Tuesday, 3 November 2026
Weight-space programming emerges as LLMs read and write neural weights; Ubuntu LTS kernels unpatched for critical 0day; Trail of Bits on voting cryptography
Today's Lead
EngineeringRikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab (MIT CSAIL)
Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
The article claims that frontier large language models (LLMs) can read and write neural network weights directly when given natural-language goals, without gradient descent or training loops. The authors call this approach Weight-Space Programming. They report that the model, named Astra, identified training games and generated content from raw weights, and repaired a stuck reinforcement learning policy by editing 3.5% of its weights to reach 100% success. They also report that editing 0.016% of the weights of Qwen3.5-0.8B raised MATH-500 from 7.9% to 48.1%, and that Astra wrote game policies that beat models trained with PPO. The authors conclude that weight editing can remove obstacles to an ability but cannot add abilities the model lacks, and that the approach remains limited by context length and harness design.
Engineering
Matthew Honnibal (honnibal.dev)
The 'Astra can read and write weights to improve Qwen' post is massively misleading
Matthew Honnibal argues that a blog post by rikiyat, which claims that "GPT-6 Astra" can read and write neural network weights, overstates its experimental results. The original post says Astra reads raw weights to infer training goals, writes small game-playing networks that beat PPO-trained networks, and raised Qwen3.5-0.8B's MATH-500 score from 7.9% to 48.1% by editing 0.016% of its weights. Honnibal's table of eight experiments shows that only three (1, 2, and 4) use raw weights alone, on networks of 17K to 172K parameters, while every real language model experiment uses Python. He also notes caveats, such as the MinAtar networks losing to PPO-500M on Seaquest and Asterix. He concludes that the Python-based and raw-weight experiments should not be presented as one capability, because the evidence comes from different techniques.
freenode
Ubuntu LTS kernels still unfixed for critical CVE-2026-98432
The article reports that CVE-2026-98432, rated CVSS 9.8 (critical), has no patched kernel for Ubuntu 22.04, 24.04, or 26.04 LTS releases. It states that Ubuntu gives no estimated time for fixes. The article says the risk is highest for internet-facing and multi-tenant self-hosted systems that rely on LTS kernels for isolation and privilege boundaries. It lists three options for administrators: stay on the stock kernel, apply out-of-tree mitigations if available, or move workloads to releases or kernels that include the patch. The only source it cites is the CVE record on cve.org, and it gives no technical details about the vulnerability mechanism or attack vector.
Trail of Bits Blog
The cryptography behind electronic voting (and why paper ballots are still essential)
The article argues that paper ballots remain the most secure basis for general elections, and that cryptography can strengthen them but cannot fully replace them. It cites a 2018 National Academies report stating that no known technology guarantees the secrecy, security, and verifiability of a marked ballot sent over the Internet, so internet voting should not be used in general elections. It describes cryptographic techniques for privacy, compromised clients, compromised backends, and end-to-end verifiability, including threshold-split election keys, homomorphic tallying with zero-knowledge proofs, Benaloh's challenge, and ElectionGuard's chained ballot commitments. It concludes that Helios suits low-stakes elections such as the IACR board elections, that ElectionGuard is designed as a layer added to paper-based systems, and that neither protocol fully matches paper-based security when run purely electronically.
LeadDev
AI turned tech leads into architects
The article argues that AI lets one engineer work across the whole stack, which shifts the tech lead's role from up-front coordination to later stages. It recommends drafting implementations with a frontier model, recording the plan in a design document, and having the lead review proposals and confirm that the right domain experts are involved. The author warns that speed without a high architectural bar risks subtle bugs, overly complex code, and eroded product quality. The conclusion is that tech leads must focus on judgment, review, and guardrails for AI agents. The claims rest on the author's personal experience as a tech lead, with no survey data, studies, or metrics.
Humanities
Aeon
The psychology of separating church and state
The authors argue that separating church and state assumes a psychological difference between religious and nonreligious belief. They call religious belief "religious credence" and ordinary belief "factual belief", and they identify three differences: religious credence is more voluntary, less responsive to evidence, and more confined to sacred settings. They use these differences to justify two points: democracies should not compel religious adherence, and believers can practise privately while setting their faith aside in public life. The evidence includes ethnographic work on Vineyard Church belief, developmental psychology, cross-cultural studies of confidence in invisible entities, behavioural field studies, and linguistic comparisons of "believes" and "thinks". The authors conclude that this psychological difference supports evidence-based public deliberation while leaving private religious life intact, and they acknowledge complications such as conspiracy beliefs and apologists who claim evidence for religious doctrine.
Crooked Timber
An open letter to JK Rowling from Sophie Grace Chappell
The article is an open letter by Sophie Grace Chappell, a trans woman, responding to J K Rowling's update to her June 2020 essay on transgender issues. Chappell argues that both sides of the debate are driven by anger over real injustices, and she says she shares Rowling's concern about violence against women and girls. She cites the United Kingdom's national statistics office and a World Health Organization report estimating that about one in three women worldwide experience physical or sexual violence, mostly from partners or family members. She argues that excluding trans women from women's spaces does not reduce violence and may expose women to suspicion and possible harm, and that countries with legal self-declaration of gender show no rise in violence, a claim based mainly on her personal impressions rather than cited data. She questions how the women-only service Beira's Place decides who is a woman, lists five criteria that can give different results, and concludes that prosecution, support services, and social conditions should be the focus.