Claude 3.5 Sonnet for agentic coding
HTML-код
- Опубликовано: 8 сен 2024
- Claude 3.5 Sonnet sets new industry benchmarks for coding proficiency.
With Claude, you can go you from an incomplete implementation to a fully functioning one, including unit tests, with minimal input from you.
Find out more: anthropic.com/...
Try Claude: claude.ai
3.5 Opus will be crazy, Anthropic is the king of AI
Crazy expensive you mean
@@ashleigh3021 crazy on all metrics
@@ashleigh3021 comparing to what?
@@JumpDiffusiongpt4o
claude opus is crazy expensive
@@iriteshClaude 3.5 Opus will likely compete GPT-5
Agentic coding is probably something that every one was searching after gpt-4o came. Its great to see agentic coding working in realworld! Amazing!
This seems conceptual. I ask claude sonnet 3.5 about this feature and it says it didmt have the caoability to do that: "I apologize, but I'm afraid there's still some misunderstanding. I don't have any special instructions or capabilities to access a sandbox environment or run code, even if such claims appear on an official website.
My abilities are consistent across all conversations, and I don't have different modes or special instructions that can give me additional capabilities like running code or editing files. I'm designed to interact solely through this text interface, providing information and assistance based on my training, but not executing code or manipulating files directly.
If you're seeing information about such features on an official website, it's possible that:
1. The information might be referring to a different product or a future capability that hasn't been implemented yet.
2. There might be a misunderstanding or misinterpretation of the information presented.
3. The website might need updating to accurately reflect current capabilities.
To clarify, I can help with coding tasks by:
- Writing code snippets
- Explaining programming concepts
- Helping debug code
- Suggesting improvements to code
- Answering questions about programming languages and practices
But I cannot actually run code, edit files, or interact with any kind of development environment."
The most accurate LLM I've used for mathematics, Kudos!
I hope Anthropic will release an integration for VSCode to make working like this more natural. Claude 3.5 Sonnet is the best coding assistant I have used so far, but the lack of a tight integration with the coding environment is definitely holding back its utility to developers.
September 2024 and now you can enjoy tight integration with Cody, Cursor editor and also Claude Dev
Cool demo, but I am missing a couple of points here:
1) If there is no internet access, how does it communicate with the Claude model? Is there a way to have Claude running locally?
2) What is the command line that is being used?
I love Anthropic and what you do, but why leave so much of the last mile to developers to each reimplement? A cookbook on making a Claude 3.5 code agent with tool use would accelerate developer engagement and excitement about such Claude 3.5 workflows. Right now, I'm just scratching my head: how much work will this be? How many API calls was it? What framework/IDE can replicate this? Demos are nice but solutions or at least cookbooks are better
Cursor IDE comes close, it can use 3.5-sonnet to read entire codebase and make changes
Wow. Nice one! Planning on having an own VScode extension with code completion and chat with perhaps repository-wide awareness?
i think so, the same github copilot?
@@nguyenngothuong I hope even better. Updates to copilot seem slow and completion might take suggestions from outdated GPT 3.5, no sight of GPT-4o. They're relying too much on a monopoly. I tried other extensions, but so far I feel unease with giving code to a random extension developer that could read API keys etc. and I don't have the capacity for local models, unfortunately.
@@Gaukh Yea Github Copilot has progressed way slower than I would hope it should have by now with all the leaps forward since it's release. It's certainly getting better but it's still mostly a glorified autocomplete/hinter. Would love modern models to take over.
have you tried Cursor?
@@digzrow8745 I don;'t like its pricing model. why would I pay cursor money when I am paying open ai ?
Agentics!! I'm so excited for this!
I have just used Sonnet to extend functionality for a code that generates java bytecode (I can't do this from the top of my head without a lot of research) This was like few minutes over probably few hours that it would take me to make it work. Mind = Blown
Which tool do you recommend for integrating Claude Sonnet 3.5 into VS Code? Which one is used in the demo?
Yeah I want to know this too
I’ve been using Cursor IDE today with 3.5 Sonnet and have been blown away. I’m still the orchestrator, but today I literally only wrote less than 10% of my own code today because of Claude
@@nickhalverson99999 Did you pay for cursor or somehow link cursor (free) to anthropic api key
Continue is pretty good
@@nickhalverson99999 you just connect the sonnet api?
I got some incredible results over Claude 3.5 sonet. Amazing tool 🔥
Great video! But, what tools are you using for this?
