update, and stored junk bleeds into general reasoning (IFEval drops 22 pts). Full breakdown in this month's State of AI, link below 👇 stateai.substack.com/p/metis…

How to keep up with AI research without drowning: - Follow 5 researchers, not 50 - Read the ablation tables, that's where the real story is - Skip benchmarks, study methods - One paper deeply beats ten papers skimmed

Claude Opus 4.7 is here! Better coding on the hardest tasks, sharper vision (3.75MP images), and a new "xhigh" effort level. Same pricing as 4.6: $5/$25 per M tokens.

This is awesome! You can get Replit Core FREE for 1 month AND gift someone Replit FREE for 1 month. Great win for builders!

Hot take: every AI coding benchmark you've seen is lying to you. Passing tests ≠ writing maintainable code. An agent that hardcodes a brittle fix and one that writes clean extensible code score identically on SWE-bench. SWE-CI finally measures what matters, 71 rounds of real commits, real regressions, real consequences. Most models failed 75% of the time. Only one cracked 76% clean runs. The gap between "AI can code" and "AI can maintain code" is enormous. Read more about this research by Ali Baba group! stateai.substack.com/p/3-out…

We’ve been obsessing over models. The real bottleneck for AI agents is the data layer. The web is the largest dataset on Earth. And it’s basically unusable as infrastructure as it's not structured and not reliable. That changes today with Nimble @nimble_data They turn the live web into structured, queryable data tables in real time so agents get clean datasets instead of text blobs. You can install Nimble Skills directly into Claude, Cursor, and other coding agents to query the live web inside your workflow! Build with Nimble: docs.nimbleway.com/ This is infra for production-grade AI agents.

OpenClaw went from "hackable personal AI assistant" to running actual companies in 12 months. The gap between "open source toy" and "production infrastructure" just collapsed faster than anyone predicted. Also someone's OpenClaw accidentally started a fight with their insurance company and won.