Months ago, me and a few buddies created https://t.co/H1xglOEIdV for a Deepmind Hackathon. We thought about making our tool into a real application once we saw just how powerful it was, but never followed through. Really cool to see someone else working on this.
Introducing CarSignal: The AI operating system for auto shops!
Getting your car fixed is a many-step process and every one of those steps is pretty annoying. It turns out repair shops find it just as annoying. CarSignal uses AI agents plus a new new OBD-II scanner to smooth out all the bumps in the road to getting a car fixed.
@avec Avec is awesome. I use it daily, and it feels nice to have more AI-primitive applications that aren't just a chatbox. Is there an ETA when the desktop app will come out?
Well deserved. One of the few tools that "just works."
At Google, Wispr got me into such a strong habit of talking to my laptop that I had to stop working at my desk and start working out of meeting rooms.
.@ExaAILabs CEO Will Bryk says we can reach the moon, but still struggle to find good information:
"In high school, I was like, 'I'm gonna start a new type of news organization,' because I have had this frustration... We can get to the moon and split the atom, and yet if I wanna understand what's going on at the border right now, it's really hard to. That to me is just an error in humanity."
"Why is everything politicized? I always thought this was ridiculous, and I don't accept it. I think that's just an error in our coordination structure. Something's wrong here."
"I think a lot comes back to our information tools. If there was some really well-funded news organization that didn't have bad incentives and was fully devoted to the truth, I think that would have a huge impact in the world."
"Me and my co-founder built a search engine in college together, way before Exa, because we were frustrated that it's so hard to find high-quality information."
@WilliamBryk on @KnuckleUpHQ
It’ll be interesting to see how open and closed-weight models diverge on memory.
With open weights, you can try creative approaches like Doc-to-LoRA or Code-to-LoRA where instead of stuffing the same documents into context, we compile some of that knowledge into the model itself.
Closed labs can do this privately too, of course. But as a user, all you'll typically see from them is bigger context and more compute.
The funny outcome may be that the less elegant approach still wins.
My hot take is that markdown files are a horrible way to structure memory.
And that's why agents make so many stupid mistakes.
They perform the same way a human amnesia and having to read a bunch of docs before doing a basic task would: poorly, cutting corners.
Intelligence is not more data.
Indexing docs in a graph (presumptuously called knolwdge graph) is just better search. It doesn't solve the fundamental problem.
The rate of progress in LLM optimization is just mind-blowing! 🤯
Running 13B LLMs like LLaMA on edge devices (e.g. MacBook Pro with an M1 chip) is now almost a breeze.
I remember when this looked like a distant future not too long ago. I feel old lol. 😅
1/ 🧵👇