They probably will. But will be caught with pants down again as sales in cities stall due to Waymo and Robotaxi. The consumer will change faster than they are expecting. Saw it in Austin this weekend.
@scobleizer
-
Driving Model Distillation Has Implications for Resource Constrained Robots
By
–
It really is amazing technology. Has deep implications for Optimus. They are practicing distillation because the robots will have shitty computers. 🙂
-
Unexpected Learning Is Described as Models Changing Their Weights
By
–
That is what I was told. I think he was trying to say they started moving their weights around. Either way they started learning in an unexpected way.
-
Robots and Brain Interfaces Could Increase Belief in Machine Consciousness
By
–
Thank you. And wait until the robots and BCIs arrive. More and more humans will believe they are conscious. Probably even if we froze all today’s LLMs. But of course they will be way better in four years than today.
-
Frequent Robotaxi Use Changes Car Ownership Habits
By
–
I begged Grant to get one, he loves it now and says he takes Cybercabs everywhere and using his car less and less. Same as what Waymo customers in San Francisco tell me. Cybercab is a bigger deal than most know about or recognize.
-
Distilled Driving Models Improve Older Tesla Cameras and Hardware
By
–
I gave my friend Jeff my car for past 10 days. He updated it last night. Now it drives very similarly to a Cybercab. It is very hard to tell the difference. @aelluswamy and his team are underrated. Think about what that means. A distilled model running on shitty cameras and
-
A Driver Reports Full Self Driving Has Become Smoother
By
–
That was eight years ago. Now FSD is way smoother at driving than I am
-
Cybercab Riders Compare Repeat Trips Favorably With Uber
By
–
I am interviewing Cybercab users. Most say they have used it several times already. Those who did their first ride say it was way better than uber.
-
SubQ Snap Makes Decisions Over Whole Large Documents Without Chunking
By
–
Most AI tools chop a long document into pieces and hope nothing important falls between them.
— Robert Scoble (@Scobleizer) 12 octobre 2026
Snap from @SubQ reads up to 384K tokens in one go. Give it 100K tokens and it makes a decision in under a second.
On a long context test at 256K tokens it scored 99.8 against Jev's… https://t.co/Snc82DgjwIMost AI tools chop a long document into pieces and hope nothing important falls between them. Snap from @SubQ reads up to 384K tokens in one go. Give it 100K tokens and it makes a decision in under a second. On a long context test at 256K tokens it scored 99.8 against Jev's
-
Commentary describes efficient iterative non LLM learning for robotics and BCIs
By
–
What he is building is an entirely new AI. Not LLM based. But, yes, it learns iteratively and quickly, with far less compute needed. Perfect for robotics and BCIs which will be scaling up as his system gets smarter and smarter.