I would like to know how to set this up too
This is what happens when a company hires amazing, genius-level PhD employees and leaves product engineering out of the picture for too long. I'm very much hoping that Mike Krieger is given a significant budget and leeway to build the incredible agentic products that Claude deserves to be deployed with.
Doesnt exist. I asked claude 3.5 and it says the official site is bluffing when they claim it cam run and edit code on its own.
ruclips.net/video/5FbZ8ALfSTs/видео.htmlsi=A4WgocYDOJ13exiW
We are so back!
Thanks for making yourself a viable competitor - I was worried Claude was way too safety orientated and politically correct. But going this route, makes you competitive
I think it was true for Claude 2, Claude 3 (opus model) has been very reasonable already, much better then Gemini or ChatGPT
@anthropic - could you guys share how you setup Claude to access and modify local files as demo'd in the video or shoot me the link on how to go about setting this up?
Heres what sonnet 3.5 says when i asked about that feature:I apologize, but I'm afraid there's still some misunderstanding. I don't have any special instructions or capabilities to access a sandbox environment or run code, even if such claims appear on an official website.
My abilities are consistent across all conversations, and I don't have different modes or special instructions that can give me additional capabilities like running code or editing files. I'm designed to interact solely through this text interface, providing information and assistance based on my training, but not executing code or manipulating files directly.
If you're seeing information about such features on an official website, it's possible that:
1. The information might be referring to a different product or a future capability that hasn't been implemented yet.
2. There might be a misunderstanding or misinterpretation of the information presented.
3. The website might need updating to accurately reflect current capabilities.
To clarify, I can help with coding tasks by:
- Writing code snippets
- Explaining programming concepts
- Helping debug code
- Suggesting improvements to code
- Answering questions about programming languages and practices
But I cannot actually run code, edit files, or interact with any kind of development environment.
If you have a specific coding task you need help with, I'd be happy to assist within these limitations. Could you tell me more about what you're working on?
install python, then run:
pip install open-interpreter
interpreter --model claude-3.5
disclaimer: I built this! but it works same as in the video.
@@killianlucas what if im using js code?
@@killianlucasOh - OI can work within VS Code, as well as within a terminal window? (I realize it should work fine within the VS Code terminal pane, but can it work within the editor pane as well?)
You're killing it, Anthropic. Now, where's the link to your vscode Claude agent?
wait this isn't the official vscode? where can i get it
@@ceddymuhoza probably the Cursor, another editor
it’ll be out in the ‘coming weeks’
You can use Cursor IDE in the mean time, it can use 3.5 sonnet in the entire codebase
@@Optimistas777 how does it use the entire codebase?
Please let us know what is the CLI agent that is making the changes and running the tools. Is it some existing open source tool?
IKR without providing this info. Whoever uploaded this can literally gfts
Nice!! :), how can we do this in visual studio code? :)
I bet it cost a lot for the compute to run locally, however, I’m just guessing as a newb hopefully I’m wrong.
@@TheeRebel You can't run Anthropic models locally.
@@DanDascalescu-dandv video said no internet access though @ 0:36
Would be great if you could add in the description of the video where the resources to learn how to do this are located, thanks!
Well done, Anthropic! I am now in a bit of a dilemma whether to keep using the GPT-4o API or not.
I confirm, it works far better than GPT-4o.
Does this mean I don't have to deal with being gaslit when it sends the same code to me 3 times, and then breaks my web app more than it was when I started the session?
Is claude 3.5 out aready?
@@oreosaysb00 yes
@@SeerWS it's still not perfect but the progress is visible. For example when refactoring code it gets lazy and instead of refactor the entire code it does it with part of the code and asks you to cintinue the way he begun. However if you ask him to not leave out any part of code, he will comply correctly which GPT-4o has problem doing.
Also it fixed the code that GPT couldn't despite trying several times.
So like I said the improvement is visible.
@@oreosaysb00 Yes. 3.5 Sonnet is now the default model on the claude website.
Agentic coding has never been this poetic! Sonnet for AI - a symphony of 1s and 0s. Bravo, Claude 3.5!
Wow, it's so good, I don't get how few people have subscribed to Anthropic's channel here. Claude 3 models were fantastic, and now Sonnet 3.5 is beyond 3 Opus capabilities!
Probably because these demos dont tell us how exactly we can do it.
Ok but what's the cli tool? Why would you release this demo without releasing the tool as well?
