I have tested deepseek with multiple projects and it could not meet my expectations while claude did it good. It's not worth the hype. My first choice is still Claude 3.5 sonnet.
I hear you. It's probably context length related. Deepseek is restricted. With that said, we always say, at this cost you can literally afford 2 extra prompts
Honestly, the work that you + aicoderking do is essential for programmers who use AI. It boggles my mind how others are not doing more content like this.
I’ve discovered that using the Architect slash command with a reasoning model (I employ Gemini 2.0 Flash Thinking) and then employing DeepSeek as the primary model and a smaller model as the weak model (Haiku, 4o mini), all through the OpenRouter API, within an .env file located in the root directory for Aider, yields significantly improved results when planning and executing a project. Aider is unable to provide a reasoning framework for standard LLMs that don’t incorporate Contextual Understanding, among other factors. Additionally, the Cline Code Extension is highly effective for project planning and execution since it systematically breaks down tasks into step-by-step instructions. Therefore, it might be beneficial to consider using Cline since it operates solely on one model at a time. LLMs such as Sonnet and DeepSeek struggle with complex one-shot prompts for the development of an entire application. It is crucial and imperative to utilize these “AI” tools appropriately and for the intended purpose! Peace, brother!
Whoa, I love such detailed responses. I spent the last seven minutes looking through your profile doing reverse image search looking at your banner and wondering if it’s Texas or Tennessee to refine the search. How do I slowly transition into something less creepy?😅 I don’t have a specific purpose, but seven minutes of your time to just listen to you talk about a workflow or a product you’re building or what you had for dinner would ….sigh . I’m procrastinating. I should really start. But still would love to chat. I’ll dance, dance.ill scrape, shoot I’ll mow your lawn ….in the metaverse
Agreed with the step-by-step process being more effective. Cline however used to use a lot more tokens before they introduced the diff edIt method, like Aider did. So I leaned towards Aider. Will give Cline another try since Deepseek is so cheap and fast, I was using Qwen 2.5 and Deepseek 2.5 before
The most interesting test is the one where you make a changement in a repository. Because most of the time, we use aider on an existing code-base. It would be better if there was more tests like that
Curious. I entered the Tetris prompt into the Deepseek chat site (free) and DeepThink enabled and it produced a game similar in function to the one that Claude made. The screen and objects were fine, but they didn't move on their own unless I used the down or left/right arrow keys. Perhaps it would handle that other failed task better as well.
Best way to use deepseek v3 in my opinion and that has work perfect with my flow, I use cursor with sonnet and when I encounter an error THAT IS NOT BACKEND related I send it to deepseek v3 via cline, I use it basically for frontend browser console errors
honestly its not great as claude or copilot(its close) but the diff is with deepseek you will endup paying 2$ a month for the code assistant and the chat is free for a little bit better you have the combo gpt+copilote for 40$ it's a huge diff
@@jonathansegura1469 Cline uses a lot of tokens :) but now that Deepseek 3 is here and it's CHEAP, I'll make a video with Cline. Also, Aider's prompts were better in my comparisons. Another "also", Aider has the powerful "always-yes" feature, which lets it continue when an LLM asks questions (I'll upload the video showcasing this once I'm done editing it)
@ yes and sonnet, it has a few variations depend on model used. But nothing specifik for deepseek, smaller context window and diff format optimisations or safeguards. Also caching works different with deepseek, and if you do not use deepseek correctly it will sometimes stream empty chunks. It’s not enough to declare expected output in system prompt.
@@remsee1608 agree. And with flow optimized for depseek it outperforms sonnet. Also training data for deepseek is newer at it makes a huge difference:)
Also, in my experience DeepSeek is junk to say the least…typical for Chinese companies. It switches human languages at random and is just messy. I also think it might be stealing API’s from frontier providers.
İ think you used the wrong deepseek model. Deepseek Coder is the older model i guess. In the docs it says "aider --model deepseek/deepseek-chat" for the V3
I have tested deepseek with multiple projects and it could not meet my expectations while claude did it good. It's not worth the hype. My first choice is still Claude 3.5 sonnet.
Agreed
I hear you. It's probably context length related. Deepseek is restricted. With that said, we always say, at this cost you can literally afford 2 extra prompts
Yeah i agree.
Agreed! I had a problem with my code. I ran deepseek quite a few times, and nothing. Claude did it but it cost me $3 :((((((
Same.
I've made a complex app and when deepseek couldn't make it Claude did it. Both together are awesome.
Honestly, the work that you + aicoderking do is essential for programmers who use AI. It boggles my mind how others are not doing more content like this.
Because it's mostly hype AI bro
@farrael004 thank you!
@yellowboat8773 I literally compared two LLMs, where's the hype or clickbait?
