@jorisgio.bsky.social
www.javierchavarri.com/llms-and-pro...
Playing the devil advocate 🧵
Playing the devil advocate 🧵
LLMs, programming languages and other tools: a guess at the future
How LLMs as coding assistants are changing which programming languages, patterns, and architectures make sense. From the rise of explicit code over clever abstractions to the shifting economics of tes...
www.javierchavarri.com
June 21, 2025 at 10:12 AM
www.javierchavarri.com/llms-and-pro...
Playing the devil advocate 🧵
Playing the devil advocate 🧵
There is likely a massive amount of information that would make a difference for LLM that are probably not part of the training set for privacy and secrecy reason. Teams/slack history of thousands of companies. Email accounts, etc. My understanding is that most training side are on public data.
June 12, 2025 at 11:42 PM
There is likely a massive amount of information that would make a difference for LLM that are probably not part of the training set for privacy and secrecy reason. Teams/slack history of thousands of companies. Email accounts, etc. My understanding is that most training side are on public data.
I guess deploying national guard to control riot is essentially a political statement more than a practical thing. It doesn't actually correlate with actual level of violence. See France, where the riot police was almost militarized, so we can claim to be civilized and not deploy the military!
June 10, 2025 at 11:20 AM
I guess deploying national guard to control riot is essentially a political statement more than a practical thing. It doesn't actually correlate with actual level of violence. See France, where the riot police was almost militarized, so we can claim to be civilized and not deploy the military!
Ai will change the world.. it is a fact. Wake up. You can have Ideal, disagree with the processing or the goal. But it is just a fact. Wake up. It will change society, it already did. Wake up. The darkest timeline is harsh.
June 3, 2025 at 10:25 PM
Ai will change the world.. it is a fact. Wake up. You can have Ideal, disagree with the processing or the goal. But it is just a fact. Wake up. It will change society, it already did. Wake up. The darkest timeline is harsh.
youtu.be/sQvLjJkSv44?...
I really enjoy listening to F1 people talk. Not that I like F1 that much but F1 is a performance orientated engineering competition. And those people usually have very interesting to say.
I really enjoy listening to F1 people talk. Not that I like F1 that much but F1 is a performance orientated engineering competition. And those people usually have very interesting to say.
Flavio Briatore | Beyond The Grid | F1 Official Podcast
YouTube video by FORMULA 1
youtu.be
June 3, 2025 at 8:01 PM
youtu.be/sQvLjJkSv44?...
I really enjoy listening to F1 people talk. Not that I like F1 that much but F1 is a performance orientated engineering competition. And those people usually have very interesting to say.
I really enjoy listening to F1 people talk. Not that I like F1 that much but F1 is a performance orientated engineering competition. And those people usually have very interesting to say.
Reposted
The Stanford researchers found that GPUs were utilizing only about 50% of their potential bandwidth due to the overhead from launching and terminating hundreds of small kernels. By consolidating these into a single large megakernel, they increased GPU utilization to 78% of its potential bandwidth!
May 28, 2025 at 1:30 AM
The Stanford researchers found that GPUs were utilizing only about 50% of their potential bandwidth due to the overhead from launching and terminating hundreds of small kernels. By consolidating these into a single large megakernel, they increased GPU utilization to 78% of its potential bandwidth!
When we talk about ai alignement, we usually talk big scale. Ethics, Asimov, the tramway problem. 2001, a space odyssey. Large goals large misalignment.
But there is a form of alignement that is much more trivial, in the human/ai interaction process, in the small interactions.
But there is a form of alignement that is much more trivial, in the human/ai interaction process, in the small interactions.
May 28, 2025 at 9:09 AM
When we talk about ai alignement, we usually talk big scale. Ethics, Asimov, the tramway problem. 2001, a space odyssey. Large goals large misalignment.
But there is a form of alignement that is much more trivial, in the human/ai interaction process, in the small interactions.
But there is a form of alignement that is much more trivial, in the human/ai interaction process, in the small interactions.
When i look at people, all i can see is people.
May 27, 2025 at 7:50 PM
When i look at people, all i can see is people.
newatlas.com/ai-humanoids...
I don't fully agree with the article. But as a dev, I find the AI chat pattern really frustrating in practice. In a way, AI does not want you to work with it. It wants to work for you, and take all the space.
You always need to be harsh and directive with it, guide it.
I don't fully agree with the article. But as a dev, I find the AI chat pattern really frustrating in practice. In a way, AI does not want you to work with it. It wants to work for you, and take all the space.
You always need to be harsh and directive with it, guide it.
AI is rotting your brain and making you stupid
How are you using new AI technology? Maybe you're only deploying things like ChatGPT to summarize long texts or draft up mindless emails. But what are you losing by taking these shortcuts? And is this...
newatlas.com
May 26, 2025 at 3:27 PM
newatlas.com/ai-humanoids...
