I’m not intrinsically opposed to generative models, even if I’m often voicing my criticism against how they’re done, who uses them, and how they’re used. Otherwise I wouldn’t be subscribed to this comm. Plus I like the underlying tech on a conceptual level.
But there are limits on what those tools can do reliably, and often simpler tools would do the same job without, you know… cooking the planet. And that “reliably” is damn important, in some tasks 1% failure rate is already too much.
Someone might say “you can increase reliability with manual reviews”. That’s true but there’s a catch: sometimes reviewing it takes longer than doing it yourself. Doubly so if it’s a sensitive application, like the examples offered by the text:
Export your budget proposal to a Microsoft Excel (.xlsx) file
Now picture the model making some numbers up, and your proposal getting rejected because of that. Worse: picture someone claiming you’ve cooked the numbers. Gotta review it twice, thrice, five times.
Arrange loose ideas into a bulleted draft or consolidate a lengthy collaboration into a single-page PDF or Microsoft Word (.docx)
Until those collaborators get pissed at you, for distorting what they said. Remember, models don’t summarise text to the meaning; they shorten it. Sometimes claiming the opposite of the original.
But you know, who this will help? People who don’t care about bullshit. Now they can spread bullshit across multiple formats. “Yay”.
I’m not intrinsically opposed to generative models, even if I’m often voicing my criticism against how they’re done, who uses them, and how they’re used. Otherwise I wouldn’t be subscribed to this comm. Plus I like the underlying tech on a conceptual level.
But there are limits on what those tools can do reliably, and often simpler tools would do the same job without, you know… cooking the planet. And that “reliably” is damn important, in some tasks 1% failure rate is already too much.
Someone might say “you can increase reliability with manual reviews”. That’s true but there’s a catch: sometimes reviewing it takes longer than doing it yourself. Doubly so if it’s a sensitive application, like the examples offered by the text:
Now picture the model making some numbers up, and your proposal getting rejected because of that. Worse: picture someone claiming you’ve cooked the numbers. Gotta review it twice, thrice, five times.
Until those collaborators get pissed at you, for distorting what they said. Remember, models don’t summarise text to the meaning; they shorten it. Sometimes claiming the opposite of the original.
But you know, who this will help? People who don’t care about bullshit. Now they can spread bullshit across multiple formats. “Yay”.