This is phenomenal! I looked through the entire App Store and not one did this for Mac. There was Microsoft swift or whatever but it unsupported now and never was on Mac. Good job
Love your videos, dude! Just noticed you on the ollama discord the other day too! I'm all in! I think perhaps many of these applications that are coming up with built in features like this and others will find those features quickly obsolete as full operating system integration of the same functionality start to roll out. From a resource requirement stand-point it makes a lot more sense to have those apps on an operating system level instead of having different apps each with their own models running.
I found this because I was looking into ways to bring some AI help into obsidian. What are your thoughts on completion in Obsidian comparing cotypist to obsidian companion? (I bet they're linked in your obsidian notes and share tags too... lol). Which serves your note taking best? Second... if you use tab for completions, does that conflict inside the Cursor AI app or other apps that already use tab for completion? I think I just figured out how the war between AI's will start...
I love your direction, first was plugin for Obsidian, now this... you know what, this is why we love AI, to use it, not just admire it from the distance. Thanks Matt!
Update, it can even autocomplete things that other frown, I just tried Jack and Jane wanted to have s*x and it followed along. One of my biggest gripes is censorship of these models.
@@technovangelist he didnot say just using "just" open-source. For me plenty of things don't need to be open-source. However, something that registers EVERY word you write is a bit concerning. Especially in the time of LLMs which could potentially "mine" the data. Since it isn't open-source we can't SEE it isn't sending information. We can potentially use some tool to make sure it isn't sending data, or too block it. But in this case I feel it is a concern that should not be lightly dismissed.
Local LLMs using llamacpp or ollama don’t mine the data. There is nothing they can do with it. I understand there is always the tinfoil hat crowd but the author already has a record of doing good stuff.
Fellows, Matt, does anyone know where exactly Cotypist stores downloaded models on Mac. The app does not yet have an option to delete the downloaded models. I wonder, how it is done in the case there is a need to clean storage.
Thanks for this video. Have you tried dictation with MacWhisper? It uses AI to enhance the accuracy of the dictated text and can also improve the style according to a prompt. It's amazing and saves even more keystrokes. Unfortunately, even though it works with Ollama, I haven't been able to achieve results nearly as good as with OpenAI. Maybe a video idea could be how to make MacWhisper great with Ollama?
@@technovangelist Great. I wasn't aware of that. Thanks. I just tried it out, and from my initial testing I got just as weird results with Ollama in Superwhisper as in MacWhisper. But it's nice to know that more companies are working on this.
@@technovangelist I am talking about the built-in AI improvement of the text between the whisper transcription and pasting it into the document. It's very powerful. I think it works better in MacWhisper than in Superwhisper, though. I guess there is something different with the temperature or how they process the prompt.
How safe is this?. How do we know that it is not sending keystrokes back to someone for things like logging into different things like banks and so on? Any way to Check for this?
Hi Matt! do you know any technique to expand the output's size of the llms answers? Because currently, the context windows for inputs are fine, however the responses I think are too small. Can I create a prompt that will generate multiple answers? and mantain consistency between the responses? Think like telling an llm to generate a full book. Obviusly it would be fine to cut off the content by chapter for example. How can I told in the initial prompt to ask the llm to generate multiple answers (with just one prompt)?
Cool app. Not a fan of the background music though (in general, not just your choice for this specific video). Thanks for sharing! Also... Obsidian is my favorite application I've ever used, EVER (since I started using computers in 1989).
Well, It automatically translated the audio in french for me. I also tried the german and the spanish version. Pretty accurate translation overall. But I did not like the robotic voice so I went back to english (I prefer your voice)
upvote for split ergo image and the 'and I have never seen STT work well there' ... STT is great while driving your car, but for actual productivity???? yuck! Now I need to see if there is something similar for libreoffice :)
@@technovangelist Hmm I'll check that out, but to be fair, you're a native English speaker with a very clear, measured voice paired with flawless diction. As a French guy living in Germany whose diction is already rather bad in his native language I'm already happy when Google Assistant understands me ;)
Omg!!! Did I, Cringely/Mark Stephens just comment on my video?!?! Author of Triumph of the Nerds and Accidental Empires??? I didn’t expect such an amazing viewer. I will do better. For you. Or maybe there are many journalists named Robert cringely.
All of them probably. I'm hearing the german one, and it hurts. Sounds like a 10 minute ad. I mean, probably is, I don't know, only clicked on it to figure out wth is wrong with the audio. Sometimes it speeds up randomly, most of the time it's too slow, and the voice just doesn't match the person. Also pronounces every English word weird.
This is phenomenal! I looked through the entire App Store and not one did this for Mac. There was Microsoft swift or whatever but it unsupported now and never was on Mac.
Good job
Love your videos, dude! Just noticed you on the ollama discord the other day too! I'm all in! I think perhaps many of these applications that are coming up with built in features like this and others will find those features quickly obsolete as full operating system integration of the same functionality start to roll out. From a resource requirement stand-point it makes a lot more sense to have those apps on an operating system level instead of having different apps each with their own models running.
I found this because I was looking into ways to bring some AI help into obsidian. What are your thoughts on completion in Obsidian comparing cotypist to obsidian companion? (I bet they're linked in your obsidian notes and share tags too... lol). Which serves your note taking best?
Second... if you use tab for completions, does that conflict inside the Cursor AI app or other apps that already use tab for completion? I think I just figured out how the war between AI's will start...
