Friday, 16 October 2026
Goodfire challenges polysemanticity orthodoxy; Meta releases Muse Spark Open; BIPM votes to end leap seconds in May 2027
Today's Lead
EngineeringGoodfire
Neural Networks Are Already Clean
Nick Cammarata argues that neural network neurons are already clean. He says polysemanticity is mostly caused by poor architecture, weak tools, and bad explanations. He proposes recursive generalization as the main explanation, not superposition. A neuron grows outward from a core concept, and its offshoots can look unrelated. He says researchers can study neurons directly, without sparse autoencoders (SAE). In a preliminary Qwen3-32B audit, about 90% of neuron poles have one main job.
Engineering
Meta AI
Introducing Muse Spark Open, The Most Capable US Open‑Weight Model to Date
Meta released the weights for Muse Spark Open, the most capable open-weight model from the United States to date. The model has 816 billion total parameters, 42 billion active parameters, and a 1M-token context window. Two Apple M5 Ultra Mac Studios can run the model with the MLX framework. Meta used Muse Code with the model to fix bugs and build features in Ice Cubes, an open-source Mastodon client. The weights use the Muse Model License, and hosted access is offered by several providers. Meta added extra safety post-training, and the full evaluation report is coming soon.
IPInfo
RPKI Now Protects 70% of Internet Routes, Up From 54% a Year Ago
The share of RPKI-valid routes rose from 54.0% to 70.3% in one year. RPKI lets address holders publish signed records that name the networks allowed to announce their addresses. China caused much of the rise, from 3.7% to 83.0% of its IPv4 addresses. The article warns that RPKI checks only the route origin, not the full path. In August 2026, a forged origin passed validation because the maxLength setting was too loose.
BIPM
Resolution 5 of the 28th CGPM (2026): On the technical actions needed to ensure the continuity of UTC
The 28th General Conference on Weights and Measures (CGPM) issued Resolution 5 on technical actions for Coordinated Universal Time (UTC) continuity. Continuous UTC takes effect on 20 May 2027. The maximum difference between UT1 and UTC will be 3,600 seconds (1 hour). This value keeps UTC continuous for several centuries. The BIPM must work with other bodies to inform users and monitor UT1-UTC.
Martin Fowler
When Software Meets Hardware
A team modernized a fleet of 800,000 battery-powered sensors in 3,000 stores. The team knew web, cloud, and mobile, but had little embedded firmware experience. It built a lab with real sensors and a virtual device emulator to test at scale. Devices must handle disconnection as normal behavior. Each device reports its firmware, health, and network state. Updates use staged rollouts and can resume after interruption.
Humanities
Aeon
How to be a (Roman) man
Edwards argues that Roman masculinity was always contested, not fixed. Cicero, the orator of the mid-1st century BCE, is her main case study. Roman men won manliness through battlefield courage, and Cicero showed that speech could win it too. Politicians' dress, gestures, and style were closely judged. After his exile and grief, Cicero wrote philosophy on self-control and virtue. Antony and Octavian had Cicero killed in 43 BCE.
Also today
Terence Tao
Why are we even doing all this?
Rikab Gambhir wrote this satirical guest post on Terence Tao's blog. The author says AI now produces many mathematics and physics results. The results look correct, but few people read or check them. The author uses irony to argue that this reduces research to a pile of unread facts.