Image: Michael Voss · CC0 1.0 Public Domain Dedication · Openverse
Two of today’s stories are about AI already in operational hands: the FAA starts advising controllers with a predictive tool, and a federal records site dropped a model it had been running in public view.
The rest are about what happens when fluent output is trusted too far, and what gets cheaper when a model stops generating language at all.
FAA to Trial AI Flow Advisor at Washington Airports
The FAA will start using an AI system to advise air traffic controllers in the congested airspace above Washington, DC, with a limited launch possible as early as Monday at the region’s three major airports. The tool, called SMART, predicts traffic flows and flags potential conflicts from airline schedules, weather, airport capacity and airspace conditions. It covers a first slice of the 29 million square miles the agency oversees, with a nationwide rollout planned later. Controllers stay in charge; the tool advises. For anyone selling predictive operations software, the signal is that a federal safety agency is buying capacity-planning models and has chosen to prove the authority split on a narrow slice before scaling it.
Source: Ars Technica
Federal Register Removes Chinese AI Search Tool
The National Archives took an AI search feature built on Alibaba’s Qwen model off the Federal Register website after users noticed the agency was running a model family the FBI had publicly accused of copying US systems. The agency has not explained the removal or said how long the option had been live. It is a procurement story as much as a technical one: model provenance is now something a public buyer can be embarrassed by. Anyone selling AI into government or regulated industries should expect questions about where weights, training data and hosting come from, and vendors who can answer them in writing hold an advantage that has nothing to do with benchmarks.
Source: Ars Technica
Fabricated Intelligence Report Nearly Triggered an Armed Operation
US aircraft were already in the air when officials learned that the intelligence behind an operation against a Chinese vessel had been invented by a chatbot. The order was called off at the last minute. Reporting describes an analyst who used the tool to combine open-source material with classified signals intelligence, then asked it again to write the result up as an official summary, which then circulated up the chain. It is the clearest case yet of the failure mode enterprises quietly underwrite: fluent output treated as verified analysis. The lesson for anyone shipping AI into a decision path is procedural, not algorithmic: label generated conclusions, and require a human gate before anything irreversible follows.
Source: TechCrunch AI
TypeSafe’s Jev Returns Probabilities Instead of Prose
A former OpenAI researcher who helped invent RLHF has shipped a transformer that classifies rather than chats. TypeSafe AI’s Jev answers only the categories a developer defines in advance and hands back a probability, so there is no generated text to hallucinate; the company bills input by the billion and treats output tokens as free. Early adopters say the appeal is price and speed: Vercel replaced a large model used for command-safety checks and reported faster results, while another team found a frontier model slightly more accurate but far costlier. Many production AI features are really classification, and this makes paying frontier prices for them a choice rather than a default.
Source: TechCrunch AI
India Extends Anti-Spam Rules to Automated Calling
India’s telecom regulator now requires caller-ID and call-management apps to forward users’ spam reports to a blockchain platform run by carriers, and it has pulled automatically dialed and prerecorded-voice calls into its commercial-messaging framework. Companies using automated calling must declare the numbers involved to their carrier in advance; undeclared calls are treated as spam and can attract a per-minute termination charge. Truecaller, for which India is the largest market, calls the data-sharing requirement a one-way, anti-competitive exchange. For teams running outbound voice agents or click-to-call in India, the compliance work has shifted from consent copy to registration and disclosure.
Source: TechCrunch AI
Sources
About the Author
