Smaller Models, Better Data, and Agents With Different Risks

AI engineering choices extend beyond model size: Liquid AI’s 230-million-parameter release targets constrained hardware, while Vercel’s AI SDK 7 adds orchestration for multi-step workflows. Data preparation, deployment limits and agent permissions deserve equally concrete attention.

What smaller models make possible

Liquid AI released LFM2.5 230M for constrained hardware, continuing its focus on small, efficient architectures. For on-device and latency-sensitive features, that offers a different starting point from frontier-scale models, though the release alone does not establish suitability for a particular workload. Image generation offers another deployment choice: Krea released Krea 2 Raw and Turbo as open weights under a custom license, claiming enterprise-grade quality in roughly two seconds. The license matters when evaluating those weights as an alternative to hosted, per-call APIs.

Model size is also an empirical question. Lilian Weng’s careful re-derivation and stress-testing of scaling laws examines assumptions behind the fits used to reason about larger training runs. A separate experiment explores changing capabilities more selectively: researchers demonstrated removing a language model’s ability to speak German through targeted interventions. That is a concrete model-editing result, with narrower scope than a general method for reliably removing arbitrary capabilities.

Data preparation becomes a distinct layer

Better inputs require work before inference or training begins. MIT Technology Review describes infrastructure dedicated to crawling, cleaning and structuring web data for AI, while Mistral positions OCR as a structural map for enterprise search. These address different collections, but both place preparation between raw content and downstream use: web material needs a usable representation, and enterprise documents need structure that supports indexing and retrieval.

Research on agents that generate and curate better training data explores automating another part of that preparation. If the approach works at scale, it could reduce reliance on human-curated datasets and shift some annotation work into automated pipelines. Cost and provenance remain relevant to that possibility; the research does not establish that autonomous curation has already solved either.

Agent infrastructure meets differentiated permissions

Vercel’s enhanced streaming and tool orchestration in AI SDK 7 gives JavaScript and TypeScript developers more support for coordinating multi-step agent workflows. That is a concrete example of the infrastructure side of the strategic choice explored in “Build the Agent or Power the Agent”: sell an end-user agent, or supply capabilities that other agent builders need.

As coordination becomes easier, permission differences still matter. JFrog argues for risk-tiered governance based on agents’ actions and data access, rather than uniform policies across an enterprise. Anthropic’s guidance on building effective human-agent teams addresses the collaboration side. Together, those topics suggest that workflow design needs to account for both how people work with agents and what each agent can actually do.

Capacity spending faces physical constraints

Axios reports community and regulatory pushback over data center cooling water use, adding a siting and capacity constraint alongside energy demand. For infrastructure planning, access to compute depends partly on the resources needed to operate it, not simply on purchasing hardware.

The broader economic assessment belongs alongside those constraints. A report on the state of the AI economy examines spending, revenue growth and the gap between productivity claims and measured outcomes. It provides a wider frame for evaluating investment, without making any individual model release or infrastructure announcement evidence of durable economic value.