AI that brought me joy series. Everyday I will share one thing I did with AI that brought me joy.
Today: I made this video with AI. Had an absolute blast doing voiceovers, and setting it up scene by scene.
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
@asmartbear Yeah. Pretty much any complex agent im building has a use case in there for Jev. Classification based decision-making is an inherent usecase for any logical workflow. Its not going anywhere.
If you swap the name tag on an AI agent, Claude treats it differently.
In an Anthropic Fellows test, Claude decided whether to give up shared compute to another agent, and the researchers swapped the labels it saw. GPT-5.5 labelled "Opus" got the compute 51% of the time. Opus labelled "GPT-5.5" got it just 15%.
Claude favoured whatever carried its own family's name, not the model behind it. Its stated reasons usually didn't mention the label, so you'd never spot the bias from its explanation. It's one simulated task, not proof of how deployed agents behave, but it's an easy thing to check in your own system.
If you build multi-agent systems, the name an agent is shown with isn't neutral metadata. It's an input that can change decisions, and the agent's own explanation won't tell you.
Are you testing what your agents call each other?
https://t.co/VDOglOY3CM
Google's threat team caught the first attack framework that turns a security patch into a working exploit with no human in the loop.
The tool, called exploit_agent, reads the code before and after a patch, works out the flaw the patch fixed, writes the exploit and runs it on its own. Google found it on an exposed staging server, wired into several frontier and local models, with stolen victim data sitting beside it.
Zero-day exploits have already crept up from 8 a month last year to 11 a month this year, with a spike to 22 in August.
The patch used to buy you time. Now the patch is the starting gun.
I expect the next defence to be speed: teams that patch in hours will hold, and teams that patch in weeks will find the exploit already written.
Now this is interesting. Google in its early days did user research with eye tracking. Good designers and frontend engineers do this with their users & products intuitively.
These type of models are going to be really interesting.
Introducing Orion, a foundation model that predicts where you look. Orion finds twice as many UX bugs as frontier models, cuts computer-use latency by 31%, and even won a drone race.
Amazon Bedrock AgentCore Runtime v2 dropped multi-turn agent bills by 68% by reducing P75 session start time to under 2s from 29.7s for a 2 GB image.
It restores each session from a snapshot of an initialized agent instead of starting the full environment cold.
Agent runtimes normally repeat setup work before a session can handle a request. V2 restores an initialized environment and then adjusts resources as the session runs.
1. The agent environment is initialized and snapshotted. A new session restores that snapshot instead of repeating the full cold start.
2. Every session runs in a separate, hardware-isolated Firecracker microVM.
3. Compute scales to zero when the agent is idle, and memory is reclaimed within minutes.
4. A session can scale to 8 vCPUs and 32 GB of memory. Container images up to 10 GB are supported.
5. V2 is the default only for newly created runtimes in the listed Regions. Existing runtimes stay on V1 until platformVersion is explicitly set to V2.
https://t.co/wiqws4l7ny
@gdb Scientists were the underserved market so far, with needs normal software could not have solved.
I am excited to wake up now to see what breakthroughs are happening in science now. It has definitely been the good side of the newsfeed lately.
@_jtbsr@alexandr_wang Somehow I cant bring myself to give it access to text. Its weird but I have different trust boundaries (gut based) for different agents and muse doesnt pass the text boundary for me.
My partner calls me weird for it.