When Claude 3.5 Sonnet Became The Better Chatbot

[@bycloudAI]

Scale your business easily now with AI/ML API to deploy any models you need! aimlapi.com/?ref=bycloud and Claude 3.5 Opus is about to be glorious

[@VigilVanDijk666]

Saw someone get it to make a 3D fps shooter, it absolutely blew my mind

[@springheeledjackofthegurdi2117]

Do you have a link?

[@Dedjkeorrn42]

I had it make a voxel engine with raytracing.

[@jaredf6205]

The only thing it needs next is to be able to have longer code and to be able to automatically read and fix bugs and errors.

[@JonDoe-uq1mk]

Probably ripped off someone's code from GitHub

[@Always.Smarter]

How?

[@yashizuko]

I am so happy I switched from chatgpt to Claude, like the amount of fucking stress I had generating and correcting prompts on chargpt was soo bad, Claude is a breeze working with, absolutely a pleasure, the same feeling I had when chatgpt 4 was actually smart and not lobotomised, and projects are amazing to manage shit codebases

[@M-dv1yj]

Claude is not much different once your level of complexity gets high enough both Claude and gpt respond the same. For mid intensity use Claude is better. But once really pushed it breaks down just the same as gpt.

[@jonathancarter5120]

@@M-dv1yj I agree with your assessment.

[@Omega-mr1jg]

I worked with claude, the prompting is easy when i switched to chatgpt to test out some other things, it blew up on my face not even chat gpt 4.0 fixed it, claude was better since haiku came out

[@SahilP2648]

You should checkout out abacus AI. I am not rate limited when using it, and it also supports team related shared RAG like claude projects.

[@bigglyguy8429]

I've had the total opposite experience. I cancelled my Claude sub as it was stressful trying not to offend the prudish thing. It was like talking to Karen, the biggest prude in the office, who's best friends with Helen in HR. I'd just rather not deal with her in the first place, than keep tip-toeing around the model's sensitivities. Heard of so many people being randomly banned that I pro-actively banned myself and stuck with GPT and running my own models locally. Claude isn't fit for purpose, unless your name is Karen... Plus GPT has web access, remembers my details and can make pics, for the same price.

[@WillHarmon]

I asked ChatGPT to make an image of a website, and then passed that image to Claude and simply asked it to build the website. I was surprised at how close it got on its first attempt. Not perfect, by far, but impressive still.

[@carkawalakhatulistiwa]

2 ai working together

[@Vysair]

@@carkawalakhatulistiwa multi-modal be like

[@captaindryvids6909]

ChatGPT Pro feels like such a rip-off with Claude Sonnet 3.5 being around and still no voice mode being released.

[@Dedjkeorrn42]

Remove your subscription

[@xthesayuri5756]

cancel your subscription! let OpenAI feel that they cant get away with cucking us and GPT-5

[@ps3guy22]

Remove your subscription

[@ckq]

I think the GPTs and integration with Dalle is it's value proposition.

[@ytubeanon]

hmm, with ChatGPT I know you can do talk and reply on their mobile app, but it's not as advanced as the gpt-4o demos... the API allows some voices for TTS, I had it create an audiobook, but yeah still not Scarlett Johansson quality

[@middle-agedmacdonald2965]

I think this hop scotch from one being the best, to another, is all by design. I mean the NSA, Apple, and Microsoft seem to think Open AI is "the one". Good for Anthropic, but obviously this will be short lived. In my opinion, these other companies are letting their users use up all of their compute, while Open AI is still the bar to beat, but they aren't "wasting" their compute on the end user.....which means they're definitely using it for advancement. I know nothing, but this seems very logical to me. If they released Sora, it'd bog their compute down, for example.

[@meh2285]

I found that for visuals that Claude produces, often taking a screenshot of what it produced and uploading it so Claude can see what it's code made usually yields better results.

[@Rortti]

Last banana was close

[@Zeni-th.]

It was a triangle

[@poisonza]

It seems like Sonnet 3.5 is trained using dialogue generated through a ReAct-CoT and DFSDT-like workflow. For example, it almost always includes phrases like "let me explain step by step" in response to general requests or questions. I can get a somewhat similar result with ChatGPT, but it requires more specific prompting.

