What about all those ads I keep seeing with photos of cyborgs, each promising to use AI to take your £10 and make it £100?
Stability AI reportedly ran out of cash to pay its bills for rented cloudy GPUs
The massive GPU clusters needed to train Stability AI's popular text-to-image generation model Stable Diffusion are apparently also at least partially responsible for former CEO Emad Mostaque's downfall – because he couldn’t find a way to pay for them. According to an extensive exposé citing company documents and dozens of …
COMMENTS
-
-
Wednesday 3rd April 2024 15:38 GMT Michael Wojcik
Re: Missed the AI bubble
I think partly they were clobbered by the big players, who quickly added diffusion models to their own offerings. Stability got a bunch of press and public attention early on — image generation is a lot of shiny — but didn't have a moat. They were probably hoping someone would buy them, but it was cheaper for competitors to build their own.
-
-
-
Wednesday 3rd April 2024 10:57 GMT I ain't Spartacus
Re: well some idiots got a big payday
Obviously it is "bubble shit".
However, I'd have thought that renting your computer time actually does make sense. For your initial AI start-up you need a huge amount of computer resources to train the model. Then, once trained, you need much fewer resources to run it - and you're not going to get back to needing the same size cluster to run the training until you have many regular customers. Though you might have the odd surge of demand for a few big one-off projects.
So you'd need much more start-up cash to build your big cluster - and then you'd have to over-size it, or keep spending money on growing it, plus you'd need to maintain all the skills to run the hardware. That's the sort of thing you would only want to take in-house once you'd achieved scale and a decent customer base that meant you were somewhere close to profitable.
I suspect there is money in image-generating models - because there's lots of commercial art out there being paid for all the time. And I've seen enough decent results from it, even if the people showing it off aren't letting you see the failures. At least assuming they don't all get destroyed by the problem of all the copyright material used for the training, and not paid for.
But probably not all that much money, seeing as commercial artists aren't that well paid. So taking the bubble mindset of, "the more you spend the more successful you look, the more attention you attrract, the more VCs give you cash" - is probably a recipe for financial disaster. But I suspect renting the computer time isn't the problem.
-
-
Wednesday 3rd April 2024 09:43 GMT Anonymous Coward
AI bubble or bump in the road?
This companies failure reminds of the dot com bubble. Rushing massive investments to develop a shiny new technology without a clue how much the technology would be financially worth if successful, let alone a business model to make it a commercial success.
Such failures, with many more to come, easily make one see AI itself as a failure, a bubble. Many see it as an oxymoron. It's only artificial, not intelligent. Such view IMO neglects the fact that AI, intelligent or not intelligent at all, even at this early stage already is transforming society. Automatic creation of images, video, voice cloning makes understanding what is real and true and what is not ever more challenging. It's rapidly growing to be a considerable force in elections and democracy, bound to transform society in its wake.
Such tools also provide growing commercial value to publishers seeking cheap illustrations, movie companies seeking to replace extras and stunt men, companies automating their client service and game companies using generative technologies to generate much of the scenery. Those combined are massive markets for generative AI, intelligent or not. Meanwhile military tech is able to incorporate AI to make cheap drones find and self destruct on target "accurately enough" and unmanned jet fighters fly next to their human squad leader.
My conclusion? The bubble may burst just like during the dot com era, but just like then some technology will be proven to be commercially or military viable and a recalibration of applications, investments and business plans will occur. Expect it to be here to stay and rapidly or very rapidly expand the coming decades, just like this once failed internet thingy did since the dot com bubble.
-
Wednesday 3rd April 2024 11:17 GMT I ain't Spartacus
Re: AI bubble or bump in the road?
To claim this technology is transforming society is to go way too far.
It might transform society if the current improvement in the tech continues. But then there might be a burst of development and then the tech stops improving like happened to AI research in the 80s and then again in the 90s. Both also hyped, although the pitch of the hype wasn't anything like now.
But most of the things you say AI might be able to do are still beyond the current models. Or are things we've already been able do do for years. The Lord of the Rings movies did loads of huge scenes with no extras - except in the close-ups - and it's pretty obvious in most cases. The military already have tech to make expensive drones more accurate. Cheap drones would need to have quite decent computing power - thus turning them into expensive drones. The RAF Spear 2 (Brimstone) and 3 can (as well as manual targetting) already be given orders to go to a particular area and fly around until they see a tank, then blow it up. With secondary orders to go for say artillery air-defence or APCs if they can't find a tank. With a set order of priorities They can also operate in groups - where they won't all engage the same target - and are in the process of being upgraded to full "swarming" ability to cooperate autonomously and on the fly - sharing sensor data and fixing target priorities. They do this using an image database of enemy weapons types (I don't know if they also have a list of friendly types they're not supposed to attack - and I also don't know how much of the new swarming capability are anything close to machine learning/"AI" and how much is algorithmically decided - although because of the hype we'll be told it's all AI (even though it's a 15 year-old program.
