Artificial intelligence still has some way to go

The fact that diners so vastly outnumber chefs at the moment

The fact that students so vastly outnumber teachers at the moment

The fact that patients so vastly outnumber doctors at the moment

All kinds of everything remind me of you

whats the next sv hype object after ai, crypto had an amazing run of doing nothing while maintaining high visibility, the metaverse was basically dead on arrival, i think ai is gonna be closer to the metaverse timeline just cause its so expensive its hard to be all were in early days while burning through billions in data bills

I'd have to live to be 600 in order to listen to all the bands I've heard about on this message board alone

btw lol i hadnt heard about the currency used in microsofts big investment

Only a fraction of Microsoft’s $10 billion investment in OpenAI has been wired to the startup, while a significant portion of the funding, divided into tranches, is in the form of cloud compute purchases instead of cash, according to people familiar with their agreement.

https://www.semafor.com/article/11/18/2023/openai-has-received-just-a-fraction-of-microsofts-10-billion-investment

Sounds like the Nvidia thing (investing in startups who use the money to buy Nvidia gpus)

https://archive.is/2024.02.07-141916/https://www.ft.com/content/e1beb7a5-6c91-4d7f-bc90-79689774881d

azure credits are funny. it means openai gets a playground for whatever they’re doing, but it’s also funny money. did they get the negotiated discount azure credits, or one-to-one dollar-based azure credits?

depending on how the plan is negotiated, the azure bucks actually go less far if you don’t use them all and pay a little money past your discounted spend. microsoft’s real skill is the deals department

I’d guess that since it’s advertised as “look how much we took in” it’s the dollar equivalent. which could be as little as half as much if you’re going by corporate bulk spending plans

https://simonwillison.net/2024/Mar/22/claude-and-chatgpt-case-study/

I recently learned that the Adirondack Park is the single largest park in the contiguous United States, taking up a fifth of the state of New York.

Unlike most parks in the United States, about 52 percent of the land is privately owned inholdings.

this isnt legit im sorry

I really don't get people who think it's more fun to badger the stupid robot into doing something that you clearly could do by yourself than to just do it, tiresome

using the robot as code google is fine idk nbd but of course the one flaw in the plan is if everyone does it then itll stop working because there wont be any websites to suck the answers from, then i guess people will have to go back to asking websites

obvs also other downsides like having no idea what the robot did which is why the story was about figuring out how to do some little unfamiliar task that didnt matter, but hey life does involve a lot of little unfamiliar tasks that dont matter and if you need to understand what the robot did having some at least seemingly working code to look at isnt a bad place to start, a bigger question i guess is if these tools will be cost effective or were just in one of those periods where everyones enjoying billionaire subsidies, kind of annoying that copilot costs money actually it should be free cmon

maybe i will unsubscribe to copilot reflecting upon it doesnt really work that good, half the time youre say trying to import a function you give it the name and one would think it would be trivial for it to complete the path but instead it just straight makes one up even tho it has access to your files it should know what exists, sometimes its pretty useful tho, and then occasionally it does something legit impressive but thats pretty rare idk

it's so weird how the things that are fundamental to normal computers eg actual hard info, are the things that AIs consistently get wrong

it is very funny that it cant do arithmetic reliably, youre a computer get it together

the black box nature of it is bizarre theres all this money and geniuses behind it but theyre still like idk maybe try asking it to make hands with five fingers

ok i will make hands with five fingers from now on *adds even more fingers*

This is niche and dorky, but as someone who works predominately with audio, I'm really looking forward to the restorative programs and plug-ins that are hopefully around the corner.

There's already been some great improvements in the last few years that make my life a lot easier for sure, but they fall down in certain ways, especially where music is involved.

I expect in 5 or more years we should be able to properly restore and improve old film/tv audio, archive recordings, tape, optical and wire etc.

i think niche and dorky is sort of AI’s sweet spot tbh. general <waves hands> magical productivity improvements less so.

film world arguing about ai will catalyze the singularity

A friend sent me MRI brain scan results and I put it through Claude.

No other AI would provide a diagnosis, Claude did.

Claude found an aggressive tumour.

The radiologist report came back clean.

