Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me.
I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA.
I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation.
Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO.
Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours.
Jensen
The only way for outsized returns in the presence of superhuman AI is to take negative/suboptimal NPV bets.
AI will take the optimal NPV bet anyway and is probably better at identifying it than humans are -> This is the baseline
To beat AI one has to take a bet with a higher ceiling and lower chance of success.
If you win that, you will look like a genius for outperforming superhuman intelligence.
Robotics is solved
VLMs have hit escape velocity:
Astra generalizes across tasks and embodiments.
-> It can do RLVR in simulations
There is still a huge gap between spatial control capabilities and knowledge in the LLM backbone
-> Backbone is able to create tasks to RL on in simulation + verify them
-> The gap will become 0 in the next 3-6 months
From there the only tasks robots cannot solve, are those the VLM doesn’t have context on.
There will still be gaps regarding speed and cost that might make sense to bridge with scaffolds in the month to come.
These scaffolds will become irrelevant as models follow their trend of more faster intelligence per cost.
@NikolausWest it depends on your definition of solved.
I believe a concept is “solved” when there’s a clear way to large societal impact.
eg the car was solved when first manufacturing for the model T was built
@NikolausWest I didn’t solve it. (microagi is not training frontier VLMs like the 5 labs that do)
It’s just a consequential action which is observable.
And again: trajectory =/= intercept
@NikolausWest I’ll happily take a bet on if we have models capable of doing economically valuable work without post-training that originate from VLMs in 6 months.
There’s a lot of well-funded companies doing it.
The question was never “if” but rather “when” and “what”.
Also, the same knowledge limitations will apply to simulators as for VLMs that can only be solved with more real-world experience.
Think that comes down to definition.
With VLMs I mean current frontier multimodal models (eg Astra). These are also trained on videos etc.
So many of the described shortcomings in the paper should not apply?
Also, the backbone at scale should be better at predicting the future. (Text should hold value for that)
@GeYunhao
@Sir_Dhruv_Da1 +1! although it feels like experiments/real-world experience might still be the bottleneck there. (models can’t simulate + verify new materials from what I know)
@Sir_Dhruv_Da1 think the comparison is correct. the change is rather that the shape of how to hillclimb robotics was unknown (architecture etc). this changed now.
perfection will take a while - although physical movements feel like they have a lower skill ceiling than coding
I’m talking about the trajectory of VLMs not the current intercept.
You can ask your current frontier model to explain tying a shoe to you and design verifiers around it -> it will do well.
Also, every single step of tying a knot is solvable by VLMs in 1/16 tries with limited DOF embodiments -> it’s hillclimbable
That’s the process I’m describing.