Digital Sovereignty is important for everyone. We should explore local models to avoid data exposure. Understanding the differences between open weights and closed systems is crucial. What steps are you taking to ensure your data privacy?
At this point, I’m testing what the open models can do. Some months ago I cancelled my OpenAI subscription, and I’m getting ready to drop Anthropic as well.
What I’m finding out is still meh… for example, the only European model charges more to EU citizens! In USD is far cheaper.
I’m also learning more about orchestration and updating my knowledge to add to my consulting work.
I love the framing of Digital Sovereignty Jose! I’ve been trying to move to a place where I can run all of my models locally and get off the token wheel. Any particular tips for this in your experience?
At this moment I’m testing different models, to see if they are good enough for some of my tasks… and I’m surprised. I would say that even Mistral, that is not ranked among the best by far… is good enough for most tasks.
I would suggest that is a good first experiment we can do.
I’m using OIllama as interface to work with the different models… lots of options and it works pretty well.
My next step is to stop paying Claude, and see if the new Sonet 5 is good enough with the free version.
I think I will use the local version of n8n for orchestration
On Perplexity... I have since stopped using it because it has become problematic in the last couple of months. There has been a lot of controversy around it, a lot of lawsuits. It's been working really poorly lately.
In my toolstack (that I use using APIs/MCPs) I use Tavily, Exa and Linkup; meaning I use them directly within Claude Code (I have a research skill)
It's pretty similar to my approach years ago regarding Google. I remember that I was one of the first beta users when Gmail came out. It was a revolution back then. Of course many Google tools followed and I adopted them all. There were no alternatives that came close. Then came the time when I realized that I'm giving away my data to Google and even paying them for doing so. So about 10 years ago, I removed all of it - went from Gmail to Fastmail, Proton and all the privacy supportive tools.
But here's the thing: I only could do this, because the alternatives had catched up. I got the same value from tools like I got from Google - with added privacy. A solid deal.
The current state of AI isn't there yet. None of the mentioned models come even close to Claude. Every time I try, it just takes 5 minutes to realize how inferior they are.
So for the time being, privacy comes second to value - at least for me.
This is probably one of the few pieces where someone digs into the models first so neatly before suggesting tools and how-tos. Great reasoning!! Thank you for being “overly critical”.
This is the link to the NotebookLM I used to do part of my research on Anthropic, and why they aren't clean at all from using their AI models to support "national security" priorities.
Digital Sovereignty is important for everyone. We should explore local models to avoid data exposure. Understanding the differences between open weights and closed systems is crucial. What steps are you taking to ensure your data privacy?
At this point, I’m testing what the open models can do. Some months ago I cancelled my OpenAI subscription, and I’m getting ready to drop Anthropic as well.
What I’m finding out is still meh… for example, the only European model charges more to EU citizens! In USD is far cheaper.
I’m also learning more about orchestration and updating my knowledge to add to my consulting work.
I love the framing of Digital Sovereignty Jose! I’ve been trying to move to a place where I can run all of my models locally and get off the token wheel. Any particular tips for this in your experience?
I’m glad you find that idea useful!
At this moment I’m testing different models, to see if they are good enough for some of my tasks… and I’m surprised. I would say that even Mistral, that is not ranked among the best by far… is good enough for most tasks.
I would suggest that is a good first experiment we can do.
I’m using OIllama as interface to work with the different models… lots of options and it works pretty well.
My next step is to stop paying Claude, and see if the new Sonet 5 is good enough with the free version.
I think I will use the local version of n8n for orchestration
Very valuable post, Jose! Keep them coming!
On Perplexity... I have since stopped using it because it has become problematic in the last couple of months. There has been a lot of controversy around it, a lot of lawsuits. It's been working really poorly lately.
In my toolstack (that I use using APIs/MCPs) I use Tavily, Exa and Linkup; meaning I use them directly within Claude Code (I have a research skill)
I’ll check those tools!
It's pretty similar to my approach years ago regarding Google. I remember that I was one of the first beta users when Gmail came out. It was a revolution back then. Of course many Google tools followed and I adopted them all. There were no alternatives that came close. Then came the time when I realized that I'm giving away my data to Google and even paying them for doing so. So about 10 years ago, I removed all of it - went from Gmail to Fastmail, Proton and all the privacy supportive tools.
But here's the thing: I only could do this, because the alternatives had catched up. I got the same value from tools like I got from Google - with added privacy. A solid deal.
The current state of AI isn't there yet. None of the mentioned models come even close to Claude. Every time I try, it just takes 5 minutes to realize how inferior they are.
So for the time being, privacy comes second to value - at least for me.
Very valid point Atmos!
Nothing makes you feel productive quite like spending three hours automating a task that took eleven minutes and happens twice a year.
Owning the stack probably has to include owning the embarrassment of deleting workflows that only existed because building them felt clever.
Daniel that’s what I call my process of learning by making mistakes 😆
Love it
This is probably one of the few pieces where someone digs into the models first so neatly before suggesting tools and how-tos. Great reasoning!! Thank you for being “overly critical”.
Thanks, Nihal!
I'm so frustrated... I closed my ChatGPT subscription some months ago, and I'm thinking of doing the same with Claude.
I kinda understand ChatGPT but why Claude, too?
Check the first link at the end of the article.
The attack on the girls school in Iran was targeted using Claude
This is the link to the NotebookLM I used to do part of my research on Anthropic, and why they aren't clean at all from using their AI models to support "national security" priorities.
https://notebooklm.google.com/notebook/6e47d247-1afc-42ec-a88a-f9d7b05e60f0
Apparently, pi.dev is very well received by the dev community as a harness/interface
Is this the service? https://hey.pi.ai/
Actually, pi.dev! Just corrected my post
It looks very interesting. Do you use it?
No, not yet. But I’ve only heard great things and it was not on your list, so.
Thank you man, I’ll add it.