Eventually everybody grew into it, and we can't imagine life without the Internet. AI is having that same moment right now. The breathless hype bombardment will eventually give way to normalcy just the same.
Of course, it's likely that we'll get past that point, where AI generated images aren't in the uncanny valley, but until then it's off-putting.
Recently I had an email from someone praising my open source work and it was clearly AI generated, which felt completely insincere. And so, misuse of AI in situations where it makes the human being that interacts with it balk isn't good either.
As you say the language is annoying: "AI powered" or similar language is just meaningless to most people. People are not turned off by features in software that 'just work' because there's some AI running quietly behind the scenes. They're put off by bold claims that don't actually materialise and the human/AI interaction-boundary that can sometimes lead to more effort on the part of the human.
AI is a technique used by software engineers, we shouldn't need to talk about it at all (except to each other and maybe VCs). We don't market our apps as "Relational database powered", in time we won't mention AI at all.
The most embarrassing recurring ad I’ve seen lately is for some kind of device management IT solution with AI generated MacBooks where the Apple logo is completely mangled. No idea how you could overlook that or notice it but somehow think it doesn’t cheapen your company’s image.
Exactly this. Any serious company would manage their brand seriously. Even if you're a one person bootstrapped startup, you should take your brand seriously and get some artwork commissioned. It isn't that expensive and makes a massive difference to how you're perceived.
Using AI generated images says to me (rightly or wrongly) that you take shortcuts. So, if you take shortcuts with your brand, what's your product like?
You pretty much have to be at the cutting edge of generative ai to generate images, etc, but it still may have a shelf life.
Sonu is an AI music creator. Regardless of music style, I can immediately identify it. There are some very obvious (to me, anyway) musical choices it makes with both notes, and their delivery.
Rick Beato (YouTube music personality, for lack of a better description) pointed out his kids can tell, but he can't. I think, at least with Sonu, the moment the sound clicks, you can't not hear it when it's there, whether the song is pop, metal, country, classical... it has that specific sound that can't be ignored when you notice it.
The other day I had someone mention to me their copywriting prompts were still looking basic.. and right away I knew his process was basic or dated.
It's clear that it has been trained on everyone's real output, ever, and their "You can't request the styles of real artists" is just a figleaf.
But the output is impressive.
People often put things like that in their prompts, and confirmation bias tells them it worked when they run it a few times and get something that sounds like The Doors, even though the meaningless bit of prompt you gave it was simply ignored. The similarity is because of your words and description of the style. Period. Writing "Yngwie" in the prompt will get you arpeggios, but they sound nothing like him. Just as another example. Because the prompt is giving the AI genre cues only, and nothing so specific as actual Yngwie arpeggios, scales and ladder riffs.
The “figleaf” is the service saying “Oh we can’t use artists in prompts, here, let me use a prompt that looks suspiciously like a set of tags from a specific music site instead” nod wink. And then the underlying model quite often fills in a well known voice and style.
And that’s if the service spots the artist was in there and substitutes anyway, which it doesn’t always.
So regardless, prompts may or may not be giving detailed Yngwie arpeggios to the model, but the model is clearly trained on them and can reproduce eerily similar music and voices to any given artist when given the right triggers.
It’s (IMHO) clearly plagiarism on a massive scale.
Great fun to play with though.
(To be crystal clear, my contention is not “they are doing some dodgy prompt stuff to make their model produce known artists”, it is that their model is very good at spitting out massively ripped-off voices and styles. It is their attempts to cover this with “we don’t allow you to specify an artist!” that are the figleaf over the nature of the model)
Sometimes it sounds like every singers voice that can fit a genre can be heard at once unless you specify it.
When it works seamlessly, it fades into the background, letting non-engineers enjoy our creations. But when it falters, it distracts and adds no value.
Our focus should be on delivering flawless creations that provide real value, regardless of the technology behind them.
Same for AI generated text: it's super easy to see when something is AI generated and follows a pattern.
Those patterns get quite tiring to read after a while. In hotel websites, for example, there's often AI text that is obviously converted from existing tabular data. I would prefer to have proper UX for the data instead of the text. With text it's harder to scan for specific information or to compare between pages.
It also is being pigeonholed into areas where it is not yet ready. There are a LOT of examples out there on that.