I annoyed the radiologists until they re-checked. They did so with 3…

— Misha Saul (@misha_saul) March 22, 2024

lol

Impressed with the vitriol my post has attracted

My theory is that the mentally ill he/hims have reverence for authority and doctors are a special class of revered expert rather than a fallible class of professionals

Or maybe trying to use tech is inherently suspicious? 🤷‍♂️ https://t.co/dBG8odC038

— Misha Saul (@misha_saul) March 24, 2024

https://x.com/nousresearch/status/1771735632035127594?s=46

xp
https://i.postimg.cc/sgVbY6j2/IMG-7757.jpg

People who can’t think think “AI” can think.

I really don't get people who think it's more fun to badger the stupid robot into doing something that you clearly could do by yourself than to just do it, tiresome

because it took him 6 minutes this way rather than an hour the other way.

i don’t rely on it to think i only rely on it to have good judgment.

(def hearing that by The Cramps)

This is niche and dorky, but as someone who works predominately with audio, I'm really looking forward to the restorative programs and plug-ins that are hopefully around the corner.

There's already been some great improvements in the last few years that make my life a lot easier for sure, but they fall down in certain ways, especially where music is involved.

I expect in 5 or more years we should be able to properly restore and improve old film/tv audio, archive recordings, tape, optical and wire etc.

― Maresn3st, Sunday, March 24, 2024 5:52 AM (two hours ago) bookmarkflaglink

Honest questions: why do these things need improving? If they're archived properly already, why "improve" them? If they're not archived properly already, what is wrong with current archival protocols? I guess I am just highly suspicious of AI "improving" things to the point of totally denuding them of their original context, which is part of what makes them what they are.

Maybe I'm in over my head here, but surely archiving and restoration are two separate, though related, concerns? I took that post to mean restoring degraded audio in music, film, etc.

This admitidely gives me some knee jerk concerns as the line between restoration and messing-with is notoriously difficult to trace in image (cf complaint about the "yellow" nature of a lot of the Cinemateca di Bologna restorations of classic cinema) and I'd assume audio as well, but tbf this is an eternal concern that predates AI.

i think that’s otm. i won’t speak for maresn3st but AI has been used in video compression (where information is necessarily lost to enable efficient storage or carriage) for some time. it is in effect restoring lost information by analysing the frames around it and filling in information. at least in part it’s how we’re able to watch high quality video. (there are countless examples like this that make the world go round, load balancing mobile phone mast data loads for example) - i’m just selecting one in my area.

because it took him 6 minutes this way rather than an hour the other way.

― 𝔠𝔞𝔢𝔨 (caek), Sunday, March 24, 2024 8:18 AM (forty-three minutes ago)

The fuck does he need the other 54 minutes for jerking his dang hog

I'm not sure I understand your post, Fizzles— I would frankly rather watch a film as it would have appeared fifty years ago than in a "gloriously restored version." Maybe I'm in the minority there, but I am often left completely cold by such restorative efforts.

sorry i wasn’t being very clear about the context. in order to store, transfer and stream media, it’s necessary to compress it, which means deliberately losing information or data from the originally produced version. this is in a sense a degradation of the media. in order to avoid this degradation being visible to the viewer, AI (and other algorithmic processes) are used to “fill in” the gaps. it’s extremely refined and sophisticated - invisible to audiences.

it is worth noting that some broadcasters don’t like having compression before a hand off. that’s mainly because there will be subsequent conversion or information-loss processes depending on where it’s being distributed or stored. the more you do it, the harder it is to reconstitute.

i was using this example as an analogy for historically degraded media.

my point is that we use AI to reconstitute media all the time. this doesn’t invalidate your point, which i think daniel is also saying, which is that questions of restoration and appreciating the degraded form etc

… questions of that sort are persistent…

(trying to post while cooking only marginally less successful than trying to post when not cooking)

I would frankly rather watch a film as it would have appeared fifty years ago than in a "gloriously restored version."

The vast majority of film restoration is aimed exactly at making it possible to watch a film as it would have appeared fifty years ago - and this is necessary because film stock degrades. I think you're misreading "restoration" here to always mean stuff like adding CGI in or Lucas messing with his old Star Wars films.

I wonder how much AI is being used in the guerrilla nerd efforts to make 4k copies of the original theatrical releases. Some of them definitely used AI upscaling from the laserdiscs, but the other nerds are hardcore in procuring buried cinema prints and scanning them frame-by-frame.

