I took the bet. The details are documented somewhere in the internal Facebook Workspace, but it had to do with whether time spent in video consumption would exceed non-consumption use cases in VR.

This is accurate xitter.cf/ssh_sharma135/status/2…

This explains a lot about Luminar.

Controversial restaurant opinion of the week: Sungho is better than San Ho Won and should have the star instead.

My old blog post that the raw image came from: eclecti.cc/computervision/re…

Throwing old projects into Opus 5.5 is fun. It was able to one-shot a solution in 30 minutes to a project I put on hold in 2012: decoding and rendering images from Lytro cameras.

You can connect to support during the ride and ask. The answer I got for messed up routing during a trip was road sections dynamically being blocked off. Probably a policy overreaction to the gridlock that Waymos triggered a few times in SF during heavy congestion.

Not a judgement on whether it was good or bad. It was net good for VR (and certainly for Oculus employees), it just has delivered poor financial RoI to date.

I just realized, in terms of return on investment, Oculus recently passed AOL-Time Warner as the worst acquisition of all time.

Very impressed!

Impressed.

Has anyone ever successfully depleted the battery on a Fluke or Mitutoyo tool? These things might be powered from another dimension.

The Mission one is still consistent and good.

Do you believe that AI will never solve PCB routing, or that some other electrical engineering domain will still be beyond that.

I'd be curious to see examples of 4 or 6 layer boards with larger QFNs or BGAs that it has successfully routed.

Certainly hope so. It seems like part of the challenge is that the scoring function (simulating the physics) is still extremely computationally expensive. However… humans are still routing these things easily without simulating the physics on anything but the most sensitive boards, so our brains are just really good at some part of this that computers aren’t yet!

Because it’s excellent at boring math, parsing datasheets, and reading existing reference designs, I’ve found Opus 5.5 to be great so far at electrical and firmware design out to a schematic and working firmware. This is much like it’s able to do a solid job at non-embedded software development already. It’s still PCB layout where it switches over to being an unsolved NP-hard problem that humans somehow excel at with just a little training.

The routing of that is about as simple as it gets, since it’s low frequency and can mostly be routed in a single layer. Pre-LLM autorouters would also handle it. Iteratively handling trace crossings across multiple layers is where both get tripped up, even without considering signal integrity, inductance, etc.

Yes, future models will certainly get there, but probably long after they already excel at every other part of electrical and mechanical design!

Playing around with Opus 5.5, it's doing a surprisingly good job of electrical and mechanical co-design, even with high voltage and antenna constraints coming into play, and then it is falling apart completely when it comes to routing.