Last week in our lab, an AI agent autonomously optimized an AI model, which we then deployed in production. 1/5

Of course, you can ask ChatGPT for advice, and it will reply with all the banalities from the web. Ask @paulg instead, and you'll get a MUCH better answer. www.chatwithpg.com/

I ask an LLM to watch all the video of @karpathy and write a blog post. Here it is: damien-henry.com/about-ai-st...

"If something is not clear, please ask" works well on humans, but even more on LLMs

It's not surprising that LLMs excel at translation since the transformer was originally designed for it. The only surprise is how well English-to-code translation works!

Anti-gravity is amazing, but sonnet 4.5 is still better at coding than gemini 3 high. (and GPT*)

When I bored, I'm asking question to my digital counterpart.

There is a big misunderstanding around "Vibe Coding" Many see it as just another gadget, a quick way to build a proof-of-concept. I think this view misses the point entirely. 1/16

Fun experiment: I recorded myself while riding my scooter to work. Then, I used the transcript and asked Jasper.ai to write a post for me. It took less than an hour from start to publication. You can check the result here: damien-henry.com/about-ai-st...

Apply gradient descent to your life: Do something random. Does it feel better than expected? Do more of it if yes; stop doing it if no. Repeat. I've been doing this all my life, and I'm one of the happiest people on earth.

I asked ChatGPT to write the story of Clipdrop, and it's not too bad! damien-henry.com/clipdrop

I've just added the latest models to this small tool, which helps in understanding the context window sizes of various LLM models. context-windows.damien-henry.com/

i just added a chatbot to my website. It's fun! damien-henry.com

Ok, so the "vibe code it, instead of paying a subscription" is getting real! I was paying $12 per month for this tools that publish a website, using @NotionHQ as a CMS. I vibe coded the equivalent in 3 evenings. More details👇

I replaced this yearly subscription in 3 evening of vibe coding. If I have time tomorrow I'll explain how!

Productivity trick: Uninstall Instagram and avoid watching Netflix. This can save the average user about 3 hours per day. It's incredible what you can accomplish with an extra 3 hours each day.

I'm using the Vercel SDK, but I'm hearing good things about Cloudflare Agent. If you read this and it's bullish about Cloudflare, I'm curious to learn more from your experience. agents.cloudflare.com/

While it's certain that all apps will eventually resemble this format (a chatbot paired with a traditional interaction surface), this particular example from OpenAI is very poor.

Just in case you missed the official launch: Latent Bridge Matching gojasper.github.io/latent-brid...

p(doom) from AI < p(doom) from nukes < p(doom) from climate.

An easy prediction regarding Sora: @OpenAI will likely adopt a business model for Sora similar to what @Google did with @YouTube: redistributing a portion of the generated revenue to IP holders who request it.

Guess what the app is!

The most underrated website on the entire internet: floor796.com From @Floor_796

The rm -rf is not a joke...

If you look closely, you'll realize that all SaaS are, in fact, "Good Practices as a Service."

Old post but still very true. The only meaningful way to think about AGI is thru cognitive cone. damien-henry.com/writings/re...

Space-time is bounded by the laws of physics. The digital space is only limited by our Imagination.

This was too easy. github.com/dh7/simple_...

I made this tutorial 9 years ago! IMO, the one about RNNs is still a good resource to get started before you jump to the ones made by @karpathy. github.com/dh7/ML-Tuto...

Algorithms describe paths between two states in the universe. Some algorithms provide directions to known states. Some algorithms are shortcuts. Some can lead to destinations not yet discovered. Nothing can change the world more than discovering a new algorithm.

@drmichaellevin makes me reconsider everything I know about life and minds. It's like watching the very forefront of human knowledge unfold, talk after talk. If your weekend is not over, watch this: www.youtube.com/watch?v=1gZ...

I just realised that context-windows.damien-henry.com/ was getting some traffic, so I ask cursor to polish it a bit.

I made some small edit to this old post that describes how to create value using AI. (MCP was not as big in Jan, so I added a small note about it) damien-henry.com/writings/th...

In a conversation, the bigger faux pas isn’t replying with ChatGPT’s words — it’s asking without having consulted ChatGPT first.

All my ChatGPT5 answers are like: Title, bullets bullets, title bullets bullets, title numbers numbers, title bullets bullets. Why does paragraph 3 have numbers instead of bullets for no reason?

Human rights aren’t optional. Join the fight. Support @amnesty ✊ 👉 www.amnesty.org/donate

If you’re tired of watching Israel trample on human rights and wondering how you can make a difference — support the fight for accountability. Donate to @amnesty today. ✊💜 👉 www.amnesty.org/donate

You could be like Jesus. You could save the world with just your love. But would you want to die on a cross with nails through your hands? The Bible makes for a rather poor motivational book. Is it on purpose?

Na! Google Cardboard will never die!

2022: “LLMs are just parrots, repeating words without understanding.” 2023: OMG! They are surprisingly good chatbots! 2024: Wait… they can code?! 2025: Agents!!!! Just connect them with tools! Anyone want to share a bet about 2026?

One of my favorite content on the web is johnhw.github.io/umap_primes/... It is a classic, but if you missed this representation of 1 million integers, you should deep dive now!

Why use a CMS if your website is entirely generated with Cursor? Why not let Cursor create your website's content as well? developers.jasper.ai/docs/jasper...

Multidimensional spaces are famously difficult to represent. Most of the work done on them involves equations, and you simply have to accept the results because, well, they are true. However, I wanted to build some intuition and see if an LLM could help.

All modern AI is based on high-dimensional spaces, so here is a collection of counter intuitive, fun facts, about them. For instance in hight-dimensional space, the volume of a unit sphere converge to zero!

Elicit is my new fav word.

When I read or listen to something, my brain automatically annotates it. Already known or expected from a source? => Boring, fast-forward. Bullshit? => Skip and distrust the source. The unclassified content that remains is the only thing I care about.

If you are designing AI Agents, you must read this! manus.im/blog/Contex...

Israel is committing genocide @theguardian.com www.theguardian.com/commentisfre...

Where are the gradients pointing to?