Coding with AI episode 1: Claude Sonnet 3.5 vs DeepSeek v3 for coding analysis and modifications

[@antirez]

In the next video, I'll show how Claude, with the right hints (and that's the interesting part! How to start the chat, what design clues to provide) single-handled (code-wise) implemented UTF-8 support inside Kilo.

[@RiccardoCosenza]

Form Campobello to Catania is a looooooong viaggiu…. 😅😅😅

[@waynetrout6977]

nice comparison, thank you for some insight into deepseek, I have not used this model yet. You may know this but for modifications i might sugest prompting the AI to utilize CLI functions for quick and accurate modification/replacement in longer code. This works well with C 3.5 I am not sure about deepseek. also, if you are having issues with the AI accepting your code submission sometimes this is due to file type just change it to .txt and it should read this just fine. example .ahk some models will not accept as a file type but will help you with the code if submitted via a .txt.

[@GianlucaB73]

davvero interessante e coinvolgente questo format... continua così!

[@halemm]

bellissimo questo tipo di video!

[@Thelegends96]

Hi Salvatore, thanks so much for this! Loving these recents videos. Some constructive feedback since you are looking to make more (please do!) - While showing how the model can fail was interesting, the back-and-forth interactivity took some of time. Perhaps you can go through previous / pre-prepared chat histories with commentary on what was your thought process? this way you get to pick interesting cases! - It would be great to see examples like using the model for code reviews, brainstorming, or handling bigger projects with long context windows and multiple files. (If you use them for these purposes) P.s I can relate with the "fotta" to make the videos, but get some rest! :p

[@antirez]

Thank you! In the next video I am doing exactly what you suggested of having a pre built history. All is still authentic but much more information dense and easier to follow :) thanks.

[@DomenicoLupo]

Another difference i saw is that DeepSeek cannot execute the code, while Claude can do it. I hope they add this feature soon. Great video thank you!

[@antirez]

Yep. In Claude they did something very smart: the model writes the code in Javascript using console.output() for output, so it can run in the user browser. No backend needed for code execution. With webasm they should be able to easily do that for Python as well I guess.

[@nonefvnfvnjnjnjevjenjvonej3384]

thank you thank you thank you for doing this.

[@ConteMascetti1971]

OOC question: Gentle words are necessary for the models ? Or its mandatory by your sicilian education (like"please")

[@antirez]

@@ConteMascetti1971 comes natural to me for some odd reason :)

[@ConteMascetti1971]

@antirez this evening i Will do a check of how much gentle words GPU time costs in a LLM Model like llama

[@ninadsachania3652]

Is it Ghostty?

[@antirez]

@@ninadsachania3652 yes

[@ciao1307]

Hi Salvatore, is there no way to remove the black bands on the sides? Good content anyway :)

[@antirez]

@@ciao1307 next time I'll make sure to use the correct aspect ratio :) thanks

[@Techonsapevole]

Anche Qwen-coder-32B non era male

[@erminiottone]

Why do you use overpriced computers like Apple?

[@antirez]

Unified RAM between GPU / CPU. Unix but without the issues of Linux.