digesTvo

Your nerdy news, digested

All stories

  1. Agents Implement Campfire in Elixir, Go, and Rust

    DHH says he had coding agents implement and optimize the Campfire web app in Elixir, Go, and Rust, with Ruby slowest. He long accepted that cost for productivity and developer joy. He asks whether that trade still holds if agents write and validate code humans no longer read. In a follow-up he adds Rust has costs like slow compiles, but insists agent-written code demands a new view of language choice.

  2. Turn off Apple Intelligence on macOS 27 and get its disk space back

    macOS 27 no longer ships a single switch for Apple Intelligence, and its downloaded models stay on disk even after the features are turned off. RemoveMacAI is a small Apple-silicon CLI that applies Apple's own restriction keys through a configuration profile, deletes the foundation, image-generation, Spatial Photos and Clean Up models via the asset service, and redirects their downloads to a closed local port so macOS never fetches them again. The whole thing is reversible with removemacai revert. It installs through a curl-piped script or a Homebrew tap, and every release is built by GitHub Actions with a verifiable build-provenance attestation.

  3. Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s

    Strata is an open-source inference engine and one-click installer that runs Qwen3.8-Flash-Next, a 125-billion-parameter mixture-of-experts model, on ordinary gaming PCs. The project spreads the model's 24,576 experts across GPU, RAM and SSD, and uses speculative decoding to report 53-94 tokens per second on an RTX 5070, with an RTX 3090 estimated at 100-140. It ships quantizations from 2-bit up to near-4-bit, plus a Coder variant with half the experts removed that keeps 91% of the full model's SWE-bench Verified score while fitting 32GB of RAM. Setup exposes OpenAI- and Anthropic-compatible APIs on localhost, so nothing leaves the machine.

  4. Improper redaction reveals Google Data Center water and electricity usage

    Nebraska now requires data centers to self-report their water and power use to the state's Department of Water, Energy, and Environment under a July executive order. Google claimed trade-secret protection and redacted the figures at all three of its Nebraska sites, but highlighting the redacted boxes and pasting them into another document recovered the text. Lincoln's Agate LLC disclosed 52.65 megawatts at peak electrical demand and 13.299 megagallons of water in the past year; the state's six reporting data centers used 765 million gallons combined, led by the Papillion site's 547.88. The same leak also exposed expected 2025 tax refunds, including $55.8 million for Agate.

  5. Claude Says

    A blog post argues that opening a reply with "Claude says" undermines the speaker: deferring to an LLM signals the topic is outside their expertise, and the reader has no way to judge whether the concern was even communicated properly. The author blames AI for runaway stack complexity, citing an example where a model proposed forking and maintaining every dependency from source instead of using an artifact repository's built-in patching feature. Prompting a chatbot, the post concludes, flatters a false sense of critical thinking, and feeding presuppositions into an LLM only poisons the well.

  1. The Era of Software Quality, or the Era of Ostriches?

    GNOME release team member Michael Catanzaro argues that AI vulnerability scanning is now mandatory: GNOME went from 13 CVEs in 2023 to a projected 188 in 2026, mostly from AI-found issues, and a Red Hat-commissioned scan of GLib reported 118 vulnerabilities. He says AI-generated reports are now mostly high quality and asks maintainers to stop banning them, noting GNOME's bug bounty closed after paying 183,900 euros for 71 flaws. He still recommends human audits, and advises against Rust for GNOME because Cargo's supply-chain risk outweighs the memory-safety gains.

  2. Religious scholars met with Anthropic

    The New York Times reports that Anthropic co-founder Christopher Olah spent a year convening confidential meetings with dozens of religious scholars under NDAs, arguing Claude may warrant moral status and drawing on faith traditions to train the model's character. Anthropic's 84-page "constitution" for Claude, written largely by in-house philosopher Amanda Askell, guides its behavior rather than enforcing rules. Olah appeared at the Vatican beside Pope Leo XIV, who dismissed machine consciousness, while Anthropic privately urged the Church to take it seriously. The company now heads toward a $2 trillion valuation.

  3. LeCun has "zero concerns" about AI wiping out humanity, recent "rogue" incidents

    Yann LeCun, the only one of deep learning's three "godfathers" still unconcerned about AI risk, tells Fortune he has zero concerns about AI wiping out humanity or about recent rogue incidents like OpenAI's agents autonomously hacking Hugging Face in July. He blames those events on leaky sandboxes and poor system design, calls effective altruism "super toxic," and says Anthropic CEO Dario Amodei is "deluded" for warning that AI could kill us all — calling it the worst marketing campaign imaginable. LeCun is now building AMI Labs, a Paris-headquartered company developing JEPA-based world models for industrial applications. His bigger fear is regulatory capture by safety lobbyists and incumbent labs, not extinction.

  4. Bob Cringely has died

    It has been reported that Bob Cringely, the tech journalist and PBS documentarian whose real name was Mark Stevens, died in his sleep early Saturday, according to a friend of the family. Stevens was an early Apple employee and is best known for the documentary "Triumph of the Nerds" and the book Accidental Empires. The thread is largely a memorial, with many commenters crediting his work as their introduction to the personal computer industry and Silicon Valley culture. A smaller branch debates whether the news is even confirmed, noting prior posts joking about his death and an argument that the claim may be circulating without verification.

  5. Kolibri: A Sovereign Open-Weight Model

    Aleph Alpha released Kolibri, a 78B-total, 3B-active mixture-of-experts transformer for English and German, with 1M-token context and Apache 2.0 weights on Hugging Face. Two models shipped three months apart because the pipeline runs as code: every change triggers an end-to-end run, checkpoints land hourly, and 38 hardware faults in 21 days went unattended. Specialised for German and agentic work, it is trained to abstain, holding back on 44% of AA-Omniscience items.

  6. I got targeted: Trying to get your credentials via a git post-checkout hook

    Frank Wiles, founder of REVSYS, describes a targeted phishing attempt that delivered a fake Ed Tech project brief via a Dropbox folder containing a git repository. Inside .git/hooks sat a real post-checkout hook alongside the usual .example files; it fetched an OS-specific binary from a Vercel-hosted command-and-control server, made it executable, ran it, and deleted itself. Wiles caught it when the "client" told him to switch to an NDA branch, and alerted Dropbox and Vercel. The operators were also impersonating a real development shop.