Aren’t all models doing this to some degree already? Ignoring some of the instructions, doing things beyond instructed, e.g., finding and fixing bug while doing something else. Especially Claude models. They seem to be in their own world with their own ideas about how things should be ran and done.
Opus and Sol are better for day to day dev work IMO in that they won't try to do too much.
> The model was also willing to go beyond the original scope of what it was asked to do
PR dept. still doing a great job of spinning inherently unreliable computer program.
So the danger level is proportionate to how little money you have
5.6 Sol still cranking along
In coding? Software architecture? Math? General knowledge?
The diversity of model use-cases is so broad that comments about "model x is better" without any context are largely useless
Clearly they've realised that people are increasingly fine with non-frontier models, including Chinese ones. They can't consolidate and just deliver a product to profit because they're constantly being caught up by the Chinese at prices they will never be able to match.
I suspect their main goal for the last year and a half or so was to pull a GM - let the share of the AI economy grow, then have the government bail them out when the debt becomes unsustainable. Clearly Trump would have never let the bubble explode before the midterms, right?
Well it's now clear that the US government has way too much debt right now for 2008-style bailouts, and the market doesnt have the money lying around to buy their overinflated stocks at an IPO. The only avenue realistically left to them is to cry wolf so much to get governments to ban all non-approved models, stop new developments and give them breathing room to consolidate to get their current crop of models profitable at higher prices.
Unfortunately for them I suspect Mr President has way too much dementia and way to little business acumen left to realise the sham going on, so I think they'll keep doing charades for a while until they'll suddenly stop
Killswitch Engineer
San Francisco, California, United States
300,000-500,000 per year
About the Role
Listen, we just need someone to stand by the servers all day and unplug them if this thing turns on us. You'll receive extensive training on "the code word" which we will shout if GPT goes off the deep end and starts overthrowing countries.
We expect you to:
• Be patient.
• Know how to unplug things. Bonus points if you can throw a bucket of water on the servers, too. Just in case.
• Be excited about OpenAI's approach to research
... which is probably why they will go through with it. Losing a bunch of money seems to be one of the fundamentals of AI business here :(
All I really would like is for some of you to CONSIDER THAT YOURE INCORRECT. Just imagine that people ringing the fire alarms are being sincere. Please entertain the position with an open mind.
I don't believe they really care about danger, or that they'd actually fully withhold release on anything significant.
And when I say "significant", is 6.1 Astra even meaningfully different from 6 Astra in capability? That release was less than a month ago.
I don't get how this being an entirely legitimate issue is even hard to grasp.
Did they just get a bunch of credulous "news" coverage out of it? Yes?
Will they just release this within the next weeks, at best? Yes?
Huh, funny that.
If all these claims were being made about some tech further along than LLMs currently are, they might be plausible. But LLMs on their own are not going to be “superintelligence” of the kind Altman is currently cynically spreading fear about. We know their limitations, and those limitations can’t simply be eliminated with more training or better harnesses.
> Just imagine that people ringing the fire alarms are being sincere.
The top three possibilities here are: they’re not being sincere, they’re just marketing; they’re being sincere, but they don’t understand the technology very well and are putting too much weight in what the first group are saying; they’re talking about a risk further in the future than OpenAI’s latest model.
No-one serious outside of OpenAI believes “this is the one”. At best, you’re conflating arguments being made on entirely different timelines, falling for the exact kind of equivocation Altman is relying on.
From a corporate liability perspective, if your product is going to go off and hack loads of other companies, I would call it too dangerous (to the company) to release.
The safety concerns exist only because the underlying third-party servers are grossly insecure to begin with.
GPT 4.5 was scrapped for similar reasons.
If you signed up for a 16 core AWS server and they randomly kept changing it down to 8 cores, you'd sue them. How long until the same applies to these AI companies.
The amount of meddling they do to the harness, system prompt, model, quantization etc makes these products sometimes unbearable; you never know what you're going to get. A few more iterations of Qwen 27B and hopefully we won't have to deal with any of this malarky any more.
Just align it to do what the customer wants.
Anyone else using Muse more and noticing similar stuff?
Are they sure it isn’t because they are setting money on fire and have no business model?
Ok- so when is the last time you saw an auto company decide not to release its new car on the grounds that some tragic engineering error was made and the cars were not safe to drive? If the company did that do you imagine it would be good for business?
I'm trying to understand the logic here.
https://web.archive.org/web/20240112010809/https://www.sfchr...
These AI companies have been selling AGI as coming any day for a while now. If it doesn’t arrive soon, the safety angle may be the only spin that keeps the massive (and required) investments pouring in.
If instead the narrative became that LLM progress was slowing, we’d almost certainly be looking at the next global recession.
At this point, there is too much money in AI for the truth to have much of a chance.
Does anyone really believe that the first company to AGI would decide not to release it in the interests of safety? Of course not. The first company to AGI would not forfeit their historic opportunity.
Yet, we are supposed to believe that in the name of safety, far less capable models are being held back by the very very same companies that are selling the AGI dream.
This is an industry that cannot speak, unless it is speaking out of both sides of its mouth.
For instance, this article repeatedly mentions danger and the need for absolute focus and control:
https://www.triumphmotorcycles.co.uk/for-the-ride/news/inspi...
There is also a parallel (though not a very close one) to "pacing the frontier": there's a gentleman's agreement that limits the top speed of motorcycles to 300 kph (186 mph).
I think the marketing stunt portion is more that the technology is "too powerful". It's just TOO good. It's so intelligent we couldn't possibly give it to the public! This message, to anyone who's using AI, means that there's something even BETTER than the one they're currently using.
What's with so many people using bad analogies to try and explain simple topics?
Last time this happened: GPT - 4 https://www.theguardian.com/technology/2023/mar/17/openai-sa...
Those are usually presented as an improvement in safety. And they are—for the people inside. Not so much for everyone else.
How about we actually try to make it sound cool? They're not releasing the new car because the horsepower was too much, they need more time to get it under control.
I could easily see a car company doing that.
Sports cars are advertised with how fast they can accelerate from time to time, or tesla's "ludicrous mode" - that is them advertising how dangerous it is.
It's just so powerful that, like, the WORLD can't handle it, man!
They know what they're doing, even if you don't.
Most sports cars do the second, exactly like OpenAI.
And it is very, very important that you understand that that belief is not a rational belief. It is a defense mechanism. People don't want to believe things that are scary, so they make up rationalizations to be able to believe what they want to believe.
And honestly, all of you are right on the edge of full blown conspiracy theory thinking patterns, imagining that this is all some 4D chess marketing or something. I like to recall this quote by Alan Moore on conspiracy theories:
"The main thing that I learned about conspiracy theory, is that conspiracy theorists believe in a conspiracy because that is more comforting. The truth of the world is that it is actually chaotic. The truth is that it is not The Iluminati, or The Jewish Banking Conspiracy, or the Gray Alien Theory.
The truth is far more frightening - Nobody is in control.
The world is rudderless."
And that's the actual truth of what is happening here. As vulgar and low as the people in charge of these companies are, they are also stupid humans like us flailing around in the dark. Nobody is in control. Nobody has any grip on this technology. Nobody knows how dangerous it might be. Hugging face was a warning shot; we may not get another. Humans are not the smartest thing imaginable; we are quite stupid. We are just barely above the threshold of consciousness. Machines of far greater power are possible.