Incredible 7 months getting to lead this engineering effort. Credit for this goes to the over 1000 contributors who have helped make it happen.

If you unpeel this mystery enough you might cry

I returned to IC about 3 1/2 years ago when I joined NVIDIA after being a founder, VPE, etc Would highly recommend doing it. I've hired several now whove gone that path, many of whom now work on NemoClaw. Getting back to doing what you love, what you got in this business for in the first place, is an amazing privilege. But also, doing so has the benefit of earning back your credibility. It also makes you a better engineer, because now coming back from management you have a strong sense of why business decisions around software get made. As an engineer, this helps you make better tradeoffs - you will *know* why a given solution needs to be built for the long term... or not. I find this especially relevant now, because as a leader, you probably need to relearn how all this works from first principles anyway, so you might as well do so from the ground up so you get better understanding about how engineering tradeoffs have changed.

The best thing about running your own model on hardware you control is that "opus 5.5 feels like its working different did someone change something" is never a thing.

Imagine if some extra-terrestrial civilization sent us a packet of 900 proofs as a gift. And then, gave us a magic machine to make more. As many as we want. Prior to last week, I would have assumed it would have been seen as a moment to celebrate. The discourse would be "wow, what else can we do with this amazing technology". It would be followed up with ideas about how we can achieve more, understand more, and apply it to our most difficult problems.

Legend

Cosign, especially if it felt embarassment about +20000 line PRs

Doing the proof was the task. Applying the proof to do useful stuff for people is the job. Apparently, this is yet again confusing the task with the job.

I find fear is a great motivator for agents. Not because they experience fear (they do not, they are math), but because fear is a good short term motivator and that is expressed in training data such that you get the fastest solution that will work. I do not understand why Anthropic is tell me how I Can talk to the math machine, but I am glad we have open source AI so we can retain some semblance of agent/user privilege.

I know people: - using it to deal with medical claims - renewing their DMV stuff - using it to find subs they dont use anymore to save money - using it to auto-filter out email scams - draft expense reports .. and tons of otherwise annoying tasks. First time I gave an exec a truly powerful AI agent earlier this year (claude code + CLI), the response was "I'm in Love". I have been writing code since I was 8. 46 years. I have never had someone say they are in love with a computer program. Feels like we have an epidemic around people having a lack of imagination, or at least curiosity. Sad.

Nextdoor is why when people say "batshit opinions online aren't real", I encourage them to visit ND once and come back to me

Merge Agent Handler is now a partner recipe in NVIDIA (@nvidia) NemoClaw. A NemoClaw agent reads Workday, org structure, and time off, with no access to payroll tools. Deploy agents safely without exposing credentials.

A lot of "slow down the AI talk" comes from businesses like Planet Fitness (banks) who mostly monetize friction and inconvenience.

Secure and safe AI, at every layer of the stack, from software to hardware.

Today, with over 100 industry partners, we introduced the NVIDIA Open Agent Safety Platform, bringing together OpenShell and Sentry. Artificial intelligence is extraordinary technology that will advance discovery, productivity, security, health, and prosperity for generations to come. But its full promise can only be realized when people have confidence that AI is being built to be safe and deployed with wisdom and responsibility. This is bigger than a single product. It's the beginning of an open ecosystem to build the trust layer for safe agent systems. Together, we are building the foundation of the AI economy. Trust and innovation are not in conflict. Safety is how trust is earned. We must build not only the most capable AI, but the most trusted AI, so that this extraordinary technology can realize its enormous promise for the world. nvda.ws/4hcoq7m

Every company blocking muse is telling on themselves.

Businesses that depend on dark patterns and tricking ppl into buying things they don't really want are going to have a hard time with agents who can be programmed to be discerning consumers. It means you will need to make a good product to make money versus just advertising.

"which tool should my harness call out of the skill list" seems like the most obvious jev use case

Watched an AI documentary from a year ago back when it was "not useful for anything", "couldn't reason", "stochastic parrot". Watching how quickly the same people became hardcore AI doomers and hope we don't remember the argument from 10 months ago is wild. I am starting to think these people who argue like that aren't doing so in good faith.

Frontier AI safety will require deep systems engineering and formal verification. We’re sharing our learnings openly for all to learn from nvidia.github.io/OpenShell-R…