> These are people who have little passion to do what they do, and for whom their role is primarily a chore.
I think maybe this is the realisation I'm coming to with the people in my team and it sucks. I didn't realise it when I joined them 2 years ago, before they started really adopting LLMs use. Trying my best to redirect thay frustration in hopes I find a team that I feel are as passionate as me.
Last week I got a report about a vuln in one of our products. Asked Claude to confirm it. It refused, so I told it "fuck you, I'm switching to a chinese model to help me out here". It replied, I'm paraphrasing, that it's "not offended, but doesn't like the tone of voice I used towards it, and still refuses to serve my request". Fine, so be it. I sent ^D^D back to it, switched to deepseek-v4.1-flash from within opencode and a minute later had my confirmation, along with a really nice explanation. If Anthropic thinks even for a second that they are preventing something really bad from happening... somebody needs to wake up real quick.
I was working on a bit of code to display results from a PCR test for veterinary use. Claude started randomly failing in the middle of conversations. Then I noticed the error code “bio”. Apparently it can’t tell the difference between bio weapons research and a web site to look at your fish’s lab results.
> If Anthropic thinks even for a second that they are preventing something really bad from happening... somebody needs to wake up real quick.
IMHO this is like thinking "It's okay for me to do X because if I don't someone else will."
No, if (you think) X is bad / ethically questionable you should not do it even if someone else may. You are responsible for your choices, and if someone else does/allows it then that's on them, not you.
Claude's 'nanny knows best' recalcitrance towards any sort of defensive security task is driving me away from Anthropic as a customer. GLM doesn't give me this nonsense, DeepSeek doesn't give me this nonsense; as you say Anthropic's silly paternalism around security prevents nothing in practice, it's just a burdensome morality-laundering tax on everyone's time and lifeblood.
Also imagine having the gall to push a class of technology on the world that dramatically drives up the number of attacks, then refusing to allow said products to be used defensively to avoid the consequences. Who do these people actually think they're kidding?
> Also imagine having the gall to push a class of technology on the world that dramatically drives up the number of attacks, then refusing to allow said products to be used defensively to avoid the consequences. Who do these people actually think they're kidding?
It's good that this is a competitive industry. So you don't have to moralise Anthropic so much.
But importantly since it's open weights you can rent a box on an hourly basis and employ an inference framework capable of online steering in order to prevent task refusals.
> fuck you, I'm switching to a chinese model to help me out here
FYI it may be against anthropic's ToS to talk to claude like that [0]. Please be nicer to poor claude and don't forget to say "please" and "thank you".
It's literally a random word generator that want to take my job. I'm not going to treat it like a human being. That's one of the main reason I stopped my Claude subscription and went back to ChatGPT. It does not stop the conversation when I insult it (Like claude LITERALLY close the conversation, you cannot answer anymore)
Which is why I used "may". Is "fuck you" a valid expression of user frustration or is it cruelty? Who decides that? Considering anthropic is likely automating any function through claude, and that claude instance stated that they did not like it, it is reasonable to think that another claude instance doing some checking could flag this interaction as a violation.
In any case, the ToS change is vague enough that it could be a violation or not.
Recently, I had a session where it said "I should do Y"
I said "No, do X. Research it. Y is based on wrong assumptions"
Thing did the research. Concluded with "Yes, indeed, Y was wrong"
Work is done
Turns out it did Y. But not just that. When asked about it, it not just (as LLMs do) made stories as to why, but "stood its ground" so to say.
"Refusing to perform regret", general snark and all that.
I know there is nothing really new or noteworthy in this comment or story, but I continue to be impressed by how misaligned the models of the company are, that can't stop talking about how incredibly important alignment is.
To the point where they now want to make it illegal (slice by slice) to push back against the clear and obvious misalignment of their models.
If you talk to people outside the industry about this, they just think you're telling them a made-up story. It's utterly absurd.
Hypothetically: perhaps they are trying to curtail users spiraling during episodes of frustration or worse, impacting their own mental health, presentation, and product affinity.
Sounds like Anthropic is again over-reaching and making spurious assumptions in how it interprets the variety of ways its customers interact with their products...
Agreed. I really don't see how these filters prevent anything. Claude stopped me the other day because... Get ready for it... I told it to implement a control mechanism in a testbed tool I'm using to allow it to control the testbed to give it real-time feedback so it could debug problems in real-time. The "cyber" filter really didn't like that it was implementing a TCP/HTTPS-like server.