[@Leonhart_93]

Yes, it requires specific prompting. So they can give Claude a specific hidden system prompt, telling it to explain it to you step by step when applicable. So it probably has nothing to do with the training, the data is too massive to ever be specific. If you ever played with local LLMs, you can give them any system prompt making them behave in a very specific way. They can even impersonate characters like that.

[@poisonza]

@@Leonhart_93 the "step-by-step" behavior of Sonnet 3.5 is likely a result of a combination of factors, including its training data, architecture, and possibly a hidden system prompt (as you pointed out). It's a complex interplay that makes it difficult to isolate the exact cause.

[@Leonhart_93]

​@@poisonza The training data is immense. Likely TBs in size. You can filter the data, but there is no way to "arrange" the final model, the trainers don't have control over specifics, it's mostly a black box. The control part comes when they set up alignment layers over the training data and then the specific hidden system prompts. I know that very well from using models with the alignment completely removed, you can get it to be whatever you want it to be and easily change it with each chat.

[@poisonza]

@@Leonhart_93 what do you mean by alignment layer? alignment is done by RLHF or DPO. There is no specific layer for this can you point out reference papers?

[@Leonhart_93]

@@poisonza How do think they censor models like ChatGPT and Gemini and everything else public facing? They only trained those once in the past, there is no way they would spend millions of dollars and days to re-traing them for each small modification. Here come the alignment layers, basically system prompts that tell the model how to answer and what they should not answer. And when you remove those layers programmatically, you get something capable of answering anything. Because under those layers, the model is unchanged.

[@Seventhsu]

I'm still waiting for Claude to be able to display LaTeX natively and do math using it. That and the lack of a mobile app are the only things holding me back from switching

[@edenassos]

Lack of a mobile app? Have you been living in a cave?

[@bananabatsy3708]

@@edenassos The mobile app for Claude is out.

[@dennisnotamenace2047]

ClosedAI is coping like crazy

[@DBonacich]

FYI: As an avid user of both the Claude Pro subscription and the Claude API, I can confirm that you can use the API even if your Pro limits are reached. The subscription and the API use two completely different rate limiting and pricing mechanisms.

[@ytubeanon]

are the quality of responses similar from the API to Pro/Artifacts?

[@4ytherium]

Prepare to get your wallet fucked by using the API

[@DBonacich]

@@ytubeanon It depends on the way you use the API. For the most part, the model is the same. I use the API for things like open-devin and TypingMind.

[@DBonacich]

@@4ytherium The Claude 3.5 API cost is honestly pretty cheap. If you're a lite user, the costs would likely come out to be cheaper than the Pro subscription. But if you're a heavy user, or you're using the API for a program that makes a ton of calls, the costs definitely add up.

[@countofst.germain6417]

He was just talking about tos

[@Guedez1]

Available to everyone unless you are a Brazilian that is, which people assume it's due to our data privacy laws

[@RonyPlayer]

Download opera, my fellow Brazilian

[@Earthball_Productions]

​@@RonyPlayer1st Opera VPN isnt a real VPN, It sends your data to china, 2nd, you need a phone number.

[@seto007]

Can't comment specifically on Brazil, but since Claude is not available in Canada, I've found that using an American VPN and processing my payment while connected to that VPN, despite the card being Canadian, allowed me to use it just fine

[@gustavodutra3633]

Brazil mentioned 🇧🇷

[@johndank2209]

Brazil's got these super strict data privacy laws, but meanwhile there's a ton of brutal killings going down every day. Kinda makes you wonder if they've got their priorities straight.

[@SangoProductions213]

I've personally had literally 0 problems with especially the latest chatgpt models, but .. I'm going to try and use both for my new project and see which helps me most. Worst case scenario is spending a bit longer on the accelerated process that ai provides. Best case would be to find where both well and truly shine.