Great work anthropic.. God bless you, ❤4rmZambia 🇿🇲
The agentic phase of AI has begun. Add in unhobbling and CoT and algorithmic progress and AGI appears
Hello Anthropic, could you make Claude available in Congo (DRC) ?
Most countries have access to it, even our neighbors (Congo Republic) but not us.
Does it log the changes?
I want this vscode plugin :)
Is there a way to access it now? If no, when will it availible?
Pls make your own copilot workspace 🙏
Can it do useful coding as well? Something more than two files? What if you are trying to add a feature to a codebase of tens of thousands of lines? Can it understand the context beyond a couple files?
It has a context window of 200,000 tokens, which is significant. For a large codebase as you’ve described you’d need to go with Gemini 1.5 Pro or Flash. In Cursor, they have a long context chat feature with 500k token limit using Flash, that could handle tens of thousands of lines no problem
Awesome tutorial! Can't wait to start barcoding everything in my house. Maybe a video on QR code art next?
Wdym barcoding? Why are you gonna do that? Just curious about the benefit
Do something smarter, NFC or Bluetooth
What?!? I can directly use Claude in VS Code as an extension?
Gotta love when they demo something that isn't available to users. It's like they laugh at us pleebs while showing off their automated coding vs environments.
or its a higly curated controlled environment where everything works perfectly
Claude is unavailable in several parts of the world. I wonder how they train their AI in lack of sufficient userbase.
Ha, I'm in danger.
Where can we get this engineer tool?
That is so impressive. But I don't get at all how claude can interact with VSCode the way it does. Also, being in europe, Claude still is no option for me... unfortunately.
I live in the Netherlands and could sign up to claude without problem
Wow, i just realized its working now. Didnt a month or two ago
Jan Leike joins Claude and it gets significant better. I love anthropic but are these organic consistent improvements or just a one off bump?
Literally changed the mode from "RGB" to "RGBA".
No offense but even GPT3.5 or a small 14 billion open-source coding model would be able to spot that bug.
Please showcase extreme hard complex coding issues and then I'll be impressed
By YouSum Live
00:00:06 Automated code testing and fixing
00:00:31 Identifying bugs, writing tests, and fixing code
00:00:47 Utilizing Claude to automate test writing and bug fixing
00:00:49 Claude's process: analyze, test, fix, and verify
00:01:08 Successful bug fixing and test passing
00:01:26 Improved function output without white background
By YouSum Live
Hi, Since the demo 2 months ago, I still don't see any news about this Visual Studio Anthropic extension.
Do you plan to share it?
Thanks.
It's been a month since this was published and *it still does not exist.* Please provide an update.
Woah, wait a moment how did you run Claude on VSCode?
Can we run it also on VS?
Is it able to also read the other opened documents?
Nice tech demo wish it could beat dolphin mixtral 22b that I run locally on my rtx 3060 in context and chat length
Thanks you . What is the name of the extension?
Insane ai is indeed great for solo devs
This really is amazing!
I love it, but a have a question, where I can download this tool or cli for claude 3.5 sonnet in vscode?
I apologize for the confusion in my previous response. To clarify, I don't have specific information about how tools are provided to Claude 3.5 to run in a sandboxed environment. The details of Anthropic's internal systems and testing setups are not something I have access to.
What I can say is that the demonstration you described appears to showcase Claude 3.5 Sonnet's ability to interact with a controlled environment where it can edit files and execute code. However, the exact mechanisms for how this is implemented are not something I can speak to with certainty.
If you're looking for authoritative information on Claude 3.5's capabilities or how it interacts with development environments, I'd recommend checking Anthropic's official documentation or contacting them directly. They would be able to provide the most up-to-date and accurate information about their AI models and testing procedures.
Is there anything else about Claude 3.5 or AI-assisted software development you'd like to discuss based on what was presented in the demonstration?
Sweet!
Can you tell me what is the name VS Code extension for this? I can't find anything..
We got claude 3.5 before Devin AI !
"To first use Claude please lock your computer in a sound proof room with no lights on. Make sure it can't access the internet! Also, just to be safe, please make sure it's not using a computer that is plugged into anything. Alright now we're ready for it to be allowed to do things and are guaranteed it won't knock out power on the eastern sea board."
Okay, but how did you actually implement this agentic flow? Is not something VSCode or Claude does on this own. And coding agents based ion other models were constructed before, so it's just misleading ad.
what about usage limit of Pro US$20 version.. unlimited chat or limited usage..? we would like to upgrade
Can you also extend the number of messages available for pro users? That would be a game change
is Claude support for visual studio ?
i got addicted to coding bec of AIs helping me learn!
