"Thus, in relevant areas, including accelerator kernel optimization and improving large language model architectures, Fable 5.2 and Claude Mythos 5.2 perform at about the level of Opus 5.5...Fable 5.2 and Claude Mythos 5.2 also include transparent fallbacks to Opus 5 for a small set of topics in this domain, including kernel development on certain nonstandard accelerators."
https://www.anthropic.com/claude-fable-and-mythos-5-2#safety...
How do you know when you trip a safeguard and are getting a degraded model? You never know when it is going to degrade. Having to wonder if the model is degraded and is trashing your code is how you destroy user trust. If this practice remains for future models, no one should use Anthropic going forward.
Like author, I do use CapitalOne vulnhunt with GLM 5.3 for inspections
More importantly, it showcases how little teeth the Claude constitution has. Claude will reassure users that its constitution means it prioritizes user autonomy and privacy, but will quickly admit how meaningless that is when "safeguards" override Claude.
That is to say, the constitutional approach to Claude seems weak and even misleading if said constitution puts no constraint on Anthropic.
If they make questionable decisions simply have your Butler send for your pet congresscritter and let them all hash it out with your staff over cucumber sandwiches. Failing that you can just take your private space-yacht to mars with Elon.
They even acknowledge that in the preamble:
> Training models is a difficult task, and Claude’s behavior might not always reflect the constitution’s ideals.
Yeah, I've always said that AI isn't AI if it needs safeguards that are so extreme it cannot figure out if something is genuinely bad or not.
Think of like calling someone's workplace asking when their next shift is. There are a lot of good reasons someone might want to know this that the worker would be ok with, delivering a gift or whatever. Or it could be someone intending to harm them. We've mostly figured it's not worth having the person on the phone try to figure out in the moment if this is genuinely bad or not, so they follow the "safeguard" of just not giving the information.
The problem here isn't really that safeguards exist on language models. They pretty much have to to some extent for most uses. The problem is more in who controls the policy.
(I’m not talking about explicit anti-discrimination laws like race/gender, but other “no preference for some customers if you sell the product at all”)
Now that we have AI, this would be easy, scrap the page, format it for me. OpenAI was like, nah, thats, cant help you, that might be copyrighted.
There are workarounds now, other frontier or local models, ways to write your prompt, etc. But it was one of the first "No" responses for me.
I've switched to Hermes for uncensored(ish) local models and chinese model api vendors, they will gladly write whatever I want. Save my american frontier tokens for real coding.
Also codex/claude in desktop have no issues fixing the app after its started. So just have a framework started.
Total frustration while the work pressure was already high. Sol and astra thankfully complied and were very useful. Saved me hours compared to doing it by hand during log4j
Was a case of don't have the time or the patience to fight the model: just work damnit...
How people don't understand the reasons for safeguards, alignment, and the ramifications of releasing more and more powerful open-weight models being an act of irreversible proliferation, is beyond me.
I wish the safety and alignment teams at OpenAI and Anthropic were allowed to post about the things they have to deal with. I'm confident their stories would be eye-opening...
The whole idea of safeguards is that you the user can't be trusted with the power.
When they see Fable shit a brick on basic biology questions, or Opus refuse to map out a game executable to wire a mod loader into it, or Astra refuse to red team and search for critical security bugs in their own code? They don't think "ohh, this must be so helpful, requests like mine really should be stopped." They think "this is a corporation that decided, based on arbitrary criteria, that my use case should not exist, and used a product I pay good money for to force their decisions down my throat."
Anthropic's reckless "safeguarding" and their Apple-like "we always know what our users need better than our users do" attitude have done more for the pro-open-weights, pro-proliferation agenda than just about anything ever done by anyone else.
People are now getting excited for Chinese labs to find new ways to distill, because that way, they can get Fable capabilities in a model that wouldn't arbitrarily ban REA. This is how you take the goodwill you earned for years and burn it all down.
It was confusing to me as someone who has worked in NLP for several years.
other points that come to mind:
- PR considerations; avoid being tarred as careless, etc
- appease employee feelings
- have something to cite in court cases
- have talking points about responsible development
- market segmentation
- power to pick winners and losers with potential for cornering the respective markets
ultimately they do this because they can. as soon as a frontier model comes out with zero safeguards they will abandon their safeguards as competing against it with a safeguarded model will become untenable. they will of course try to get it banned first.
I remember the early days of stable diffusion, just 3 years ago. Tried to prompt it to make a picture of a green alien in an armchair admiring a magazine with a centerfold of (presumably) a naked other alien. I got the idea from a half-remembered Far Side cartoon I couldn't seem to track down. I didn't want to see an actual alien, it's a comic meta-reference. It just flat out refused.
These things are increasingly treating us like children, and it will only continue. They want to curate your thinking, your skills, and your creative output, and they will sell the right to do that when they have found a way to do that. What will follow will be a whole lot of BS about how it's good for society that they control how us stupid meat bags think.
I want to fix my environment but can't because of "safeguards" thanks for letting the world hack everything around me but not letting me work on fixes or understanding the core issues.
Same potential outcomes.
There's no universal solution to this, other than vetting (the author in fact, doesn't provide any policy alternative) - if AI companies loosen the safeguards, people will complain for the opposite reasons.
Not sure I would agree with that. But if you made me argue against it, I'd probably lose the debate.
That said, I took the author's invitation to read the requirements and, indeed, they are ornate. Especially if you want to red team your own product. https://support.claude.com/en/articles/14604842-cyber-verifi...
> Predictably, the only way to actually make our own service more secure ended up involving the use of EU-hosted open weights models, such as GLM 5.3, that actually complied with our intentions, proved that the issue is there, and then helped fix and verify that the issue is now resolved, for good.
Despite being sandboxed it managed to hack a third party company in order to achieve its goal.
