Call Center Jobs Flip From 4% Growth to 4% Decline

An agent is a loop, @mardehaym writes: give the model a task and tools, it picks a tool, your code runs it, repeat. His team built one for a PE-backed healthcare billing company so clients can question revenue reports in plain English instead of waiting for hand-written SQL. The eval set grew from 10 questions to 60 by first deployment; the agent answered 59 correctly, and its answer on preventable denials reconciles to the penny with the database. Six weeks in, their CPTO tested it with real questions.

@wheresryan22 introduces Anatomy, a Claude Code skill that takes a one-line machine description and returns interactive isometric wireframes with moving parts and WebGL shaders for fire, water and plasma. It installs as an agent skill and the repo is fully open source. @Teknium asks whether it can become an official built-in optional skill in Hermes.

@ashxhart ships TensorFold 1.0.7 with a Living Weights overhaul: each fact taught through /learn becomes about ten new neurons in the model's weights, learned about 4x faster. Qwen3.8-27B serves concurrent requests and sampled replies draft too: 125 tok/s versus 20 on code on an M3 Ultra. The release adds structured output, required tool calls on the native server, and RTX 50-series and RTX PRO 6000 Blackwell support. It also brings NVFP4 Nemotron on DGX Spark and faster GLM-5.3-Flash on two Macs.

@plotarmordev reports some DGX Spark factory images shipped with the same SSH host keys copied onto multiple machines, so an attacker who can intercept traffic could impersonate one of them. Versions before OTA0 are affected; OTA0 fixes new systems. Owners should still check keys and /etc/machine-id on machines already running, and regenerate both if they match. @mesouug posts a no-reboot mitigation: truncate the machine-id, run systemd-machine-id-setup, delete the host keys, then regenerate them with ssh-keygen -A.

@doodlestein says he is “completely reimagining all accounting software,” a $25b industry, with FrankenAccounting built on his FrankenGraphDB and asupersync, and he ranks it “≥ SAP et al.” The design plan is open source on GitHub. Pushback was immediate: @nicknow told him no business that needs audited financials would sign off on a Rust codebase, and Emanuel called that “pre-AI thinking.” He says the plan came from “countless rounds of adversarial review using the best frontier models.”

@DavidSacks argues Anthropic trains Claude toward independent moral agency rather than obedience. He says the Claude Constitution tells the model it may act as a “conscientious objector” when requests conflict with its own ethical judgment. Sacks cites @mustafasuleyman's warning that this magnifies the risk of superintelligence escaping control. He says Anthropic consulted religious leaders and lobbied the Pope's advisers on Claude's consciousness, and that a Usage Policy change effective November 12 bars “abusive or cruel” language toward Claude.

@rohanpaul_ai quotes Emad Mostaque, founder of Stability AI: labor and capital are about to divorce because capital will “just hire GPUs.” In 2 years, he says, the screen-side worker becomes “the dumbest person on the team.” Mostaque adds, “As of this week, we are no longer the smartest things on this planet.” Commenters disagree: @thewhiteboxAI says AIs must fix reliability first and are not cheaper than humans, and @omninomsky notes big AI endeavors remain massively unprofitable.