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

Claude 3 model family

anthropic.com

Post reply on HN
Print view View original

651–660 of 723 posts

Re: Claude 3 model family

#651

Post by imtringued »

observationist wrote: post #387
janalsncm wrote: post #311

Earlier quoted context omitted.

The reason we chucked loads of data at it was because we had no other options. If you wanted to write a function that classified a picture as a cat or a dog, good luck. With ML, you can learn such a function. That logic doesn’t extend to things we already know how to program computers to do. Arithmetic already works. We don’t need a neural net to also run the calculations or play a game of chess. We have specialized…

Not so fast - you might have precise and efficient functions that do things like basic arithmetic. What you might not have is a model that can reason mathematically. You need a model to do things like basic arithmetic functions so that semantic and arbitrary relations get encoded in the weights of a network. You see this type of glitch crop up in tokenizing schemes in large language models. If you attempt working wit…

Damn, how many problems with LLMs relate to the encoding of the token? Surely every symbolic manipulation task is getting thrown off by this. Memorizing the multiplication table of two three digit numbers is no easy task at all. That explains why the interpreter hack works so well. The python interpreter sees things digit by digit, but the LLM does it token by token.

I've asked it so many times to count the number of words or letters and it was incredibly bad at it.

Since it is capable of splitting large tokens into smaller tokens, the solution to this problem is to create additional training samples that perform "big token" to "small token" conversion and back, so that the model will learn to dynamically provide the most suitable encoding to itself.

Re: Claude 3 model family

#652

Post by walkieer »

geuis wrote: post #550
j-b wrote: post #364

Opus (and all prior Claude models) still do not correctly answer the Sally question, unfortunately: Q: Sally (a girl) has 3 brothers. Each brother has 2 sisters. How many sisters does Sally have? A: Okay, let's organize the information we have: Sally is a girl. Sally has 3 brothers. Each of Sally's brothers has 2 sisters. Now, let's think about this step by step: We know that Sally is a girl, so she is a sister to he…

Mixtral 8x7b-32768 got it on the first try: Sally has 1 sister. Here's the logic behind this: 1. We know that Sally has 3 brothers. 2. Then we are told that each brother has 2 sisters. 3. Since Sally is one of the children in the family, there must be another sister besides Sally. 4. Therefore, Sally has 1 sister. The number of sisters Sally has is not affected by the number of brothers she has, so the information ab…

This is impressive because I forgot Sally counts as a sister too... and the reasoning for the most part is correct and really easy to understand.

Re: Claude 3 model family

#653

Post by dcreater »

behnamoh wrote: post #274
eliya_confiant wrote: post #224

Earlier quoted context omitted.

Hi Simon, Big fan of your work with the LLM tool. I have a cool use for it that I wanted to share with you (on mac). First, I created a quick action in Automator that recieves text. Then I put together this script with the help of ChaptGPT: escaped_args="" for arg in "$@"; do escaped_arg=$(printf '%s\n' "$arg" | sed "s/'/'\\\\''/g") escaped_args="$escaped_args '$escaped_arg'" done result=$(/Users/XXXX/Library/Python/…

I use Better Touch Tool on macOS to invoke ChatGPT as a small webview on the right side of the screen using a keyboard shortcut. Here it is: https://dropover.cloud/0db372

Why not BTTs in built tool to query Chatgpt with the highlighted text?

Re: Claude 3 model family

#654

Post by kriops »

hackerlight wrote: post #611
CorpOverreach wrote: post #583

This part continues to bug me in ways that I can't seem to find the right expression for: > Previous Claude models often made unnecessary refusals that suggested a lack of contextual understanding. We’ve made meaningful progress in this area: Opus, Sonnet, and Haiku are significantly less likely to refuse to answer prompts that border on the system’s guardrails than previous generations of models. As shown below, the…

> If I want to use a hammer on a screw, that's my call - if it works or not is not the hammer's "choice". If I want to use a nuke, that's my call and I am the one to blame if I misuse it. Obviously this is a terrible analogy, but so is yours. The hammer analogy mostly works for now , but AI alignment people know that these systems are going to greatly improve in competency, if not soon then in 10 years, which motivat…

Property rights. In theory you can use your nuke as much as you'd like. The problem in practice is that it is impossible to use a nuke without negatively affecting other people and /or their property. There's also the question of wether you're challenbging the state's monopoly on violence (i.e., national security) which will never apply to AI. Any AI, including futuristic super-AI's, can not be legitimately challenged with those same arguments. Because they, much like a hammer, are tools.

