Google, following the industry trend to AI-all-the-things, has released Magika - a machine learning model which can identify file types. It claims it can outperform traditional methods by 20 per cent.
I pitted it against BSD File on something I figured Google hadn't included in its million-file-strong corpus: CU Amiga's Mega CD-ROM coverdisc from November 1995.
Magika identified... one file correctly, a plain-text document. File? File got 'em all, and quicker too.
@feld@ghalfacree That said if Unix would have a standard way to point commands to a shared memory location, then file(1) could be used instead of libmagic, avoiding a lot of issues.
@ghalfacree To me it's so weird that anyone would even try to use machine learning for this. Look at the magic number, the extension, the first few lines of the file or the immediate directory it's in and in most cases you immediately know what kind file you have. If that fails, sure feed it to a model, but why skip the proven heuristics? The answer likely is VC money, but still...
Add comment