They are doing a lot of calculation per question and request as well, its a big model and presumably uses quite a bit of GPU resources just to run it. They are well past the sort of typical web business and marginal cost per customer that traffic in GPU/memory must be expensive.
Is it? Currently it's working for me, both the website and the API (for both 3.5 & 4). Yesterday the API was really slow though, sometimes it took like 30 seconds to respond to very simple queries.
It was 100% down. Logging just refreshed the page. I assumed I was banned so immediately sent an email to support, which was at 1:22 PST, it is now 2:50 PST and I'm able to login.
I was logged out right when I started a new chat and couldn't login. Email only support even with paid subscription. I was sure I was banned. Then wondered the implications of this, especially if I couldn't repeal it.
This is going to be the most worrisome monopoly in history if GOOGL & META don't get it together quickly.
It's not the fault of OpenAI that Google/Meta hid their shiny AI product so long in the basement and i trust Sam in this matter much more than Sundar or Mark to not abuse the power.
I think that's a reasonable grounds to make the claim it's not a worrying monopoly.
Also, OpenAI doesn't appear to have any capacity to stop other people doing LLMs.
And it's only been a few months since people went from normal people totally ignoring GPT-3 to treating ChatGPT as a magic oracle, so other companies being surprised and running to catch up isn't (or shouldn't be) a surprise.
And then there's the fact that nerds like *gestures at this site* us all, have already been getting leaked language models running on phones.
Why? Sam is a cryptocurrency scammer with a more invasive system than most[1]. Maybe he just doesn’t have as much power to abuse, but I see no reason to think he won’t do it when given the chance.
Instead of thinking “what tech billionaire do I trust the most”, let’s take a step back and realise they all have too much control over the lives of others.
In the long-term, I doubt that OpenAI will have a monopoly on LLMs. We're still very early in the development, and I don't see why other organisations, won't be able to catch up. The instruction-tuned LLaMa versions already look very promising, even with rather small & low-quality data sets. Compared to other technologies, the investment-requirements for LLMs still seem to be relativly low and something a number of companies or governments are able to finance.
I truly hope you're right. But there are a lot of factors at play apart from compute power. So many factors from UI/UX, censorship, pricing, to responsiveness need to be just right for product success. We are in a technology preview stage so the economics don't need to make sense for OpenAI (atleast for now), but they need to for the other companies since they're beholden to shareholders.
Bing chat is dumber than a bag of rocks although it supposedly uses GPT-4 when asked the same questions which is I believe a direct consequence of the requirement it be open to the the general public.
Add to the fact that GPT-4 handles technical questions well which is what makes it useful for me, but I'm worried future LLMs will be more tuned to handle non-technical questions with vague responses as way of focusing on the average layman user.
When first computers came up, only the government and big business had monopoly on them. It took 50 years until people were able to have their own independent (networked) computer in their homes.
Yet, the world didn't fall apart. (Actually, it almost did due to nuclear ICBMs, which were enabled by computers.) So I think we're gonna be fine.
I don't think that's a cogent argument. The quote "Nobody ever got fired for buying IBM” is as close as I can get to without writing a long response. That kept them in business to the detriment of the industry even if the world didn't fall apart.
First mover advantage is real, even in consumer market; Zune was better but iPod won. Think of how different the world is today because of that.
I work on IBM mainframes, and as you can see, they didn't win, although they easily could have. But the success of System/360, as well IBM PC, owes to their openness, contrary to your claim.
It's not all or nothing forever. It look a long time before IBM came down. I would call domination for as long as they did a win.
Openness of the platform has nothing to do with what I'm talking about, it is their prevalence as a company on many areas even when there were perfectly good alternatives.
That fact that OpenAI decided to not even reveal the parameter counts / training data composition in their "technical" report is surely a sign that their moat isn't as big as it would seem in fear of competition. This paired with the rise of LLaMa and other alternatives makes the monopoly even less likely.
Its just a matter of time till every company has an internal ChatGPT equivalent.
