ilovefood

You're right about understanding. Where we might disagree is the conclusion. I think it's possible to build the understanding without having to type out syntax by hand (debugging, writing tests, etc). Maybe we're missing much better verification tools? LLMs will likely play a big part in those too (explain the codebase, walk me through A, B, C etc)

I'm really open to feedback. I checked your past comments and you posted:

> AI is good at coding if there's an oracle. If the system is ancient, unreadable, untestable, that's exactly the opposite. It won't get the exact set of corner cases.

I sort of mention this in the article, so I'm sure we're somewhat aligned on the core. How would you have worded things?

Agreed, which is why I also believe "token usage" and similar proxy metrics aren't the right ones for what's next.

Makes sense, perhaps I'll post the raw notes + more material with the next article, prior to any edit. I like to use local models, even if they are not that powerful precisely because of that, I'm forced to put in more thought and it's my signed off article in the end.

Thank you for the suggestion.

A genuine mistake. I make these posts for myself, mainly to structure & share my thoughts. Let me know if you find other errors, I appreciate the feedback.

Yes correct, Gemma. Will correct it shortly.

You are correct, my bad. I meant to write Gemma 4. I'll update.

I can't change my name, but happy to change the hero picture :))

I've written this myself, Gemini 4 did a bit of editing:

> What follows is a cleaned-up version of notes I accumulated over the past year. Gemini 4 helped with the editing.

The image is made by an AI image generator on fal.ai. It's better I spare you all my design skills :)

Thank you very much for the kind words :)

Super cool and well done. They are much better in 3D! :)

I made a similar experiment a while ago and randomized the parameters. Given it's difficult to stumble on a stable arrangement, I turned it into a small game to find pretty ones: (big disclaimer: this involves NFT tech, please skip if you're against that sort of stuff) https://karimjedda.com/symmetry-in-chaos-my-first-generative...

Sharing a small experiment that somewhat attracted a lot of attention yesterday, perhaps you'll like it too.

Thesis: Many devices or apps listen in while not in use. We're assured these only do topic matching (ie: identifying keywords) but who knows.

Observation: Today, generative AI can be used to generate a bunch of things, text, code, video but also audio. So why not bid local algorithms against remote algorithms?

Solution: I put my phones in a metal box i'm dubbing "Gaslight Garage", where a speaker plays randomly generated Elevenlabs audio to the phones. The pipeline is fully automated and sometimes hilarious. While white noise generators and other things are detectable, I believe this is much more difficult to detect on-device.

I've yet to see the results but maybe it'll give the one or the other a few ideas. This is pretty much in line with my local LLM extensions (on my blog) to use local-AIs as countermeasures against remote AIs, they love data, right?

Enjoy your Sunday!

Great write-up! Gives me a few ideas for a governance bot that I'm working on. Thanks for sharing :)

Same problem here :)

I fully agree. It's difficult though because I genuinely believe that the solution space overlaps with cryptography, which is quickly discounted as viable option because it is now laden with negative connotations.

Fair point and it's still a domain of research. The GNU project works on something quite cool I think, called Anastasis. If we solve these things & make them easy, it's a win.

Vault doesn't use emails, usernames or anything like that. It's just shared as an example of a signer. It doesn't have to be that app per se, but it's the closest example I know of.

I can’t tell, I don’t spend too much time on LinkedIn. What makes it feel that way to you?

Whatever option gets us closer to portable and less centralized digital identity, data and payment is a win in my book.

You're 100% right, it's about trust or lack thereof. Thank you for the Star Trek references. :D

Thank you for sharing your perspective, I'll read your article. As long as something good comes out of it for everyone I'm happy. Have a good day as well!

Wrote it myself with some help to formulate things properly. Sorry it came accross as AI-written.

The last thing isn't a wallet, it is a cryptographic signer.

Almost exactly the same score :)

iroh is fantastic tech.

I attended Rüdiger's (N0) workshop 2 weeks ago at the web3 summit in Berlin and was left super inspired. The code for building something like this is available here https://github.com/rklaehn/iroh-workshop-web3summit2025 and I highly recommend checking out the slides too :)

I did something similar a while ago, perhaps the writeup could be useful to you: https://karimjedda.com/llms-in-the-middle-content-aware-clie...

If that's the case then I'll try the platform out :) I want to finetune Codestral or Qwen2.5-coder on a custom codebase. Thank you for the response! Are there some docs or infos about the compatibility of the downloaded models, meaning will they work right away with llama.cpp?

Looks pretty cool, congrats so far! Do you allow downloading the fine tuned model for local inference?

It isn't, the blog was just a high level recap of what I built, as I unfortunately don't have enough time to turn it into a product that can be used.

I'm very happy to see that someone is building it though, I really think that personal AIs/LLMs are the future.

Looks great :) it's exactly what I've built a year ago: LLM in the middle / client side content filtering using LLms.

- https://karimjedda.com/llms-in-the-middle-content-aware-clie...

For solo projects, my code is the project planning. I use a bunch of #TODO or //TODO in different files throughout each project's codebase, that a small Python script picks up and compiles into a single HTML file.

As I'm working on a lot of projects, the important thing for me is keeping it easy to start/stop working on a project, and make sure that when I have time to make a pass on a project I don't have to also deal with the additional headache of keeping the codebase and my todos in sync.

Not sure if this is really helpful though. Good luck!