A hobby of mine is using AI to find bizarre ways to code things that aren't the normal way of thinking of the problem. Obviously it isn't practical, but it is fun. I most recently got ChatGPT to generate Pi from a random length string of 5's like this:
~~~
from math import sin, radians
def fives(repeat: int) -> float:
repeated = 1 / int('5' * repeat)
value = sin(radians(repeated)) * (10 ** (repeat + 2))
return value
Yes, it uses a repeating string of 5's to generate pi. Call it with "fives(18)" as an example. This is a cleaned up version of what ChatGPT gave me, but all the same, it works. A string of 5's can give you progressively more accurate estimations of pi - more 5's, more accurate pi.
Bizarre question to ask an AI, but playing like this is how eventually discoveries are made, since it is capable of coming up with working answers.
(Edit: corrected markdown copy-pasta weirdness. You can copy/paste but have to remove the extra indents HN adds. It's normal python, you know how to do it.)
This code just outputs almost-zero. It clearly just sets `repeated` to a number very close to zero, takes the sin of it - which will also be close to zero - and then multiplies that by a small number.
I don't think this is how discoveries are going to be made.
What am I missing here?
from math import sin, radians
def fives(repeat: int) -> float:
repeated = 1 / int('5' * repeat)
value = sin(radians(repeated)) * (10 ** (repeat + 2))
return value
Cut/paste that into your python repl, and make sure the lines are intact without the HN extra indents. Change the 18 to any int you want (except it overflows at 308... easy to fix, but not worth bothering).
It might be close to zero... but it's much more like 3.14... when you run it.
But using `sin`, and `radians` means that you equation to generate pi already has pi in it.
You can do the same with 9s, for the same reason you might be picking up on if you don't trivialize that there is another pi in the mix. You can't escape it when you're talking about circular things, regardless.
Every proof or non-coincidental estimate of pi (ie: not 22/7) out there has pi embedded in the proof somehow. Dismissing it is like saying the Pythagorean theorem is bollocks because they also add up to the same number of degrees. Even though you're describing it in lengths, the degrees don't disappear. This code is very much like that. Just not triangular.
It's not (quite) as trivial as just copying the pi.
sin(radians(repeated)) * (10 ** (repeat + 2)) = radians(repeated) * 100 * 1.8 / repeated = repeated * pi / 180 * 180 / repeated = pi
So this program is basically an equivalent of
multiply_the_argument_by_pi(555) / 555
This was a result of converting degrees to radians. There is another way as well.
from math import sin, acos
def nines(repeat: int) -> float:
repeated = 2 / int('9' * repeat)
value = sin(repeated) * acos(repeated) * (10 ** repeat)
return value
nines(18)
Took a while to convince me, but you got there.
value = sin(repeated) * acos(0) * (10 ** repeat)
The problem with AI blogspam (and AI image spam) and is that most of it is not even interesting to begin with: the desired result and the prompts are already terrible, so even if it was custom drawn or written by a professional, it would be crap. With AI, because of the lower barrier to entry, people don't care about spending a few cents building garbage. This is why I doubt more quality AI will help. :/
The professionals using AI (and even then, a minority of them) actually spend the time to ensure they're getting quality results. Even AI requires some massaging to get what you actually want. It's a tool. Not a miracle. A carpenter doesn't just bang a nail once and expect things to hold together.
I'm not a particularly good coder. I'm just interested in the oddities that can be shown through code, and am just (barely) savvy enough to see it through. I only had a vague idea how you could get pi from a string of 5's, but it wasn't very hard to get ChatGPT to code it for me once I described it.
The same can be done with a randomish length string of 9's, but I'll leave that for others to discover.
If someone can't be bothered to write it, why would someone be bothered to read it?
100% agree and wish I’d also made that point in my post! It’s like it’s taken the average of human creativity, so everything is average. The unrelenting blandness is definitely offensive.
Having said that, it is exactly why training an AI on AI output results in utter crap. This has been known since before AI, so it's funny to see it arising in this industry.
We're not sick of AI. We're sick of it being used for the wrong things.
Also, random idea for the company name; EnnuAI. A play on "ennui" promoting something like "hey, we do the tedious and boring stuff so you don't have to!"
Not much of a marketer myself, but thought I'd throw it your way.
Because a lot of us are addicts.
I saw an interview with Tony Iommi (Black Sabbath) and right behind him was a rack with a Pod Pro amp simulator. I recognized it because I have one, and indeed, you can easily dial in his tone and typical amp sounds. But I'm left wondering how many Sabbath fans noticed that he's using digital modeling these days, or if they can even tell when he started doing that.
People don't like change. It takes a lot of marketing hype to create motion for change to happen. And most people won't even notice it happen. Just like with Sabbath albums.
Messenger-style chatbots abound today. They're everywhere. As well, you have Twitter and Youtube bots that are the grown up version of the irc bots from days gone by.