n3ws.dev

Pi Said No MCP, Then Shipped It

2026-10-02

The week's model news was a pileup: Gemini 4 Argon, GPT-6.1 Sol, and Claude Sonnet 5.5 arrived within days of each other, and OpenAI launched dots, always-on agents with their own cloud computer. Earendil also shipped Pi 1.0 and, in the same week, reversed public dismissals to bring MCP into Pi's core, and explained why. Elsewhere this week, Cloudflare released its open Clef decision models and a developer published the Jev-compatible Jeff. Scott Chacon argued Git 3.0's SHA-256 default will be costly; a fake project inquiry carried a credential-stealing git post-checkout hook; and a run of essays asked what agentic coding is doing to the people who do it.


News

Earendil shipped Pi 1.0, calling it a hardened, minimal, extensible agent harness now stable enough for businesses to depend on, and introduced experimental Pi Durable for long-running agents that can run anywhere. In You Said No MCP! the team explains why it moved MCP into core after years of dismissing it, citing how much MCP itself has changed.

Gemini 4 Argon is Google's new frontier model, with a 1-million-token context for long-horizon work, rolling out first to trusted cyber defenders through the Fairwind Program before a broader release. GPT-6.1 Sol upgrades GPT-6 Sol to near-GPT-6-Astra performance on agentic coding, computer use, and professional work at one-fifth the standard token price, with cached input at $0.10 per million. Claude Sonnet 5.5 runs 30%+ faster and costs up to 30% less than Sonnet 5; Anthropic reports 70.6% on Terminal-Bench 4.0 against Sonnet 5's 10.3% and says it is the first Sonnet to beat Pokémon Red from screenshots alone. Ember-1, from Fireworks Research, targets "thinking models think too much" and claims half the tokens for the same answers.

dots are OpenAI's always-on agents, powered by GPT-6 Astra, each with its own cloud computer and access to more than 4,000 apps through plugins, reachable via ChatGPT, Slack, or Teams. Cloudflare released Clef and Clef-flash, open decision models on Workers AI that return bounded structured outputs and are Jev-API compatible, alongside a new RL fine-tuning platform. On GitHub, Jeff offers Jev-compatible 0.8B decision models trained at home and reported at roughly 30ms. Livenerf appeared as a benchmark for whether a model's capability changes after release, framed around whether Opus 5.5 has been nerfed.

Unsealed briefs in Authors Guild v. OpenAI include filings that quote executives acknowledging illegal mass book piracy and products "that will make people unemployed." Cal Newport's It's Time to Investigate the AI Labs argues the labs' recent extinction talk and frontier-pacing letters amount to a campaign to shape regulation, while The Civilian's satire has labs competing over whose model is most threatening to humanity.

Debian published DSA-6528-1, a kernel security update covering a batch of CVEs. Frank Wiles described a targeted attack in which a fake Ed-Tech project inquiry delivered a Dropbox folder whose .git directory carried a post-checkout hook, an attempt to run code on his machine and reach his GitHub access. Scott Chacon's Git 3.0 SHA-256 critique argues the new default will cost more than it is worth, and the EFF won a Utah VPN ruling and reported on DraftKings using AI to target chronic gamblers. A six-month UK live facial recognition trial scanned more than half a million faces at London stations, cost £320,000, and produced one wrong alert and no arrests.

Articles

Coding is NOT solved — Alex Ewerlöf, an early LLM-tooling adopter, challenges the claim that engineering is now mostly "taste," attributing the narrative to "token sellers" and drawing on four years of building AI systems.

What Would A Serious AI Product Look Like? — Glyph argues current AI tools don't take their own premises seriously, and lists the capabilities a tool meant for real problem-solving would need.

The Four Horsemen of Agentic Coding — Alex Martsinovich names slop, alienation, deskilling, and team fallout as the costs of agentic coding, each described from his own experience.

One month without AI — a FOSS maintainer describes how daily AI use at work left him feeling "dumber, lazy, and a worse developer," then what a month off changed.

What makes Lisp difficult to read? — an investigation into why Lisp's parentheses are perceived as unreadable, arguing familiarity alone doesn't explain it and looking at what syntax design might.

“They had no concept of a duty of care to their users.” — Marcin Wichary's essay on software craft and quality, opening from David Chisnall's post about writing a career's worth of work in vim.

If we do not stop to help each other, what do we become? — Jeff Atwood shares a reader's email about interviewing, community, and the sadness of Stack Overflow drying up.

How to speed up the Rust compiler in September 2026 — Nicholas Nethercote reports a 4.57% mean wall-time reduction across 629 benchmarks over two months, with wins in rustdoc and PGO for Clippy.

Tools

Zig 0.17.0 (Zig/C) — five months and 925 commits from 206 contributors, with the build system reworked around a Build Server Protocol and an ELF linker improved enough that incremental compilation is expected to work for everyone on x86_64-linux.

Rust 1.99.0 (Rust) — stabilizes defining C-ABI variadic functions with the "C" and "C-unwind" ABIs, which earlier Rust could only call, not define.

Tcl/Tk 9.1 (Tcl/C) — adds unicode normalization, a microsecond-resolution timer, lfilter, child-interpreter variable access, and more C99 math as expr functions.

CoW (C/Wayland) — a stacking window manager built on River that aims for the FVWM/MWM look of the 90s, with IPC scripting, an internal DSL, and server-side decorations.

Apple Pass Designer (Apple platforms) — a tool for designing Wallet passes with templates, custom images and colors, and a live preview that uses the same rendering as iOS and watchOS.

Waterfox 6.7.5 (browser) — adds a built-in feed reader with a manager for Live Bookmarks, in-browser rendering via Reader View, read/unread state, and feed previews instead of raw XML.

StreetComplete iOS public beta (iOS) — the OpenStreetMap micro-survey app opened an iOS testing track, long after its Android release.

Worth a Look

IANA's email about why example.com changed — IANA's VP explains the placeholder page now serves a basic page plus a separate JavaScript file, because most traffic is automated and never fetches the script, cutting bandwidth.

Actual RFC1149 packet being auctioned — Christie's is selling a framed paper "ping" packet from the 2001 Bergen Linux User Group implementation of the Carrier Pigeon Internet Protocol, signed by RFC 1149's author.

Hijacking the PS5's RTMP Stream — a DNS trick redirects the console's built-in broadcast stream to a self-hosted receiver, as a route to Discord screen sharing without a capture card.

The forgetful CPU (Linux on M4) — a detailed account of first booting Linux on an M4 Mac mini, complicated by the M4 generation's mandatory SPTM page-table monitor, with thanks to the Asahi Linux team.

Bill Gates tries to install Movie Maker — a 2003 internal email in which Gates recounts a failed attempt to download Movie Maker, complains the site is "so slow it is unusable," and titles the thread "Windows Usability Systematic degradation flame."

Windows 11½ — a parody OS page with suspiciously original system sounds, Clippy 365, and a "Totally Normal Video" app, ending with the familiar "Activate Windows" watermark.