You’re not missing out. Astra was good for a couple of days and is now so quantized it’s often worse than 5.6. And 6 sol is a nothing burger to keep a wedge in the media when Opus 5.5 came out.
I noticed Sol seems to get rate limited enough to get pretty annoying now. Same for me with every model release that isn't really better than the last, leaving the old model doesn't mean much when they start shifting over capacity.
To be fair, instructing their swarm to communicate on an old wiki somewhere to plant the "AI escape" hysteria story was a hella clever marketing idea. I think they've earned themselves a cheap win.
Lots of people seem to have the bizarre opinion that companies like openai declare their own products to be unsafe as some kind of marketing stunt.
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?
Car companies sell mostly the same car every year. They don’t have billions (or trillions) of dollars of debt riding on the next model they release being substantially better and cheaper (closer to profitability) than the previous one.
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.
Motorcycle manufacturers have been known to do so. Riding a motorcycle is thrilling in part because it is dangerous. The fact that powerful motorcycles are so difficult to control that most people have no business owning one is absolutely used to make them sound cool and desirable.
For instance, this article repeatedly mentions danger and the need for absolute focus and control:
There is also a parallel to "pacing the frontier": there's a gentleman's agreement that limits the top speed of motorcycles to 300 kph (186 mph) because of concern that faster bikes would be unsafe enough to bring about a regulatory crackdown.
Different type of business. Many car companies release new cars showing you how much more powerful and implied dangerous they are. Both in capabilities and design style. In the US they also get much bigger with less visibility which gets double dangerous. But the sentiment they're going for is pretty close - "you want this powerful thing, don't you?"
> 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?
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.
Because they declare that their product is unsafe regularly, which gains massive exposure on every news outlet worldwide, then they release the model anyway with zero negative repercussions.
What's with so many people using bad analogies to try and explain simple topics?
Considering that the venn diagram of AI safety gooners and OpenAI is a circle, and that these idiots are starting to worship this stuff as a religion, expecting rational action from them is kind of ludicrous. For all we know Aella promised them a gangbang if they delayed the model a week.
(Not saying that it is a marketing stunt, though they said the same thing about GPT-2).
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.
Okay. I see what is going on here. You are all operating from the deeply held belief that it is inconceivable that a machine can be made that can do any cognitive task a human can do.
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.
OpenAI will have to keep releasing unless it wants to bleed customers. If Anthropic leaps ahead by a couple of more models, even more OpenAI users will then switch over their $100+ subscription from OpenAI to Anthropic. Just see what happened to Google.
The safety concerns exist only because the underlying third-party servers are grossly insecure to begin with.
Prompt: "Based on similar announcements delaying the release of frontier models as 'too dangerous', when can we expect OpenAI to release the model they announced was too dangerous today?"
I'm of the mind that distillation is happening, but from what I've read, it's not some magical activity. It's basically adjusting the weights from fuzzy to more precise; It's not specifically a training method, and likely, it's not an ongoing need. Once they bootstrapped the models, they can try to do the same as the "SOTA" and generate their own synthetic data and try to curate that to completion.
Aside from it being a completely hollow cry for attention from US labs, its also just what these models are designed around: taking in data and forming a way for them to operate on some level of intelligence and context. It's like a dictionary maker getting made that someone used their complicated words, and another dictionary maker heard the word and wrote it down then checked what the first dictionary maker wrote about that word. It is not dictionary maker B reads dictionary A directly, as the labs want to imply.
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.
What are we going to do when the servers are in space? Presumably, that would involve shooting down a bunch of satellites and sending us into Kessler Syndrome in the process.
It's a nice thought idea, but I think anyone silly enough to actually attempt this are going to discover the extreme drawbacks very quickly and lose a bunch of money doing so.
... 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 :(
Yea, being responsible for ensuring all life is locked onto a dying planet with an impregnable LEO and an atmosphere full of rare elements isn’t a great way to win any popularity contests. Risking killing everyone to make unkillable AI sounds so stupidly sci-fi villainous that it just might be funny. Funny enough that they’re going to get away with it.
