AI News, page 2 of 6

AI and software engineering, as it happens.

Block open-sources Berd, a Tauri desktop app for AI agents built on Goose

Block has released Berd, an open-source desktop app for working with AI agents built on Tauri 2, React 19, and the Goose backend, and the repository has picked up 868 stars in its first month.

Open source

Cole Medin's AI Software Factory turns GitHub issues into merged pull requests with no human review

Cole Medin has released AI Software Factory, a repository that takes GitHub issues in and ships merged pull requests out with nobody reading the diff, using a doctor script that refuses to raise the autonomy dial until specific evidence exists.

Dev tools

DonSeTch gives AI agents keyless web fetch, search, and crawl through one Rust binary

DonSeTch is a Rust binary that gives any MCP-capable coding agent web fetch, search, and crawl through three tools with no API keys, and the repository has picked up 582 stars in its first month.

Open source

Armature measured 16,893 coding agent sessions to see which tools they pick

Armature ran 16,893 sandboxed sessions across Claude Code, Codex, and Cursor on 75 repositories to see which third-party tools each agent installs, and the three agents agreed on the same pick in only 42 percent of categories.

Productivity

A game developer used Claude Fable to reverse-engineer his own 1993 Amiga assembly, byte for byte

Rabah Shihab ported his 1993 Amiga game Babylonian Twins to Godot by having Claude Fable read the original 68000 assembly and rebuild 72,758 lines across 26 files into a byte-identical binary.

Productivity

Wake pulls sessions from 15 coding agents into one searchable desktop app

Wake is a new MIT-licensed desktop app that indexes coding agent transcripts from 15 tools including Claude Code, Codex CLI, Cursor, and Gemini CLI into one searchable interface, and the repository has picked up 731 stars.

Dev tools

ARC Prize runs GPT-6 Astra on ARC-AGI-3, and swapping the harness moves the score from 17.5 percent to 99.9 percent

ARC Prize published cost and score numbers for OpenAI's new GPT-6 Astra on ARC-AGI-3 across two harnesses and six reasoning levels, and the same model spans from 17.5 percent to 99.9 percent depending on the configuration around it.

Models & agents

npm makes multiple trusted publishing configurations per package generally available

A single npm package can now carry more than one trusted publishing configuration, so a maintainer no longer has to keep a long-lived token for the release paths that OIDC could not previously cover.

Dev tools

Cloudflare launches Vulnerability Discovery and Remediation, using WAF traffic to rank findings and propose patches

Cloudflare opened early access to Vulnerability Discovery and Remediation, an invitation-only service inside Managed Defense that pairs source code review by OpenAI Daybreak models with live WAF and traffic data to rank findings and draft both a code patch and an edge rule.

Infrastructure

terminal-code puts a real VS Code window inside the terminal by piping code-server's frames through the kitty graphics protocol

An open-source project called terminal-code, from zenbu-labs, embeds a real VS Code window inside a terminal by combining code-server with a browser that renders through the kitty graphics protocol, and has picked up 1,868 stars since it was created on 8 August.

Open source

GitHub Copilot will retire Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code, and Claude Opus 4.7 across every Copilot experience on 2 October

GitHub Copilot will remove Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code, and Claude Opus 4.7 from Copilot Chat, inline edits, ask mode, agent mode, and code completions on 2 October, with named replacement models for each.

Dev tools

Shopify says it cut its Sidekick GraphQL agent's system prompt from 6,000 tokens to 1,500 using learned gist tokens, and freed 14 percent of the GPUs serving that agent

Shopify says its production Sidekick GraphQL agent now runs on a 1,500-token compressed prompt learned by knowledge distillation, with a 4:1 reduction, no drop in prediction quality, and 14 percent fewer GPUs serving the same traffic.

Infrastructure

Google's John Mueller says recovery from an algorithm-related ranking drop can take many months even after the cause is fixed

Google's John Mueller told a site owner on Reddit that fixing what triggered a large algorithmic ranking loss does not bring rankings back quickly, and that recovery from a broader spam update can take many months.

Go-to-market

A B2B citation study says 88 percent of the sources AI engines quote are pages the brand controls

A study of 7,387 citations that ChatGPT, Perplexity, Claude, and Gemini gave to 170 B2B evaluation prompts found that product pages accounted for 24.1 percent of the sources, and Reddit, YouTube and forums together for 4.2 percent.

Go-to-market

GitHub CLI Linux signing key expires on 5 September, and only one rotation command fixes it

The PGP key that signs GitHub CLI Linux packages expires on 5 September 2026, and any apt or rpm install that has not picked up the replacement key will fail signature checks the moment the next release ships.

