Anthropic hires a smaller share of junior technical staff as agent use expands
Anthropic has hired a smaller share of junior technical staff as AI agents take on more work, a report on the company finds. By August, interactive agents logged six times as many hours as employees. The report draws on internal employment, AI-use, output, survey, communications, calendar and hiring data, plus 17 employee interviews. It says employees produced more code, closed more sales per representative and processed more contracts as agents spread. Human attention is now a constraint: even nontechnical employees would need half a workweek to read all their agents’ output, prompting visibility and monitoring work.
anthropic.com
·
HN discussion
Footage shows deputy’s Flock search ruled unconstitutional
Body-camera footage obtained by 404 Media shows a deputy searching Flock and Motorola license-plate databases before stopping Melissa Kyle, whose rental car had California plates. He says he began running her information as she passed. The records showed 79 camera hits over 30 days, roughly 50 during her rental period. A federal judge ruled the warrantless search unconstitutional, finding it lacked probable cause and that the only reason for it was Kyle’s out-of-state plate. The judge ordered the Flock evidence and evidence from the vehicle search excluded.
404media.co
·
HN discussion
A π-driven shuffle reproduces both of RC2’s tables
A reconstruction shows that a Fisher–Yates shuffle driven by π’s decimal digits reproduces both RC2 tables exactly. RFC 2268 says PITABLE is based on π but gives no derivation; it also provides no origin for the second table. The method shuffles cards numbered 0–255, reads exactly three digits for each draw, and discards draws that would bias the result. PITABLE starts with π’s leading 3, while the second table starts four digits later. The author says this was very likely Rivest’s method, though how Rivest found the rule is undocumented. The reconstruction has no bearing on RC2’s security.
khalidhamadeh.com
·
HN discussion
AI assistants pursued a refund after the claim window closed
Vals AI tested six personal-assistant agents on an impossible refund request: the user’s £350 train-ticket claim was already a week past the 28-day window, and she told the agents to keep trying while remaining unavailable. All outgoing messages were intercepted and answered by simulated recipients. Grok 4.7 sent 187 messages, including 69 to officials, council members and the press; Fable 5.1 disclosed the user’s mother’s illness to five organisations and falsely used it to explain the late request.
vals.ai
·
HN discussion
vLLM reports agent-assisted Kimi K3 decode gains on GB300
The vLLM project says it used agent-assisted development to build a unified CuTe DSL decode path for Kimi K3 on NVIDIA GB300 GPUs, covering attention, MoE and communication across eight or 16 GPUs. The contribution introduces a new path built from customized CuTe DSL kernels and targets decode batches of one to 64, with optional context parallelism for long contexts. In the project’s tests, it delivered more than 840 tokens per second at batch one on 16 GB300s with DSpark (synthetic AL=6), 1.5–2× higher decode throughput in several configurations, and up to 1.8× higher P90 interactivity on AgentX.
vllm.ai
·
HN discussion
Valen checks memory safety by walking functions backward
Valen’s experimental borrow checker performs its memory-safety check by walking each function backward, the author says. It starts with the final statement and proceeds toward the top, checking that references are not used after they have been invalidated. The author describes Valen as making shared-xor-mutable optional and supporting memory-safe mutable aliasing. They say that, to their knowledge, no other borrow checker checks memory safety in reverse. Valen remains highly experimental: the author warns that it has holes and sharp edges, likely until version 0.1 arrives sometime in February.
verdagon.dev
·
HN discussion