hoss1474489

Bifrost has been decent, despite being partially shareware

I recently implemented OpenPilot for my vehicle and used it on a 9 hour road trip.

Due to the nuances of my vehicle, the built-in hands-on-wheel nag was still active and required (small) driver wheel inputs every 10 seconds to prevent the beeps. Just keeping both hands on the wheel was enough to pacify the nag. But this forced me to keep attention on the vehicle, more than OpenPilot's face direction detector. Anecdotally I am certain this kept me more engaged than if I could have gone hands off.

The net effect was I had more capacity to be aware of what was going on around me on the road, AND I didn't feel nearly as fatigued at the end of the drive. Keeping the car between the lines consumes a little bit of cognitive energy constantly, which compounds over a long drive. Freeing up that load allowed me to pay more attention to what was happening on the rest of the road. And the mental burden was still half as much--I felt like I had driven maybe 4-5 hours instead of 9.

I have some ideas for how I might disable the built-in nag, but I'm pretty sure doing so would make my car less safe. The current configuration might already achieve the safety maximum for ADAS based on front-facing cameras only.

It increases situational awareness while reducing fatigue, and I think that anything further I try to do will adversely affect one or both of those effects.

I’m working on integrating with a Jeep Renegade which was designed on the Fiat small-wide platform, and the CAN bus architecture is a non-starter for mainline OpenPilot support. OpenPilot requires every signal needed from/to the car is accessible by tapping at most two CAN buses, both of which must be accessible on the cable to the lane assist camera behind the rear view mirror.

On the Renegade, the adaptive cruise control activation and control signals are on a separate bus from the steering control signals, and the ACC signals don’t get up to the lane assist camera at all. OpenPilot support on this vehicle will require tapping CAN buses in two spots on the vehicle, and this is not a Comma-supported configuration. They want to be able to make a single interposer cable that plugs in between the lane assist camera and your car, and if that’s not possible they will never support your vehicle.

To attempt to support my car I am working on a gateway device that will bridge information from the powertrain bus to the fusion bus so it will be available at the camera (which is on the fusion bus and the chassis bus). But this implementation, which may not actually work but even if it does, will never be accepted upstream. And safety-wise, that’s probably the right choice for the project.

The sad part is I think the only signal that I need that I don’t have at the camera right now to make a safe implementation is ACC engagement. I’ve scrutinized every bit of every message on the chassis and fusion buses available at the camera and there is simply no signal that indicates if ACC is engaged.

One bit is all I need to get lateral control working. Well, assuming that once I start injecting control signals there aren’t other issues, that is. And that is not probably a safe assumption.

I see a Security button in the what’s new box in the Codex section of the ChatGPT website. It appears to allow me to run vulnerability scans against my connected GitHub repositories.

Direct link: https://chatgpt.com/codex/cloud/security

It’s a slow burn, but if you keep using it, it seems to eventually catch fire as the agent builds up scripts and skills and together you build up systems of getting stuff done. In some ways it feels like building rapport with a junior. And like a junior, eventually, if you keep investing, the agent starts doing things that blow by your expectations.

By giving the agent its own isolated computer, I don’t have to care about how the project gets started and stored, I just say “I want ____” and ____ shows up. It’s not that it can do stuff that I can’t. It’s that it can do stuff that I would like but just couldn’t be bothered with.

The only fully-functional stack currently available requires Python >= 3.8, which is the main limitation to where it will run. But there’s still a lot you can do with that!

Restarting the machine got Cowork working for me.

I noticed a couple hanging `diskutil` processes that were from the hanging and killed Claude instances. Additionally, when opening Disk Utility, it would just spin and never show the disks.

A restart fixed all of the problems including the hanging Cowork tab.

I found a feedback link in a dismissible banner on the Cowork tab. Then the clock is running to fill it out and submit it before Claude crashes.

I haven’t found any place to do that.

Beachball of death on “Starting Claude’s workspace” on the Cowork tab. Force quit and relaunch, and Claude reopens on the Cowork tab, again hanging with the beachball of death on “Starting Claude’s workspace”.

Deleting vm_bundles lets me open Claude Desktop and switch tabs. Then it hangs again, I delete vm_bundles again, and open it again. This time it opens on the Chat tab and I know not to click the Cowork tab...

> there wont be a guarantee that they will never lose their jobs, they will continue to live on the wobbly and uncertain foundation

The people who lose their jobs prove this was always the case. No job comes with a guarantee, even ones that say or imply they do. Folks who believe their job is guaranteed to be there tomorrow are deceiving themselves.

In my frustration I didn’t even think to look if there was a third party solution.

Looks like that would do the one thing I need, and I’m finding the grass to be just as brown with Razer Synapse is it is with LogiOptions.

Always hated Logitech Options bloatware, why should the software drain my battery, phone home, and require screen recording permission just to emulate a keystroke when I push a button on the mouse?

I enjoyed my trip to Micro Center today to finally ditch Logitech after those buttons stopped working. Put up with Options for over a decade because at least it did the one thing I needed.

Better than you’d think; worse than you’d hope.

I like this. I more generally look for reduces chaos.

I’ve seen the pursuit of disambiguation employed to deadlock a project. Sometimes that’s the right thing to do—the project sponsor doesn’t know what they want. But many times the senior needs to document some assumptions and ship something rather than frustrating the calendars of 15 people trying to nail down some exact spec. Knowing whether to step on the brake or the gas for the benefit of the team and company is a key senior trait.

This is a yes, and to the article; building without understanding the problem usually will increase chaos—though sometimes the least effort way through it is to build a prototype, and a senior would know when to do that and how to scope it.

GPUs in 16x slots is still important for LLM stuff, especially multi-GPU, where lots of data needs to move between cards during computation.

Wow, 595 USD is insanely expensive for literally half a keyboard.

Effort is the algorithm. (Presentation on learning in the age of AI by the Veritasium guy) https://youtu.be/0xS68sl2D70

Explicit and obvious encoding in rules isn’t what makes something systemic.

Make sure to clean everything that was inside the fridge before restocking or you reset your progress

Wow. This is so simple it blows my mind. It’s obvious, now that I see it. Suddenly I simultaneously realize how much I make it suck to talk to me and how easily I could change that. Thanks for sharing.