[@TheSmokeWatcher]

Results? I prefer Claude

[@doingtime20]

I've been using sonnet for my web dev job and it works great. I'm working double the speed because of it. Granted I don't do super complex programming, but still it's amazing and pretty accurate. I also find myself bouncing off ideas with Claude, as if he was a coworker. We discuss the pros and cons of different approaches and such.

[@OperationDarkside]

If Claude Sonnet 3.5 can write decent GLSL code, I might be able to continue my game project.

[@AbelShields]

I used Claude 3.5 to give coeing style feedback on a pwrsonal project and it made some really good suggestions for how to clean up the style and organise it in a more modular way, it was really impressive.

[@rockochamp]

and don't forget the new project feature, which is actually great and helped me gigantic hugely mega wooow with my master thesis...

[@ytubeanon]

how did it help, were you collaborating with other people?

[@reipuradinata8766]

I'm not sure Claude good enough yet. Here's what I asked: Me: someone said "Traps are full of androgen receptors", is that true? Claude: That statement is not accurate. "Trap" is a slang term sometimes used to refer to transgender women or crossdressers, which many consider offensive. Regardless of gender identity or expression.... Me: WTF, traps refer to trapezius muscle Claude: I apologize for the misunderstanding. You're absolutely correct, and I thank you for the clarification. In this context, "traps" indeed refers to the trapezius muscles. Let me address the statement with this correct interpretation: "Traps are full of androgen receptors" is generally true.... Maybe I'll try again in 6 months.

[@hannariel]

All chatbots are indocrinacted by the Qu33r theory. All of them was made with these ideologies in mind.

[@sofia.eris.bauhaus]

a square (typically with diagonal sides) is commonly called a "diamond" so i don't think it should be blamed for that. i think when generating code, you want to err on the side of small size that can later expanded upon, rather than go for a maximalist interpretation of the task at hand.

[@ShowmanSharma]

Need some suggestion here, please. Claude doesn't currently connect to the internet, right? So from a utility standpoint, don't you think ChatGPT 4o (pro access) is more useful than Claude 3.5 (pro access)? I have a master's thesis to work on and a SaaS product to build. Please help me make a decision.

[@countofst.germain6417]

ChatGPT is probably the better bet, Claude doesn't search the net, ChatGPT has far more features and you hit the limits with Claude far far easier

[@SandTiger42]

I kept waiting for the video to get to the roleplay portion because I read "chatbot". My bad. My ideal AI can not only help me with productivity tasks, but alo put me in a book adventure. I want it to create a novel where I get to direct the protagonist. I think the main use for RP AI now is for weird smut and girlfriends/boyfriends. I'm ok with that if it gets me to my choose your own adventure chatbot.

[@ytubeanon]

I heard of a technique to write a novel where you use A.I. to write one paragraph and then you write the next and alternate back and forth

[@cajampa]

Have you explored Koboldai and Silly Tavern?

[@SandTiger42]

@@cajampaKobold uses some pretty small and puny LLM's. There are much better LLM's out there to use. They just cost either a lot of money to host, or crazy hardware to selfhost. I do use ST, and I plug in API's from various LLM's depending the use. But we're still years away from having novel length and novel quality choose your own adventures.

[@cajampa]

@@SandTiger42 What are you talking about dude. The only limit is the hardware you have available. So where do you get this BS that Kobold is a fixed thing of small and puny LLM. It is just like most other open source implementations of llm hosting. You can use what ever llm you have available on what ever hardware you have available. And most of the time your money you can and want to spend on it is the limit.

[@SandTiger42]

@@cajampa I just checked out the github page for a tiny bit. They mainly only list 7b and 13b models. Not my fault they aren't very clear what their program does. *shrug*

[@AndyIsHereBoi]

I thiught this was a fireship video by the thumbnail

[@big_mac_love]

Wow, what a neat tool! 🤩😯 - - That would be something a bot would say, so I just say the following: I yet got not thaaat much experience with using Sonnet3.5 in my workflow (E2E test generation). The little I tested with it I would say that code quality is not better than what gpt4o returns, but at some point its hard to compare it by code quality since its a weird subjective metric... for that I used anthropic's API, but the webUI feature where it runs code automatically seems really cool tho... as long as it doesnt run malicious code for a reverse shell automatically or shit like that xD

[@dv_interval42]

I have never hit a message limit on Claude. I strictly follow the "new-thread" guidelines and non-repeating pasted content parts. I think they base their "message limits" more around tokens processed rather than pure messages like OpenAI does.