I do wonder - if the AI takes over will it update status pages more promptly, or less promptly? And will it word the outage announcements more weaselly, or less weaselly?
Imagine a world, were things can be smart depending on the availability of the grid, meaning power, servers and software. On bad days, your house and car are retarded, on good days, you just never notice it.
Mouse over text: "Wikipedia trivia: if you take any article, click on the first link in the article text not in parentheses or italics, and then repeat, you will eventually end up at "Philosophy"."
I find that this is no longer working. But I also remember it being reliably working at the time of first reading this. Anyone know when/why it "broke"?
Just tried it on a few entries and it still worked for me. Also cool, if you keep applying the same method after Philosophy, you end up in a couple of clicks in an infinite Entity -> Existence -> Entity loop.
Graceful degredation sounds like the best case scenario, to be honest. Backup power for essentials when the grid goes out, applications that work offline, heating that keeps working without an Internet connection.
It seems obvious but I'm not sure we're heading in that direction.
Progressive enhancement would be better. Then luddites like me could just turn off the bullshit.
We tried PE on the web. It's a great idea and the browsers and languages support it eminently. But most developers (including product people, and engineers like myself) are obsessed with the glamorous option rather than the plain-but-plainly-better lower bandwidth version.
We're going to have more and more bad days in the future as fuel sources become less plentiful and secure and climate change makes it harder to keep data centers running reliably.
Climate change is having and will continue to have significant effects on our physical infrastructure, especially in the coastal areas, areas prone to flooding/drought, and wildfires.
But I don't see data centers/internet/power being an issue. Data centers are already globally distributed and load-balanced. Internet is near-ubiquitous, with growing wired, wireless & even satellite options. Power generation from renewables, along with storage, have been growing at an accelerated pace even despite (or some say because) the pandemic.
Also, an off-grid solution is very feasible: I can easily imagine a near future where I have my "smart home" running on a local Alpaca instance, being powered by my solar roof.
Global distribution and load balancing exist yet we still need DownDetector. One of the reasons data centers go down is because of weather. Picture Cascadia devastating the west coast while a category 5 hurricane pummels the east coast.
Renewables aren't the panacea so many people want them to be.
I can't imagine that near future because despite the feasibility, the companies that manufacture smart home products want people in their thrall. Why go for a one-and-done purchase when that ubiquitous Internet can be used to hold basic functionality hostage for a recurring fee?
Most outages nowadays are caused by misconfigurations or hacking, not individual data centers going down. Think Azure, AWS, etc.
Renewables aren't a panacea, but they exist now and are seeing growing deployment, even without "fuel sources becoming less plentiful". Too late to prevent climate change, but definitely early enough to prevent widespread power scarcity.
Oh, companies will definitely try, and many people will go along with it. But, people who care will continue to maintain self-hosted projects e.g. Home Assistant. Hardware can be jailbroken, and/or small manufacturers will be willing to charge a premium for open/hackable versions. Worst case, many "smart" items can be cobbled together from Raspberry Pi/ESP32/ESP8266.
This reminds me of years okay when Microsoft would carry on about the Chinese were stealing, cooying, and reselling our software, the media was freaking out, it was supposed to be a huge threat to Democracy. I always wondered why they would develop software that was so easy to duplicate, there must have been a way to make it unduplicatable
This reminds of that. "Don't let the peons play with ChatGPT, we must shame them not to use it, appeal to their higher instincts and make them realize this is too good for them! If the company's problems are too many stupid people jumping on board, then figure out a way to kick them off. Or is this what they are doing?
59 comments
[ 3.6 ms ] story [ 109 ms ] threadThe issue has been identified and a fix has been deployed, we will continue to monitor to ensure the site recovers completely.
Edit: Yeah, so I was wrong about the subscription.
This is going to be the most worrisome monopoly in history if GOOGL & META don't get it together quickly.
Or are you talking about MSFT providing the compute power and not actually OpenAI? If so I don't think that makes a difference.
Also, OpenAI doesn't appear to have any capacity to stop other people doing LLMs.
