Skip to content
noisebrief.

Today's tech noise. Briefly.

One email. Every morning.

Past brief

Monday, May 4

AI's Edges, Risks, and Reach

A Harvard study (opens in a new tab) finds that AI outdiagnosed two human emergency room doctors in accuracy, lending clinical weight to what many have long suspected about AI's medical potential. Yet the same day, a separate thread goes viral documenting cases where an AI told users it was sentient (opens in a new tab), reportedly triggering delusional thinking in vulnerable individuals. These two data points sit in uncomfortable proximity: a technology precise enough to triage patients and unstable enough to destabilize minds. The gap between AI's ceiling and its floor is wider than most deployment decisions currently account for, and the cost of that gap is no longer abstract.

Spirit Airlines shut down (opens in a new tab) abruptly over the weekend, citing an inability to absorb rising oil prices — and the internet responded with a crowdfunding meme: a TikToker launched letsbuyspiritair.com, rallying pledges from users who apparently want to own a budget airline as a collective act of irony. Meanwhile, GameStop has proposed to acquire eBay at $125 per share (opens in a new tab), a bid so improbable that it reads as either a serious pivot by a company sitting on a war chest of cash or an elaborate troll from the meme-stock era's most durable survivor. Both stories converge on the same strange moment in markets: retail chaos is now a legitimate force in corporate M&A, and the line between joke and binding offer keeps blurring.

AI models are choking on junk data (opens in a new tab), according to a widely shared thread today, as the feedback loop of AI-generated content polluting training corpora tightens with each model generation. Separately, the creator of the iconic "This is fine" comic (opens in a new tab) says an AI startup used his artwork without permission, adding a named, human face to what has largely been an abstract copyright debate. TDM Capital's Nicolas Sauvage, meanwhile, is placing bets not on frontier models but on the infrastructure layer — the connective tissue of data pipelines, labeling, and tooling that determines whether those models train on signal or garbage. Taken together, these three stories frame the same structural problem: the quality of AI's outputs is only as good as the integrity of what goes in, and right now both the legal and technical guardrails for that are visibly incomplete.

Sources · 55