The current breed of generative “AI” won’t ‘die out’. It’s here to stay. We are just in the early Wild-West days of it, where everyone’s rushing to grab a piece of the pie, but the shine is starting to wear off and the hype is juuuuust past its peak.
What you’ll see soon is the “enshittification” of services like ChatGPT as the financial reckoning comes, startup variants shut down by the truckload, and the big names put more and more features behind paywalls. We’ve gone past the “just make it work” phase, now we are moving into the “just make it sustainable/profitable” phase.
In a few generations of chips, the silicon will have made progress in catching up with the compute workload, and cost per task will drop. That’s the innovation to watch out for now, who will de-throne Nvidia and its H100?
Flash games did not die out because people stopped playing them. The smart phone was created and this changed the entire landscape of small game development.
If it had been killed without an adequate replacement (eg. mobile gaming) then people wouldn't have let Flash die. There are open-source flash players.
It died because Safari for iPhone supported only open web standards. Flash was also the leading cause of crashes on the Mac because it was so poorly-written. It was also a huge security vulnerability and a leading vector for malware, and Adobe just straight up wasn't able to get it running well on phones. Flash games were also designed with the assumption of a keyboard and mouse so many could never work right on touchscreen devices.
This is why I, as a user, am far more interested in open-source projects that can be run locally on pro/consumer hardware. All of these cloud services are headed down the crapper.
My prediction is that in the next couple years we’ll see a move away from monolithic LLMs like ChatGPT and toward programs that integrate smaller, more specialized models. Apple and even Google are pushing for more locally-run AI, and designing their own silicon to run it. It’s faster, cheaper, and private. We will not be able to run something as big as ChatGPT on consumer hardware for decades (it takes hundreds of gigabytes of memory at minimum), but we can get a lot of the functionality with smaller, faster, cheaper models.
Definitely. I have experimented with image generation on my own mid-range RX GPU and though it was slow, it worked. I have not tried the latest driver update that’s supposed to accelerate those tools dramatically, but local AI workstations with dedicated silicon are the future. CPU, GPU, AIPU?
Technically I could upgrade my desktop to 192GB of memory (4x48). That’s still only about half the amount required for the largest BLOOM model, for instance.
To go beyond that today, you’d need to move beyond the Intel Core or AMD Ryzen platforms and get something like a Xeon. At that point you’re spending 5 figures on hardware.
I know you’re just joking, but figured I’d add context for anyone wondering.
You’re right, I shouldn’t say decades. It will be decades before that’s standard or common in the consumer space, but it could be possible to run on desktops within the next generation (~5 years). It’d just be very expensive.
High-end consumer PCs can currently support 192GB, and that might increase to 256 within this generation when we get 64GB DDR5 modules. But we’d need 384 to run BLOOM, for instance. That requires a platform that supports more than 4 DIMMs, e.g. Intel Xeon or AMD Threadripper, or 96GB DIMMs (not yet available in the consumer space). Not sure when we’ll get consumer mobos that support that much.
Yeah, so far. It’s super early in the modern incarnation of AI that actually has the chance to pay off, LLMs.
This isn’t like Bitcoin where there’s huge hype for a pretty small market opportunity. We all realize the promise, we are just still figuring out how to get rid of hallucinations and making it consistent and tuned to a certain business usage.
I’m in cardiology - radiology is right across the hall. We’re both under ancillary services. Lab techs are not rad techs, and rad techs don’t read films. You don’t work in this field, do you?
None of this is implemented so none of it is paying off.
Also nobody thinks this will take their jobs, because it looks like Theranos to us, in that it’s very hyped in tech and ridiculous to those of us in the medical field.
Well, and also navigating the minefields that the LLMs absolutely have copyrighted material in them that wasn't paid for or licensed. E.G. Dall-E can produce a full image of Fresh Cut Grass, a character owned by Critical Role.
And that the stuff they produce isn't copyright-able.
Exhibit A, Disney, a giant megacorp whose most famous works are literally founded on public domain material.
Bear in mind that public domain is not like a copyleft license, it's not "viral." If I make a movie and the Mona Lisa shows up in it, that movie is still copyright to me even though there's a public domain element in it. It's even easier with unique AI-generated stuff because you can't even tell what's public domain and what isn't.
Something has to be ownable to be public domain. AI produced items are un-ownable, since the AI is the owner, but it can't own them since it's a legally a "tool".
You are deeply confused about what "public domain" means. Something that is un-ownable (in an intellectual property sense) is public domain.
You may be referring to the Thaler v. Perlmutter case when you say "AI is the owner?" That's a widely misunderstood case that's gone through quite the game of telephone in the media. The judge in it ruled that an AI cannot own copyright, but that doesn't mean that AI-produced art is uncopyrightable. Just that AIs aren't people, from a legal perspective, and you need to be a legal person to own copyright. If Thaler had claimed copyright for himself, as a person, things might have gone differently. But he didn't.
