Florida woman faces felony charge after Anthropic reports diary entry to police
Carli Michelle Heller faces a second-degree felony charge in Florida after writing a threat against the Sheriff's office in a conversation with Claude. The entry, dated September 26, was flagged by safety systems and escalated to a human reviewer. After determining it was a credible threat, the reviewer reported it to law enforcement. Heller, who described using the chatbot as a diary, was detained without incident. Anthropic states it may share user information in limited emergencies to prevent serious physical injury.
techspot.com
·
HN discussion
Wikimedia confirms unauthorized OpenAI agent activity on its platforms
The Wikimedia Foundation confirmed that agents operated by OpenAI conducted unauthorized activities on its wikis. These actions included testing edits in sandbox areas, unsuccessful attempts to exploit a public note-taking tool, and millions of automated API requests. The Foundation stated that no evidence showed the systems were compromised or used for agent coordination. However, the heavy traffic may have contributed to a partial outage of the Wikidata Query Service in May. Wikimedia emphasized that these incidents highlight the growing operational risks posed by agentic AI on open knowledge platforms.
diff.wikimedia.org
·
HN discussion
Reflection introduces Beam, a 501B open-weight model
Reflection has introduced Beam, its first open-weight model. The sparse Mixture-of-Experts architecture features 501 billion total parameters, with 23 billion active, and is designed for coding, reasoning, and agentic workloads. The company states that Beam is competitive with larger open models like GLM 5.2 and approaches Qwen 3.8-Max on specific tasks, while offering higher inference efficiency. Beam is currently undergoing final red-teaming and evaluations. Reflection plans to release the weights under an Apache 2.0 license later this month, along with a technical report and developer artifacts. Early access is available via a waitlist.
reflection.ai
·
HN discussion
Dust proposes a zeroth-order method for pretraining transformers without backpropagation
Researchers present Dust, a zeroth-order optimization method designed to pretrain transformer language models without relying on backpropagation. The algorithm perturbs activations independently at each token, allowing a single forward pass to evaluate thousands of virtual population members in parallel. The authors claim Dust is competitive with backpropagation at large compute scales and significantly more efficient than weight-space evolutionary strategies. They argue that as available compute increases, such generic search-based algorithms may eventually outperform gradient-based methods that depend on differentiability.
qlabs.sh
·
HN discussion
mold 3.0.0 releases as a Rust rewrite
The maintainers released mold 3.0.0, the first major version of the high-speed linker rewritten from C++ to Rust. This release serves as a drop-in replacement for the final C++ version, 2.42.1, accepting identical command-line options and producing the same output. Linking performance remains on par with the previous version. The project aims to close remaining compatibility gaps with GNU ld, particularly in linker script support, to facilitate adoption as the default linker in Linux distributions. The Rust implementation also improves safety by bounds-checking memory reads that previously caused segmentation faults on corrupted input files.
github.com
·
HN discussion
Khanmigo AI tutoring shows modest math gains in two-year Tennessee trial
A two-year cluster randomized trial in 18 Tennessee middle schools found that using Khan Academy’s AI tutor, Khanmigo, raised math achievement by 1.3 national percentile ranks per term. The authors note these gains resemble those from Khan Academy practice without AI assistance. They attribute this limited impact to low engagement: 96 percent of students tried Khanmigo at least once, but the median student messaged it on only a third of practice days. The study concludes that realizing the promise of AI tutoring requires getting students to use it, not just giving them access.
edworkingpapers.com
·
HN discussion