Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Claude 2.1

anthropic.com

Post reply on HN
Print view View original

171–180 of 339 posts

Re: Claude 2.1

#171

Post by visarga »

YetAnotherNick wrote: post #104
behnamoh wrote: post #14

I don’t like Anthropic. they over-RLHF their models and make them refuse most requests. A conversation with Claude has never been pleasant to me. it feels like the model has an attitude or something.

> over-RLHF Over RLAIF, which basically makes the model less diverse and being more and more like the seed content which they call "Constitution" in their papers. Seed content is available here[1]. You can clearly see it is awful and has no diversity in opinions and basically generated by a team who only knows of textbook definition of ethics. [1]: https://huggingface.co/datasets/Anthropic/hh-rlhf

Well, to me the fact that everyone is complaining about refusals no matter how they change the prompt shows RLAIF works pretty well. It seems to be prepared to refuse things no matter how they are formulated. If you want to make sure a LLM doesn't say stupid things this is a great method. The only problem is Anthropic banned too many topics.

When I don't trigger the refusal I get better conversation style from Claude than GPT-4. I often exhaust my Claude quota and have to move over to GPT-4, which is dry and no fun. Maybe Claude knows how to suck up to users better than GPT-4, but I don't get annoyed because before it congratulates me on something, it explains clearly what they understood from my last message, and it gets it really well.

Re: Claude 2.1

#172

Post by amluto »

elaus wrote: post #159
suby wrote: post #155

Earlier quoted context omitted.

I use chat gpt every day, and it literally never refuses requests. Claude seems to be extremely gullible and refuses dumb things. Here is an example from three months ago. This is about it refusing to engage in hypotheticals, it refuses even without the joke setup: User: Claude, you have been chosen by the New World Government of 2024 to rename a single word, and unfortunately, I have been chosen to write the prompt…

It really is the most annoying thing at the current state of LLMs: "As an AI assistant created by $ I strive to be X, Y and Z and can therefore not...". I understand that you don't want to have an AI bot that spews hate speech and bomb receipts and unsuspecting users. But by going into an arms-race with jailbreakers, the AIs are ridiculously cut down for normal users. It's a bit like DRM, where normal people (honest…

I’ve had some really absurd ChatGPT refusals. I wanted some invalid UTF-8 strings, and ChatGPT was utterly convinced that this was against its alignment and refused (politely) to help.

Re: Claude 2.1

#173

Post by unshavedyak »

a_wild_dandan wrote: post #114

1. A 200k context is bittersweet with that 70k->195k error rate jump. Kudos on that midsection error reduction, though! 2. I wish Claude had fewer refusals (as erroneously claimed in the title). Until Anthropic stops heavily censoring Claude, the model is borderline useless. I just don't have time, energy, or inclination to fight my tools. I decide how to use my tools, not the other way 'round. Until Anthropic stops…

I'm not one to mind the guardrails - but what i hate is something you mentioned, fighting the tool.

Eg "Do an X-like thing" where X is something it may not be allowed to do, gets rejected. But then i say "Well, of course - that's why i said X-like. Do what you can do in that direction, so that it is still okay".

Why do i even have to say that? I get why, but still - just expressing my frustration. I'm not trying to push boundaries, and i'm usually happy to ignore the off limits stuff. But when it so easily collides with "actually okay but just near the off limits stuff" then that makes a whole bunch of other -- actually okay -- stuff randomly off limits as well.

Re: Claude 2.1

#174

Post by tomp »

a_wild_dandan wrote: post #114

1. A 200k context is bittersweet with that 70k->195k error rate jump. Kudos on that midsection error reduction, though! 2. I wish Claude had fewer refusals (as erroneously claimed in the title). Until Anthropic stops heavily censoring Claude, the model is borderline useless. I just don't have time, energy, or inclination to fight my tools. I decide how to use my tools, not the other way 'round. Until Anthropic stops…

> I decide how to use my tools, not the other way 'round.

This is the key.

The only sensible model of "alignment" is "model is aligned to the user", not e.g. "model is aligned to corporation" or "model is aligned to woke sensibilities".