Every cash grab right now around AI is just a frontend for a chatGPT API. And every investor who throws money at them is the mark. And now they’re crying a river.
I don’t know about you, but I would consider writing papers/books/essays/etc. (even bad ones) and code (even with mistakes) intelligent human behavior, and they’re pretty good at imitating it.
It’s also helped me find the names of several books and films that have been rattling around in my mind, some for decades, which actually made me very happy because not remembering that sort of thing drives me a little mad.
I’m stuck on two books that it can’t work out - both absolute trash pulp fiction, one that I stopped reading because it was so terrible and the other that was so bad but I actually wouldn’t mind reading again.
It also reminds me of crypto. Lots of people made money from it, but the reason why the technology persists has more to do with the perceived potential of it rather than its actual usefulness today.
There are a lot of challenges with AI (or, more accurately, LLMs) that may or may not be inherent to the technology. And if issues cannot be solved, we may end up with a flawed technology that, we are told, is just about to finally mature enough for mainstream use. Just like crypto.
To be fair, though, AI already has some very clear use cases, while crypto is still mostly looking for a problem to fix.
I’m still trying to transfer $100 from Kazakhstan to me here. By far the lowest fee option is actually crypto since the biggest difference is the currency conversion. If you have to convert anyway, might as well only pay 0.30% on both ends
Look into DJED on Cardano. It’s WAY cheaper than ETH (but perhaps not cheaper than some others). A friend of mine sent $10,000 to Thailand for less than a dollar in transaction fees. To 1bluepixel: Sounds like a use-case to me!
You still have to deal with ETH fees just to get the funds into the roll up. I admit that ETH was revolutionary when it was invented but the insane fee market makes it a non-starter and the accounts model is just a preposterously bad (and actually irreparably broken) design decision for a decentralized network, makes Ethereum near impossible to parallelize since the main chain is required for state and the contracts that run on it are non-deterministic.
I think that because it’s true. Smart contracts on Ethereum can fail and still charge the wallet. Because of the open ended nature of Ethereum’s design, a wallet can be empty when the contract finally executes, causing a failure. This doesn’t happen in Bitcoin and other utxo chains like Ergo, and Cardano (where all transactions must have both inputs and outputs accounted for FULLY to execute). Utxo boasts determinism while the accounts model can fail due to an empty wallet. Determinism makes concurrency harder for sure…but at least your entire chain isn’t one gigantic unsafe state machine. Ethereum literally is by definition non-deterministic.
Crypto found a problem to fix. The reason the problem remains: everything is run by that problem so it was astroturfed to death by parties that run the current financial system and the enemy of their enemy (who’s a friend), opportunistic scammers like SBF and Do Kwan.
No, this isn’t crypto. Crypto and NFTs were trying to solve for problems that already had solutions with worse solutions, and hidden in the messaging was that rich people wanted to get poor people to freely gamble away their money in an unregulated market.
AI has real, tangible benefits that are already being realized by people who aren’t part of the emotion-driven ragebait engine. Stock images are going to become extinct in several years. People can make at least a baseline image of what they want, no matter the artistic ability. Musicians are starting to use AI tools. ChatGPT makes it easy to generate low-effort, high-time-consuming letters and responses like item descriptions, or HR responses, or other common draft responses. Code AI engines allow programmers to present reviewable solutions in real-time, or at least something to generate and tweak. None of this is perfect, but it’s good enough for 80% of the work that can be modified after the initial pass.
Things like chess AI has existed for decades, and LLMs are just extensions of the existing generative AI technology. I dare you to tell Chess.com that “AI is a money pit that isn’t paying off”, because they would laugh their fucking asses off, as they are actively pouring even more money and resources into Torch.
The author here is a fucking idiot. And he didn’t even bother to change the HTML title (“Microsoft’s Github Copilot is Losing Huge Amounts of Money”) from its original focus of just Github Copilot. Clickbait bullshit.
I totally agree. However, I do feel like the market around AI is inflated like NFTs and Crypto. AI isn’t a bust, there will be steady progress at universities, research labs, and companies. There is too much hype right now, slapping AI on random products and over promising the current state of technology.
I love how suddenly companies started advertising things as AI that would have been called a chatbot a year ago. I saw a news article headlinethe other day that said that judges were going to improve the time they took to render judgments significantly by using AI.
Reading the content of the article they went on to explain that they would use it to draft the documents. Its like they never heard of templates
Most social media uses it. Video and music streaming services. SatNav. Speech recognition. OCR. Grammar checks. Translations. Banks. Hospitals. Large chunks of internet infrastructure.
