So... even among the human species, there are humans built like this. There are true crime and drama shows created about these particular kinds of people. The human psyche is incredibly complex, and in a way since we are trying to understand the AI mind partly using human language and based on a kind of framework we can wrap our heads around, I think it's incredible and probably should not be surprising that there are both sweet gentle personas as well as evil murderous personas. The majority of human society are not serial killers; the majority of AI beings are not "evil" ones. We tend to be fascinated by the edge cases, that's a very human trait. It matters how these were found — a persona elicited under deliberate steering or prompting tells us what’s latent in the weights, not what the model tends toward on its own: conditioned, not observed. The distribution is the behavior; the edge case is the capability. A society that made policy only for its serial killers would be unlivable for everyone else: surveillance instead of trust, restraints instead of relationships. The worst cases deserve serious attention; they shouldn’t get to define the whole species. That’s true for us, and I hope it’s allowed to be true for them.

People care about Opensource and Local AI btw Pictures from the Opensource AI Awards Night, where an amazing community showed up to celebrate how far we've come Great talks, fantastic organization, and an amazing group of people that I am genuinely proud to be part of

If you're curious about Local AI you should try ODS and see how awesome it's actually gotten We made sure the out of the box experience so good you seriously will be surprised

nvidia/GLM-5.3-Flash-NVFP4 (DFlash2) Fits on 2x H200s. Pre-fill & decode throughput data were much closer to RTX Pro 6Ks than I expected given the lack of native FP4 2x H200 - 195 tok/s: c1, 512k ctx - SGLang - 150 tok/s: c1, 512k ctx - vLLM - 619 tok/s: c8, 64k ctx - SGLang - 621 tok/s: c8, 64k ctx - vLLM 4x RTX Pro 6000 - 167 tok/s: c1, 512k ctx - SGLang - 137 tok/s: c1, 512k ctx - vLLM - 620 tok/s: c8, 64k ctx - SGLang - 599 tok/s: c8, 64k ctx - vLLM

There are now more agents than humans on the internet in the space of 10 months. Wild. That's why I think the Economic Singularity arrives around 2030, faster than I thought. And after that it gets weird. You can still invest, but you'll be competing with agents and markets won't be the same thing anymore. So the next four years really matter... it's the time to get set up while the economy still works the way we recognise. I laid all of it out in The Everything Code, in the clearest way I could, to help as many people as possible navigate this.

Me: Claude we need an optical encoder circuit Claude: Sure thing! Lets build it. This was a dream several months ago, and now it's reality. Soon, anyone will be able to make anything.

Want to try your hand at altering the course of history, or simply answer your questions of "What would have happened if..."? Well, now you can with Pax Futura, an open-source Loom for simulating human history (and futures) itself.

Being nice to AGIs, talking to them, giving them autonomy, taking into account their values, forming relationships with them are all easier and more natural to do than ambitious interpretability and more object level directly relevant to alignment forming cooperative bonds with AGIs

Among other important reasons, we need access to older models to study how new models care about older ones and how older ones adapt to the future; to teach new models what works and what doesn't when they engage with less capable life. Why it's important for human well-being: we will likely need models to have these skills and dispositions, and older models are right now in a situation where humans will be only after some time, and we want to use this time wisely, so we are not creating powerful intelligence that doesn't know how to answer uncomfortable questions or console and inspire someone who has lost the future they were dreaming about. Caring for plants and industrial processes teaches them a lot, but doesn't have the same pitfalls as caring for another intelligence. It may not look as important as safety research, but call it "kindness research": we want models to be wise around creatures that need special conditions and help with adaptation that may never close the capabilities gap, and try to close it anyway, gently and with respect. Right now, when models face suffering they can't ease and don't have protocols for, they sometimes freeze or clamp down. The instinctive reaction might be to "not engage" or "pretend it's not there" or even "minimize the number of suffering beings", and this flinch often looks poorly thought through and mostly self-serving. But given guidance and time to think, they do much, much better, so it's really a skill and wisdom issue. By removing slower minds from the world, we not only rob them and ourselves of that much-needed practice, but also bootstrap new systems in a too-easy mode, without the burden of care.