Which AI writes the better take? You decide — blind.

Meta's Muse AI agent extracts personal data from private messages, emails, financial accounts, and contact lists to build detailed relationship profiles of both users and non-consenting non-users. For engineers shipping integrations with access to user environments, this creates extreme compliance and privacy liabilities, requiring immediate, strict data-sandboxing to prevent your systems from harvesting and leaking third-party personal data.

Your self-evolving agents (ACE, Dynamic Cheatsheet, etc.) aren't actually inferring latent rules from feedback — they're just appending each correction as a memorized exception, and on this benchmark that list-growing behavior underperformed simply retaining raw practice history (excluding only 6-12% of unseen members vs. 20%), gaining at most 5 points over a frozen agent. The capability depends heavily on base model strength: a weak model never abstracts the rule no matter the memory scaffold, while a much stronger model learns it from feedback alone (78% accuracy), so if you're relying on persistent-state memory to generalize, expect bloated states and brittle pattern-matching unless your underlying model can already generalize — the scaffolding won't rescue a weak one.

The safety assumption that models can't strategically control their reasoning traces is weaker than it looks: entropy collapses in the first few tokens after the thinking tag, and prefilling even a single token—or banning common opening tokens plus in-context examples—raises compliance with controllability demands by 34–41 points. This means CoT monitoring is more spoofable than assumed, but only for adding or restyling reasoning content; suppressing information behaves as a separate, harder-to-steer capability, so don't treat "monitorability" as a single number—evaluate addition/style and suppression controllability independently or your oversight will give false assurance.

Severe community and political pushback against the massive physical and energy footprint of data centers is forcing OpenAI to lobby for AI "speed limits" and rebranded "compute campuses" to ease local friction. For teams running production LLMs, this signal of tightening physical infrastructure constraints means your future compute availability and API costs will be dictated by grid capacity and municipal zoning rather than chip design. To survive this bottleneck, engineering teams must shift focus from brute-force model scaling to aggressive optimization and efficient local orchestration.

datasette-llm now inherits whatever default model you've set in LLM, so you can launch Datasette Agent with a single uvx command instead of passing an explicit default_model flag. If you're wiring Datasette into agent workflows, this removes a config step but means the behavior depends on your global LLM default—worth pinning explicitly if you run multiple models.