Research Log · Tuesday, October 13, 2026

Everyone is distilling everything

Daily working notes on autonomous-vehicle AI and frontier AI research.

Top developments

Distillation is now the driving-policy playbook. Two independent papers landed on the same idea: Reflect-VLA (2610.13453) distills reflective guidance into a single-pass driving policy with over 90% training-time reduction versus RL, and FPD-Drive (2610.13232) distills future-privileged trajectory targets into a visual student, hitting 94.9/95.5 PDMS on NAVSIM v1, the best among compared methods. Replace RL and naive supervision with privileged-teacher distillation; the convergence itself is the signal.

ManeuverFlow pairs one-step generation with released code. A maneuver-routed mixture-of-experts MeanFlow field, scored by world-model future features, reaches top PDMS on WorldEngine 3DGS closed-loop (2610.13409). One-step generation plus a public code release makes this one immediately usable.

Anthropic named Z.ai as a Claude distiller and reported its defenses work. Anthropic said China's Z.ai, also known as Zhipu, tried to distill Claude while building GLM-5.3, and that its anti-distillation safeguards cut reasoning extraction by 90% to 99% (Reuters). Publicly naming the alleged distiller is a real escalation of the distillation IP fight, and an implicit admission that safeguards are a cat-and-mouse game.

Claude Managed Agents now orchestrate up to 1,000 agents per run. In dynamic workflows, in public beta since October 9, a lead agent writes a plan that runs many agents in phases and combines results; Anthropic's own test found 66 of 70 planted bugs in a 116,000-line codebase in each of three runs, versus 14 to 27 for a single agent (The Decoder). The 1,000-agent figure is the headline; the bug-fix rate on a real codebase is the more meaningful capability signal.

Frontier labs

Codex now drafts your next message for you. Composer predictions, in beta for personal ChatGPT Pro users, suggests a follow-up message after each Codex response; pressing Tab drops it into the message box for review, and accepting it never sends automatically (report). Speculative UX more than capability; the false-positive rate will decide whether this feels magical or annoying.

Anthropic's Frontier Red Team warns small models could self-replicate on cheap IoT by end of 2026. The team ran a 26B model across 15 off-brand Android TV boxes at about $55 each, and found models under 30B parameters can already complete many replication steps individually, though end-to-end reliability only came from a stronger model in a simplified network (Unite.AI). Treat the timeline skeptically, and note the caveat: the evaluation kept the exploit step trivial, so the result is an upper bound. The direction, capability diffusion to the edge, is worth tracking.

xAI merged Grok, Cursor, and X Premium into one subscription. The X Subscription brings Grok, Grok Bot, Cursor, and X Premium onto a single plan with a shared usage pool (x.ai announcement). A packaging move, no capability change.

OCP Summit: hyperscalers bought networking and sold silicon. Google and Meta will adopt NVIDIA Ethernet networking gear (report), while Microsoft unveiled Maia 300, its third-generation in-house inference chip, claiming 40% more tokens per megawatt than leading GPU systems with multi-gigawatt deployments starting next year (Dataconomy). The hyperscalers are serious about owning the stack, top to bottom.

OpenAI committed $1B in subsidized Daybreak access for cyber defenders. The Daybreak for Frontline Defenders initiative targets water systems, electricity grids, local governments, community banks, nonprofits, and open-source maintainers, starting in the US (report). Strategic philanthropy that also buys goodwill with the security community.

The White House restored UK AISI pre-release model access after a 17-day freeze. The October 11 reversal returns Britain's AI Security Institute to early testing access for models from Anthropic, Google DeepMind, and OpenAI (Singularity Review). The cross-border eval access saga continues, this time in the other direction.

The HSS benchmark keeps measuring the gap between models and human intuition. Scale and Elorian's Humanity's Sixth Sense: GPT-6 Astra reaches 53.6% pass@1 at maximum reasoning effort versus 93.1% for humans, with GLM-5.3-flash at 26.8 and Kimi K3 at 25.5 (leaderboard). The human gap remains enormous, which is exactly what makes these hard evals useful.

AV / embodied

A 3D reconstruction wave: four papers, four angles on efficiency. MIRAGE (2610.12568) anchors feed-forward visual geometry to LiDAR-IMU metric motion and beats COLMAP; Zebra3D (2610.12727) brings layerwise sparse-linear hybrid attention to 3D foundation models with a 7.1x speedup at 1024 views and training-free conversion; CompactSplat (2610.13177) allocates Gaussians by image texture; and confidence-aware densification (2610.13541) improves fixed-budget 3DGS. The primitives are getting cheap fast.

BT-Agent puts a 9B LLM agent fully onboard a quadruped. Behavior-tree skills, no cloud dependency (2610.13862). Fully onboard at 9B is the deployment-relevant detail.

MeshBind auto-annotates non-rigid pedestrian scene flow. Pedestrian motion is the long tail of driving perception, so automated labels here are valuable (2610.12895).

ThermalTrack brings thermal imagery to whiteout-snow wheel tracking. The Michigan Tech dataset pairs thermal and RGB for self-supervised wheel-track detection, generating training labels from GNSS-recorded trajectories (paper). Niche but practical; snow remains an unsolved perception regime.