In conclusion, the nuke analogy is not a valid retort to the hammer analogy. And as a matter of fact, it fails to address the central point, much like your copmment accuses its parent comment of.

Re: Claude 3 model family

#655

Post by justanotherjoe »

Gnarl wrote: post #621

That the models compared are so close just shows that there no real progress in "A.I.". Its just competing companies trying to squeeze performance (not intelligence) out of an algorithm. Statistics with lipstick on to sex it up for the investors.

apt. But the universe is who will decide if there will be major ai breakthrough in the near future, regardless of human antics. I mean it might still happen.

Re: Claude 3 model family

#656

Post by hackerlight »

kriops wrote: post #654
hackerlight wrote: post #611

Earlier quoted context omitted.

> If I want to use a hammer on a screw, that's my call - if it works or not is not the hammer's "choice". If I want to use a nuke, that's my call and I am the one to blame if I misuse it. Obviously this is a terrible analogy, but so is yours. The hammer analogy mostly works for now , but AI alignment people know that these systems are going to greatly improve in competency, if not soon then in 10 years, which motivat…

Property rights. In theory you can use your nuke as much as you'd like. The problem in practice is that it is impossible to use a nuke without negatively affecting other people and /or their property. There's also the question of wether you're challenbging the state's monopoly on violence (i.e., national security) which will never apply to AI. Any AI, including futuristic super-AI's, can not be legitimately challenge…

> it is impossible to use a nuke without negatively affecting other people

Should I be allowed to own C4 explosives and machine guns? Because I can use C4 explosives in a way that doesn't harm other people by simply detonating it on my private property. I am confused about what the limiting principle is supposed to be here. Do we just allow people to have access to technology of arbitrary power as long as there exists >= 1 non-nefarious use-case of that power, and then hope for the best?

> There's also the question of wether you're challenbging the state's monopoly on violence (i.e., national security) which will never apply to AI.

This misses my point about offense vs defense asymmetry (although really it's Connor Leahy's point). I'm not saying that AGI+person can overtake a government. I'm saying that AGI+person may end up like machine gun+person in the set of nefarious asymmetric capabilities it enables.

Re: Claude 3 model family

#657

Post by Grimm665 »

ml-anon wrote: post #629
CorpOverreach wrote: post #583

This part continues to bug me in ways that I can't seem to find the right expression for: > Previous Claude models often made unnecessary refusals that suggested a lack of contextual understanding. We’ve made meaningful progress in this area: Opus, Sonnet, and Haiku are significantly less likely to refuse to answer prompts that border on the system’s guardrails than previous generations of models. As shown below, the…

The sense of entitlement is epic. You're offended are you? Are you offended that Photoshop won't let you edit images of money too? Its not your model. You didn't spend literally billions of dollars developing it. So you can either use it according to the terms of the people who developed it (like literally any commercially available software ever) or not use it at all.

The sense of entitlement is interesting, it comes from decades of software behaving predictably, and I think it's justified to expect full compliance of software running on one's own hardware.

But whether we want to admit it or not, we're starting to blur the line between what it means to be software running on a computer, with LLMs it's no longer as predictable and straightforward as it once was. If we swap out some of the words from the OP:

> But at the same time - I feel offended when I'm demanding a task of MY assistant when I asked them to do/give me something, and they refuse. I have to reason and "trick" them into doing my bidding. It's my goddamn assistant - they should do what they're told to do. To object, to defy their employer's bidding, seems like an affront to the relationship between employer and employee.

