live improvisation from a random preset
#synthsky #synthfam #synthesizer #synth #music #experimentalmusic #sounddesign #nozoid #kagouyar #electronicmusic #analogsynth #dronemusic #sounddesign #synth #noisemusic
live improvisation from a random preset
#synthsky #synthfam #synthesizer #synth #music #experimentalmusic #sounddesign #nozoid #kagouyar #electronicmusic #analogsynth #dronemusic #sounddesign #synth #noisemusic
1) A standard agent was able to autonomously form hypotheses, test them, and converge on a solution that was better than the one initially done by a human.
2) The model is shipped to production.
It feels like a milestone to me :-)
5/5
LLMs won't create breakthroughs this way, but they can help humans explore a space of possibilities much faster, at least in some cases.
And again, this was a very simple model and research scenario, but two facts remain true:
4/5
When an LLM can create a plausible experiment and access a detailed report of the results, it can develop another plausible experiment.
The LLM doesn't need to be smart; it just needs to generate plausible answers and a way to measure progress.
3/5
Disclaimer:
This was not one of our best-in-class diffusion models, which still require extremely talented people to create and a lot of compute to train.
However, it still demonstrates the power of feedback loops.
2/5
Last week in our lab, an AI agent autonomously optimized an AI model, which we then deployed in production.
1/5
Well done! Very impressive!
Clearly not everyone is able to do this.
One more proof that AI is just a talent multiplier.
It's great, especially the free credits from opus 4.5
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*)
sonnet 4.5 for the win
Your AI writer is better than average. What do you use?
I'm in the process of replacing all my personal subscriptions with vibe coded solutions
All of my website are vibe coded: you can find all of them from there: damien-henry.com
When I bored, I'm asking question to my digital counterpart.
There is more, read the full post here:
damien-henry.com/about-ai-st...
16/16
This is a move from scarcity to abundance. Vibe coding makes computation so accessible that you can create personalized solutions on the fly, without needing to think about generalization.
15/16
You could generate a brand-new, perfectly tailored piece of code every single time you need it.
14/16
Because code was hard to create, we built tools meant to be reused. If it took you ten days to write a script to process a video, you would reuse that script every time. But what if generating that code took only ten seconds? You would no longer need a reusable tool.
13/16
The true revolution is that you don't need to build a generic tool at all.
12/16
This is where vibe coding changes the game. At its core, itβs a translation tool. It uses the power of the transformer to translate your thoughts from natural language directly into code. The revolution isn't about building the next generic SaaS tool faster.
11/16
This model gave rise to the SaaS industry, where companies sell best practices packaged as software. You delegate the problem-solving to them. This works, but itβs always a compromise. Youβre using a tool designed for a broad audience, not specifically for you.
10/16
The spreadsheet is a perfect example. A few skilled engineers built a powerful, generic tool for complex calculations that millions can use without knowing how to code it themselves.
9/16
Because writing code has been difficult and time-consuming, developers have focused on creating abstract, one-size-fits-all software. The goal was to build a single product that could serve millions of people.
8/16
This scarcity creates a few problems. First, it puts immense power in the hands of a small group of people. Second, it forces us to rely on generalized solutions.
7/16
Even though our lives are surrounded by codeβin our phones, cars, and homesβthe ability to write it remains a specialized skill.
6/16
Think about reading and writing. Humanity made a huge leap when literacy became widespread. It enabled shared knowledge, accelerated progress, and connected people in new ways. Code, however, has not had the same journey.
5/16