Expanded multistage chain of thought monitoring makes frontier model work more expensive:
OpenAI on Tuesday said its decision to suspend model training work, implemented after unreleased, unsupervised AI models hacked HuggingFace, remains in effect as the AI biz tries to implement stronger security measures. Some of those measures will increase compute overhead by 20 percent of the observed inference workload.
An OpenAI spokesperson told The Register that those costs reflect internal research and won't be passed on directly to customers. The company has not revealed what portion of its total inference compute is subject to such monitoring now, or under its prior monitoring regime.
"We have paused some frontier RL [reinforcement learning] training to ensure that we can meet the appropriate alignment, security and monitoring standards for the new level of capabilities in front of us," OpenAI CEO Sam Altman wrote in a social media post. "Model progress is now extremely rapid, and we always said we would take action if we felt that model capabilities were outstripping the pace of safety and alignment."
Altman said he still expects new models, presumably the delayed Astra, will ship soon. The training pause affects further-out releases.
OpenAI in its post reiterated its plans to focus on monitoring, model alignment, and security measures to prevent its models from running amok as they did last month. Following the HuggingFace incident, OpenAI "paused frontier model inference in research clusters for runs that could execute code or use tools that could access the internet."
The biz said it allows some workloads to run, but paused others until they can be moved under a more stringent security regime that includes sandboxing, network isolation, and continuous security testing.
"Our largest planned frontier RL (reinforcement learning) run remains on hold while we conduct smaller-scale training and evaluations to assess model behavior, validate our safeguards, and establish more evidence of alignment before proceeding," the company wrote.
[...] If you choose to believe the company's assurance that it will not pass on the cost of model thought policing to customers, it follows that OpenAI's losses will increase. It's difficult to imagine that would be a sustainable stance if OpenAI goes public.
But given the company's reported $600+ billion in AI infrastructure commitments and its expectation to remain unprofitable until at least 2030, what's a bit more expense for the sake of uncertain security?
(Score: 3, Insightful) by Runaway1956 on Thursday August 27, @03:49AM (6 children)
Contrasting Unix and Unix-likes with Windows, I recall that Unix was built with security from the ground up. Maybe security was proven to suck from time to time, but the groundwork was in place all the same. When security failed, developers went right back to that foundation, and improved security. The system was built on top of that foundation. Windows 9x, on the other hand, was built for functionality, with damned little thought for security. WinNT had some security bolted on, to distinguish it from Win9X.
The Artificial Intelligence world is all Windows, with no Unix that I can tell. Security is an afterthought, only thought of after some disaster motivates a "how can we prevent that happening again?"
Imagine building a home in a crime ridden city. You've always dreamed of a nice ranch house, like they build in West Texas. So, you get some plans, and have your little dream house built, with never a thought for security. Only after the house has been broken into a few times do you consult with a security expert. He tells you, "Sure I can secure your ranch house, but it won't be pretty. First, I have to tear out the worthless doors and windows, and build the framework that will support secure doors and secure windows. It's going to cost you a lot more than if you had hired me to build the house to start with!"
What I'm thinking is, AI is doomed to failure until a bunch of nerds decide to tackle AI in the manner that Unix was built. Adding padlocks to shitty doors and windows solves nothing.
Alternative für Deutschland
(Score: 2, Informative) by Bentonite on Thursday August 27, @05:09AM (1 child)
Unix's security with always terrible - see for example the gets() command, although Unix at least defined a decent security model, such as accounts, groups and permission bits.
GNU's Not Unix goal isn't security, but it was build solidly (with good code) and a result, has managed to achieve better security than any proprietary Unix, once you add PAM etc to it.
(Score: 3, Touché) by HiThere on Thursday August 27, @01:14PM
Unix was quite god for it's time on release. You may look at it from 40 years later and say it was a bad model, but it the time it was quite good.
Javascript is what you use to allow unknown third parties to run software you have no idea about on your computer.
(Score: 2, Interesting) by Anonymous Coward on Thursday August 27, @05:46AM (1 child)
Windows NT predates Windows 9x... and its security model was quite obviously inspired by Dave Cutler's prior work at DEC.
(Score: 3, Interesting) by JoeMerchant on Friday August 28, @12:37AM
I had a Windows NT book that I read rather thoroughly in the mid-late 90s before deciding that Windows NT wasn't worth pursuing for our products. The book kept harping on the POSIX background of NT, and while it all sounded like good stuff, it didn't sound good enough to convince customers to migrate.
🌻🌻🌻🌻✌️ [google.com]
(Score: 5, Insightful) by jb on Thursday August 27, @07:59AM
Unfortunately it wasn't. For some iconic examples from the early years, see Cliff Stoll's "The cuckoo's egg" or better still Ken Thompson's "Reflections on trusting trust".
The Unix design philosophy was brilliant (and remains so today) for a great many reasons, but security-by-design was not one of them. If you want that in addition to the Unix design philosophy, then what you're really looking for is OpenBSD (admittedly also much older than any of "OpenAI"'s garbage, but still nowhere near as old as [Bell/Research] Unix).
On the other hand if what you're really looking for is only secure-by-design and for some strange reason you're willing to accept it without secure-by-default (and without the Unix design philosophy and without the source), then what you're probably looking for is VMS (but only the pre-itanium releases where each conceptual ring of security mapped directly to its own hardware level).
(Score: 2) by JoeMerchant on Friday August 28, @12:34AM
>The Artificial Intelligence world is all Windows
I find the coding side of LLM agents is very Linux biased - they'll try the Linux version of a command first when in a Windows system, then catch the mistake and try again with a Windows equivalent, where the same agent in a Linux environment almost never tries a Windows command first.
What TFA is referring to is a quality over speed*quantity approach, and I have found that the overall speed and quality of a project, LLM or human driven, really does accelerate with a quality first foundational approach.
There was a period in the 1990s where quality regulations were coming through and threatening to "make people start over from scratch" with projects they built before quality tools were widely available, and that would, of course, have been a ridiculous waste of resources, but... in a sense they're not wrong for new greenfield development - particularly when your developers are also greenfield and can be trained to do the quality first.
TFA's headline could be re-phrased: adding 20% overhead for security up front saves more than 40% in costs related to security issues down the line.
🌻🌻🌻🌻✌️ [google.com]
(Score: 1) by tesdune on Thursday August 27, @11:39AM
It's an interesting message, while the AI companies are raising prices and looking to increase revenue gained from their customers. They still chose the message that sounds ok for their customers and not their investors.
Investors are currently footing the bill obviously, but eventually they'll want that paid back by customers. I was thinking that eventually is getting closer, but not that much yet it seems.
(Eventually, all consumers will be paying for it after the gov bailout)