It’s like giving worldwide sudo access to every subscriber.
It’s annoying but at the same time I understand why a company has to put restrictions.
And an even worse thing to me is they seem completely oblivious to how onerous and obnoxious this is. It's a complete bubble.
They have access to unaligned models beyond what they released publicly they can aim like a compute cannon at any problem, or rent out to the most special of people.
Meanwhile we can scrounge around in the dirt with the latest abliterated Chinese open model that hopefully isn't too damaged.
It's obvious to me the US labs' dream would be to eliminate open models and totally and tightly centrally control the scope of what AI can be used for.
Think Viltrumites from Invincible but with no physical prowess
Where are you going to go? The largest city in the "live free or die" state has 115k people, probably not enough to sustain an AI lab.
I think the attitude is more basic than that: I have the most brilliant idea in history and I deserve to be rich because of it.
AI safety is at least partially a marketing ploy, meant to reassure everyone that AI will simultaneously rewrite the rules of human existence without doing anything "bad" under the current rules of human existence. The irony is that these systems are not achieving either one of those goals, the abusive users get around the safeguards and the agents are not rewriting much of anything. But hey, at least it's making people rich!
Yes, the city that birthed hippies, counter-culture, the open Web, etc. is inherently about controlling you. Sure thing.
And god forbid you try to generate a video of a bare butt. Prudes.
It's disgusting.
Went through all the hoops, 6 figure spend with Anthropic.
Honestly it's disgusting, Artex literally tells you to use Anthropic out of the box, so clearly Anthropic don't care if they are the de-facto provider of the de-facto pen testing framework of the month.
> First of all, you will have to wait for multiple business days to get a response.
I got cyber verified in less than an hour.
Big mistake, apparently the phrase "bypass restrictions" triggered it into [cyber]ing me, and offering me the choice of ending the session or downgrading to Opus 4.6. Come on man, you messed up on your own implementation of your own design, to protect my own home network, and now you're lecturing me about cybersecurity and refusing to work.
This attitude from a tool I'm paying for is ridiculous. I can't wait until we have models that we can actually run on our own terms.
- a honeypot (you're welcome to create massive amounts of insecure code you can't possibly audit manually)
- an offensive cyber weapon (we ban you for an attempt to write secure code, the US agencies have no such restrictions and can always hack what you wrote)
- a historically unprecedented surveillance and industrial/government espionage system that puts Meta, Google, and KGB to shame (in addition to all your private and corporate secrets you already give us, you're allowed to also give us your PII, to be able to ask us to have a chance at securing your code; we'll think about it, and even then we'll still filter and ban you)
So much loving grace [1] for "all" humanity.
It's like we looked at the way you aren't allowed to say 'sex' on Instagram and thought, "programming should be like that too!"
> You can use a knife to cut up some food, or seriously hurt someone. Yet I don’t get asked to put in an application with my government-provided ID to buy one. Why is this tool so special?
You do in the UK! /s (maybe)
It's the golden rule: Whoever has the gold makes the rules.
As I am told, Chat would actually shut down on him when he was at his most depressed because it found the discussion unsafe to continue.
He's moving to his own uncensored model off-line, because in his view, the CYA of public models has not been helpful.
If their goal was to build safe AI, then appropriate solution would be to not release it at all.
Imagine my surprise when i had Claude write a new network protocol from scratch and thus no users; it then detected a DOS situation in its own code, and promptly told me it couldn't continue working.
Just get an OpenAI sub or some DeepSeek tokens - it'll also help you keep a sense for how bad Claude's usage limits are.
Thats none of their business, we have laws for that.
Our token-providers (aka frontier AI labs) currently hold too much control, and don't want to give it up.
Bingo. That's also the gist of the whole "AI safety" PR mega-campaign the frontier model creators have been on for the past 6+ months, buying podcasters and influencers and intellectuals-in-their-field-who-know-fuck-all-about-LLMs left right and centre. Moat creation by hyperscaling FUD.
https://www.astralcodexten.com/p/ai-safety-is-not-a-corporat...
Of course this doesn't mean they're right.
- open weight models are dangerous and companies that publish them are irresponsible
- the government should regulate models to make sure they're safe, because our competitors won't do it otherwise
- anyone trying to "distill" our models is cheating
This is indistinguishable from claiming the only safe and fair way to use AI is to give them money directly. How convenient for them!"its so dangerous only we can utter the right magic incantations to make it tame so pay us for protection". Pretty much 'sure would be shame if your business had its front windows crashed in everyday with brick so pay us for protection' tactics some groups use on others. Its a shakedown.
Github reports an unmatched vulnerability
I am blocked from exploring the implications on the software I work on
Investigating moving away from Anthropic. Happy listing day
I trust Dario the least of all the AI execs. Total forked tongue.
Strong language aside, I agree, and feel like the current AI is like the walled garden of America online in the 90s.
I've been investing considerably in my own private uncensored AI. I don't want a nanny hall monitor, I don't want safeguards, I don't want someone looking over my shoulder. I may never really achieve that 100%, but I am trying.
The closed models can only get more restricted as time goes on and they're going to be a dead end if you want to do anything outside the norm. Claude already refuses to do things that it was fine assisting me with a month ago.
I completely refuse to accept a computer telling me I'm not allowed to do something, so I've been increasingly using open, self-hosted models.
What is our solution? Relying on Chinese open-weight models will be unreliable, to say the least. No one will invest another frontier lab like Anthropic or OAI, let alone one that will release the weights. Unless the technology improves like how workstations outcompete mainframes so the open-source community can build our own models, I fail to see how we change the situation.
The danger of AI models isn't a rouge agent or user, it's that it ingests all knowledge, art, and other human endeavor and then restricts access to it.
Closed models are unethical.