Exactly. I can no longer solve them. I always use the audio now.

LLMs used as chatbots are fine. LLMs hooked up to harness agents that can perform actions is a different matter.
Since the agents are not legal entities, legal responsibility falls on those who operate them, ie. corporations or individuals.

If corporations aren't getting criminally charged, and no one is personally going to jail, perhaps everyone who is running agents should be setting up a corporation. I bet there is a market for liability insurance for unintended agent actions, too.

Liquor sales are not allowed after 2am in California. Bars close at 2am as a result. There is an increase in drunk driving accidents in the following hour as a result. I got a power outage at 3am Sunday morning. I knew it had to be another drunk driver knocking down a power pole. The final e-mail from PG&E after the outage resolved confirmed that it was due to a vehicle.

Not again with that disproven shit.

SkyNet.

And never put it back. It's unconscionable. These features will not be back. I'm sticking with Musescore 3.

I have 4 roof sections on different orientations with 70 PV panels, and a lot of excess energy. No batteries, though, so I can't really run the local AI on renewables all the time. Trying to rectify the battery situation this year.

I ran out of my ChatGPT Codex $100/month sub for the week. They still let me use Codex 5.3-spark. I tried to use it today for some really basic stuff. It's a complete P.O.S., not capable of making even minor bug fixes without adding regressions. These were simple changes to a dashboard I use to monitor my local inference box.

I have been using Qwen 3.8 27B, and it is rather amazing. NVFP4 quant, on 4 x 5060 Ti 16 GB. Running with SGlang, 1M context. It chews input tokens at around 2000/s, and spits out around 100/s. Not as fast as the cloud models, but very usable. I do need to relocate it due to heat/noise, though. Mainly noise. I like heat.

Yeah. PCIe is a bottleneck for inference, and bandwidth is measured in multiple GB/s. My LAN runs on 10 Gbase-t, but still not the same range.

My TR PRO 3955WX idles at 240W. 22W in suspend mode. 8W when shut down. With the quad 5060 Ti 16 GB GPUs, it gets quite loud during inference. Even my Sony WH1000MX3 noisr canceling headphones are no match for the fans.

I think I will relocate it next to my heat pump water heater in the garage. Maybe it can recover some heat. But it is mostly about noise. WOL will not work on Wifi, though.

Systemctl hibernate + BIOS setting to restart after power loss + smartplug is likely the right solution.

Which he has done too, by ending USAID.

Can't wait for 6G to still not reach.

Facilitating the spread of deadly illnesses cuts down on vulnerable people who may be collecting social security benefits, medicare or otherwise. Killing them is his way of cutting expenses, and thus, the deficit. All part of project 2025.

EOM

Easy - it will just spray the unsafe keyword all over the place.

This is so wrong. I speak from experience working on many httpd and libraries written in C/C++/assembly. There is a ton that can gp wrong. A lot has to do with input validation and undefined behavior. These problems can lead to crashes and remote code execution vulnerabilities. I'm not a Rust advocate. I still code in C/C++. But this is reality.