(Adding Stellarst0rm’s example because I somehow forgot it https://www.reddit.com/r/funny/comments/6gzfra/damn_wish/)

Do you notice how the original sketch version has very different facial expressions for the punch line (if simplistic sketch art) and the AI doesn’t clearly convey anything by the younger’s facial reaction because the model paints the same “angry/alarm” pattern on both characters? Why is the younger vaguely angry in the AI version? You can make something up like “that’s just the style” but…

There is no answer because there is no style because nobody made this. It’s just a copy missing a piece of the humor the source had that didn’t make it through as statistically relevant even though it DOES contain data for the punchline to observers. Do you see how the next model training on this will actually lose something that’s a concrete part of the joke? Let’s ignore “soul” and say that it dropped the ball on nonverbal communication in a way that’s meaningfully worse to the viewer.

It’s true that people will have a hard time articulating why modeling art and language are worse than actually making them, but I don’t think their concerns are unfounded. It’s slop in my opinion because it loses to napkin art despite having vast resources applied to it. I understand how the difference may seem like a grumpy nitpick but the models can’t train off their own output without a buildup of convolutional traces that poison their outputs eventually and break them according to current studies. It’s clearly missing SOMETHING intangible and whatever people call it, it’s real.

I should have read more carefully, thanks for catching! I’ll edit and add something on topic.

“New York Life all said they’ve never hired prompt engineers, but instead found that—to the extent better prompting skills are needed—it was an expertise that all existing employees could be trained on.”

Are you telling me that the jobs invented to support a bullshit technology that lies are themselves ALSO bullshit lies?

How could this happen??

Original comment:

I agree with all of this under the caveat that your DM interpretation of the “what constitutes interaction” in 5e can also have more influence on your success than clever or clumsy play. DMs I’ve played with have varied from very loose discussions about their concerns for balance to self crippling houserules that wind up with illusionist NPCs failing to do anything because of a broomstick and blanket being flailed at them by a fighter. Specifically speak with the DM running the game and come to a common understanding before you start. Find out how interactions and passive checks will work against your illusions ahead of time in that game!

Edit because I’m turning into a lich and forgetting which editions we’re talking about:

Major Illusion, Hallucinatory Terrain, and Greater invisibility are amazing staples of illusion and you can look to Shadow Evocation and Shadow Conjuration as patches for any missing features a wizard should have in your party. Other than that I think… take standard wizard spells and ban schools based on your party if you have the chance to plan with them. Also consider the synergy that lies (enchantment) can have with misdirection and see if you can fill that role or support your local lying bard/rogue/beguiler.

Illusion for gnomes in 3.5 invariably leads to talking about their shadow magic prestige class from Races of Stone and the ability to change what disbelief entails as well as cast evocation and conjurations as shadowy illusion that become more and more real as you progress. This is amazing for high powered campaigns but personally I recommend house ruling metamagic level out of the %realism formula. This prestige class is in no way necessary though as illusion rocks and gets license to make memorable scenes in ways other spell schools don’t!