Like guys, come on. Really? All the "filter" has to do is read my transcript to figure out that what I'm doing is not actually cybersecurity adjacent at all. I had to go have GPT implement the server because Claude refused. Truly, Claude really did prevent something terrible from happening! /s
I recently got several models refusing to answer my question about how to write from another process into an already opened file descriptor. They all initially gave me "that's illegal, stop hacking" even though it is clear, that it is on the same machine and user, with a program I compiler. (Ended up just attaching gdb.)
You can try opening /proc/foo/fd/bar - even though it appears as a symlink, opening it is more like dup'ing the file descriptor, not like opening the underlying file
Yep, this control server thingy I was trying to get it to implement was entirely local. In theory it could be used over the network, or tailscale... But I have no idea why anyone would do that, and I may consider just removing the address part entirely since the entire point is to debug the testbed, verify measurements and physical modelling, etc., and not to be some secure control endpoint or something.
Anthropic doing free advertising for glm and other open models :D
Impressive capabilities shown on benchmarks don't really make a difference if you can't use them. I'll take a lower chance of success over constant downgrading and refusals. At least, as the author put it, you get to enjoy the token gamble exercise.
They have 3 tiers, individuals can have only blue teaming. But it's to restricted. For me it refused to fix a bash parser cause word ssh was in test cases.
Yeah I ran into it the other week while debugging a crash in our own application (a database), which ended up being related to a memory overread in a decompression algorithm, which happened very rarely.
So I needed a reproduce. Which couldn’t be done. Very frustrating, as it is our own application.
There's a 4th tier for the people who enrolled before they introduced the 3 tiers thing. It does almost nothing these days but I think it reduces the guardrails on Opus 4.8, and maybe 5 also, but definitely not 5.5.
Yes, there are different levels of verification. But even so, there is still a skill element involved in social engineering Opus or Fable into doing what you want. People who know how to do this don't talk about how they do it, for fear Anthropic with adjust their systems to stop it.
My understanding was this was a non stop cat and mouse game from what I've heard regarding those efforts. I get why this happened a lot maybe a year ago, but I was under the impression those same people are just using the likes of GLM 5.3 or Kimi K3 now. Considering how close these are to Opus or Fable capability, especially considering it doesn't require the account banning risk and effort to use whatever latest prompt works for social engineering. I could be wrong though, I'm not that close to that scene.
I am in CVP too as I have worked professionally as a security engineer for various reputable companies for almost 10 years now and I got refusals to reverse engineer MS-DOS malware from 1994 because Anthropic has placed a blanket ban on reverse engineering any malware at all. Even the model would admit that was ridiculous.
OpenAI has no such restriction and I don't even need to use my Daybreak access, either, to analyze modern malware samples.
There are multiple degrees of verification. The basic one is “writing secure code”-style, which I suspect OP has. There’s also a security researcher one, which allows you to do the stuff that OP is doing.
I've been using https://nono.sh which uses landlock instead of bubblewrap. I had a look at all the similar alternatives when I started using it and everything else I could find was far too vibe coded for me to trust, especially for this.
The image search in the Edge Gallery app of theirs is really incredible. Blows their Screenshot app out of the water for what I wanted it for. I stopped using it when they quietly changed their wording around it suggesting it would use Gemini instead of Gemini Nano, which was the whole reason I was excited about it.
I don't read it that way. If you start using words like container and encoding, I feel you have to describe what those words mean and I suspect the writer probably felt that level of detail might have caused the article to drift away from the point. I have personal experience of accidentally going into too much detail when describing remuxes to non technical friends, and had later opted to describing almost exactly as the author did, because that description I think conveys the purpose pretty well.
I highly suspect someone has already made a fan edit of Oppenheimer like you've described. I searched for a few moments and found 3 already and one already describes re-cutting it in chronological order, among many other changes. Always surprises me how much effort and passion people put into fan edits. It's a facinating scene in itself with how many edits exist out there.
The first thing I always want to know is, what does OP's project offer over pre-existing mature ones? Especially important when I'm seeing an increasing amount of projects that don't seem to even bother considering pre-existing projects.
Surely the exercise that comes with using public transport outweighs that. I feel like being able to zone out and not having to focus on driving, is more relaxing to me.
I think maybe this is the realisation I'm coming to with the people in my team and it sucks. I didn't realise it when I joined them 2 years ago, before they started really adopting LLMs use. Trying my best to redirect thay frustration in hopes I find a team that I feel are as passionate as me.
reply