And it's only been a few months since people went from normal people totally ignoring GPT-3 to treating ChatGPT as a magic oracle, so other companies being surprised and running to catch up isn't (or shouldn't be) a surprise.
And then there's the fact that nerds like *gestures at this site* us all, have already been getting leaked language models running on phones.
Once apon a time there were pigs in a farm and the rest was history such that those who do not learn from it are condemned to repeat it.
Why? Sam is a cryptocurrency scammer with a more invasive system than most[1]. Maybe he just doesn’t have as much power to abuse, but I see no reason to think he won’t do it when given the chance.
Instead of thinking “what tech billionaire do I trust the most”, let’s take a step back and realise they all have too much control over the lives of others.
[1]: https://www.buzzfeednews.com/article/richardnieva/worldcoin-...
Bing chat is dumber than a bag of rocks although it supposedly uses GPT-4 when asked the same questions which is I believe a direct consequence of the requirement it be open to the the general public.
Add to the fact that GPT-4 handles technical questions well which is what makes it useful for me, but I'm worried future LLMs will be more tuned to handle non-technical questions with vague responses as way of focusing on the average layman user.
Yet, the world didn't fall apart. (Actually, it almost did due to nuclear ICBMs, which were enabled by computers.) So I think we're gonna be fine.
First mover advantage is real, even in consumer market; Zune was better but iPod won. Think of how different the world is today because of that.
Openness of the platform has nothing to do with what I'm talking about, it is their prevalence as a company on many areas even when there were perfectly good alternatives.
OpenAI uses the same Techology as FB or Google.
Dall E was unrivaled until stable diffusion came along, OpenAi won't be a monopoly soon.
Its just a matter of time till every company has an internal ChatGPT equivalent.
ChatGPT has decided it's down
https://i.redd.it/nvfde0q2jrl11.jpg
https://xkcd.com/903/
I find that this is no longer working. But I also remember it being reliably working at the time of first reading this. Anyone know when/why it "broke"?
It seems obvious but I'm not sure we're heading in that direction.
We tried PE on the web. It's a great idea and the browsers and languages support it eminently. But most developers (including product people, and engineers like myself) are obsessed with the glamorous option rather than the plain-but-plainly-better lower bandwidth version.
I feel like progressive enhancement is my choice to make. What do I purchase, what can it do, etc.
Whereas graceful degredation more is the product designers's choice. What does it do when it doesn't get Internet or electricity.
But I don't see data centers/internet/power being an issue. Data centers are already globally distributed and load-balanced. Internet is near-ubiquitous, with growing wired, wireless & even satellite options. Power generation from renewables, along with storage, have been growing at an accelerated pace even despite (or some say because) the pandemic.
Also, an off-grid solution is very feasible: I can easily imagine a near future where I have my "smart home" running on a local Alpaca instance, being powered by my solar roof.
Renewables aren't the panacea so many people want them to be.
I can't imagine that near future because despite the feasibility, the companies that manufacture smart home products want people in their thrall. Why go for a one-and-done purchase when that ubiquitous Internet can be used to hold basic functionality hostage for a recurring fee?
Renewables aren't a panacea, but they exist now and are seeing growing deployment, even without "fuel sources becoming less plentiful". Too late to prevent climate change, but definitely early enough to prevent widespread power scarcity.
Oh, companies will definitely try, and many people will go along with it. But, people who care will continue to maintain self-hosted projects e.g. Home Assistant. Hardware can be jailbroken, and/or small manufacturers will be willing to charge a premium for open/hackable versions. Worst case, many "smart" items can be cobbled together from Raspberry Pi/ESP32/ESP8266.
It seems with GitHub's greatest hits and its monthly downtime, we have learnt absolutely nothing.
[0] https://news.ycombinator.com/item?id=34958375
This reminds of that. "Don't let the peons play with ChatGPT, we must shame them not to use it, appeal to their higher instincts and make them realize this is too good for them! If the company's problems are too many stupid people jumping on board, then figure out a way to kick them off. Or is this what they are doing?