I don't fully agree with the article. But as a dev, I find the AI chat pattern really frustrating in practice. In a way, AI does not want you to work with it. It wants to work for you, and take all the space.
You always need to be harsh and directive with it, guide it.
I don't fully agree with the article. But as a dev, I find the AI chat pattern really frustrating in practice. In a way, AI does not want you to work with it. It wants to work for you, and take all the space.
You always need to be harsh and directive with it, guide it.
The 3 phase of a street photography session :
1. First, you are not 'in it'. You are in normal mode. But it is a sport, you need to warm up.
So you need to just shoot. Those pictures will suck but it does not matter. Break the glass, start looking, drop the social anxiety.
1. First, you are not 'in it'. You are in normal mode. But it is a sport, you need to warm up.
So you need to just shoot. Those pictures will suck but it does not matter. Break the glass, start looking, drop the social anxiety.
May 25, 2025 at 5:40 PM
The 3 phase of a street photography session :
1. First, you are not 'in it'. You are in normal mode. But it is a sport, you need to warm up.
So you need to just shoot. Those pictures will suck but it does not matter. Break the glass, start looking, drop the social anxiety.
1. First, you are not 'in it'. You are in normal mode. But it is a sport, you need to warm up.
So you need to just shoot. Those pictures will suck but it does not matter. Break the glass, start looking, drop the social anxiety.
Claude feels like the best model for code in practice. Especially because it is really good at tool uses. But when you look at performance per dollar, it's not clear. It's insanely expensive compared to Gemini or gpt 4.1.
It seems the max 100$ subscription is a better deal than pay per use.
It seems the max 100$ subscription is a better deal than pay per use.
May 23, 2025 at 2:12 PM
Claude feels like the best model for code in practice. Especially because it is really good at tool uses. But when you look at performance per dollar, it's not clear. It's insanely expensive compared to Gemini or gpt 4.1.
It seems the max 100$ subscription is a better deal than pay per use.
It seems the max 100$ subscription is a better deal than pay per use.
C++ build system and package management is so fucked up that even llm can't figure this shit out. 4h in this mess, kill me please
May 22, 2025 at 9:52 PM
C++ build system and package management is so fucked up that even llm can't figure this shit out. 4h in this mess, kill me please
The more I think about valve, the more I wonder. How can they be so consistently good with so few people ?
May 20, 2025 at 4:03 PM
The more I think about valve, the more I wonder. How can they be so consistently good with so few people ?
I really dislike Adobe, but lightroom and Photoshop are a very good examples of how to introduce IA in a product successfully.
Gradually, in the existing workflow, solving actual problems, not too disruptive. Claude code and cursor. Are also very good examples of such success.
Gradually, in the existing workflow, solving actual problems, not too disruptive. Claude code and cursor. Are also very good examples of such success.
May 19, 2025 at 12:41 PM
I really dislike Adobe, but lightroom and Photoshop are a very good examples of how to introduce IA in a product successfully.
Gradually, in the existing workflow, solving actual problems, not too disruptive. Claude code and cursor. Are also very good examples of such success.
Gradually, in the existing workflow, solving actual problems, not too disruptive. Claude code and cursor. Are also very good examples of such success.
Engineers and scientists do not change the world. Engineers build the tech. Engineers build the foundations.
The world change when product people use the tech to meet a purpose.
Tim berner lee didn't change the world, but the web did.
Tony fadell and Scott forstall didn't, Steve and the iphone did
The world change when product people use the tech to meet a purpose.
Tim berner lee didn't change the world, but the web did.
Tony fadell and Scott forstall didn't, Steve and the iphone did
May 17, 2025 at 9:17 PM
Engineers and scientists do not change the world. Engineers build the tech. Engineers build the foundations.
The world change when product people use the tech to meet a purpose.
Tim berner lee didn't change the world, but the web did.
Tony fadell and Scott forstall didn't, Steve and the iphone did
The world change when product people use the tech to meet a purpose.
Tim berner lee didn't change the world, but the web did.
Tony fadell and Scott forstall didn't, Steve and the iphone did
Programming languages have a 'personality', in part coming from the original goal of the language authors but also from the community that builds around it.
When you choose a programming language it is probably more important to think about that personality than about the set of features it offers.
When you choose a programming language it is probably more important to think about that personality than about the set of features it offers.
May 17, 2025 at 6:08 PM
Programming languages have a 'personality', in part coming from the original goal of the language authors but also from the community that builds around it.
When you choose a programming language it is probably more important to think about that personality than about the set of features it offers.
When you choose a programming language it is probably more important to think about that personality than about the set of features it offers.