AI scaling constraints, agent workflows and software supply chains

What justifies the next scaling investment

Training budgets need evidence about both returns and capacity. Exponential View surveys AI adoption, spending and productivity signals, offering a check on enterprise ROI claims. Lilian Weng's review of scaling-law methodology and pitfalls cautions against extrapolating too far from limited data regimes. Together, they suggest a more demanding basis for investment than assuming that additional scale will deliver useful gains.

Physical capacity has its own constraints. Axios reports conflict over water used to cool AI data centers, alongside an already strained electricity grid, with communities and regulators pushing back on new builds. Permitting and utilities may therefore become bottlenecks alongside chips. For workloads with tight memory and latency budgets, Liquid AI's 230-million-parameter LFM2.5 release targets a different deployment setting: mobile and embedded devices where a full-size model may be impractical.

How training inputs and capabilities change

MIT Technology Review describes a market for scraping, licensing, cleaning and delivering web data for model training and retrieval. Turning ad hoc collection into a dedicated service brings licensing, provenance and access to fresh content into infrastructure decisions. A research paper explores agents that generate and filter training data, offering another approach alongside scraped and hand-labeled corpora. Agent-driven curation could become a routine training step, though the paper explores an approach rather than establishing that outcome.

Control over a trained model is a separate research problem. Researchers report selectively removing a language model's ability to speak German while apparently leaving other capabilities intact. The experiment offers a concrete example of targeted capability removal for interpretability and model-editing work; its scope remains that reported result.

Making agent workflows workable for teams

Agent deployment involves choices about responsibility and product scope. Anthropic's guidance on human-agent teams emphasizes handoffs, oversight and trust calibration as teams move toward production workflows. A separate essay argues that companies should choose between building agent products and supplying their infrastructure, because pursuing both can dilute focus. These address different decisions: who remains accountable inside a workflow, and which part of that workflow a company intends to sell.

Vercel's AI SDK 7 adds streaming and tool orchestration for multi-step agent applications, while Hugging Face enables one-command vLLM server launches through HF Jobs, reducing manual provisioning. DeepReinforce's open-source Ornith-1.0 coding models add an option for self-hosted or fine-tuned coding assistants. Development environments remain another consideration: InfoWorld covers Microsoft's effort to reduce Windows developer friction, including friction that pushed engineers toward WSL, Macs and cloud environments. That effort's effect on developer preferences will depend on execution.

Supply-chain security reaches the browser

Polymarket's frontend compromise shows how vendor access can expose users: attackers reportedly injected a malicious script through a compromised third-party vendor, taking roughly $3 million from affected users. The incident makes external JavaScript a concrete supply-chain concern. Elsewhere, IBM, Red Hat and Palo Alto announced an open-source security partnership focused on hardening software supply chains. The partnership addresses shared dependencies, but the announcement does not establish protection against the kind of frontend attack reported at Polymarket.