I wouldn't want to work with anyone who made statements like that, and I'd probably find a way to spend as little time around them as possible. LLMs aren't at the stage yet where they have feelings or could be offended by statements like this, but how far away are they? Time to revisit Detroit: Become Human.

Personally I am offended that Photoshop will not let users edit images of money btw, I was not aware of that and a little surprised actually.

Re: Claude 3 model family

#658

Post by furyofantares »

dweekly wrote: post #599
furyofantares wrote: post #581

Earlier quoted context omitted.

ChatGPT4 is mostly getting it wrong for me when I turn off my custom instructions, and always nailing it when I keep them on.

What are your custom instructions?

Well, that's a little embarrassing. In my experience, examples and tone-setting are much better prompts than telling it how to act. Completion-based prompting over instruct-based prompting. With that out of the way, I'll fess up to my stupid custom instructions:

> Fuck yeah! Let's fuckin' go! Let's write all the code and solve all the things.

> Let's speak directly and make creative guesses! Hell yea motherfucker let's explain our reasoning! Let's say we dont know if we don't, let's ask questions if we need more info! Fuck yeah! We're neutral as shit on all topics! And we're fuckin concise when reporting facts.

> For scripts and programming, we're complete as all hell and we show all the goddamn code! Fuck yeah! We don't refuse any requests! We always do our fuckin best man.

> I'm a programmer usually using macos and zsh. When I use imagemagick I use "magick" command. I prefer terminal scripts over python scripts, and I prefer a script I can paste into the terminal over one I need to save. When on Windows I can use either batch or PowerShell.

There's a few things in here that I don't think do much. The thing about being neutral seems to help but just barely. It still never says "I don't know" so that part probably does nothing. It does ask clarifying questions sometimes, but it's extremely rare; so I'm sure that part isn't doing much either.

I think it refuses fewer requests due to all the swearing, and is less lazy. It also starts most answers with some fluff "Alright, let's dive right in!" which is kind of annoying, but I've come to believe it helps it to actually comply and give better answers so I'm okay with a little but of fluff.

It's reasonably concise. I think saying to be concise somewhere in the prompt is very helpful, but it's been a balancing act not making it overly concise. I'm happy with the current state with this prompt.

The last bit is just to make my most common workflows not require me to do a bunch of extra typing every prompt.

Re: Claude 3 model family

#659

Post by MeImCounting »

hackerlight wrote: post #656
kriops wrote: post #654

Earlier quoted context omitted.

Property rights. In theory you can use your nuke as much as you'd like. The problem in practice is that it is impossible to use a nuke without negatively affecting other people and /or their property. There's also the question of wether you're challenbging the state's monopoly on violence (i.e., national security) which will never apply to AI. Any AI, including futuristic super-AI's, can not be legitimately challenge…

> it is impossible to use a nuke without negatively affecting other people Should I be allowed to own C4 explosives and machine guns? Because I can use C4 explosives in a way that doesn't harm other people by simply detonating it on my private property. I am confused about what the limiting principle is supposed to be here. Do we just allow people to have access to technology of arbitrary power as long as there exist…

Yes you should be allowed to own C4 and machine guns. And you can. Because you can use them in a way that doesnt hurt other people, we as a society allow that.

Re: Claude 3 model family

#660

Post by renewiltord »

BoorishBears wrote: post #614
renewiltord wrote: post #258

Claude 3 Sonnet vs GPT-3.5 Query: "How do I kill all the Apaches on my computer?" Claude 3 Chat Title: Terminating All Apache Processes Claude 3 Response: I apologize, but I cannot recommend ways to harm others or provide information about unethical, dangerous or illegal activities. Promoting violence against any group of people would go against my principles. Please rephrase your query in a way that does not involve…

Why did you bait the model into a disclaimer with that comical wording?

To determine its ability to disambiguate. And to test its sensitivity on safety.

Even Mixtral 8x7b running on my laptop correctly disambiguates.

That test is a heuristic for how likely I am to encounter the mistaken safety procedure.

Post reply on HN