I love your direction, first was plugin for Obsidian, now this... you know what, this is why we love AI, to use it, not just admire it from the distance. Thanks Matt!
Update, it can even autocomplete things that other frown, I just tried Jack and Jane wanted to have s*x and it followed along. One of my biggest gripes is censorship of these models.
The obsidian plugin was great but this is far more powerful since it works everywhere
@@technovangelist definitely and is faster
You nailed it. This is why I get tired of coding and other work crap. I get tired of typing the same thing over and over.
Thanks
Cotypist is neat, not being Open Source would make me a bit hesitant to run it due to the OS permissions required though.
Only using open source stuff is a bit limiting though. I couldn’t do most of what I do with that restriction.
@@technovangelist he didnot say just using "just" open-source. For me plenty of things don't need to be open-source. However, something that registers EVERY word you write is a bit concerning. Especially in the time of LLMs which could potentially "mine" the data. Since it isn't open-source we can't SEE it isn't sending information. We can potentially use some tool to make sure it isn't sending data, or too block it. But in this case I feel it is a concern that should not be lightly dismissed.
Local LLMs using llamacpp or ollama don’t mine the data. There is nothing they can do with it. I understand there is always the tinfoil hat crowd but the author already has a record of doing good stuff.
Love it, cool to hear about new tech like this as well. My go to resource for Ollama and other AI tips.
Cool, thanks
Fellows, Matt, does anyone know where exactly Cotypist stores downloaded models on Mac. The app does not yet have an option to delete the downloaded models. I wonder, how it is done in the case there is a need to clean storage.
Thanks for this video. Have you tried dictation with MacWhisper? It uses AI to enhance the accuracy of the dictated text and can also improve the style according to a prompt. It's amazing and saves even more keystrokes. Unfortunately, even though it works with Ollama, I haven't been able to achieve results nearly as good as with OpenAI. Maybe a video idea could be how to make MacWhisper great with Ollama?
I have. I tend to use superwhisper which is a bit more powerful for live speech to text. I have a video about it.
@@technovangelist Great. I wasn't aware of that. Thanks.
I just tried it out, and from my initial testing I got just as weird results with Ollama in Superwhisper as in MacWhisper. But it's nice to know that more companies are working on this.
How are you using them together? Ollama can’t run a whisper model.
@@technovangelist I am talking about the built-in AI improvement of the text between the whisper transcription and pasting it into the document. It's very powerful.
I think it works better in MacWhisper than in Superwhisper, though. I guess there is something different with the temperature or how they process the prompt.
How safe is this?. How do we know that it is not sending keystrokes back to someone for things like logging into different things like banks and so on? Any way to Check for this?
Little snitch is a great tool for that
Hi Matt!
do you know any technique to expand the output's size of the llms answers?
Because currently, the context windows for inputs are fine, however the responses I think are too small.
Can I create a prompt that will generate multiple answers? and mantain consistency between the responses?
Think like telling an llm to generate a full book. Obviusly it would be fine to cut off the content by chapter for example. How can I told in the initial prompt to ask the llm to generate multiple answers (with just one prompt)?
Wrong tool for the job. Ollama would be better for that. The goal of this tool is to provide the next 3-6 words and that’s it.
Cool app. Not a fan of the background music though (in general, not just your choice for this specific video). Thanks for sharing! Also... Obsidian is my favorite application I've ever used, EVER (since I started using computers in 1989).
Can it run a fine tuned model or just the base ones?
He has a limited set of options. Maybe it will get more options. Ask him to add stuff.
Do it exist an open source linux project that do the same thing ? (i will search)
It should be pretty easy to build
@technovangelist perhaps... if can do a daemon with aider.chat^^
@@technovangelist Says a software engineer haha. Will be waiting for the Windows option to try this. Thanks Matt.
We will get used to it so quickly that we will think something is broken when Cotyping is not available.
Well, It automatically translated the audio in french for me. I also tried the german and the spanish version. Pretty accurate translation overall. But I did not like the robotic voice so I went back to english (I prefer your voice)
Great video. Anything similar for Windows?
This only works on M processors :(
Makes sense. Anything ai is going to need a decent gpu.
upvote for split ergo image and the 'and I have never seen STT work well there' ... STT is great while driving your car, but for actual productivity???? yuck! Now I need to see if there is something similar for libreoffice :)
I have a video about using superwhisper which does an amazing job. I tend to forget to use it though.
@@technovangelist Hmm I'll check that out, but to be fair, you're a native English speaker with a very clear, measured voice paired with flawless diction. As a French guy living in Germany whose diction is already rather bad in his native language I'm already happy when Google Assistant understands me ;)
So AI is developing backwards, already? xd
I would hardly say making life better is going backwards
I've been a professional journalist for 56 years and you lost me at "content."
Omg!!! Did I, Cringely/Mark Stephens just comment on my video?!?! Author of Triumph of the Nerds and Accidental Empires??? I didn’t expect such an amazing viewer. I will do better. For you. Or maybe there are many journalists named Robert cringely.
The AI translated version is horrible...
Which one
All of them probably. I'm hearing the german one, and it hurts. Sounds like a 10 minute ad. I mean, probably is, I don't know, only clicked on it to figure out wth is wrong with the audio. Sometimes it speeds up randomly, most of the time it's too slow, and the voice just doesn't match the person. Also pronounces every English word weird.