This is why sideloading addons is so important. They’ve recently removed the bypass-paywalls-clean addon too.

On the desktop version you can easily sideload addons but on the mobile version they forbid this :(

The question is always: What do you want to use it for?

When raspberry started the landscape was very difficult. Small computer boards were expensive, now there’s the N100 if you need a tiny cheap computer. Microcontrollers were really dumb and unconnected, now there’s the ESP32 which has WiFi and Bluetooth and decent performance. Right in the middle of this wide spectrum is the raspberry pi and its clones.

This is a very different situation than in the introduction era where PCs were heavy and expensive and microcontrollers were dumb. There was a much wider niche for the raspberry then. For a small server I would now get a $100 N100 from aliexpress. For embedded electronics I would grab a $10 ESP32. Only in the middle is the raspberry pi, but the problem is, it’s only in the middle in terms of performance, not price. A raspberry pi with case, PSU, storage etc costs more than a decked out N100, while actually being slower.

The only remaining usecase I see for a pi 5 would be an electronics project where you need some more compute than a microcontroller can provide, like some machine vision project. Otherwise:

• Do you want to make some electronics IoT thingy: Get an ESP32
• Do you want a small light computer or server: Get an N100

Yes I was just writing that, I would love to see more integrations that can talk against ollama.

One thing I’d love to see in Firefox is a way to offload the translation engine to my local ollama server. This way I can get much better translations but still have everything private.

If you had a visual disability you would certainly think otherwise.

Or maybe Affinity Designer? I bought that a few years ago for Mac and it was really good.

I don’t think it will.

Microsoft’s endgame is being the lord and master of AI. AI thrives on knowing more data about the user. What good is an assistant if it doesn’t know your habits, your wishes and desires, your schedule and your attitude towards each person in your life?

This is not really a feature primarily aimed at helping the user directly (even though it’s currently marketed as such), but to have the AI build up a repository of knowledge about you. Which is hopefully used locally only. For now this seems to be the case, but knowing Microsoft, once they have established themselves as the leading product they will start monetising it in every way possible.

Of course I’m very unhappy with this too. I’d like to have an AI assistant. But it has to be FOSS, and owned and operated by me. I don’t trust microsoft in any way. I’m already playing around with ollama, RAG scripting etc. It won’t be as good as simply signing up to OpenAI, Google or Microsoft but at least it will be mine.

Yep this is one of the reasons I kept deleting my account even before the whole spez drama.

Too bad they don’t do OpenPGP like Yubikeys do. I still need that even more (much more!) than Fido2. Sites are so slow adopting Fido2.

I don’t use it for email but I use it for SSH and my password manager (“pass”). And yes I know SSH can use Fido2 natively as well but there’s many embedded SSH daemons that don’t support that yet.

Luckily Yubico is still around but I’m betting on them going down the drain (subscription models etc) soon because they were taken over by a venture capital firm :(

I didn’t think it was super creepy but I thought the voice was so overly enthusiastic and overacted and soooo sugary. bleh.

This won’t work for me unless that can be customised and toned down a lot.

The audio from the AI also seemed to cut out a lot during the demo. So it does appear like no shenanigans to me.

It depends on your prompt/context size too. The more you have the more memory you need. Try to check the memory usage of your GPU with GPU-Z with different models and scenarios.

Hmmm weird. I have a 4090 / Ryzen 5800X3D and 64GB and it runs really well. Admittedly it’s the 8B model because the intermediate sizes aren’t out yet and 70B simply won’t fly on a single GPU.

But it really screams. Much faster than I can read. PS: Ollama is just llama.cpp under the hood.

Edit: Ah, wait, I know what’s going wrong here. The 22B parameter model is probably too big for your VRAM. Then it gets extremely slow yes.

Training your own will be very difficult. You will need to gather so much data to get a model that has basic language understanding.

What I would do (and am doing) is just taking something like llama3 or mistral and adding your own content using RAG techniques.

But fair play if you do manage to train a real model!

Yeah she clarified that literally, it’s not linked in the article.

https://twitter.com/RealSexyCyborg/status/1677480809450835969

I can’t find the source of her saying it was about the IME thing but I recall reading that from a person close to her. She had just raised it before all this happened. Edit: Oh wait, that’s here: https://skepchick.org/2023/08/maker-naomi-wu-is-silenced-by-chinese-authorities-and-why-i-blame-elon-musk/ (This was linked on wikipedia)

And yes she’s a great person, she was often criticised for being a CCP stooge but that was BS. She was as outspoken as one can be being in China (and unfortunately, clearly a bit more than that).

This is something that the amazing Naomi Wu brought up for years before, and was ordered to stop publishing by the local government. It was about the same thing. It’s sometimes misrepresented as being about Signal, but her point was: There’s no point in using a secure messaging app like Signal if your keyboard (IME) leaks everything you write! So she was making the exact same point as in this article.

I really miss her content. 😢

The LLMs for text are also based on “theft”. They’re just much better at hiding it because they have a multitude more source material. Still, it does sometimes happen that they quote a source article verbatim.

But yeah basically they’re just really good copy/paste engines that work with statistical analysis to determine the most likely answer based on what’s written in basically the whole internet :P It’s a bit hard to explain sometimes to people who think that the AI really “thinks”. I always say: If that were the case, why is the response to a really complicated question just as fast as a simple one? The wait is just based on the length of the output.

In terms of the “theft” I think it’s similar ethically to google cache though.

A lot of translation and summarisation. ChatGPT is extremely good in absorbing a whole mix of comments in different languages and summarising them in English (or whatever other language).

For programming I don’t use it so much anymore because it hallucinates too much, calling APIs that don’t even exist. And when I lower the temperature the output is too sparse.

I’m also trying to build an assistant that can also communicate proactively (I intend to auto-prompt it when things happen and then evaluate if it should cause a message to me). But I need to get a local LLM going for that because running that through the ChatGPT API will be too costly.

Also, a replacement for some of my web searches. Sometimes I just want to know something and it’s refreshing that it can give me an answer (even though it does need to be validated, it’s much easier to do that when you know what you’re looking for!)

That’s true. They actually stopped supporting Nginx recently which really bothered me too because I want to keep using self-signed certs (my server is only reachable internally and I do not want to expose it to the internet). And the new server they use (I forgot which) didn’t really have that option. So right now I’m locked out from updating until I fix that.

And yes it is totally feasible to use upstream! Not a problem at all.

I would recommend to use the dockers though, as the whole debian thing becomes a bit of a mess with different python requirements for some of the bridges. I tried that in a long forgotten past and there is a reason I’m trying to forget that 🤭

Like you I know the ansible playbook has its limits (for example one other thing I run into is that I want to run several instances of the same bridge to bridge eg. 2 whatsapp accounts!) but I do think docker is the way to go. I’m interested to hear how you’re faring though as it’s a long time ago since I tried that.

True, nobody should ever have billions. There’s simply no need for that much money, you can’t ever use it up.