[@paulzhu7694]

In fact, after using GPT-4 and Claude 3.5, the biggest impression is that GPT-4 seems like a blockhead, while Claude 3.5 feels a level above GPT-4 in terms of comprehension abilities in various aspects. It feels more like a true AI

[@gamershd6225]

I think imma hold out till gpt5

[@lumiey]

bruh it took 2 prompts for claude to copy someone's shader code (1st prompt was kinda nice tho)

[@Laszer271]

For LLM to work with data (like in .csv format), it shouldn't look at the data directly, it should interact with the data through code. That's how openai does this and I would bet that's how it works for claude too. Because of that 50 or 5000 records shouldn't make the difference, it all comes down to df.head(), df.groupby() and df.sum() anyway.

[@micbab-vg2mu]

Claude 3.5 is amazing:)

[@trench01]

i cant get on it. also any updates on alpaca ?

[@TheMattaBase]

Projects are available for Pro users not just Teams.

[@CMak3r]

I found at least 5 repositories with 2D subway surfers game mechanics

[@ckq]

The limit is pretty fair (I think higher than GPT 4os limit)

[@blackpiller3777]

no its not

[@picklechill]

Oh, hey! Thanks for breaking this down.

[@Kutsushita_yukino]

the response sounds so robotic though but pretty smart

[@Matt-bp5vy]

First comment not by bot!

[@Kynatosh]

Still waiting on a good AI that can know my entire codebase and help me code. If anyone knows any, please do tell

[@alansmithee419]

Will do as soon as one exists in, say, 2025/6?

[@danacosta5237]

There are some already like codeium (free), they still have the same issues that they all do have with hallucinations. Codeium's autocomplete can speed you up sometimes i've found especially with boilerplate

[@Khari99]

I've been using Claude Opus and Sonnet for a while now. Sonnet is great but model size definitely still matters. Sonnet falls apart when it comes to long complex prompts. Opus on the other hand handles complexity better than any other model I've ever used. For most things Sonnet is fine though. I can't wait until Opus 3.5 drops

[@Dannnneh]

They let *everyone* use it? Even those in Europe like me?

[@gaggix7095]

Not having a problem using it in the EU

[@NighttimeJuneau]

@@gaggix7095And you just signed up like that, no VPN, with a European state credit card? Because it’s been restricted for a very long time.

[@Kotakill]

I am a power user of both. I didn’t like that Claude’s PRO subscription limit it waaay lower then ChatGPTs. Also I really don’t understand why Claude needs to write fixes to its code into a separate file(windows) each time, when I first saw this feature I assumed it will make changes in the same window. Also Claude does not accept a lot of file extensions. I was furious when it refused to accept my .cs files that were accepted fine by ChatGPT

[@ytubeanon]

agreed about usage limit, it sucks that Artifacts works by reviewing the entirety of everything written before every time it updates something

[@cajampa]

I read here in the comments to this video that the limit is separate on the API side of things. So if you need more you can always use it that way.

[@nuvotion-live]

I’m not paying for Claude unless they give the same amount of usage as ChatGPT. It’s way way too little. I’m coding, and if I’m working for hours I never want to hit a limit. And never have for ChatGPT Teams. I do use the Claude API for integration in my editor though.

[@murillodaniel9208]

Shame I can't use Claude since there is no phone option of my countrie

[@TheSonOfDumb]

This is what happens when you don't waste time having Kenyan contractors RLHF your LLM to death.

[@Thomas-vm1jy]

Is anyone else suspicious about how Claude just stole someone else's code and treated it like AI generated it? Like how much of Claude is generating code with there AI model versus Claude using stolen code as a base and using AI to tweak that code instead of generating it fully.