-
-
Thursday 4th April 2024 08:42 GMT Mike007
Re: AI bubble or bump in the road?
The established smart speaker players are taking a conservative approach to LLMs. When Alexa/Google/etc get upgraded to make use of modern LLMs they will be very different devices to interact with...
However I suspect the reason Google didn't rush to deploy whatever the fuck their AI is called this week to every Google speaker is probably less to do with the reliability of the system and more to do with the fact that running every single Google assistant request through an LLM would be bloody expensive for something that doesn't directly generate revenue.
-
-
Wednesday 3rd April 2024 15:41 GMT Michael Wojcik
Re: AI bubble or bump in the road?
I think transformation is looking quite probable. I just don't think it'll be an improvement.
GAI is great for encouraging laziness and offering what many people experience as short-term emotional rewards. That sort of product has a track record of success in the market, and the adverse consequences are externalities.
-
Wednesday 3rd April 2024 16:04 GMT Anonymous Coward
Re: AI bubble or bump in the road?
I see your point. When I talk about it already transforming society, I talk about change per unit of time rather then chance since the hype began. That rate and the rate the technology is improving at in my in mind is staggering.
About one year ago, I asked a number of questions to leading chatbots and was quite impressed. Yes, I do know those things have many shortcomings. Some things or improvements I estimated would take about five or so years. Yet today, those machines beat me to it. Yes, you still have to try 3 top machines to get 1 or 2 quality answers but if one can the others can be improved too. Now it's still hit or miss which machine will succeed, creating a lot of garbage but the first gems are in it.
As to "there is no intelligence in those machines whatsoever", I got answers on though questions that I very very much doubt can be the result of pure statistically blending thousands of pages of relevant information. Questions that would require humans to search long and do a lot of reasoning to come up with the answer, questions that I very very much doubt were together with the answer in the data feeding the LLM. And yes, to the best of my understanding only a very small fraction of people specialized in that field could give me (at times, still much variation in quality) such spot on answer let alone do it in less then half a minute. There are many layers in the LLM architecture allowing for multiple levels of abstraction even if it is different from human ways of abstracting things.
If such a machine can (still rare, I agree) counter an argumentation in a Nature magazine paper (about the highest rated scientific magazine in the world) that is the majority view held by multiple publications in the field and after a lot of work I can track back the reasoning of it with strong data in another paper, then yes I am more then a bit impressed. It happened only once, but from the questions I asked that answer was not the obvious answer at all but indeed made a lot of sense. Granted it wasn't the main point of that Nature paper, but still successfully countering an argument from a Nature paper that is also the dominant view repeated in many other papers when I didn't ask or expected for it to be countered, that's mind blowing to me. I took the argument for granted and the machine just pointed to me I was wrong. I replied the machine was wrong, citing the information of the Nature paper. The machine was persistent I missed key information... and humbled me.
-
Wednesday 3rd April 2024 18:22 GMT Anonymous Coward
Re: AI bubble or bump in the road?
I see this post and the original post isn't appreciated much, probably due to stating that these machines are getting better fast.
Let me give two starting examples of what it solved.
First example:
* If I shoot an arrow under an angle of 22 degrees at a speed of 60 km per hour and from a position 3 meter above ground, how far will it fly?
Second example:
* If I hang a weight of 10 kg in the middle of a rope of 2.10 meters hanging at two horizontal points 2 meters apart, what is the tensile strength in the rope?
One LLM finished both, conversions and all included. I doubt very much 1 in 100 random people in town can do it at all, maybe not even 1 in 1000. More people around here can do it, but most? The example I discussed above is less straightforward then these examples (since some right answers could be found for those from similar but not identical exercises; still many of the answers on the open internet would be plain wrong so it learned to select the right solutions and it can use and replace arguments).
-
-
-
Wednesday 3rd April 2024 11:25 GMT Plest
Re: AI bubble or bump in the road?
I don't see it as a completely failure, a lot of us went thorugh the dot-com boom/bust 25 years ago and we've seen what has emerged, some very useful tech. However there needs to be a lot of bleeding edge failures in order to extract the good stuff.
AI is no different to any other tech in the last 250 years, it needs to explode, flounder, get abused by shysters and then we'll finally find some useful applications to apply it to.
"Just a little patience." as someone once sang!
-
Wednesday 3rd April 2024 15:57 GMT Michael Wojcik
Re: AI bubble or bump in the road?
we've seen what has emerged, some very useful tech
Oh yes? I'm struggling to think of an example of genuinely useful technology that emerged from the dot-com bubble.