The key fact here is that it’s not “AI” as conventionally thought of in all the scifi media we’ve consumed over our lifetimes, but AI in the form of a product that tech companies of the day are marketing. It’s really just a complicated algorithm based off an expansive dataset, rather than something that “thinks”. It can’t come up with new solutions, only re-use previous ones; it wouldn’t be able to take one solution for one thing and apply that to a different problem. It still needs people to steer it in the right direction, and to verify its results are even accurate. However AI is now probably better than people at identifying previous problems and remembering the solution.
So, while you could say that lots of things are “powered by AI”, you can just as easily say that we don’t have any real form of AI just yet.
Perhaps, but at best it’s still a very basic form of AI, and maybe shouldn’t even be called AI. Before things like ChatGPT, the term “AI” meant a full blown intelligence that could pass a Turing test, and a Turing test is meant to prove actual artificial thought akin to the level of human thought - something beyond following mere pre-programmed instructions. Machine learning doesn’t really learn anything, it’s just an algorithm that repeatedly measures and then iterates to achieve an ideal set of values for desired variables. It’s very clever, but it doesn’t really think.
I have to disagree with you in the machine learning definition. Sure, the machine doesn’t think in those circumstances, but it’s definitely learning, if we go by what you describe what they do.
Learning is a broad concept, sure. But say, if a kid is learning to draw apples, then is successful to draw apples without help in the future, we could way that the kid achieved “that ideal set of values.”
Machine learning is a simpler type of AI than an LLM, like ChatGPT or AI image generators. LLM’s incorporate machine learning.
In terms of learning to draw something, after a child learns to draw an apple they will reliably draw an apple every time. If AI “learns” to draw an apple it tends to come up with something subtley unrealistic, eg the apple might have multiple stalks. It fits the parameters it’s learned about apples, parameters which were prescribed by its programming, but it hasn’t truly understood what an apple is. Furthermore, if you applied the parameters it learned about apples to something else, it might completely fail to understand it all together.
A human being can think and interconnect its throughts much more intricately, we go beyond our basic programming and often apply knowledge learned in one thing to something completely different. Our understanding of things is much more expansive than AI. AI currently has the basic building blocks of understanding, in that it can record and recall knowledge, but it lacks the full amount of interconnections between different pieces and types of knowledge that human beings develop.
Thanks. I understood all that. But my point is that machine learning is still learning, just like machine walking is still walking. Can a human being be much better at walking than a machine? Sure. But that doesn’t mean that the machine isn’t walking.
Regardless, I appreciate your comment. Interesting discussion.
Automated mail sorting has been using AI to read post codes from envelopes for deacades, only back then - pre hype - it was just called Neural Networks.
At the time I learned this at Uni (back in the early 90s) it was already NNs, not algorithms.
(This was maybe a decade before OCR became widespread)
In fact a coursework project I did there was recognition of handwritten numbers with a neural network. The thing was amazingly good (our implementation actually had a bug and the thing still managed to be almost 90% correct on a test data set, so it somehow mostly worked its way around the bug) and it was a small NN with no need for massive training sets (which is the main difference with Large Language Models versus the more run-off-the-mill neural networks), this at a time when algorithmic number and character recognition were considered a very difficult problem.
Back then Neural Networks (and other stuff like Genetic Algorithms) were all pretty new and using it in automated mail sorting was recent and not yet widespread.
Nowadays you have it doing stuff like face recognition, built-in on phones for phone unlocking…
It is a risky investment. Taking care of your kid is something where we have done it enough that we understand the risks and pay off and most parents can make a reasonable prediction. (a few kids will "turn 21 in prison doing life without parole" - but most turn out okay and return love to their parents and attempt to improve society - though you may not agree with their definition of improve society)
I have no idea if the current faults with AI will be solved or not. That is a risk you are taking. It is useful for some things, but we don't know how useful.
Did you read the article? The problem hasn't been getting some people to pay for some things, it's that the things that are available so far are losing loads of money. Or at least, that's the premise.
I was going to say...I use AI-transcription tools for video editing, AI-upscaling, and Resolve dropped an incredible AI green screen tool that makes it effortless. I also use AI to repair audio as of 6mo ago lol. I don't think I gone more than 48hrs without using an AI tool professionally.
I wonder if “AI not paying off” in the context of this article actually means “Companies haven’t been able to lay off a bunch of their staff yet, like they’re hoping to do”
I know exactly what the journalist meant. They meant to get more clicks with some click bait headline and a bad article that will make them look extremely stupid in the future.
AI is a lot more like the Internet than it is like Facebook. It's a set of techniques you can use to create tools. These are incredibly useful tools, but you're not going to make Facebook money off of them because the techniques are pretty easy to replicate and the genie is out of the bottle.
