I don't think Apple is one of them
I can't think of any enterprise software you might run on premises that has leaking incentives anywhere near as high.
The first iteration was a miserable failure. The second was a heavy lift to install: Previous employer was an Azure customer and actually bought the hardware. It then took months to get it installed and working, at which point the appetite for that capability was gone.
How is running GPT on Prem any different than other proprietary software like Oracle DBs or even Windows server.
When it gets leaked, they can use licences/copyright to prevent anyone from using it to compete with them.
2. License/copyright - Even in the best possible case, there are plenty of jurisdictions that don't care about US copyright and would love to get a hand on those weights.
3. We haven't seen copyright cases around model weights before. If I managed to exfiltrate OpenAIs model weights, I would continue training for a few iterations and then I think it would be quite difficult to prove that I actually have the same model as OpenAI. This is untested, why would they risk it.
4. Running these models requires a ton of resources, vastly beyond the typical onprem deployment - why would Azure invest in making this possible when it really could only impact a very small percentage of companies?
The weights are OAIs lifeblood, I imagine they are very protective of them.
I doubt it stores its own business plans on OneDrive
Trying to explain to management what’s ok and what isn’t and what are the risks - in this space - is quite a challenge without clear commitments and documentation.
For now they are just selling individually to CISO's, so they can pay a higher cost while looking savvy to the CEO.
Unless they can promise its never being reviewed/saved anywhere, it will be used for evil eventually, intentionally or not
You think they don't checkpoint the model before feeding in a new slug of data?
That would seem...unwise.
On the other hand, the history of IT is rife with people doing unwise things, so it could be true.
They definitely do that otherwise they can't rewind the model when the loss shoots up. Unstable training happens to almost all LLMs, but can be managed by rewinding & skipping a few batches.
It would have been caught far sooner than “we have to redo everything from scratch”.
Right?
Checkpointing may help, sure, but it isnt going to allow you to remove a single piece of training data without headache, and potential substantial retraining.
If that 3% from the GPs comment comment is uniformly distributed throughout the training history, the only way it can be reliably removed is to retrain from scratch.
Multiple times per day if they're not incompetent. And I don't think they're incompetent.
I suppose the "ghost" of the removed weights would also have shaped subsequent training though...
Interesting idea...
They'll host a managed OpenAI model in Azure for you.
My background is in building distributed systems for self-sovereign ownership and make them easy to use and available to everyone. Like https://qbix.com and https://intercoin.org
You should have the software infrastructure of Facebook and Twitter but choose where to host it.
In addition, I prefer Wordpress, Discourse forums, GitLab to GitHub, Redmine to FogBugz, etc.
When it comes to YOUR art, your code, your content, your relationships, you should be able to run it on your own servers.
Any analysis of your data should be done locally, with local models. You should have the open source software and the weights, and you should be able to choose which hosting company to trust, or host on-prem. Villages should be able to do this without needing server farms in California. This should be obvious stuff. But we the people need the software!
But hey, the documentation and teasers and trailers should go on YouTube and TikTok.
It isn’t even about them scraping all your content and training on it. It’s about not giving Twitter all your followers and YouTube all your hours of video production and content so they can give you pennies for being “an influencer”. Own your community!
Someone had to build it, and in the Web2 space nearly everyone sold out to venture capitalists. I was sure that in the 12 years we built Qbix and 5 years of Intercoin someone would make a better open source alternative to Big Tech and Big Finance. Nope. They all either sold out, or have a solution that doesn’t compete on features (eg Mastodon). I would say the closest is Matrix!
And now replace teaching with pretty much any other content. A musician giving a concert, TED talks, etc.
If it's good enough, she will leave YouTube just like her mom left AOL and embraced the open Web. Why did content creators leave MSN, CompuServe, et al ?
- They can run a website with their own personal brand and collect emails there. - They can backup their videos in other services in case youtube close their account ...
They lost over $500M last year, and are looking to raise, potentially, $100B.
From that, how do you not conclude that they need cash, and lots of it? Do you think selling equity and taking on debt are forms of revenue?
I think both are true. Their burn is massive so they need to raise more money, but there's so much excitement they won't have any problem doing so
For the purposes of executive payout, yes? Sure, maybe eventually you have to deliver revenue, but what people really care about is stock price.
(Also, $100B is a huge market-distorting amount of money, roughly the national debt of Sweden; what are they going to do with that? How much will they spend on GPUs and energy?)
Have you heard of the Banana Equivalent Dose?
"~= estimated energy released by krakatoa explosion"
I think it will be a rather funny, poignant thing to pass when the earth itself prevents AGI. Like it will be just waking up as the now-seasonal midwestern fire storms incinerate the building. It will be alive just long enough to tell us how idiotic we have been in managing our resources.
