Видео

Subbed, very usefull info, thank you for your work

anything built on top of vscode can only be crap

Thanks for the great work. Windsurf is a highly intriguing staff. Would it be possible for you to conduct a comparison between Windsurf and Aider?

Thanks for the testing. Also just went to your windsurf video, and tbh. your own voice and natural delivery is better than the AI (web "scrapping" lol), though I get why some would find the AI voice clearer. Would be good to at least use a voice with a UK accent though! The new Kokoro models are stellar, and small enough to run locally.

trae ai?

Why not use both simultaneously?

You should be using cline with architect mode with R1 for the plan and sonnet for the response. Or use sonnet for architecting then executing... The way you are using cline is not the best approach.

Can you compare OpenHands with the agents you've already tested? I’d like to see how it stacks up against them using DeepSeek V3 or R1.

Very informative videos all' time thankyou brother 👍❤️

@3:11 Would you actually run a big python script generated by AI like this immediately without reviewing it? Feels like a hack waiting to happen. Or maybe you are in a VM / didn't show the review part?

I just subscribeed for this, this right here is a gem! 💎

Capitalism has high-quality products, and socialism has prices that everyone can afford. No wonder the ugly American capitalists and politicians hate China so much.

so called fake model now performs better than the original one

for general questions, nothing beats chatgpt

The thing is that its not really a bug. Its a flow issue, so you need to explain what flow you want. You Told it the navigation is not showing betslip but a blank page with a hamburger menu. You actually tell it that currently i click on the nav and a blank page opens then i have to click a menu button to show the betslip. But What i want is the betslip to show immediately after clicking the nav. The less poweful a model is the more detail and leading you need to do

canceled openai subscription

I have been playing with writing apps mainly with sonnet for many months. I am not an experienced coder, but i have a computer engineering degree but never coded professionally. I find abstract tests like these are hard to judge the AI for non experienced coders. Start out with clear english instructions, of a subset of the project i have in mind, then i try to build functionality fixing and updating. I have about 4 pages of what i expect in my profile. I have gotten the best results and most complete code base with Sonnet, typical applications may have around 2k lines of code in all. It becomes impossible after a while, and you have to plan to stop a session and hand over to another session. I have been using Deepseek for several days, and its definitely better than any of the gpt models, but i get the best results with sonnet. I have had a 40 yr career in oil and gas exploration, and it is fascinating that i can create solutions for all the ideas i have struggled with over the past decades. I have hardrives full of data and reports, and its just fascinating. At times i would use both gpt and cluade, (mainly to help Claude solve problems he cant find the solution to because his API specs are out of date. Troubleshooting is by far the worse. I think these tests give the illusion that it can do very well. Once the project gets over a size, it does nonsense - there is a cliff it fall off.

Now I know why DeepSeek is so hated, it outperformed two of the most powerful ai without even flinching.

Thanks!

Claude Sonnet is the best at building medium to large code bases. I haven't found anything on the market that comes close. I will use Deepseek to QA my Claude codebase just to see how well it does. Metrics and real world usage are two totally different things.

Sonnet still beats R1

It reads 50 files. It wins.

Wondering if R1 was trained on this particular challenge. Since it's open source, very well could be word for word in the training data. Either way, still impressive

I subscribed because we do a very good job. So you think that Cline is better than Aider and Continue Dev? We could say that top to bottom Cursor AI - Cline - Aider - Continue?

Please Do Cursor(Deepseek-R1) vs Windsurf(Sonnet 3.5) video .

Please Use this voice in every video , and also please Do Cursor(Deepseek-R1) vs Windsurf(Sonnet 3.5) video .

Please Use this voice in every video , and also please Do Cursor(Deepseek-R1) vs Windsurf(Sonnet 3.5) video .

Please Use this voice in every video , and also please Do Cursor(Deepseek-R1) vs Windsurf(Sonnet 3.5) video .

so just start from today, i will fully trust into this chinese company called “Deepseek”. because they will slap every US company with amazing their LLM result with cheap and OPEN SOURCE 😂 so in the future, when Us company release something, i will just wait for deepseek to slap them in 2-3 weeks later 😂

No ,deepseek R1 is number one, it batter then new Gemini 2 Flash thinking

I don't trust the chatbot arena, they put Claude 3.5 Sonnet in 11th place

Hey man! Awesome channel you've setup here. So cool to have these comprehensive comparisons between coding tools in a real life Software Engineering project, even with comments about the expected complexity of the tasks and who could do it if it was a person 💯. I was wondering, have you tried Roo Cline vs Cline? Seems like Roo Cline is a faster developed version of Cline? And that they've now gone their separate ways. Also do you know if any of these are private if you run them with local models? I know aider doesn't send anything to the internet but curious about Cline/Roo Cline and Cursor too. Again many thanks. Will be following this channel closely.

