Maligned #4 - Regulation Moves Faster Than Expected

A few things moved this week that deserve your attention.

EU AI Act enforcement is actually happening

The first enforcement deadlines under the EU AI Act are hitting, and the scramble is on. Companies that assumed this would get delayed or watered down are now realizing they need to classify their AI systems, document their risk assessments, and demonstrate compliance. The consulting firms are having a field day. My take: the regulation is imperfect and parts of it are poorly conceived, but the direction is right. If you’re deploying AI in high-stakes settings, you should be able to explain what it does and why. That shouldn’t be controversial.

Google’s Gemini lineup refresh

Google updated its Gemini model family with improved performance across the board. The interesting bit isn’t the benchmark improvements, it’s the pricing. They’re positioning aggressively against OpenAI and Anthropic on cost-per-token, especially for high-volume enterprise contracts. The model quality differences between the top three providers are narrowing to the point where pricing and integration ecosystem are becoming the primary differentiators. This is what commoditization looks like.

Synthetic data quality is becoming a real problem

Multiple teams have published findings showing that training on synthetic data, if not carefully curated, leads to model collapse. The models start losing diversity in their outputs and converge on a narrow band of behaviors. This matters because synthetic data was supposed to be the solution to data scarcity. It still can be, but only with careful filtering and quality control. Cheap synthetic data pipelines that just generate and train without human review are producing worse models, not better ones.

Apple’s on-device AI push continues

Apple’s developer preview of their expanded on-device model capabilities is getting positive early feedback. The performance-to-size ratio is impressive, and the privacy angle gives them a genuine differentiator. Running useful AI locally without sending data to the cloud is something a lot of enterprise customers want, especially in regulated industries. Apple is positioned well here, though their usual pace of iteration means it’ll take a while to see the full picture.

See you next week.

Maligned - AI news by Mal