eoskx

Congrats on the v5 release! Glad to see the tokenizer rewrite.

Thanks -- I'll have to start using xcancel.

Sorry for the X link, but relevant

Glad to see this post didn't get flagged like the one that was posted yesterday on a similar topic about ICE data mining and user tracking.

https://news.ycombinator.com/item?id=46748336

This has now been flagged twice today when it has a legitimate tech/privacy angle. Unbelievable.

Yep, I could totally see that at this point.

As someone who has been out of Java for close to 10 years now, you certainly could do without Mockito, but you'd be writing a lot of boiler plate code repetitively. There's also the case of third-party libraries that you don't control and Mockito has decent facilities for working with those, especially when you're working with a codebase that isn't pure DI and interfaces.

What's also telling is how long the bugs stayed around, too. They were reported on Reddit and Apple's forums for awhile with various workarounds, like deleting the phantom entries from the SQLlite database manually and doing some other gymnastics like removing the other devices from iCloud in hopes that everything would sync up nicely. No one at Apple had the time or took the time to chase down the bugs. In a world of Claude Code or Codex you would think they would have at least tried a cursory "fix this".

On a related note, maybe one of these days iCloud will have a force sync option that tells the other devices to trash their copies vs having to remove all devices and re-add to get everything coherently synced.

My issues with Safari have mostly been iCloud-related. The latest one being the iCloud tabs SQLlite database getting corrupted constantly and keeping stale tabs around that I have long closed. 26.2 seems to have fixed it, but it was around at least since Sonoma. I've had similar issues with Reading List, where again, the database gets corrupted and changes that I've made to Reading List get reverted. It is just little stuff like this that adds up and creates poor UX.

I thought Liquid Glass was cool & interesting when I first saw it in the Developer releases, but I find myself yearning to go back to Sequoia. Hopefully, Apple decides to go back to "simple" soon.

I've had similar observations with different behaviors in Safari and Finder. One would think the quality of Apple's software would be increasing with the usage of Swift over Objective-C, but the opposite seems to be true.

Some more history you might find interesting - Thomele & Mayerl designed the FCU and P250 that would ultimately be the foundation for the P320. Thomele then went onto some of the design work for the P365 from my understanding while the P320 was going through trials and having issues. He's been in numerous depositions along with some of the US engineers that report to Thomele now like Sean Toner.

That's not correct. Sig Sauer USA's VP of Engineering Adrian Thomele was originally an Sig Sauer AG employee, who owned a company that was acquired by Sig Sauer AG. There are other examples of engineering management propagating from the German/Swiss entities and sharing IP between one another. Secondly, Sig Swiss AG recently announced that they would be manufacturing the P320. The German entity was shuddered to decrease manufacturing costs and effectively offshore to Sig Sauer USA, who is now also offshoring more manufacturing operations to India. I think there is much less independence than you might think between the entities.

Please stop repeating this incorrect corporate structure. While there are still 3 entities, they are all now owned by the same German holding corporation, L&O Holdings, which is owned by 2 German individuals. The holding company now includes Sig Sauer USA, Sig Sauer AG (Swiss), and Sig Germany.

I do agree with your assessment of quality between the companies, though.

Despite the name of the website, it is focused on journalistic aspects of the firearms industry, but point taken.

That company structure is no longer quite the case. L&O Holdings (Germany-based holding company controlled by 2 German individuals) owns Sig Sauer USA, Sig Germany, and Sig Sauer AG.

Not even that - the only difference between the Sig M17/M18 handguns that the US Army uses and the consumer P320 is a manual safety. Otherwise, they are indistinguishable.

Keep in mind, too, that the US Army does NOT want this FMECA document released. From the article that is cited the US Army's project manager & legal counsel gave this response to help Sig justify keeping the document sealed: > The Army position would be to oppose the distribution to the public of the > FMECA document as it potentially reveals critical information about the > handgun (design, reliability, performance, etc.).

Also, does not help that the US Army does NOT want this FMECA document released. From the article that is cited the US Army's project manager & legal counsel gave this response to help Sig justify keeping the document sealed:

> The Army position would be to oppose the distribution to the public of the > FMECA document as it potentially reveals critical information about the > handgun (design, reliability, performance, etc.).

All 3 entities - the German, Swiss, and US entities are owned by a German holding company L&O Holdings, but yes, the basis of the 226 was designed by the original Swiss entity.

It appears based on some other court documents that Sig with the P320 intentionally excluded a trigger tab safety based on marketing decisions to be competitive, which every other striker-fired handgun has included. That along with some other issues appears to be the basis for where the P320 design went wrong.

There's another article that is cited where Sig with someone else was apparently developing the fixes to resolve these issues years before the gun was actually tested with the US Army, but didn't deploy the fixes until they were pressured:

https://practicalshootinginsights.com/a-year-before-the-army...

Here are some follow-ups to the article cited. Pretty wild what Sig's response has been:

https://practicalshootinginsights.com/the-document-sig-sauer...

https://practicalshootinginsights.com/sig-sauer-vp-consumer-...

https://practicalshootinginsights.com/new-unsealed-sig-sauer...

https://practicalshootinginsights.com/a-year-before-the-army...

According to a court exhibit, Sig Sauer identified several deadly risks with its P320 pistols as early as 2017.

Just curious what your issues with Triton were. We've done OK with it using it to serve LLM models w/ a classifier head via HF Transformers pipeline & Flash Attention 2, as well as serving text generation models with the vLLM back-end.

Someone commented below that with enough batchnorm/layernorm/etc. and/or gradient clipping you can manage it, but BF16 just makes life easier if you can live without some precision.

Somehow missed this from WWDC23, but it looks like Sonoma will add support for bfloat16 with Metal, and there's an active PR to add support with the PyTorch MPS back-end (PR #99272). Since M2 added bfloat16 support at the hardware level, I'm assuming this will only be supported on M2 Macs.

That maxed out Mac Studio M2 w/ 192GB of memory now looks more appealing...