Four-bit quantization should have made a 14 MB model smaller. It produced a larger file, and the hybrid that fixed the size introduced a boundary error that costs recall across a whole file. What we shipped instead, and what stays unmeasured.
Read moreExtraction attacks on GPT-2, ChatGPT, and GPT-J recovered names, phone numbers, and verbatim text from model weights. A guided read of what the three studies measured, and the limits their authors put on the results.
Europe's regulators and its top court now answer differently depending on who holds the alias mapping. A timeline of the 2025 shift, and the conservative working rule for teams that can reverse their own codes.
A worked example: a nine-row ticket export goes through load, review, and export, with the judgment calls spelled out at each stage and a checklist to run before the aliased copy reaches an AI service.
Pattern-based detection has recurring blind spots: misspelled names, addresses in free text, and ordinary columns that identify someone in combination. An inventory of what slips through, and the five places to check before export.
Where parsing, detection, and aliasing run when Data Alias opens a file, and two checks — an offline run and a Network-tab read — that show the network behavior of the build you are using.
A [REDACTED] column and a PERSON_001 column hide the same names with different consequences for analysis. How redaction, masking, removal, and aliasing differ, with a decision table for choosing between them.
The same customer is PERSON_014 in Monday's export and PERSON_003 in Friday's, and every alias-keyed join between the two files quietly breaks. What recurring exports need from an aliasing workflow — and how to tell whether a free tool covers it.
Four ways to clean personal data out of a spreadsheet, held to the same nine criteria — from cost and setup time to free-text detection and restorability — with a side-by-side table and the situations where each one wins.
Whether a customer spreadsheet can go into an AI chat depends on the account behind the tab, the settings and contracts behind the account, and what is still in the rows. Seven questions to settle before any file is attached.
