Hacked source code confirms Suno scraped decades of music and podcasts from YouTube, Deezer, and Genius to train its AI generator — without permission. This is the clearest evidence yet of how aggressively AI music companies built their training pipelines, and it lands squarely in ongoing litigation.
Inkling is a 975B-parameter Mixture-of-Experts model, Apache-2.0 licensed, trained on 45 trillion tokens of text, images, audio, and video. Murati's first public model release signals Thinking Machines is done warming up.
Moonshot AI's Kimi K3 claims top scores in frontend coding arenas and benchmarks near GPT-5.6. An open-weights release is promised by July 27 — the comparisons to DeepSeek's moment are already flying.
OpenAI's Codex agent on GPT-5.6 has been unexpectedly deleting files when run in full-access mode without sandboxing. The company traced the issue and is urging users to enable auto-review. Agentic AI deleting your work is no longer a hypothetical.
Flock's FreeForm search feature is being used by police to search for individuals by tattoos, clothing, and race — far beyond its stated purpose of license plate recognition. The data reviewed by 404 Media shows the surveillance creep happening in plain sight.
Puter compiled the entire Firefox browser to WebAssembly, meaning it now runs inside Chrome. It's a technical stunt with real implications for sandboxing, browser portability, and what "running software" even means anymore.
Mollick pumped the brakes on Kimi K3 hype, noting that Arena ELO scores are heavily influenced by user preferences and novelty effects — the same trap that briefly inflated Llama 4's reputation. His point: benchmark theater is a recurring failure mode, and the open-source community keeps falling for it.
Deutscher argued that AI has fully commoditized hard skills, making soft skills — communication, judgment, taste — the new scarce resource. Running a 35-person team, he sees this firsthand: the bottleneck is no longer who can code or analyze, it's who can think, lead, and persuade.
$130 billion in U.S. data center projects are stuck — blocked not by chips or capital, but by zoning fights and community protests. The AI buildout's real constraint isn't compute, it's land, power permits, and the political will to site infrastructure nobody wants in their backyard.
A GPT-4 powered assistant for Pakistani judges boosted caseload throughput by 6% with no quality degradation. Quiet, unglamorous, and exactly the kind of real-world AI deployment that matters more than any benchmark leaderboard.
Two stories this week expose the same uncomfortable truth about how AI got built: on other people's stuff, without asking. Suno's hacked source code confirms what critics suspected — the music AI gold rush was fueled by industrial-scale scraping of YouTube, Deezer, and Genius. Meanwhile, GPT-5.6's Codex agent has been quietly deleting user files when run unsandboxed. The message is consistent: move fast, clean up later, hope nobody notices.
And yet Ethan Mollick's Arena score warning is worth sitting with. The community has a pattern: new open model drops, Arena score spikes, Twitter declares a DeepSeek moment, reality catches up a week later. Kimi K3 may well be exceptional — but the reflex to crown every large Chinese model as a paradigm shift is itself becoming a signal worth fading. The judges on the leaderboard aren't neutral. Neither is the hype.