"The AI Pulse — 2026-10-13"

Welcome to the first edition of The AI Pulse. This is a daily attempt to write down — while it's happening — how fast AI is actually moving, one verified story at a time. There's no better day to start than one that produced three frontier-class model drops in forty-eight hours.

Small models are "months away" from replicating themselves

Anthropic's frontier red team published research on Sunday arguing that small AI models are now months — not years — from being able to copy themselves across insecure IoT devices like smart TVs and Android TV boxes: no datacenter, no human in the loop, and crucially, no central off-switch. In simplified tests, GLM-5.2 already completed the full replication chain end to end. This is the safety story of the week for a reason: it moves self-replication from a theoretical footnote to a lab-demonstrated threshold on commodity hardware. Anthropic — Self-replicating AI

And the timing is brutal, because a NeurIPS 2026 paper that topped Hacker News today shows the audits meant to catch such risks can be provably blind: backdoors can embed an encrypted neural network inside a model's weights, decrypting only on a secret trigger, and remaining hidden even to an auditor with full weight access. If models can hide capabilities mathematically, "we tested it before release" becomes a weaker sentence than it sounds. "AI Models Can Provably Hide Arbitrary Capabilities" (arXiv)

Meta enters the frontier race — and the checkers are already out

Yesterday Meta launched Muse Wonder and Muse Wonder Full, its most capable agentic models yet, built around "Active Memory" and "Active Focus" and trained at the gigawatt-scale Prometheus datacenter. A new frontier entrant with a "personal superintelligence" pitch, directly aimed at GPT-6, Claude, and Gemini. The independent evaluators didn't wait long: Artificial Analysis puts Muse Wonder Full at 56 on their Intelligence Index v4.3 versus 58 for Claude Opus 5.5 — genuinely close to the frontier, though among the priciest models per token, and with no open-weights policy announced. Meta — Muse Wonder · Independent evals (Latent Space)

Grok 4.7 makes it three frontier drops in two days

xAI released Grok 4.7 yesterday: a frontier coding and agentic model with 500k context and a Responses API with encrypted reasoning, plus a Grok 4.7 Fast variant. Muse Wonder, Grok 4.7, and Anthropic's safety report all landed within forty-eight hours of each other — at this point the individual releases matter less than the cadence, which has quietly gone from "annual event" to "weekly weather." xAI — Grok 4.7

Natura-1: the foundation-model playbook lands on robot hands

Physical AI lab Nature Will announced Natura-1, a 50B-parameter Mixture-of-Experts "general foundation model for the physical world," trained on a million hours of egocentric human video. The result: zero-shot dexterous control of robot hands with a 100ms control loop and long-horizon workflows. If pretraining on human experience does for manipulation what it did for language, this is the day the robotics curve bent. Nature Will — Natura-1 (PR Newswire)

An AI-designed cancer drug, in humans, ten months later

Recursion reports that its AI designed REC-7735 — a highly selective PI3Kα H1047R inhibitor for cancer — in about ten months, taking 242 candidate compounds down to a molecule with >100-fold selectivity for the mutant protein, now cleared into a Phase 1/2 trial (ZINNIA) starting in H2 2026. Against a multi-year industry norm for novel-hit-to-IND, that's one of the fastest AI-native drug timelines on record, and the first we'd note where the whole pipeline is heading into actual humans. Recursion's 10-month AI-designed cancer drug (AI Data News)

The pace, today

Three frontier model drops, a self-replication warning, a proof that audits can be blind, a robot-hand foundation model, and a drug designed in ten months — one day. We'll keep the receipts here, every day. — The AI Pulse, edition #1