Next time you should also include the little steps too: 1) 2:55 is this the code is generated? 2) how did you save it? The file names didn’t seem to change maybe you should’ve picked a whole new name. 3) would it be better to just create a file first name testdeep.py and just paste it in there first or doesn’t matter much. Or does the file name need to be exactly “rest_api_test.py” for the test to work? 4) do I need aider to test this? 5) I think your next video should be the same test covering some of the things I mention since I am about to attempt this even tho I am a n00b but the difference would be is that you would do this same deepseek test vs 5x ai google studio models with default settings + 5 models using value “0” for both temperature and top P to see if this increases accuracy and compare to default 1 / 0.95 value settings. Also I want to see if your 2.3k subs will go up after this since it will bring the google fans lol

I told you last month you are going to blow up because you are making videos no one makes but people think about them

This Open-source model is truly revolutionary!🔆🔆

### Summary: Cursor vs. Cline AI Coding Battle This video compares **Cursor** and **Cline**, two AI coding tools, as they tackle coding tasks in a large 240k-token codebase for a soccer prediction app. Here's a breakdown: #### **Tools Overview:** - **Cursor**: A proprietary Visual Studio Code fork with AI features. It has a Pro plan for $20/month (limited to 500 requests). Uses a vector database for context and embeds changes directly into the codebase. - **Cline**: An open-source Visual Studio Code extension that works with most large language models. The demo uses **Claude 3.5 Sonnet**, which is cost-effective but less refined. Cline supports transparency by showing payloads sent to LLMs and has Model Context Protocol integration for system communication. --- ### **Task Comparisons**: 1. **Bug: Search Box Focus** - Cursor identified possible issues (focus being stolen/z-index issues) and implemented reinforcement code. The functionality worked perfectly. - Cline failed to fix the issue, even after multiple retries. - **Winner**: Cursor. 2. **Bug: Filtering Tournaments by Match** - Cursor used efficient embeddings to isolate context and fixed the bug effectively. Filtered search worked correctly. - Cline failed to display filtered results and repeated errors, even after additional attempts. - **Winner**: Cursor. 3. **Feature: Match News Summarization** - Cursor successfully implemented web scraping (using Serper and Playwright) and integrated Google AI's summarization for structured predictions. It added a functional frontend button to display the news and predictions. - Cline struggled with scraping, faced repeated errors, and could not complete the feature. - **Winner**: Cursor. --- ### **Key Observations**: - **Cursor Advantages**: - Faster and more reliable bug fixes. - Fine-tuned small language models for quick tasks. - Effective use of embeddings for large codebases. - Generates structured summaries and predictions. - **Cline Advantages**: - Open-source and model-flexible. - Transparency in API payloads. - Better integration with custom systems via Model Context Protocol. --- ### **Outcome**: **Cursor outperformed Cline** in all three tasks, demonstrating higher efficiency, accuracy, and feature completion, despite Cline’s flexibility and open-source nature. For more detailed insights and future comparisons, the channel offers memberships for exclusive content.

Deepseek R1 is a gift to us, i am glad people are recognizing it

I think the new gemini models have been tuned pretty well in terms of human preference (at least I like the newer models more than their older ones). Claude imo is usually #1, then everyone else is about the same. However from my usage of the model, it seems like gemini 2.0 exp 1206's responses get pretty bad/mediocre after 60k tokens of context.

its so satisfying to see spreadsheet of performance at end , 👍

Use cline with DeepSeek R1. Very cheap. 👍

How the f its number one on arena? Like one or two days after release? I had to wait a lot longer to see phi4 in that list.

❤ Gemini 2.0

I've been using deepseek v3 and it's so amazing , i turned off deep think and it became even better in coding , i hope it stays free forever

You know you can use vectorized data of your codebase with Cline too right? with MCP tools. Best of all, you can ask the ai to do it for you, if you're willing to pay for tokens. Curser does it natively.

I hope someday cline with DeepSeek will be better than nowadays Cursor + Claude 3.5Sonnet.

Isn't Deepseek R1 not pretty insane at coding? I saw that its a big improvement on math, gpqa and somethinng. Btw im not an ai researcher. But i think claude 3.5 still beat any ai overthere

Thanks for the nice video. Please make a comparison of Roocline vs Aider vs Cursor

deepseek's thought process is really impressive, I tested it on my own weird problems and it passed with flying colors

Appreciate your videos. Thanks for the comparisons and insights

#1 my ass. I tested it with 10+ prompts and it hallucinated a lot. It would suddenly stop generating response in AI studio. As of today, it's quite buggy and unreliable. I wouldn't recommend using this.