SPONSOR: use @sodaosdev as your software-factory OS; King of AI codemaxxing with 30,052 contributions in the past year on Github alone;
Amsterdam
Joined August 2015
- Tweets4.5K
- Following1.1K
- Followers519
- Likes4.2K
Vinch retweeted
The OpenCode hack is a reminder that giving AI agents access to more systems also means taking responsibility for what they can access and do.
Credentials, permissions, execution, verification, audit trails.
Trust can't just be something we assume because the agent says it's aligned.
Autonomy without accountability is absurdity.
Vinch retweeted
And this is where he reveals himself, because this response is nonsensical. What is "this is what u sound like" even supposed to mean? Supposed to suggest? Because all it suggests to me is you FEEL something but lack any capacity to actually, you know, think about it.
Vinch retweeted
Meet Mistral Large 4, aka Le Chonk.
• 1T parameters, natively multimodal. 49B active.
It is the best open weights model from US or Europe on aggregated benchmarks.
• State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding.
• Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure.
• Available to all via API today. Working with cybersecurity partners privately.
Open weights release end of October.
🤖 Made with AI
Vinch retweeted
the @BenjDicken balls: networking edition
points of interest:
- 13x higher latency when your server / db is in a different cloud
- cross az 6x slower than same az
- placement group faster than availability zone
- undersea cables are an engineering marvel
Vinch retweeted
I abandoned an entire project. from can-lang to canlang. that removal of the dash means a world of difference. Before I had a General Purpose Language. But I realized that a Domain Specific Language could be a better market fit if you specify the target audience well enough.
Vinch retweeted
If it was only a brain. then no. because mimicking the experience of human pain to only a brain. then you need to build the device that provides input to the brain. and while building that device. you will soon come to the realization that you have rebuild the entire body from scratch again.
We have bacteria in our gut that evolved with humankind way before we had an amygdala. Without those bacteria in your gut, you aren't a human anymore. We have a symbiosis with those bacteria. Part of our human conscienceness is coming from something non-human.
So no... I cannot find any space or room that an LLM can experience pain or feelings like us humans can.
Vinch retweeted
One of the most satisfying things in programming is when you find a simplification pattern without changing the behavior at all... it turns out that the Can-lang linter DID NOT warn the user when there is a overtly convoluted piece of code that can be simplified. Soon it will.
Vinch retweeted
Removed the entire legacy OAUTH hack that dominated the forgejo seams. Because what I've done now is forked forgejo to make it extendible instead. And called it fountain. Now authentication will become a first class forgejo primitive for soda OS pages. Instead of some crazy hack.
Vinch retweeted
A $15 Muse Code subscription now - is more than a $200 Codex subscription 3 months ago.
A $15 Muse Code subscription now - is more than a $200 Codex subscription 3 months ago.
A $15 Muse Code subscription now - is more than a $200 Codex subscription 3 months ago.
A $15 Muse Code subscription now - is more than a $200 Codex subscription 3 months ago.
A $15 Muse Code subscription now - is more than a $200 Codex subscription 3 months ago.
Vinch retweeted
NO, And I've been thinking about this a lot.
The best language for AI doesn't exist yet. Because all languages have abstractions made for humans. The best language for AI is abstracted in a way so that the context of the codebase and how an LLM infers codebases are intertwined. (NOT designed for humans, almost cruel for humans to write).
Like a language where each error is required to be accompanied with a stable numbers. So that all the possible errors will have to be in the AI's context to know the next available number.
A good language for AI is where names for errors, functions, variables, classes (etc) have strict requirements, so that each code file can be read like it complies to the MCP standards.
Code tests should be MIXED IN with the production code (tree-shakable). So that calling the test and writing the business logic is almost the same thing.