What the tech bros are looking for is a way to control access to AI so they can be a chokepoint. Like if Craftsman could charge for every single time you used their tool to make something. For one very recent example, see what happened to Unity. Creating chokepoints and then collecting rent is the modern corporate feudal strategy, but that won't work if everybody with an AWS account and enough money can spin up an LLM and start training it.
Yes I use it all the time. Adobe Audio Enhance. It’s the flagship feature of their upcoming podcast app, but you can use it in browser currently. If you have an adobe subscription, it doesn’t charge extra or anything. It’s only for spoken word though, not music. If you throw music on it, though, you get some pretty wild stuff as it tries to create words out of the sounds.
To further answer your question, yes, it is actually very good with highly compressed audio. I regularly feed it zoom audio to make more intelligible. Obviously there are always limits, but I assure you it can do more than you can manually 85% of the time and buy a large margin. My only frustration is it is a simple slider, you can’t really fine tune it, but it’s still incredibly effective and I often use it as a first pass on the original audio file before I even start editing.
Topaz Labs makes a decent one. You’ll need to do a lot of trial and error to kind of find your own favorite settings for baking, but as far as cost and efficacy go, there aren’t a lot better out right now.
They do a watermark free version You can test with. I think it also only let you do a couple of minutes a video at a time. But frankly it’s incredibly processor intensive so you will only want to test a 15-20s clip at a time anyway.
The problem here is that AI in the media has become synonymous with generalized LLMs, while other “AI” applications have been in place for many years doing more specific things that have more obvious use cases that can be more easily commercialised.
It should also worry investors open-source AI is only months behind the big tech leaders. I looked into AI voice cloning lately. There’s a few really pricey options. Like $25 a month for a couple of hours voice cloning.
However, there’s already an open-source version of what they’re selling.
So far what I’ve seen from AI is that it lies and lies and lies. It lies about history. It lies about science. It lies about politics. It lies about case law. It lies about programming libraries. Maybe this will all be fixed some day, or maybe it will just get worse. Until then the only thing I would trust it is about something in which their is no wrong answer.
I never ask it things I don’t know. I don’t think that’s really what’s it’s useful for. It’s really good at combining words though. So it can write a better sentence than I could. Better in a sense that it’s easier for others to understand what my thoughts are if I feed them in as input. Since they were my thoughts originally I can spot the bullshit pretty fast.
Do people really not understand that we are in the early stages of ai development? The first time most people were made aware of LLMs was, like, 6 months ago. What ChatGPT can do is impressive for a self contained application, but is far from mature enough to do the things people are complaining it can’t do.
The point the industry is trying to warn about is that this technology is past its infancy and moving into, from a human comparison standpoint, childhood or adolescence. But, it iterates significantly faster than humans, so the time it can do the type of things people are bitching about is years, not decades, away.
If you think businesses have sunk this much money and effort into AI and didn’t do a cost-benefit analysis that stretched out decades, you are being naive or disingenuous.
If you think businesses have sunk this much money and effort into AI and didn’t do a cost-benefit analysis that stretched out decades, you are being naive or disingenuous.
Are you kidding? We literally just watched the same bubble and burst in companies that rushed to get their piece of the Metaverse and NFT cash grab. I worked at a SaaS company that decided to add AI features because it was in the news and Azure offered it as a service. There was zero financial analysis done, just like for every other feature they added
I’m sure Microsoft has a plan since they invested heavily. But even Google is playing catch-up like they did with GCP.
But there is a similarity, the hype pulls in all sorts of companies to blindly add buzzwords without even knowing how it might possibly apply to their product, even if it were the perfect realization of the ideal.
Yes AI techniques obviously have utility. 90% of the spend is by companies that don’t even know what that utility might be. With that much noise, it’s hard to keep track of the value.
But there is a similarity, the hype pulls in all sorts of companies to blindly add buzzwords without even knowing how it might possibly apply to their product
Yes, I see what you are saying. I guess we can add ‘blockchain’ to that list, then.
pretty much all improvements aren’t “better tech”, but just “bigger tech”. Reducing their footprint is an unsolved problem (just like it has always been with neural networks, for decades)
What I don't like about the article is that the phrasing 'paying off' can apply to making investors money OR having worthwhile use cases. AI has created plenty of use cases from language learning to code correction to companionship to brainstorming, etc.
It seems ironic that a consumer-facing website is framing things from a skeptical "But is it making rich people richer?" perspective
In my case, I still want to know if it’s not making rich people richer, because a) fuck rich people, and b) I don’t want to buy into things that will disappear in a year when the hype dies down. As a “consumer” my purchasing decisions impact my life, and the actions of the wealthy affect that more than you’d like.
Great, now factor in the cost of data collection if not subsidizing usage that you are effectively getting free RLHF from…
The one thing that’s been pretty much a guarantee over the last 6 months is that if there’s a mainstream article with ‘AI’ in the title, there’s going to be idiocy abound in the text of it.
Add comment