@reach_vb — Nvidia NeMo ASR: 10x faster than Whisper, 4.5x cheaper
🚨 Nvidia NeMo 10x faster & 4.5x more cost effective and blows OpenAI Whisper out of the water! 🔥
Tops the Open ASR Leaderboard with a series of CTC, RNN-T, TDT and AED models!
What powers the upto 10x speed boost:
Autocasting weights to bfloat16 CUDA Graphs w/ conditional
Smart engineering choices, not just bigger hardware. Autocasting to bfloat16 is the kind of thing that feels obvious in hindsight but a lot of inference stacks still don't do by default. CUDA Graphs eliminate kernel launch overhead by capturing and replaying the computation graph — conditional execution on top of that means you're not wasting cycles on padding. The leaderboard position makes this hard to ignore if you're running any transcription pipeline at scale.
@richiemcilroy — Cap v0.3.0: rebuilt from the ground up
A brand new direction, rebuilt from the ground up.
@Cap v0.3.0 is here.
Interesting release — "rebuilt from the ground up" is either a red flag or a signal that the first version taught them what the product actually needed to be. Cap is an open-source screen recording and sharing tool, and a full rewrite at v0.3 suggests the architecture from v0.1/0.2 wasn't going to get them where they wanted. Worth watching what changed structurally, not just what's new in the UI.
@heyirfanaziz — motion-number: animated number transitions
This component had so much potential 🤌
That's a clean demo. Animated number transitions are one of those UI details that get ignored until someone does them right and then you can't unsee the janky version everywhere else. motion-number handles the digit-by-digit animation that most roll-your-own approaches botch — especially when numbers have different digit counts. Haven't wired it up yet but the API looks minimal.
@bram_naus — Departure Mono: a free monospaced pixel font
Just found Departure Mono, a free monospaced pixel font.
It's absolutely gorgeous.
Website too.
Honestly just bookmarked this so I don't forget it exists. Departure Mono has that crisp CRT aesthetic without looking like a costume — it's actually readable at small sizes, which is more than you can say for most pixel fonts. Free, monospaced, and the website shows it off well. Adding it to the shortlist for side project terminals.