The fuck does he need the other 54 minutes for jerking his dang hog

― G. D’Arcy Cheesewright (silby), Sunday, March 24, 2024 12:03 PM (thirty-five minutes ago) bookmarkflaglink

he can use it to contemplate if a park thats half not a park is really a park

Thanks for clarification and edification, Daniel and Fizzles.

Heya, so quite a lot of my work is spent tidying up poorly recorded audio (or degenerated/archive audio) to make it more discernable and pleasing, it is mostly spoken word.

To address your point Table, I think the only kind of media that might *need* to be restored using machine learning/AI/neural network software would be something that has degenerated past the point of being listenable or was recorded with faulty equipment or involved some kind of operator error. It may have been poorly archived, or it could perhaps have been stored correctly but still be substandard quality.

There will surely be a level of futzing around by nerds with things in the future, same as we're experiencing with film and TV images.

This might be somewhat dry so apologies in advance, but one kind of plug-in that I use a lot right now builds back in lost, or never captured, frequencies to an audio clip (while cleaning up background noise and some other balancing issues) the AI component definitely has a ways to go before being a magic bullet solution, you can easily end up giving someone a wicked lisp or making them sound robotic and weird.

It's also not good with musical elements, but I'm sure this will get better in a few years, I'm just keen to find out how much more we can improve these things.

What I was getting at with my post was the potential to improve, for instance, an old movie made in the 1930s that might have a very bad quality optical soundtrack.

It wouldn't be equivalent to a modern recording, but you could subtly build in low and high frequencies that would smooth the overall sound out and improve the listening experience, same with any archival material really, old radio recordings, historical documents, forensics etc.

Here is a link to a before and after clip to demonstrate, as you can hear, it's still very basic, but it was almost impossible to do to this level even a few years ago - https://we.tl/t-b7Q6juy3ds

I appreciate that, Maresn3st. I guess that some of what we're talking about, and what that clip demonstrates, is an idea of "integrity." I don't know much about this stuff, so while I agree that the adapted version is more clear, I was left wondering: what was the motoric noise in the first clip? Where was the original recording made, and in what context? If such a recording is clear enough to be transcribed, which the example you give is, then why does it need to be "cleaned up," so to speak? Maybe I'm just a glutton for punishment or something, but I think that the weird motoric rumbling in the original is *part* of the context of the recording— why get rid of it?

i was watching a film from '39 over the weekend and the things that needed doing to it were a) adjusting the position of all the frames so there was less jitter and b) fixing the slight changes of brightness in each frame. and i think this is the kind of thing that could be done by an algorithm (and neither of which would really destroy any data from the original the way, say, colourisation does)

XP - In the context of my work, very simply to make it easier to hear and sometimes to help bed it into something like a radio documentary or podcast.

In this instance you could argue that it doesn't have to be cleaned up, sure. However, for me, static background noise or artefacts of a primitive recording process aren't necessarily integral or precious, and in that small clip you can discern more of the voices in the background that were masked by noise.

It was an outdoor wartime recording of Neville Chamberlain, at an airfield iirc.

Got it, thanks!

I think that what I worry about isn't necessarily what you're point to or doing, which makes sense to me. What I worry about is that these cleaned up versions will become the "standard" versions by which certain events are known or available in the archive, which to my mind goes against the spirit of a lot of archival practices. That is, the adapted version has its functionality that is important, and the original has its functionality that is important, but I worry that the two will become confused, or even that the latter will be lost and discarded.

Absolutely yeah, I know my view is a little basic I guess because it's ultimately threaded into my work practices.

But I understand how future technology will affect the integrity of archive media, we already have to be careful with AI assist software as it can very easily tip the balance and make a voice sound like a different person, heck I've even heard it sneak in extra syllables in more extreme cases.

On the upside, I think what the people who write this software are reaching for is aligned with what Daniel was talking about upthread, the restorative aspect, but yeah, it may tip over into something else as archive media becomes ever more malleable.

gotta make sure you dont motion smooth buster keaton

the worst thing copilot does is create these phantom imports which you dont even notice cause theyre at the top of the page not where youre currently working, oh bool from sharp thanks thats a huge help seems super real

https://i.imgur.com/R0AV6rQ.png