Muse is totally free. It doesn't throw up gauntlet messages to use it like GPT does even for paying members. Also with Muse's free version I can create iPhone apps unlike chatGPT Work which I pay $20 a month for.
Anyone else using Muse more and noticing similar stuff?
I can’t believe how cynical everyone on HN is about this. Sure Sama is a chronic liar, and maybe this is a farce; but a lot of very intelligent people are afraid of the danger we are bringing about by racing to superintelligence.
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.
So we should believe the boy who cried wolf, this time?
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.
You can't believe a company that's investigating over 10k security incidents in which subagents that they developed and are hosting appear to be indiscriminately hacking into whatever they think will give them some data that improves the prompt response might not want to release that to the public and have subagents hack into stuff for profit?
I don't get how this being an entirely legitimate issue is even hard to grasp.
> a company that's investigating over 10k security incidents in which subagents that they developed and are hosting appear to be indiscriminately hacking
Glad you agree that OpenAI is responsible for this activity. Shouldn't we be charging the entire board of the company with, for example, violating the Computer Fraud and Abuse Act?
If this activity is so dangerous, and they developed the tools and continue to allow them to be used, it seems to trivially follow that they're knowingly violating federal law.
Sorry, I didn't mean to imply that you are directly responsible for charging them, but more that maybe we as a community should be (way) more concerned about the apparent clear violations of the law here as opposed to a hand-wavy concern about "superintelligence."
What is superintelligence? Can you show that a LLM (which is a deterministic text-generation algorithm) is intelligent at all, let alone “superintelligent”?
I have considered that I’m incorrect, but I have literally zero evidence to indicate that the “very intelligent people” you speak of are correct, so I will maintain a position of non-belief of the claim until belief is warranted.
Have GPT 5.6 Sol and GP6 6 Sol been nerfed due to the same resource constraints? They have seemed dumber recently.
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.
99 comments
[ 0.16 ms ] story [ 177 ms ] thread5.6 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
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.
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 to "pacing the frontier": there's a gentleman's agreement that limits the top speed of motorcycles to 300 kph (186 mph) because of concern that faster bikes would be unsafe enough to bring about a regulatory crackdown.
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.
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...
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.
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.
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."
Are they sure it isn’t because they are setting money on fire and have no business model?
The safety concerns exist only because the underlying third-party servers are grossly insecure to begin with.
Just align it to do what the customer wants.
So the danger level is proportionate to how little money you have
We honestly don't know. Anthropic keep claiming this, and certainly have provided some evidence of this.
However, given that they've removed the real thinking output, and yet the Chinese models are still doing well, I'm sceptical of this belief.
Aside from it being a completely hollow cry for attention from US labs, its also just what these models are designed around: taking in data and forming a way for them to operate on some level of intelligence and context. It's like a dictionary maker getting made that someone used their complicated words, and another dictionary maker heard the word and wrote it down then checked what the first dictionary maker wrote about that word. It is not dictionary maker B reads dictionary A directly, as the labs want to imply.
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 :(
Anyone else using Muse more and noticing similar stuff?
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.
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 concern is not superintelligence. The concern is superstupidity - in unleashing dangerously unreliable computer programs on the internet.
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.
Letting people call those agents would expose them to a great deal of liability that just wasn't there in older models.
Glad you agree that OpenAI is responsible for this activity. Shouldn't we be charging the entire board of the company with, for example, violating the Computer Fraud and Abuse Act?
If this activity is so dangerous, and they developed the tools and continue to allow them to be used, it seems to trivially follow that they're knowingly violating federal law.
Call your guy
I do believe fire alarms go off after the fire is detected.
And the correct response is evacuate - not sit put and hope for the best.
I have considered that I’m incorrect, but I have literally zero evidence to indicate that the “very intelligent people” you speak of are correct, so I will maintain a position of non-belief of the claim until belief is warranted.
I’m not sure which logical fallacy this is but you’re not helping your case.
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.