Episode 196 - Gemini is a surveillance nightmare
Added 2024-06-07 14:49:34 +0000 UTCGoogle's LLM Gemini is a massive step back in privacy practices. This is how terrible the AI data collection can get.
Comments
You are absolutely correct, although some useful models can run on a phone. I generally don't believe the AI hype, so I am not betting my future on it. However, I still want this tech to belong to everybody, whether they can run it locally or not. It would allow for more players to balance the scales, you could have publicly or community funded cloud-based LLMs that people with less powerful devices could still access in a permissionless manner.
The Hated One
2024-06-14 13:30:21 +0000 UTCI will look into Apple AI. Google is also doing most of their processing on device, it's called Federated Learning and it's open source. Haven't looked into Apple Intelligence but I would assume it's gonna work very similarly. Apple doesn't have its own chatbot, so it's not 1:1 comparable to Gemini for instance. But what Apple is doing doesn't seem that new to me.
The Hated One
2024-06-14 13:28:00 +0000 UTCRegarding Running LLMs locally on your machine: I think a very small percentage of the population has devices that are even capable of running those models. And even if, they're not useful at all. I saw some demonstrations of on device LLMs running on Linux, but none of them were good enough to do actually productive work, I regularly use OpenAi's GPT 4. Especially for proof reading texts or repetitive tasks it's really useful. I'd say it has already saved me months of work. That time i can then instead use to focus on stuff that actually matters to me. For language learning for example it's amazing, I can ask in which context to use specific words or constructions, it's like a personal tutor which would otherwise not be affordable for me. this is not possible with current on device models. Current on device AI models are also not multi-modal as far as I know. The amazing thing about Chatgpt is that it can understand images. Instead of manually inputting vocabulary to learn into my vocabulary app, i now simply just take a picture of my hand written notes and then make chatgpt create a typed vocabulary list for me. I really hope for a future where we can just use on device AI which never connects to the cloud, but right now it's really far from being ready for everyday use. They need to become a lot more efficient and our devices need to become more powerful. I see that for your usecase it's not that useful because research is really sensitive and i also wouldn't rely on an AI to inform myself on a topic or trust it to summarize a text about an important topic. But right now i think for most people AI is a chance to save a lot of time doing repetitive, boring tasks, and the benefit of that outweighs a lot of the cons if they use it in the right way. So until on device AI is ready, i think it would make more sense for people to still use those services, but protect their privacy instead of going out of their way to make some half baked software work on their devices. For example: you can use ChatGPT without an account and connect using tor (at least as of today they let you do that without any problems). If you are conscious of what information you are inputting and don't enter anything personal, it's a pretty acceptable from a privacy perspective in my opinion. So until on device AI is ready i think it would be helpful if you tell people that if they're gonna be using these services nonetheless how to best protect themselves.
CopyCat
2024-06-13 11:15:40 +0000 UTCI would be really interested in you analyzing Apple's new Apple Intelligence, which they announced at WWDC. I read their article on Apple's Security Research Blog and it seemed pretty interesting. Of course it is still proprietary and they want us to blindly trust them, but it actually seems a lot better than what Google or OpenAI are doing. Apparently a majority of the processing is done on device. If you could find out which requests are actually on device and which are send to the cloud, that would be interesing. Also I would like to hear your opinion on "Apple Private Cloud Compute". Again seems a lot better than what the competition is doing, but i would love to see a deeper analysis. Mental outlaw already published a video on it, but a lot of the information was factually wrong and he didn't do any research on it at all.
CopyCat
2024-06-13 10:54:44 +0000 UTC