Dev tools

CLion 2026.2.2 ships an agent skill that debugs ARM Cortex-M hard faults for you

JetBrains has added an agent skill to CLion 2026.2.2 that reads ARM Cortex-M fault registers, walks the stacked exception frame, and points at the code that caused a hard fault.

Dev tools

Hugging Face shows RL over taste is workable, training a coding model to paint watercolours with TRL and OpenEnv

Hugging Face published a training recipe that uses TRL and OpenEnv to teach a coding model to paint watercolours in JavaScript, judged by a pairwise vision model and a preference scorer instead of a verifiable metric.

Models & agents

Kotlin Toolchain 0.12 adds multiplatform library publishing, Wasm apps, and CLI hot reload

JetBrains has shipped Kotlin Toolchain 0.12 with multiplatform library publishing, a preview of Wasm applications, and Compose Hot Reload from the command line.

Dev tools

GitHub reopens Copilot Business and Enterprise signups and requires payment before a seat unlocks

GitHub is reopening Copilot Business and Enterprise signups for credit-card and PayPal customers, and from 1 October every assigned seat is billed upfront before the user can use Copilot.

Dev tools

SEOs observe Google's AI Mode citing fewer sources after the switch to Gemini 3.8 Flash

SEO researchers Gagan Ghotra and Glenn Gabe say Google's AI Mode is returning far fewer links or citations since it started using Gemini 3.8 Flash, based on side-by-side comparisons against the default model.

Go-to-market

A 350M model beats its own JSON baseline by 14 points after 100 fine-tuning steps

Hugging Face published a walkthrough showing Liquid AI's 350M model gaining 13.9 points on JSON output compliance after 100 GRPO steps on 500 samples, cheap enough to run on a free Colab GPU.

Models & agents

Nvidia releases PAIR, a free tool that turns your RTX PCs and Macs into a local AI cluster

Nvidia announced PAIR at IFA 2026, a free open-source tool that discovers RTX PCs, DGX Spark units and Apple Silicon Macs on a local network and routes AI inference to whichever machine has capacity.

Infrastructure

GitHub Copilot content exclusions are generally available in the Copilot app and CLI

GitHub has moved Copilot content exclusions to general availability in the Copilot app and CLI, letting Business and Enterprise admins keep named files out of the model's context.

Dev tools

Nvidia agrees to buy Hugging Face for $12.9 billion, and says the platform stays open and multi-cloud

Nvidia has agreed to acquire Hugging Face for $12.93 billion, and says the platform will keep its brand, stay multi-cloud, and continue supporting open weights across the ecosystem.

Infrastructure

SEO Landing packages a technical SEO discipline as an agent skill for Claude Code, Cursor, Codex and six other coding agents

SEO Landing is an agent skill that gives Claude Code, Cursor, Codex and six other coding agents a technical SEO checklist targeting a 100/100 PageSpeed score, a 15 KB JavaScript budget, and Core Web Vitals inside the green thresholds, at 136 stars in its first week on GitHub.

Dev tools

NorthCinder ships an open-source MCP server for AI shopping agents that compare across Shopify, WooCommerce, eBay, Etsy and Amazon and require buyer approval before any purchase

NorthCinder is a local-first MCP server that lets an AI agent compare products across five marketplaces and requires human approval before a purchase goes through, at 1,219 stars in its first 17 days on GitHub.

Open source

Hugging Face open-sources Funes, an agent memory layer that beat a written handoff by 8x on its own benchmark

Hugging Face has published Funes, an Apache 2.0 memory layer for coding agents that indexes past sessions locally and lets a new session recall passages from them, and reports it was 8x cheaper than a written handoff on one of its own benchmark tasks.

Open source

Google AI Mode shows almost none of the same products as the shopping carousel, and picks pricier ones

A new study from Productrise finds that only 1.28 percent of the products shown in Google's shopping carousel also appear in AI Mode for the same query, and when they do overlap, AI Mode shows prices that are on average 21.6 percent higher.

Go-to-market

Reverify pairs an AI model with a pure-Python RE toolkit and marks every claim VERIFIED or REFUTED against the actual bytes

A new open-source RE toolkit, reverify, uses deterministic parsers and emulators as the judge for AI-proposed hypotheses about a binary, and has 662 stars in the three days since it was created.

Open source

Rakazo, an open-source alternative to Grok Bot, reaches 1,822 stars in 21 days with bring-your-own-model teammates

Rakazo is a self-hosted platform for persistent AI teammates with bring-your-own-model credentials and isolated computer sandboxes, and has picked up 1,822 GitHub stars in the three weeks since it was created.

Open source

Archive

179 items published so far, by month.

All months