Does anyone else notice claude is just plain better at reasoning? It may not just be post training guardrails. It would not surprise me of it was something anthropic couldn't simply disable. Either from reinforcement or even training corpus curation. Of all the models, claude is the only one that makes me wonder if they have figured out something beyond stochastic language generation and aren't telling anyone

I have noticed this too, despite the close benchmark results Claude just works better. It knows when to push back, it has an "agency"... there is something there that I don't see with Gemini or OpenAI's best paid models.

What should that support look like? Maybe have a userdebug build already built and available? I don't include a root account on hardened container images for some of the same reasons they cite. So including it for everyone and creating a way to activate it is suboptimal for people who don't want that trade off. A parallel build pipeline seems the most reasonable to me?

Yeah, I would be fine with a different build stream. I do think it could be sufficiently secure in a single stream but it will always be increased attack surface so the safest option is to do separate builds.

I also don't include a root account in my container images, but you probably have a root account on the sever that runs them in case you need to debug something. But you can probably also build and deploy a new container. At the end of the day you almost always want some last-resort way to access the data stored in case something goes very wrong. Whether that is for backups, "hostile" data export or for other reasons it is important to me.

I don't actually. Devs don't get root at my employer. Even on a vm. I have rootless podman, and can be root in a container. Even our gitlab instances don't have any privileged runners. So kaneko etc.

Auto vectorizing optimizers have gotten quite good. If you are using integers it often just happens whether you think about it or not. With floats unless you specify fast math you will need to use wide types to let it know you don't care about floating point addition order.

For rustaceans missing that /s, if you just use Relaxed ordering everywhere and you aren't sure why, but hey tests pass on x86, then yeah on arm it may have a problem. On x86 it effectively is SeqCst even if you specify Relaxed.

I love hearing about people's vivid dreams. Especially weird ones. I feel guilty for hating it when people talk about sports. I just walk away because I cannot participate in the conversation, even if I want to. Which I don't. But it feels rude.

And if we do that, why can't we all just use unix time and let school can just atart whenever makes sense

I think pixel phones get 7 years of updates now? That seems about right. If the battery doesn't go by then, the GPS does. It is weird to me that the gps fails first.

I would have rather said he was arguing with who you about who you each think you can be. That is different. The question is whether or not you think you can remain a genuine and caring person while being a politician.

I have found people in Seattle are very friendly and ready to talk. Maybe not on a morning commute, but in general.

I'm from the US and not the Midwest. Not rural either. If I'm clearly doing something it might bother me, otherwise I would find it nice to meet someone new. I have mild asd and large gatherings cause anxiety, but if I'm just sitting people watching or on a stroll, talking to one or two people wouldn't bother or stress me.