[@TheSmokeWatcher]

99% of code is built on already existing code and coding theory. You’re asking an artificial intelligence, that was trained on solely preexisting data, to create something completely original…. Ethics don’t apply

[@onlyyoucanstopevil9024]

Yeah

[@RedOneM]

0:50 nothing is uniquely synthesized by an AI model...

[@powervidsful]

Code doesn't get stolen just copy and pasted.

[@ainet8415]

Good job 🎉🎉🎉🎉❤❤❤

[@Mobay18]

I would highly advice against uploading anything to this. Your files are not private.

[@peanut7970]

but can claude 3.5 play chess?

[@starpause]

huhu i got it 😋😋

[@Arcticwhir]

wow yeah, tired several prompting tricks - even uploaded pic of a banana for reference, it couldnt draw an svg of a banana correctly.

[@Purpbatboi]

for free- IF you live in the US if not then not free

[@gaggix7095]

I'm not in the US and it's free.

[@Pawnsappsee]

Its free😊

[@Purpbatboi]

@@Pawnsappsee not*

[@FinalMorningstar]

​@@Purpbatboiit is lol. im not in the US, and also if this was such a huge problem, just use a vpn...

[@illyaeater]

where is gemini bruh 2 mil context code executor filter disableler juiced out the wazoo

[@HasimFN]

gpt has less context but you can just start new conversation. claude u need to wait 5 hours

[@desmondmailswhite8023]

"... allows everyone to access their latest model" No they don't. It's not accessible in China, Russia, Iran, Cuba, Venezuela, Iran, Afghanistan, Ethiopia, Egypt, Myanmar (maybe somewhere else, lazy to check them all), which is almost two billion people. It's not like I care, they can restrict it all they want. But it's not pleasant to hear myself being excluded from "everyone on earth".

[@Alice_Fumo]

While your statement is true (I assume), you could probably use a VPN to access the model as a free user at least? As a european I used to not be able to subscribe to Claude, but was happy to find this was changed in the meantime. Free tier isn't enough for a power user, but it might be for someone who only occasionally throws some queries at a language model throughout the day. I don't know your use cases.

[@Earthball_Productions]

You are required to​ use a phone number @@Alice_Fumo

[@nathanb011]

You can't really get past import restrictions unfortunately

[@gaggix7095]

China (1.5 bln people) would block the site regardless

[@brg2147]

I also like gemini 1.5 pro for coding, much better than gpt

[@bikkihsn]

Until Claude is censored i would never pay for it, such a waste of potential

[@willienel4906]

Fireship rip off 😂

[@thetabletopskirmisher]

I'd use Claude in a flash if it wasn't so useless for creative writing. I tried using one of the Claudes to help me with worldbuilding for my fantasy novel. I asked it to describe the Pale Man in Pan's Labyrinth for me. It refused. 'I dont feel comfortable bla bla bla'. Useless!

[@NormTurtle]

your late dude

[@JohnnysaidWhat]

chatgpt has become absolute trash. There is a new king.

[@override367]

Claude is an absolutely worthless chatbot lol, it can't do anything other than G rated fiction. You can't even use it to like, create D&D characters, it says that it's "potentially harmful power imbalances" to create characters, just absolutely worthlessly hobbled

[@aidajam3294]

It is garbage. I have really long experience using different GPTs (in programming). Claude (at least Sonnet) is making up stuff, therefore you should pay more attention how precise it is. In the end you will spend more time. On contrary ChatGPT (4o) is verbosy (even with your custrom instruction) but more accurate. IMHO, it is better to get annoying verbosy results instead of keeping constantly in eye on result's precison.

[@4.0.4]

That fractal animation stolen from @kishimisu was pretty blatant! In fact, I wonder if Claude learned it from the many people who uploaded their versions of it (myself included) to various online code repositories / showcase sites.

[@mkultravictim0]

30 mins no likes, bro fell off

[@human_shaped]

It's not "codes" it's "code". It is source code, it is not source codes. If you don't want to sound clueless, start saying "code" even in the plural context.

[@jld-ni3vf]

Chill out man, its only a word

[@freedom_aint_free]

Yes, but it's far more nannying and woke for sure!