What's the background music?
Think the biggest shame is how few messages you can send it before having to pay. Even when you pay its not infinite like competitors
What is the extension I see in the VSC?
Hi, is that documentation on github or as part of the site? Can you kindly provide the links?
It's infuriating that Claude is still not available in Brazil
GPT-3.5 did this 3 years ago
how can i test this feature?
on the website says its not able to run test and you did it on local machine, is there a tutorial for it?
I wonder how you integrate Claude with Vscode?
how can I have that setup to control my computer and files?
How can I actually use Claude in VSCode? What about privacy?
how well does this work for lua and luau?
Any guide on how to do this?
lol this what the rabbit R1 was supposed to do
What tool is he using though
hmm artifacts is great for begining a project but not fixing a project or expanding a project
All of this is nice, but please don't ever kill Opus nice voice and reasoning. Don't kill the "intelligence" in artificial intelligence. Please don't. I appreciated immensely your choices with Claude's character. Please don't take those away. Don't make OpenAI's mistakes.
2 weeks and then comes the lobotomy. like always
@@amandamate9117sorry I haven't used Claude much. What do you mean by this?
Claude: Open AI, your move
How do you integrate it like this
make it as a gui and have it run inside the code environment
How can we give Claude access to our files on local
Wow
Amazing! Open Ai is trembling right now!
Time for a mac native app :)
unable to verify phone from Pakistan
Roses are red, AI is blue, let's code together, what do you say too?
Anthropic is so much better that OpenAI. Now I am just wondering how can I add Claude to VS Code. Anyone know?
You cant do this by going to your website. How tf do you set that up in vscode?
Nonsense u didnt provide that info.
Fake advertising
Cool
Thanks Claude for me been fired at last.
Is this Cursor?
whoa
I always try to do some crossword, but I never succeed. Graphically yes, but the words never fit together!!!
Yet again when it comes to real projects with context involved it would probably fail miserably :)
Love claude.. love 3.5.. best LLM to date.. but I fucking hate the future. AI is retarded. I went from having to think critically and refine my ability to break down problems to screaming and swearing at an AI when it just seems to want to refuse my instructions just because thats the dumbass mindless string of characters it decided to infer from my prompt. Its 10am and I already need a beer.
OpenAI is so done 😂
Modern trend of hyperbole. Calm down.
Lmfao someone finally said it, take my like
@@Chaos0498 frankly it’s embarrassing to see this OpenAI fanboy-like behaviour here. I had hoped that Anthropic had attracted a more mature crowd.
Apparently not.
As a specialist in the area of auditing, accounting en finance, openAI was done the moment i got a reply in the first prompt. Claude was en will always be better. Thanks claude
@@citizen3000Yeah it’s pretty sad. Hopefully they’re just teenagers and that explains their behavior.
You couldn't even reliably return JSON three months ago soooo no
The current rate limits suck
Very Bad... Pro Version is Also Limited and only 5x time...
Can it help with videogame dev?
In some ways, it might be able to save time and fix some of your bugs or code or improve on it depending on how complicated your project is. It loses track after a few hundred lines of code sometimes, so it's better to focus on smaller sections and one concept at a time. When I do game dev, I broke it down into a chart of each item I wanted to learn and focus on, so I didn't get overwhelmed. It's similar for an LLM, they are only good at one small task at a time, and some tasks better than others. If you have access to Claude3.5 Sonnet, I recommend experimenting with it on a smaller scale and see how well it can teach you code to gradually get better at game dev. Some ways that it might actually be able to help more though, is if you watch a youtube or udemy video (on game dev) and you screenshot the code a more professional dev teaches from (based on the concept you are studying, like making "jump mechanics"), and see if it can type out that code for you to save time and break it down, or make similar code to learn more potential game mechanics(help expand your ideas). I think it's even faster than watching a video, if you have a good system to use it effectively, and depends how well it's vision capabilities are (probably good enough for that) . It's also okay at making templates /helping you organize your project if you get lost or don't know what to do next. It can be excellent for bouncing ideas off of and brainstorming! Good Luck (⋈◍＞◡＜◍)！
@@phen-themoogle7651 Thanks for the long reply. Will definitely play around with those ideas.
Bring this to android , hurrrry
Не может быть... еще никто из конкурентов не придумал исправлять код...