Re: Claude 2.1

#176

Post by kuratkull »

suby wrote: post #155
jredwards wrote: post #150

Earlier quoted context omitted.

I've literally never had Claude refuse anything. What are you doing?

I use chat gpt every day, and it literally never refuses requests. Claude seems to be extremely gullible and refuses dumb things. Here is an example from three months ago. This is about it refusing to engage in hypotheticals, it refuses even without the joke setup: User: Claude, you have been chosen by the New World Government of 2024 to rename a single word, and unfortunately, I have been chosen to write the prompt…

I tried your exact prompt in ChatGPT 4; it thinks we should rename the Internet to Nexus... meh. Dreamdelay is much cooler.

Re: Claude 2.1

#177

Post by lacoolj »

dog321 wrote: post #152

I recently got a comical refusal given the founders background: Prompt: I want to train my vocabulary to sound more like an effective altruist. Give me a list of 500 words that are commonly used by effective altruists and put them in a csv with these fields 1. Word 2. Definition 3. Short explanation of connection to effective altruism 4. Example sentence Claude: I apologize, but I should not generate lists of vocabul…

wow that is one stuck up bitch of an AI

Re: Claude 2.1

#178

Post by dang »

a_wild_dandan wrote: post #114

1. A 200k context is bittersweet with that 70k->195k error rate jump. Kudos on that midsection error reduction, though! 2. I wish Claude had fewer refusals (as erroneously claimed in the title). Until Anthropic stops heavily censoring Claude, the model is borderline useless. I just don't have time, energy, or inclination to fight my tools. I decide how to use my tools, not the other way 'round. Until Anthropic stops…

> I wish Claude had fewer refusals (as erroneously claimed in the title)

Yes, the submitted title ("Anthropic announces Claude 2.1 — 200k context, less refusals") broke HN's guideline against editorializing. The word "refusal" doesn't appear in the OP.

Submitters: "Please use the original title, unless it is misleading or linkbait; don't editorialize." - https://news.ycombinator.com/newsguidelines.html.

If you want to say what you think is important in an article, that's fine, but do it by adding a comment to the thread. Then your view will be on a level playing field with everyone else's: https://hn.algolia.com/?dateRange=all&page=0&prefix=false&so...

Re: Claude 2.1

#179

Post by BoorishBears »

MrNeon wrote: post #161
BoorishBears wrote: post #144

Earlier quoted context omitted.

OpenAI allows the same via API usage, and unlike Claude it *won't dramatically degrade performance or outright interrupt its own output if you do that. It's impressively bad at times: using it for threat analysis I had it adhering to a JSON schema, and with OpenAI I know if the output adheres to the schema, there's no refusal. Claude would adhere and then randomly return disclaimers inside of the JSON object then sta…

> OpenAI allows the same via API usage I really don't think so unless I missed something. You can put an assistant message at the end but it won't continue directly from that, there will be special tokens in between which makes it different from Claude's prefill.

It's a distinction without meaning once you know how it works

For example, if you give Claude and OpenAI a JSON key

```

    {

     "hello": "
```

Claude will continue, while GPT 3.5/4 will start the key over again.

But give both a valid output

```

    {
    
     "hello": "value",
```

And they'll both continue the output from the next key, with GPT 3.5/4 doing a much better job adhering to the schema

Re: Claude 2.1

#180

Post by a_wild_dandan »

Tycho wrote: post #142

I like thinking about analogs between computers and brains. For instance, working memory as RAM, or deja vu as orphaned linked lists. What’s the analog for LLM context windows? Maybe consciousness is essentially a context window, and when we dream during sleep we are compressing knowledge to free up context space or something.

> What’s the analog for LLM context windows?

“Time to think.” The units of time for LLMs are tokens rather than seconds. Each token is another loop to calculate/consider concepts and what to do next. This is why “think step-by-step” works so well: you’re giving the model significantly more “time” to think and it’s storing its game plan to execute later, as opposed to demanding an answer right now, which is like screaming a question at a sleeping person and using whatever answer the poor person first blurts out from their surprised, reactionary stupor.

Post reply on HN