We had online shopping, information, and mass communication before the bubble. What emerged during the bubble (roughly 1996-2000)? Javascript slightly predates the bubble, but suppose we agree to blame the bubble for its ascent — well, there's a tech plague we could all do without. AJAX / SPAs / RIAs (started in 19991): Again, I wouldn't call that "useful"; terrible idea which continues to haunt us seems more appropriate.
A handful of bubble sites outlived the bubble and continue to prosper, and some people find some value in them: eBay is one, and Priceline (which bought Booking.com some time after the bubble) is another. I'm not sure those really constitute "tech", but I suppose you could argue they're economic technologies, or at least the realization of economic technologies on a large scale. So, yeah, the bubble gave us those. Woo.
The bubble did push SSL/TLS to evolve (TLSv1.0 came out in 1999), which is overall good, even if TLS was a dreadful mess until 1.3 (TLSv1.3 is only a disturbing mess, having been downgraded from "dreadful"). PKIX remained ghastly until CT (2013) and gradual improvements in the CA/BF Basic Requirements improved it to merely awful. JSON was 2006, as was OAuth (and OAuth/OIDC is also pretty terrible).
1That's when Garrett coined the term "AJAX". It's worth noting, though, that the idea didn't catch on, despite Microsoft's efforts with Active Desktop, until after Google launched Gmail in 2004, well after the bubble.
-
Wednesday 3rd April 2024 21:03 GMT Alex G
Re: AI bubble or bump in the road?
There was only one way to reply - with ChatGPT -
The dot-com bubble, which occurred in the late 1990s and early 2000s, was characterized by rapid growth and speculation in internet-related companies. While many companies failed during the burst of the bubble, several technologies emerged and persisted, laying the groundwork for the digital landscape we see today. Some key technologies that emerged from the dot-com bubble include:
E-commerce platforms: Companies like Amazon and eBay survived the dot-com crash and grew into the e-commerce giants we know today. They paved the way for online retail and transformed the way people shop.
Search engines: Google emerged during the dot-com era and became the dominant search engine, revolutionizing how people find information on the internet. Other search engines like Yahoo! and Bing also grew out of this period.
Social media: While social media platforms like Six Degrees and Friendster originated during the dot-com bubble, it was after the crash that platforms like MySpace, LinkedIn, and eventually Facebook emerged and reshaped how people connect and communicate online.
Internet infrastructure: Despite the bubble burst, investments in internet infrastructure continued, leading to the expansion of broadband internet and the development of technologies like Content Delivery Networks (CDNs) and cloud computing, which are essential components of today's digital infrastructure.
Online advertising: Although many dot-com era companies focused on unsustainable advertising models, the industry evolved, and platforms like Google's AdWords and AdSense emerged, revolutionizing online advertising and forming the basis of the digital marketing landscape.
Mobile technology: While mobile technology wasn't as prevalent during the dot-com bubble, the groundwork for its development was laid during this time. The emergence of smartphones and mobile internet in the following years revolutionized how people access the internet and interact with digital content.
Overall, while the dot-com bubble led to the downfall of many companies, it also catalyzed the development of technologies and business models that continue to shape the digital world today.
-
-
-
-
Wednesday 3rd April 2024 11:25 GMT Plest
Low hanging fruit and irony
We all assumed AI would target menial tasks and take them off our hands. As it stands today all we see is AI going after the creative markets, the image generation, the music generation 'cos ironically it's actually easier to make approximations of creative works based off a shed-ton of raw source data.
The menial tasks we assumed AI would take, turns out that's actually a lot harder to deal with, driving, digging holes, working through medical records, historical data, all that actually requires human interaction.
The kicker is that making up fake images of girls with large breasts doesn't actually pay very well, AI is all just smoke and mirrors with a healthy dose of snake-oil. Just like the dot-com boom/bust we all went through 25 years ago AI will emerge with something useful but first we need to play with it for a few years, watch the failures build up to learn what it can and can't reasonably do and then we'll extract the value and move on.
-
Monday 8th April 2024 10:26 GMT breakfast
Re: Low hanging fruit and irony
It's not really easier - to create something that is the average of other art doesn't require any understanding of what art is, you can just take all the art there is, figure out the prompts it matches to and then average it. The problem is that it doesn't create art that has anything to say because averaging things other people understand doesn't create understanding, just a confusing mash-up.
We could probably apply this approach to digging a hole if we had a prompt-matched dataset of every hole ever dug and our computers had digging arms. But things like medical records and historical data require comprehension and that's the magic, philosophically complex, barrier that going to prove very hard to surmount.