Quick Hits — August 13, 2026
An evening catch-up: Suno grows into a full DAW, OpenAI publishes hard data on how companies actually use enterprise AI, a 941MB model takes on shell commands, and Apple's first US-built AI servers roll off the line in Houston.
Suno Studio 2.0 turns the AI music maker into a chat-driven DAW — Premier subscribers can now direct the studio like a bandmate, creating instruments, vocals, and even custom plugins from plain-text requests. The upgrade adds MIDI import, recording and editing, stem separation, automation curves, and unrestricted 32-bit/48 kHz multitrack export; plugin creation doesn't burn credits yet, though Suno warns a credit system may come later. It's a bet that paying professionals, not spam-flooding amateurs, are the future — CEO Mikey Shulman just imposed download limits to curb mass distribution of AI songs on streaming platforms, and the company is still fighting a German court ruling that its training violates copyright law. The take: AI music is graduating from prompt-to-song parlor trick to a serious production tool.
OpenAI, Columbia, and Wharton published the largest telemetry-based look yet at how companies actually use enterprise AI — a working paper built on ChatGPT Enterprise records covering more than 1,700 organizations and 17 million messages. Four findings stand out: enterprise output tokens grew roughly sevenfold between June 2025 and March 2026, with about half of that growth coming from firms that had already adopted; early adopters skew larger, more valuable, and more R&D- and SG&A-intensive; usage spans every job function, but early-career workers send eight to nine more weekly messages than the average active user; and the work is broad — documentation, technical writing, and message drafting dominate volume rather than any single workflow. The take: adoption is spreading fast but unevenly, and whether scattered individual use becomes real organizational capability is the open question.
A developer trained a 1.5B-parameter model to write shell commands so he'd stop googling tar flags — and it runs on a plain laptop CPU in about half a second. The fine-tuned Qwen2.5-Coder-1.5B, trained on 125,000 natural-language-to-command pairs and quantized down to 941MB, scores 0.620 on the InterCode-ALFA benchmark — ahead of the untuned 7B version (0.613) and within striking distance of GPT-4o (0.73). Weights and code are out under Apache-2.0, complete with a safety checker, because as the author notes, it will absolutely write a command that wipes your root directory if asked. The take: the most mundane coding chores are where tiny local models are quietly becoming good enough.
Apple's first US-built AI servers are rolling off the line at its new Houston Advanced Manufacturing Center, opened this week by Tim Cook and Commerce Secretary Howard Lutnick. Cook said the facility has "shipped the first advanced AI servers off the line," with Mac Mini production to follow later this year, and Apple has spent hundreds of millions on the site — part of a push that saw it source more than 20 billion chips from US suppliers last year. The take: Apple is weaving its AI infrastructure into its American-manufacturing narrative, a useful shield with tariff negotiations still ongoing.
Sources: Suno · The Verge · OpenAI · Hacker News discussion · Reddit discussion · nl2sh weights (Hugging Face) · CNBC · MacRumors
