this guy looks mildly grumpy he was summoned for no reason

it's online, but two versions out of date; i fell off from midjourney after v6 and for sure none of the stuff about CLIP encoding or multi-prompting still applies. probably other portions don't as well. but at the time i was proud of how i got to some examples here
docs.google.com/document/d/1...

i'm assuming this person is referring to image gen, in which case they definitely haven't touched it themselves. any halfway specific ask gets quite difficult! i wrote a friggin 60-page guide to midjourney because i was tired of watching people in their discord bang their heads against the wall

mildly cynical read: sergey, like many smart people too past their prime, has been one-shot (probably many-shot, tbf) by gemini
extremely cynical read: he massaged the prompt or just lied that it picked the "young woman engineer" he already had his eye on

glossing "don't sleep with sources" as "chastity" shunts the basic ethics violation out the window! women WHO ARE JOURNALISTS absolutely "owe society" that much, because of the second part, not the first

the rug-pull that udio did does not augur well. i'm glad i snatched the ~1 promising draft waveform i got out of em

Also, we've all seen that these influences are as quick to recede as they are to occur. Trump will also fly the blue and yellow for a little while after every Zelensky meeting before inevitably drifting back to team Putin. It's hard to get durable, tangible results from the flattery.

there are plenty of AI startups with no VC subsidy! midjourney famously had no outside investors until the meta licensing deal this year, which is why they're a bit pricier than other image gens. lots of smaller companies are self-funded and turn modest profits off niche fine-tunes or good GUIs

the chinese government isn't subsidizing US operating costs at all! the models are released open-weight and are hosted by numerous third parties who just buy marginal server time (the cost of which will nosedive further if the bubble pops and we're left with excess servers)

i'm not really here to defend CoT prebakes but taking your inference trends off that obscures the fact that the basic pretraining stack has been ablated to very low margins and is still very useful for many mundane asks

you can 🙉 if you wanna but plenty of companies are profiting on inference alone with chinese/open-source models (its a commodity that will consistently support *some* businesses, just probably not as load-bearing macroeconomic trends)

it's pretty clear to me, between investors and users, which class deserves more to get it good and hard