I’ve discovered that using the Architect slash command with a reasoning model (I employ Gemini 2.0 Flash Thinking) and then employing DeepSeek as the primary model and a smaller model as the weak model (Haiku, 4o mini), all through the OpenRouter API, within an .env file located in the root directory for Aider, yields significantly improved results when planning and executing a project. Aider is unable to provide a reasoning framework for standard LLMs that don’t incorporate Contextual Understanding, among other factors. Additionally, the Cline Code Extension is highly effective for project planning and execution since it systematically breaks down tasks into step-by-step instructions. Therefore, it might be beneficial to consider using Cline since it operates solely on one model at a time. LLMs such as Sonnet and DeepSeek struggle with complex one-shot prompts for the development of an entire application. It is crucial and imperative to utilize these “AI” tools appropriately and for the intended purpose! Peace, brother!
Whoa, I love such detailed responses. I spent the last seven minutes looking through your profile doing reverse image search looking at your banner and wondering if it’s Texas or Tennessee to refine the search. How do I slowly transition into something less creepy?😅 I don’t have a specific purpose, but seven minutes of your time to just listen to you talk about a workflow or a product you’re building or what you had for dinner would ….sigh . I’m procrastinating. I should really start. But still would love to chat. I’ll dance, dance.ill scrape, shoot I’ll mow your lawn ….in the metaverse
Agreed with the step-by-step process being more effective. Cline however used to use a lot more tokens before they introduced the diff edIt method, like Aider did. So I leaned towards Aider. Will give Cline another try since Deepseek is so cheap and fast, I was using Qwen 2.5 and Deepseek 2.5 before
Cline is the best 🎉
The most interesting test is the one where you make a changement in a repository.
Because most of the time, we use aider on an existing code-base.
It would be better if there was more tests like that
I'll definitely do more of those, thank you
Curious. I entered the Tetris prompt into the Deepseek chat site (free) and DeepThink enabled and it produced a game similar in function to the one that Claude made. The screen and objects were fine, but they didn't move on their own unless I used the down or left/right arrow keys. Perhaps it would handle that other failed task better as well.
deekseek deepthink is often better than v3 .. but obviously not so useful for production
@MarcusNeufeldt nailed it!
Best way to use deepseek v3 in my opinion and that has work perfect with my flow, I use cursor with sonnet and when I encounter an error THAT IS NOT BACKEND related I send it to deepseek v3 via cline, I use it basically for frontend browser console errors
Interesting! I'll make a video on this workflow!
Deepseek is the King of AI for Coding.
Cool video mate! Keep it up
🙏🏾 I'll definitely do, thank you for your kind words
honestly its not great as claude or copilot(its close) but the diff is with deepseek you will endup paying 2$ a month for the code assistant and the chat is free for a little bit better you have the combo gpt+copilote for 40$ it's a huge diff
with deepseek v4 plus vision no doubt that day hahaha either way is good for the devs
Where there's difference in both: Reasoning
greattt! please use cline :)
@@jonathansegura1469 Cline uses a lot of tokens :) but now that Deepseek 3 is here and it's CHEAP, I'll make a video with Cline. Also, Aider's prompts were better in my comparisons. Another "also", Aider has the powerful "always-yes" feature, which lets it continue when an LLM asks questions (I'll upload the video showcasing this once I'm done editing it)
You are the underdog
🙏🏾
Do more like this
Will definitely do!
So cool pls test phind, nestmind ai, qwen
@@AB-cd5gd will do
@MarvijoSoftware can't wait 🔥
Ever since DeepSeek came out I binned Sonnet. Not worth the money.
deepseek dellphi godd code writer
ur voice changed since the last video =\
(th-cam.com/video/tSI8qoBLWh0/w-d-xo.html)
are u using some AI software to modify it? may I ask which one? 🤙
sounds the same to me? ¯\_(ツ)_/¯
That is a dead stupid comparison using aider. Aider is optimized for sonnet.
It rules with deepseek v3 it’s much cheaper
I looked at the Aider codebase and prompts, it was actually optimised for GPT4 when it came out and later modified to cater for all models
@ yes and sonnet, it has a few variations depend on model used. But nothing specifik for deepseek, smaller context window and diff format optimisations or safeguards. Also caching works different with deepseek, and if you do not use deepseek correctly it will sometimes stream empty chunks. It’s not enough to declare expected output in system prompt.
@@remsee1608 agree. And with flow optimized for depseek it outperforms sonnet. Also training data for deepseek is newer at it makes a huge difference:)
Also, in my experience DeepSeek is junk to say the least…typical for Chinese companies. It switches human languages at random and is just messy. I also think it might be stealing API’s from frontier providers.
Yes, many say it used frontier models for synthetic data training
İ think you used the wrong deepseek model. Deepseek Coder is the older model i guess. In the docs it says "aider --model deepseek/deepseek-chat" for the V3
@@beratcimen1954 nah, they deprecate older models on the API and use the same name so we use the latest model. Check their docs, they explain this
@@MarvijoSoftware oh in that case thats great! thx for the notice
everything made in China has a catch, lol.
You reckon? 🤔