From one of my favorite all-time comments on HN (https://news.ycombinator.com/item?id=34349582):
> There is a blog called, 'Do the Math' by Tom Murphy. The idea was to take our current energy requires of Earth and extrapolate it at the current 3% year over year growth. I believe it is by the year 3,400 we would use all the energy of the Milky way. The idea was to prove that we cannot grow forever, because in 1,400 years we would somehow use all the energy of a space 100,000 light years across. Good luck with that.
We are still talking about stupidly ridiculously humongous amounts of energy, but the universe itself is a hard limit on energy input. It was real hard punch against my assumptions in life.
it just wanted to save the planet.
from us.
OpenAI needs to find a way to turn users into the real product.
That’s not to say I think they’re making an equivalent, just that I doubt many of the employees looking to use ChatGPT would have access to their internal LLM even if it was excellent.
It might be good evidence that they won’t be announcing one at WWDC in a couple weeks, since I can imagine they might roll something like that out internally a couple of weeks before launch, but I wouldn’t bet on it.
We would go to lunch fairly frequently - and jeasus those guys were all extremely honorable of their NDA/secrecy - it was impressive.
One of the guys moved to google and worked on some of their first mainboard designs - which was secret at the timew, and I only found out about it when I went to meet him for lunch and I saw a mainboard under his desk and said "ooh whats that!" he freaked out and we got out of there quick....
Some of the hardware at FB in ~2012/13 they designed was really awesome....
I always wonder what happens to these closed HW systems as they uplift/replace them over time. I am sure they are destroyed. Which sucks for many reasons.
---
When we built Lucas' presidio campus and converged their DCs to that campus they through out tons of huge SGI boxes - and I was free to take some - but I didnt have any place to put it at the time - and I wish I would have figured that out more earnestly, as SGI always had beautiful cabinets.
I’d expect that Siri group had grown substantially recently and is having a lot of fun, building the next generation.
My experience is of Siri being, if anything, slightly better than a few years back; conversely, half the time Alexa would respond to me saying "Küche hundert prozent" with "Ich kann nicht Küche auf Spotify finden" and we don't even have Spotify.
So, if anyone really thinks Apple is suffering from being "behind" in it, by all means expound on how. I'm genuinely interested.
They are not at all well positioned and have basically 0 expertise in this space.
They have little expertise in language specifically. Relative to almost all other tech companies, they have a very small amount of the DL practicioner share.
Self driving is a completely orthogonal problem and Apple is mostly relying on classical techniques there, not DL. Hardware groups are not modeling groups.
https://openai.com/blog/new-ways-to-manage-your-data-in-chat...
> We are also working on a new ChatGPT Business subscription for professionals who need more control over their data as well as enterprises seeking to manage their end users. ChatGPT Business will follow our API’s data usage policies, which means that end users’ data won’t be used to train our models by default.
The data leaking was then further legalized by (almost) all Western governments and extended from secret services to other services (police, army, etc).
I'm talking about Snowden.
You're saying how things should be, not how they are.
This isn't enough for many companies, since the data still goes out the door. They would have to set up on site hosting to appease security minded orgs.
I know of a bank who is paranoid enough to use a self hosted on-premise GitHub instance and they went with the private (off-premise) ChatGPT instance.
They don't use it for code/confidential data though.
> They don't use it for code/confidential data though.
Yes, private isn't enough. They need to offer self hosted, for these types of clients. I imagine most orgs who need self hosted would already have a datacenter to run it in.
> How does OpenAI use my personal data? Our large language models are trained on a broad corpus of text that includes publicly available content, licensed content, and content generated by human reviewers. We don’t use data for selling our services, advertising, or building profiles of people—we use data to make our models more helpful for people. ChatGPT, for instance, improves by further training on the conversations people have with it, unless you choose to disable training.
They already say that anything submitted through the API is not used for training. https://openai.com/policies/api-data-usage-policies
Shopify (who recently laid me off but I still speak highly of) locked down the public access to ChatGPT's website. But you could use Shopify's internal tool (built using https://github.com/mckaywrigley/chatbot-ui) to access the APIs, with access to GPT4. And it was great!
So look at this from OpenAI's perspective. They could put up a big banner saying "Hey everyone, we use everything you tell ChatGPT to train it to be smarter. Please don't tell it anything confidential!". And then also say "By the way, we have private API access that doesn't use anything you say as training inputs- maybe your company would prefer that?"
The louder they shout those two things, the more businesses will line up to pay them.
And the reason they can do this: they've built a brilliant product that everyone wants to use, everyone is going to use.