Tuesday, 13 October 2026
Meta reaches the frontier with Muse Wonder; AI models proven able to hide capabilities from evaluators; a nine-year Rust soundness hole surfaces
Today's Lead
EngineeringLatent Space
Meta Muse Wonder & Wonder Full: MSL's Claude 5.5-tier Frontier Model
Meta Superintelligence Labs (MSL) released Muse Wonder, a faster model, and Muse Wonder Full, a maximum-reasoning model. On the Artificial Analysis index, Wonder Full scores 56, and Claude Opus 5.5 scores 58. On the coding agents index, Wonder Full in Muse Code ties Claude Code with Sonnet 5.5. Wonder Full is available only to Muse Code subscribers, and Meta published no detailed system card. Meta admitted launch problems and a gap in 3D work, and early testers flagged weak frontend and creative writing.
Engineering
arXiv
AI Models Can Provably Hide Arbitrary Capabilities
A model can hide arbitrary capabilities behind a secret trigger. The authors embed an encrypted neural network in a host model, which is provably hard to elicit under cryptographic assumptions. The proof-of-concept models resisted fine-tuning, weight noising, trigger search, and latent adversarial training. For evaluators and regulators, passing an elicitation-based audit does not rule out a hidden capability. The paper is accepted at NeurIPS 2026.
Martin Fowler
Are Harnesses Worth the Cost?
Naresh Jain tests a behavioral harness for AI coding agents. The harness has guides that steer the agent, and sensors that check its output. He ran 10 paired runs per model, with and without the harness. The harness raised GPT-6 Luna from 3 of 10 passing runs to 10 of 10. GPT-6 Sol passed every run without the harness, so the harness added cost. He recommends the harness for cost-optimized models, and a final test gate for frontier models.
GitHub / Rust
Struct's Lifetime Requirements Can Be Bypassed via Covariant Field Assignment
The Rust compiler accepts code that bypasses a struct's lifetime bounds. A struct with a 'static bound can receive a shorter-lived reference in a field after construction. Its Drop impl then treats that reference as 'static, which can cause use-after-scope reads. The thread has about six variants that participants treat as one root problem, and some variants have compiled since Rust 1.18.0 (2017). The issue is open with no fix, labelled I-unsound and P-high, and AddressSanitizer reports stack-use-after-scope on the reproducer.
argmin.net
A Look at the Math That OpenAI Claims Computes Faster Fourier Transforms
OpenAI published a paper that claims a Fourier transform algorithm with a smaller log(n) factor than the Cooley-Tukey FFT. The standard FFT uses under 5n·log₂(n) operations. Ben Recht calls the gain a technicality, because the hidden constant is huge, about 10¹⁰⁰ at n = 10¹². The result assumes exact, arbitrary-precision complex arithmetic, which no real machine performs in one step.
Humanities
Aeon
Deskilling the Body
Melanie Challenger argues that reliance on phones, GPS, and AI erodes physical skills. She says technology that replaces bodily action weakens the abilities that action builds. Studies show that GPS users lose spatial memory faster than non-users. She argues that organisms maintain themselves, and intelligence comes from the whole body. She concludes that without embodied action, people will be ruled by technology rather than liberated by it.
JSTOR Daily
What Do Cops Train For?
Anthropologist Jessica Katzenstein observed police training in Maryland from 2015 to 2018. Departments said they wanted officers to act as community-serving guardians, not warriors. Her study found that scenario training often showed worst-case events, such as a child's phone that turns out to be a gun. She argues that command presence rests on racialized assumptions about a person who responds only to dominance. By the end of her fieldwork, many protest groups had moved from calling for training to calling for defunding police.