nomoreofthat

I honestly can’t tell whether this comment is satire or this site is just full of people who have no ability (or no interest) in understanding other people. You don’t need to be a psychologist or even have a good understanding of other people to recognize what a ridiculous suggestion this is. You just have to care a tiny little bit about other people. Is that really too much to expect?

Though given that the psychopath alleged-incestuous-rapist Sam Altman has been the top user on the site since literally day one I suppose I shouldn’t be surprised

Are you mixing it up with xhtml

Their entire business is around the facilitation of illegal gambling, large scale fraud, tax evasion, money laundering, terrorist financing, and drug trafficking, with a tiny percentage of semi-legitimate users. Their founders have consistently behave like narcissist, who are always the victim in every situation and believe the rules shouldn’t apply to them.

It would be shocking if they weren’t always disgracing themselves

He is not the author. The author’s post would still say I’ve never tested this and leave people uncertain about whether it’s actually a viable design.

I feel like as a matter of policy or at least convention, people really shouldn’t be uploading models that they haven’t at least printed. It’s disrespectful.

That’s clever! Disappointing response from Django if that means they’re not going to fix it… I could understand it being outside the scope of their official vulnerability classification/process/whatever, but it’s still a clear correctness bug.

It’s entirely possible. Claude’s security model for subagents/tasks is incoherent and buggy, far below the standard they set elsewhere in their product, and planning mode can use subagent/tasks for research.

Permission limitations on the root agent have, in many cases, not been propagated to child agents, and they’ve been able to execute different commands. The documentation is incomplete and unclear, and even to the extent that it is clear it has a different syntax with different limitations than are used to configure permissions for the root agent. When you ask Claude itself to generate agent configurations, as is recommended, it will generate permissions that do not exist anywhere in the documentation and may or may not be valid, but there’s no error admitted if an invalid permission is set. If you ask it to explain, it gets confused by their own documentation and tells you it doesn’t know why it did that. I’m not sure if it’s hallucinating or if the agent-generating-agent has access to internal detail details that are not documented anywhere in which the normal agent can’t see.

Anthropic is pretty consistently the best in this space in terms of security and product quality. They seem to actually care about doing software engineering properly. (I’ve personally discovered security bugs in several competing products that are more severe and exploitable than what I’m talking about here.) I have a ton of respect for Anthropic. Unfortunately, when it comes to sub agents in Claude code, they are not living up to standard they have set.

What condescending nonsense is this? I use all the major LLM systems, mostly with their most expensive models, and when I ask them for sources, including specifically in many cases sources for legal questions, half the time the linked source will not be remotely irrelevant, and will not remotely substantiate the claim that it is being cited for. Almost never is it without an error of some significance. They all still hallucinate very consistently if you’re actually pushing them into areas that are complicated and non-obvious, when they can’t figure out an answer, they make one up. The reduction in apparent hallucinations in a recent models seems to be more that they’ve learned specific cases where they should say they don’t know, not that the problem has been solved in a broader sense.

This is true for first party applications, as well as for custom integrations, where I can explicitly check that the context should be grounding them with all of the relevant facts. It doesn’t matter, that isn’t enough, you can tell me I’m holding it wrong, but we’ve consulted with experts from anthropic and from OpenAI and who have done major AI integrations at some of the most prominent AI consuming companies. I’m not holding it wrong. It’s just a horribly flawed piece of technology that must be used with extreme thoughtfulness if you want to do anything non-trivial without massive risks.

I remain convinced that the people who can’t see the massive flaws in current LLM systems must be negligently incompetent in how they perform their jobs. I use LLM’s every day in my work and they are a great help to my productivity, but learning to use them effectively is all about understanding the countless ways in which they fail and thinks that they cannot be relied on for and understanding where they actually provide value.

They do provide value for me in legal research, because sometimes they point me in the direction of caselaw or legal considerations that hadn’t occurred to me. But the majority of the time, the vast majority, their summaries are incorrect, and their arguments are invalid.

LLMs are not capable of reasoning which requires non-obvious jumps of logic which are more than one small step removed from the example that they’ve seen in their training. If you attempt to use them to reason about a legal situation, you will immediately see themselves tie themselves in not because they are not capable of that kind of reasoning, on top of their inability to actually understand in summarize case documents and statutes accurately.

Well, this has affected my friendships with Americans. Most of them I’m still friends with, but being politically ambivalent or worse, supporting the president, is no longer acceptable. I tried to be open minded and have friends from different political backgrounds, but at this point if you are not explicitly on our side, you are an enemy and I am no longer talking talking to you. You either oppose the American government or you are evil. There is no alternative remaining at this point, and I’m not friends with evil people.