this post was submitted on 04 Nov 2023
1088 points (95.5% liked)
Memes
45727 readers
894 users here now
Rules:
- Be civil and nice.
- Try not to excessively repost, as a rule of thumb, wait at least 2 months to do it if you have to.
founded 5 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
I didn’t, mind providing links?
Yep. AutoGen + MemGPT (+Locally hosted models) https://youtu.be/VJ6bK81meu8?si=mGnvMTJsLn_vMvRb
Basically, a small company of self-refining LLM prompts that output meaningful results + a robust memory management for more long-term back and forths. Instead of "one input, one output. Next"
Another example: https://youtu.be/5Zj_zstLLP4?si=nHu4vHwidRmvuViY
I can share more examples and papers if desired.
On the robotics front, the focus is still on training custom models for given actions. Which is having some success: https://youtu.be/Jy3zjXK4ao4?si=yFdqnl8z9Z8Becsc
https://youtu.be/WlIYa3lH5UI?si=FQSZAm44h3FuuCoR
I'm convinced these "hivemind agents" will pass custom model training soon
I'm interested!!! I'm SUPER interested!!!
Wow! I honestly can’t keep up with this stuff anymore… It’s insane how fast it’s advancing.
I do it for a living and have given up being up to date on all the new shit going on. Billboard gets the W for now. LLMs cannot build houses in isolation.
What everyone is missing is that they don't have to. The right LLM with the right question can output a meaningful "decision" or "judgment call". That's all you need. Ask the right series of questions. We'll call it "thinking". I really believe that well is pretty deep. Will it be the first version of "AI"? Maybe. Maybe not. But it's gonna be a big milestone that is going to soon fuck up everything.
I can't wait :)
I'll be convinced they can fully replace most trade work when they figure out fine motor control and finally build servo or hydraulic systems that don't act all janky and with slop a mile wide. When they can strip a wire, and then terminate it into a screw terminal, and then install an outlet in the wall. All with one robot, using each tool as needed as finely tuned as a human would do it. And also being able to adapt to different situations on the spot. For instance "shit, the hole for the outlet overlaps a stud, wtf do I do to fix this" type stuff.
From what I've seen even from the best like Boston Dynamics, there are still many decades to go before we have fully capable robotic trade workers.
Many decades? Try five years.
Big lol. Doing unattended trades work is practically the definition of general AI, something we don't see happening any time soon.
Build prefab RTMs in a factory? Today, if the desire was there. Design the house around the line and build it like a car.
Run a new circuit from the basement to an upstairs bedroom, in an old house with weird idiosyncrasies? Not in our lifetimes. The combination of mapping, movement, intuition and the fact that something is guaranteed to go wrong and likely require rethinking the whole job makes this a very hard problem™.
Believe me if someone can invent a robot that can navigate a lumpy, rat infested crawlspace and install pipe/wires/insulation the apprentices of the world will be eternally grateful
If Tesla has given up on fully self driving cars, wherein driving is a much simpler mechanical activity to replicating the full breadth of human construction tasks, then I don't see how people are expecting tradecraft to get replaced by Mr. Fixitron anytime soon.