Google debuts Gemini 4 Argon, Netlify cuts edge latency, and mathematicians challenge AI lab practices

In this issue

Google announces Gemini 4 Argon for coding and cyber defense
Pi adds native MCP support after re-evaluating the protocol
Netlify replaces V8 isolates with Firecracker MicroVMs for edge functions
PSSA: A non-transformer language model written in Rust
Mathematicians urge AI labs to stop testing advanced problems on proprietary models
TLA+ cannot verify properties that lack a logical formula
and 4 more in brief

Top stories

Google announces Gemini 4 Argon for coding and cyber defense

Google has announced Gemini 4 Argon, a new frontier model designed for software engineering, enterprise knowledge work, and cybersecurity defense. The model is currently rolling out to a limited group of trusted cyber defenders through the Fairwind Program. Google states it is following a phased release approach and engaging in the U.S. government’s voluntary pre-release access process. Before broad availability, the company plans to gather feedback from early testers to refine safety guardrails. Argon will launch with introductory pricing of $2 per million input tokens and $10 per million output tokens.

blog.google HN discussion

Pi adds native MCP support after re-evaluating the protocol

Earendil has integrated Model Context Protocol (MCP) directly into the core of Pi, reversing previous public dismissals of the standard. The team explains that modern MCP has evolved significantly, making it a better fit for Pi's architecture than it was a year ago. This integration also enables a new feature called Codemode, which allows for more structured tool execution within a JavaScript sandbox. By embedding MCP in the core rather than treating it as an extension, Pi aims to improve tool composition and align with emerging patterns for intelligent tool discovery. This move reflects a strategic shift to actively shape how MCP functions in smaller agent harnesses.

earendil.com HN discussion

Netlify replaces V8 isolates with Firecracker MicroVMs for edge functions

Netlify has rebuilt its Edge Functions infrastructure, replacing hosted V8 isolates with Firecracker MicroVMs running inside its own edge network. This architectural shift reduces median warm invocation latency from 25–40ms to 5–6ms, a fivefold improvement. The new system also accelerates log delivery and enhances security by isolating workloads in dedicated microvirtual machines. The change is already serving production traffic with no required migration steps for developers.

netlify.com HN discussion

PSSA: A non-transformer language model written in Rust

PSSA is a small language model that uses a plastic state-space architecture instead of a transformer. It processes tokens sequentially through a recurrent state-space layer, maintaining a bank of episodic memories for lookup. The model rewrites part of its own weights during the forward pass. Written from scratch in Rust without PyTorch or TensorFlow, it learns faster than a transformer at matched parameters and generates text about twelve times quicker on the same CPU.

github.com HN discussion

Mathematicians urge AI labs to stop testing advanced problems on proprietary models

A group of mathematicians published recommendations urging AI labs to cease testing advanced mathematical problems on proprietary models inaccessible to the broader scientific community. The authors state they do not endorse this practice and ask labs to stop. They argue that human understanding of mathematics remains paramount and that current norms require authors to verify and take responsibility for their work. The recommendations, informed by over 600 community responses, call for responsible release of significant results and support for developing human understanding, including funding for workshops and expository writing. Labs should also provide broad, equitable access to publicly available models to prevent a two-tier system in the field.

agmai.org HN discussion

TLA+ cannot verify properties that lack a logical formula

A TLA+ educator argues against the narrative that formal verification will solve agentic software development. While the tool excels at checking safety and liveness invariants for concurrent systems, it has inherent limits. Specifically, TLA+ cannot verify properties that cannot be expressed as logical formulas or those requiring quantification over sets of behaviors, such as proving a game is winnable. The author notes that while workarounds like self-composition exist, they are complex and often produce messy models that do not accurately reflect the actual system.

buttondown.com HN discussion

In brief

Western labs adopt DeepSeek KV cache optimizations. The author argues that Western AI labs are now adopting DeepSeek’s KV cache optimizations rather than merely distilling models. This shift is evidenced by sharp drops in cache-read pricing for recent Claude and GPT releases, reflecting the integration of DeepSeek’s architectural advances.

insufferable.dev HN discussion

Magnitude launches as an open source inference engine that tunes kernels locally. Magnitude is an open source inference engine that compiles and tunes kernels on the user's specific hardware. The project claims this approach allows open models to run up to 2x faster than llama.cpp. It supports Apple Silicon, NVIDIA, AMD, and CPU-only environments.

github.com HN discussion

AlphaPixel releases Slughorn, a C++20 implementation of the Slug text rendering technique. AlphaPixel has released Slughorn, an MIT-licensed C++20 implementation of the Slug GPU text rendering technique. The library targets OpenGL, Vulkan, WebGPU, and DirectX, and includes Python bindings. It follows Eric Lengyel dedicating the Slug patent to the public domain in March 2026.

alphapixeldev.com HN discussion

Amutable introduces Quarry for distributing immutable Linux images. Amutable has introduced Quarry, a TUF-based toolchain for distributing immutable images. The system relies on static hosting for payloads while keeping update logic on the client side. It supports per-machine repositories and cross-repository links to manage signed data ownership.

amutable.com HN discussion