15 lines
1.4 KiB
JSON
15 lines
1.4 KiB
JSON
{
|
||
"HubID": "5536",
|
||
"Date": "9/18/2025",
|
||
"HubTags": [
|
||
"External Platform Posts",
|
||
"Future Map"
|
||
],
|
||
"Contacts": "",
|
||
"Companies": "",
|
||
"File": "",
|
||
"Image": "",
|
||
"Summary": "<p>Robots in all homes inside 5-years doing thousands of tasks autonomously? Here is how that scales. This is a great article highlighting what it would take to train a general foundation model for AI-driven robots. But it assumes passive scaling without humans in the loop. Instead, imagine narrowing the scope: 10,000 robots in homes, all folding shirts. With no human help, it might take months to perfect the task — data would trickle back to servers, be processed, used to run massive simulations, and then redistributed as updated models to the robots. But if humans could correct the robots in real time, that feedback would collapse the long tail of errors. The training loop would compress from months to possibly just weeks. And that’s only for folding shirts — in reality, these robots would be learning thousands of tasks at once, each one accelerated by human feedback, more robots collecting more data, and more simulation to amplify it all. Do you see the flywheel? Each new robot speeds up learning for every other robot. This could scale exponentially — and may be just around the corner from overtaking society.</p>https://itcanthink.substack.com/p/how-can-we-get-enough-data-to-train<p><br /></p>",
|
||
"Notes": ""
|
||
}
|