Excited try it, just checked out your plugin on vscode 🙌

Wonderful work! Switching to twinny today.

thanks for your contribution to the community

And thank you for making Twinny! ❤ Exactly the balance I have been looking for. Twinny is probably what most people want - they just haven't heard of it or tried it. Works perfectly with llama3 in ollama locally. No mandatory logins/accounts. No paywalled "Pro level feature" and so on. It is 10/10. Please organize a team around your vision for this extension - a no-nonsense, locally or API-hosted AI code completion. Again - Thank you for your contribution 🙏

works great! But how do I connect it to my AI server? my AI is on another machine and i cant get it to connect to it after changing the settings to the IP and port.

Agreed! Very funny.

I know nothings going to happen at the end but I watch it anyway for some reason.

Sweet. Maybe I should stand there staring at the camera for 4 hours and just rake in the ad revenue.

Haha it’s my favourite

Infinite money glitch, you have found.@@technovangelist

Thanks so much. I used to drive through Italy on my way to Konstansz from Amsterdam. There are more direct routes but I loved the drive. Also did a few cheap weekends in Rome and then two weeks in Florence and Cedri in Tuscany. Hoping to take our 5 year old daughter to see Italy soon. My wife did a semester in Rome while working one her architecture masters.

I think I saw the Twinny prompts say to respond in whatever language the current file is set to in vscode .

That would be smoother than my current system of having to remember to choose language manually. Guess I have to test out Twinny.

Yup. Richard let me know it’s meant to be your twin developer

Glad to hear that!

Then it sounds like twinny and its chat is perfect for you. The ability to generate large sections and preview it before adding to your code is fantastic.

@@technovangelisti'll definatly give it a whirl! Thanks :D

Pieces? Never heard of it. Hmm snippets with ai. Interesting.

@@technovangelist has quite a number of supported ides, local + cloud models, attachments and storage

Tried them, but they are being way to cryptic about how they work. Seems super sketchy, deleted.

@@technovangelist Hey! I'm Mason from the Pieces team and apologize that we came across as cryptic to you! I'm not sure exactly what you are referring to as cryptic but really would love your feedback on our product as a whole. We try to be extremely transparent through livestreams and open source meetups but something obviously fell through the cracks here and this isn't how we want our brand perceived by anyone haha If you'd be open to it, I'd love to hop on a call and chat sometime about your thoughts.

I think it had to be coincidence. I didn’t do anything. But it would be cool if it did.

I can’t stand dark mode. Hurts my eyes to look at for long.

Ollama has been available on windows for 3 months or so. Turn on Microsoft’s own feature WSL2 and then install Ollama.

It’s often just as capable without the massive privacy and security risk. A huge number of companies ban use of tools like copilot

I definitely need some windows content. Thanks.

But I think the developer of twinny does it on windows

It wasn’t when the video was made

I was using something that did it but can’t remember what it was

Yes but it’s not fully supported and the tweets in the video are from the CEO saying how it may not be offline and it will probably involve a subscription and syncing with online services.

It has worked on windows for 3 or 4 months

I just installed it on WSL2 under Windows 10 and ran Mixtral on it. It look over a day to update this sh*tty Windows so that my GPU became available under WSL. It took like 5 minutes to install Ubuntu and another 2 minutes to install Ollama. Also even the quantized version of Mixtral is still 26GB. That Ollama can load it so quickly and it performs so well that it's totally usable seems like magic to me as my GPU only has 24GB.

Yeah if you don’t have the memory in gpu it’s going to be slow. Mixtral with the context is going to need 30-40gb and there isn’t much that can be done there. Stick with models that perform well on you hw.

@@technovangelist No, it works as I said, to me it's a little miracle as I had nearly given up on any MoE model. GPU memory of the 3090 in this system was mostly filled, CUDA usage surprisingly low but the model responds quicky enough that a fluid chat is possible, I've no idea how they made it working so well. I had a model with like 27 or 28GB recently running on LM Studio and had to offload some layers so it could fit into the VRAM. It was 60 layers in total or so, YI 6bit If I remember correctly. I could load up to 54 or so but it still was painfully slow. I chatted with Mixtral for well over an hour, the text generation didn't slow down much and it completed my question pool like no other model before, quite insane I have to say, it answered questions that no other local model could solve so far, including the YI model I got barely running. The context window size seemed to be a lot lower than 8k though, but I'm not sure. I haven't figured out yet where all the settings are in OLLAMA. This time I prefered to just enjoy the model running so well out of the box without me having to juggle with ChatTemplates and all kinds of parameters. :-)