Executive Summary
The tools for running private AI keep getting easier to use. The supply of models and the political cover behind them are getting less reliable. Willison's datasette-llm release shows that any OpenAI-compatible endpoint, local or hosted, can now be swapped by changing a single URL. Lambert argues that Jensen Huang's exit takes away the main funder and political protector of US open-weight models, and that the most capable open models are still Chinese. Jones's take on the FAA ground stop is speculative, and the AI link rests only on timing. It still shows the kind of event that sets regulation in motion: an AI tool goes live and an outage follows. Taken together, the advice is to make the model the most replaceable part of the stack, and to invest in governance and rollout discipline because the regulatory climate is about to tighten.
What Changed
- The interface is settled. LM Studio, Ollama and llama.cpp (now v1.0.0) all speak the OpenAI chat-completions protocol. Willison runs a SQL-querying agent on a local Qwen 27B with one
uvxcommand, with no API key, no SDK and no cloud. - The US open-weight sponsor is gone. Lambert expects Nvidia to spend roughly 24 months stabilizing. That ends the funding he calls "flooding the zone" and the executive who stepped in when open models came under political pressure.
- The capability gap has widened. Thinking Machines' Inkling 2 Small is the best recent Western open model, and it is still far behind Kimi K3.1 Pro. A Zhipu GLM update is expected.
- There is a public example of AI change risk. A first-of-its-kind AI routing tool reportedly went live about 12 hours before a nationwide ground stop. No causal link has been published.
Cross-Expert Synthesis
Willison and Lambert describe two layers of the same stack moving in opposite directions. The protocol layer is commoditizing: the endpoint is a pseudo-standard and switching costs are close to zero. The model layer is not. Lambert argues compute scarcity keeps open-weight serving margins high, and that vertically integrated closed labs win while it lasts. The practical point is that an enterprise can switch models cheaply but cannot easily get a good model to switch to. Model portability is the hedge against Lambert's supply risk. It does not solve the problem.
Lambert and Jones meet at the regulatory trigger. Lambert expects a major security incident tied to a popular Chinese open-weight model, calls it a matter of "when, not if", and notes that media timing alone brought Huang down. Jones shows how quickly the public pins an outage on AI before any evidence exists. Taken together, the enterprise exposure is less about model quality than about being the visible deployer when the story breaks, especially if the deployer is running a Chinese-origin model, behind an opaque integration, in a critical workflow.
There is also a tension. Willison's pattern makes it trivial to point an agent at whichever open model scores best, which today means Chinese weights. Lambert's analysis says that is exactly the deployment state-level rules are likely to target first.
Enterprise Implications
The Private AI pitch gets stronger on architecture and weaker on supply. Clients can build sovereign, on-premises agent stacks cheaply, but the model roadmap behind them is less predictable than it was a month ago. Four implications follow:
- Model provenance becomes a governance attribute, not just a benchmark line. Clients need to know the origin of every model behind every endpoint, because regulators will ask.
- An intermediate model-routing layer, like
datasette-llmsitting between plugins and models, is the right control point: one config, one audit surface, one place to swap a model out when a regulation or an incident requires it. - Local inference removes the risk of data leaving the building, but not prompt injection. An agent with database credentials is still an exfiltration path.
- Calling a tool "advisory" does not isolate it. Integration with legacy systems is where failures happen, as Jones argues, and the reputational damage lands on whoever shipped last.
What To Do About It
- Make OpenAI-compatible
api_basesupport a procurement requirement. Any AI tool that cannot point at a self-hosted endpoint is a lock-in flaw. - Run a same-day local pilot on client data with a mid-size open model and read-only credentials. It turns the Private AI conversation from slides into evidence.
- Build a dual-sourced model register. For each workload, list a Chinese-origin option, a Western open option and a closed fallback, and test that switching between them is just a config change.
- Require an integration map, a kill switch tested under load, shadow-mode exposure and peak-period change freezes before any AI touches an operationally critical path.
- Customer question: "If a regulator or your board banned one model family tomorrow, how many config files would you change, and how long would it take?" Most clients can't answer, and that gap is the engagement.
Risks and Blind Spots
The FAA link is unverified, and acting on the AI-blame narrative before the post-incident report would be a mistake. Use it as an example of a failure pattern, not as evidence. Willison's release is alpha and comes with no data on quality, latency or security for the local model. Lambert's supply pessimism depends on two assumptions: that hyperscalers stay out as replacement champions, and that compute stays scarce. If either breaks, the open-weight outlook improves quickly. The biggest blind spot is that the market for enterprise post-training and "ownership of intelligence" has not been measured, because the most valuable use runs out of view on private infrastructure. That is the bet Private AI rests on, and nobody here has priced it.