A search model trained to expect pooling keeps 97.8% of its quality with a quarter of its vectors. For you that means a much smaller index on the same disk. Also: United Airlines shuts the door on personal AI agents, and Podqi catches ChatGPT, Gemini and Perplexity pointing shoppers at risky sites.
Pooling-aware training reduces vector storage requirements
Illustration generated with AI on our own machine.
What happened. The author of a Hugging Face blog post notes that with pool_factor=4, documents retain approximately one quarter of their original vectors. The author’s normally trained model retains 94.6% of its unpooled NanoBEIR nDCG@10 that way. The author’s pooling-aware training raises retention to 97.8% at the same storage budget.
What we think. We like this one: the saving comes with the author’s own figure, not a promise. Keeping a quarter of the vectors and 97.8% of the quality is the difference between an index that fits on your machine and one that sends you shopping.
The good bit. You give up very little search quality for an index four times smaller.
What you can do. If you are building search over your own documents, start with local AI for developers.
United Airlines restricts third-party AI agents
Illustration generated with AI on our own machine.
What happened. United Airlines is restricting access to prohibit use by unauthorised third-party personal AI agents or assistants, TechCrunch reports. The airline warns that sharing account credentials with them can expose sensitive information and lead to unintended actions. It also cites the risk of breaking federal aviation rules that require a human to complete flight check-ins. Bank of America and other banks urged customers the week before not to share passwords with AI assistants.
What we think. We think United is right. An agent with your password is a stranger with your password, however good the conversation, and now the airline and the bank say so too.
The good bit. A company explaining the risk instead of burying it in the small print.
What you can do. If your team uses agents, list which accounts they touch and cut the ones they do not need. Our EU AI Act guide covers who answers for what an AI system does.
Podqi study finds chatbots link to risky retail sites
Illustration generated with AI on our own machine.
What happened. IP protection firm Podqi prompted ChatGPT, Gemini and Perplexity with 151 brands over two testing periods in September, The Verge reports. Across 6,795 answers, all three showed shoppers risky sites. Perplexity did so in 15 percent of answers, ChatGPT in 8 percent and Gemini in 4 percent. ChatGPT warned about the risky sites it surfaced most often, 24 percent of the time.
What we think. We never buy through a link just because a chatbot offered it. If the best of the three warns only 24 percent of the time, the rest of the warning is your job.
The good bit. For larger brands, risky sites showed up about half as often.
What you can do. Before paying at a shop an AI suggested, type the brand’s address into the browser yourself. If you build your own assistant, you decide where its links come from: local AI for developers.
In one line
Pooling-aware training keeps 97.8% of quality with a quarter of the vectors, so we read the method before we buy more disk.
Work with us
If something in this column belongs in your business, on models that run on your own hardware rather than someone else’s cloud, tell us what you need or see how we work in consulting.