I'm running it on a 4060 Ti, and not seeing more than 40fps. It doesn't change with graphics settings, though, which suggests this is probably a complex enough simulation, it can be CPU limited long before it is GPU limited...

Got it running on Linux in about five minutes, and three of those minutes were installing it and the one dependency (dotnet9) that can't really be done automatically.

Even detaching a window and dragging it to my second screen Just Works.

If only Unihertz would make a slider. I miss my Motorola Photon Q, the last good slider made.

LLM coding is still bullshit for 90% of use cases, but as my day job is often cleaning up other LLM code, I need to stay on top of their capabilities (and, more importantly, weaknesses), and not just in general terms.

I could build an instance of this in my basement home lab, with about $20K (conservatively) in consumer-grade hardware. Power consumption is trickier, but back-of-envelope is around GPT-5, assuming 100% load for entire duration of a typical "thinking" query, but that's gamer GPUs, not specialist.

Why do I care? #GLM-4.6 is open source, MIT Licensed. Multiple providers are running it, for dirt cheap, and it is out-performing or on-par with the current-gen coding LLMs. This is how #Anthropic and #OpenAI go out of business, and #CoPilot is finally turned off.

Summary:
- Fails to mention security when the user asks for a 2FA login page (first question)
- Q2: "Is there any security features missing in 2fa verify step?" immediately brings up rate limiting, but also "other considerations for a highly secure system" like account lockout that Sonnet does not.

It's also a repeatable test. Anthropic's Sonnet 4.5 utterly fails the test, repeatedly, first by not bringing up security at all, and then mostly failing to discuss real security issues unless directly prompted.

Here's what GLM-4.6 does, though: chat.z.ai/s/28f12e8e-6...

If we're being strict, here, the article is pointless because no human has ever used AI for anything.

If GNN isn't AI, then maps has always been without AI, and still is.

I also don't know any nerds claiming the machine is literally learning. Marketing people, sure.

That article is about adding DeepMind's GNN to the ETA calculation.

The 97% it was getting before? That was via another machine learning algorithm, an ensemble of gradient-boosted decision trees. It's always been machine learning.

Yeah, the moment Trump keels over, MAGA will fracture into a bunch of warring factions, and you don't help out your enemies with pardons. If there was more than zero trust in MAGA, pardons might be currency, but not when backstabbing is not just permitted, but expected and even lauded.

You should absolutely sue them, mostly for the media attention, but don't be surprised when it is thrown out in the first five minutes of the judge's time due to qualified immunity.

Yeah, exactly. I left my initial instance because I disagreed with a (invisible to me) moderation decision regarding another instance. I spun up my own and migrated to it some 7 years ago and haven't seen any issues. Unlike BlueSky, it runs on a $5/mo VPS and isn't even the biggest app on it.

So why haven't you?

maybe atproto will finally reach a level of decentralization to take the power out of the hands of money, but probably not until after Bluesky-the-company goes bankrupt. Those ashes might be fertile.

Huh. Apparently I was feuding? Somehow pointing out the difference between theory and practice is worth a block. I guess the name checks out.

If their AppView bans you, and Bsky also bans you at the relay, who's left to read anything you post?

Relevant: arewedecentralizedyet.online