AI — Page 75

The BabyAGI creator's May 21st arXiv paper argues for agents built around append-only logs, replay, forking, and lineage.

Phargo's public scoreboard shows 3,844 upstream PHP tests passing, with the human operator using the corpus as the code reviewer.

The SMS outdoor assistant now renders wildfire and trail overlays by letting the model queue layers while server-side code handles GeoJSON.

Levy Street's browser game is open source, token-linked and now being used to test AI agents inside a shared MMO world.

The X demo claims the game was produced in two prompts, with sprites, sounds, graphics and music generated at runtime from C++ code.

The open-source fork runs the 2003 RTS on Apple Silicon, iPhone and iPad, but users still need their own EA game assets.

A new motion asks a judge to widen discovery beyond AI work that resulted in consumer-facing images or videos.

The public Ovasabi repo pairs Go, Rust/WASM and TypeScript with agent rules, performance gates and a license that limits AI training.

Anthropic has not picked a chip use case, server design or production path, but its Samsung discussions show how Claude demand is pulling the lab toward silicon.

The ex-Google AR/VR PM says IronSight was made from two pairs of smart-glasses footage after a few hours of experiments.

Arthur Mensch's Paris lab says the Lean 4 model solves 587 PutnamBench problems and found five previously unreported bugs.

The six-person inference startup says MI355X tuning got GLM-5.2 to 2626 tok/s/node, with Vercel and OpenRouter now carrying the model.

GitHub is rolling out Kimi K2.7 Code to paid Copilot users, with enterprise access gated behind admin policy controls.

A thin July 3rd post points to a real agent-builder problem: Fable 5 is powerful, expensive, and easy to waste without tighter context design.

The Wharton professor gave Anthropic's model Unity and MCP access, then asked for a WebGL FPS from one high-level prompt.

The Nature study tested Merlin across 752 tasks and released code, weights and a de-identified abdominal CT dataset.

The Paris startup's $1.03B seed round is a March bet on world models, robotics and the limits of language-first AI.

The benchmark claim gives Zuckerberg a talking point for Meta's $125 billion-plus AI infrastructure buildout, but Watermelon is still unreleased.

Meta's founder told employees the company's AI reorganization has not yet delivered the payoff behind its 2026 spending surge.

Vincent van der Meulen's prompt turns a model into a manager over Devin sessions, issue trackers, Slack and review bots.

In a newly posted Oxford lecture, the Sapiens author argues that language models are built for the systems that run money, law and trust.

Craig Campbell turned a metal-detecting itch into a six-figure historical maps startup. His Threads archive reads like a live founder diary: outages, customer calls, pricing panic, AI agents, SEO cliffs, and the strange little wins that compound.

Isaiah Taylor's Ward 250 powered a Blackwell PC, but the larger 30 MW data-center plan still has unproven economics and few public terms.

The reported proposal would put public ownership into OpenAI's cap table as Washington presses AI labs on safety, jobs and upside.