@fchollet Curious about the epistemological side of this.
We know reasoning can operate over abstractions once they're available, but why do you think:
Symbolic learning can discover the right abstractions efficiently from experience?
What changed?
Why it didn't work before?
I have been trying to build memory for agents and realised - human memory is build with goal/purpose - you can't just create a .MD file or a graph or entities - that's just summarisation.
@antirez Beautiful.
This is the first time experiencing a local agent doing something end to end. I wish one more iteration where this starts to work on my 64 GB machine - I am hopeful this will happen.
@karpathy I have been working of a version of this - that runs completely on the browser. No Obsidian, Only personal knowledge, no internet scraping, quite fast and a probably less intelligent than what you have here
@t_blom This is amazing.
One thing I personally like is - push to talk - so that the AI doesn't interrupt me - In this case it didn't even if I took long pauses.
I like this visualization. But maybe worth noting that the circle is really more like a quite high-dimensional sphere—and most of the volume of a high-dimensional sphere is concentrated near its boundary.
@CloudflareDev@DeepgramAI Trying to use this - haven't been very successful - Back to Gemini Live. Would be helpful to have a demo repo which does end to end voice agent using flux on cloudflare workers
@vaibhavshn@CloudflareDev@Cloudflare Lovely.
If only someone creates a demo for IOS/Android. All realtime SDK are filled with Daily or LiveKit.
Would love to use Cloudflare given everything is already possible using Durable Objects
Here's how Gemini 2.5 Flash-Lite writes the code for a UI and its contents based solely on the context of what appears in the previous screen - all in the time it takes to click a button. 💻 ↓
@kyutai_labs This is absolutely brilliant.
I have been trying Gemini Live and GPT-realtime but they are too costly and voice is not natural enough for casual talks.
How big are these models. Will you also release docs for how to selfhost?