@stopachka Someone made a fairly polished app a year ago following the NTNU textbooks (Zhong Chinese) but it really depends what level you want to study.
@antranapp Preview video looks nice!
I went to the Apple store and discovered a bug if your App Preview orientation is portrait and your images are iPad landscape it squishes them into the portrait sizes.
Gotta update I guess.
@VicVijayakumar@devahaz I have such a soft spot for mid-century engineering that still functions today but I am not trusting dated gas regulator designs or striking a match every time I want to fry an egg.
@illyism@Charles_SEO Ahh I just spent a month in Ko Samui. I love the low season. "Don't come it'll be raining all the time" great, I can stay inside and code all day guilt-free.
I think STQ 2.75bpw is better optimized for coding tasks, and the engine is a bit faster.
The 2.0 version was a noticeable improvement in quality. And it's a little faster on the M1 (this uses the engine changes, 6-bit trunk & STQ codebooks).
I believe that move from 8-bit to 6-bit is where the decode boost comes from.
2.75bpw is also the right size that I can set wired_limit_mb=55000 instead of 59000 and have headroom to run Chrome, Pi, iOS simulator, etc.
I'd switch if you want to optimize your machine to run QFN on a regular basis.
M1 Max 64gb, @Beamsters1 DeepSeek-V4.1-Flash-1017-Sushi-STQ-1.75bpw 10+ tok/s, streaming experts (none disabled), TOL=0.2. 3 hours 9m wall clock.
Perusing the readme I discovered the engine was ready to stream weights for any model, so I had to try it.
@coolnalu Clear:both is burned into the back of my eyelids but I'm probably never writing css again in my life, unless society breaks down and the only way to save humanity is to center a div on a windows xp workstation or something.
It is an inexpensive and fun way to run real models. Qwen Flash Next 3.8 2.75bpw STQ is very usable for single agent coding sessions on this machine. Especially if you can use it as a server, and literally run nothing else on the machine.
But it is a budget option. It's old now. It doesn't really run 27b very well, or Muse Glimmer (dense models).
The M1 Max has one trick: 400gb/s memory bandwidth.
When I benchmark or run things like this I am literally running nothing else on the machine. I shutdown spotlight search via a startup script. No browsers open. Low memory footprint agent harness. It's not usable while its decoding.
But in the grand scheme of things it's a great machine for every non-AI task. Runs blender effortlessly, does video editing, decent gaming via Crossover, wine, etc.
I don't think you're going to get better performance running asahi linux, the bottlenecks aren't really software anymore they're physics.
Qwen 4 is supposed to ship something architecturally similar to Qwen Flash Next 3.8.
KLD at that level (to 1.75bpw not full weights) is 0.008, top-1 97.7%. Dropping experts breaks the model pretty quickly, but tolerance gives you proportional tradeoff between speed and quality. Picked a safe middle ground after a few tests and that yields ~13 tok/s sustained.
IA seems to have always been a struggle. It's better but a quick win would be quick search. It doesn't go deep enough in getting you from a single zone page -> setting.
I.e. I know I want to modify "always use https" but where the eff is that setting now, after 2-3 IA reorganizations, new products were added, etc?
I could type it into Cmd+K or Quick Search, but it only has an index for the level of hierarchy above that (Edge Certificates in this case).
Cars are a fun hobby. It's more fun when you have constraints, and don't blow a ton of cash on an obviously amazing supercar, etc.
I haven't touched a BMW in years, but there's something satisfying about learning the ins and outs of a particular platform and tuning it, modding it. BMW S54 was (is?) an amazing engine, I loved it.
The S58 is supposed to be one of those really interesting and fun platforms to get familiar with as well.
Yeah I already thought that was a possibility before all of this. If a frontier lab screws up who is it most valuable to, as an acquisition, a potentially vertically integrated Nvidia or a company already trying to run an AI lab. Figured they'd pay top dollar for talent or the competing lab acquires the talent. Just fun ideas though.
I watched Pikmin Bloom take off in Taiwan, it's incredibly popular as is Pokemon and Pokemon GO. Earlier this year, I saw people flying to Japan in groups to empty out Pikmin merch from gashapon machines and stock machines in Taipei. It's nuts.
But Nintendo has had a presence in Taiwan for years, lots of marketing spend here. They're pretty savvy and seem to understand the domestic market. I expect they will do something to resolve the situation, and I wonder if this is Scopely Explore's mistake instead. Which is bewildering because this might also be where the largest Pokemon Go events outside of Japan take place, it is also insanely popular and Niantic had a presence here.