Fabio de Miranda

MACHINES: A HISTORY OF ELECTRONIC MUSIC - OUT NOW! I'm delighted to share that my new book is out now in the UK and Australia, with more countries to follow. A journey through the instruments, sounds and stories that shaped electronic music. Order now: lnk.to/MachinesJean-MichelJa…

Os uniformes da seleção nos anos 1970 eram mais legais

Essa é outra noção que passou. 2 anos atrás a gente falava em subsídio de tokens. Mas de lá pra cá otimizamos MUITA coisa, desde quantização, MoE, KV cache, prefix/prompt cache, MQA/GQA/MLA, PagedAttention/FlashAttention, chunked prefill, speculative decoding, etc etc Parece que inteligência nível GPT-4 caiu da faixa de USD 37/mi de tokens pra USD 0.10 agora. Chequem o BenchLM Token Price Index e Brief Script AI Price Index. Topo dos modelos da OpenAI: GPT-4 era 37.50, GPT-4o foi pra 7.50, GPT-5 foi pra 3.44. Entendam, já conseguimos rodar bons modelos em 32GB VRAM (importante é a banda, não a qtde de memória). Eu chutaria que hoje não tem mais subsídio e na verdade alguns cobram mais caro (por conta de sticky prices).

When I said hunch could learn in real-time I wasn't kidding. It's like teaching a kid! Behind the scenes, @Cloudflare’s Clef is helping refine the local model so it learns in real time, while @googlegemma’s EmbeddingGemma2 runs live on the CPU. 🧵

Fine-tune and run your Jev at home on your MacMini 👇

bro how is a 1.7B model beating models 2-5x its size across phones and laptops 😭. at this point Liquid AI is just bullying Gemma and Qwen .

It's been a while since I showed any Jev stuff but I've been cooking on a way to use @GoogleDeepMind's EmbeddingGemma2 as an insanely fast real-time /v1/systemone endpoint (20ms on text, 100ms on images). It's trainable in real-time, can distill an upstream teacher model, and is multi-modal by default. Hard to believe a @googlegemma model of this size can do so well with the right harness!

Oh my. Best open OCR model by a mile. ... and Baidu took the KV cache for hundred-page OCR from 47GB → 1.77GB. Unlimited OCR 2 is only ~3.3B parameters and it's designed to parse documents 100's of pages long without memory and decode latency growing as the document gets bigger. The secret sauce is SR-SWA. Instead of every new token repeatedly attending across an ever-growing document, it keeps a fixed-size set of useful text/visual references. According to Baidu ... 🧠 Peak KV cache → 47.0GB → 1.77GB 📉 ~96% reduction ⚡ per-step decode → up to 7× faster 🚀 16-doc throughput → 360 → 1,315 tps 👈👀 Now look at Baidu's 100–300-page benchmark MarathonDocBench: Unlimited OCR 2 → 77.1 Kimi-K3 → 33.7 Gemini 3.8 Flash → 28.6 Qwen3.8-Max → 22.9 Of course, this is vendor-reported, so independent reproduction matters. Baidu also reports that shorter documents get even more accurate Unlimited OCR 2 → 84.6 Qwen3.8-Max → 78.3 Model Stats 🧠 ~3.34B parameters 💾 ~6.7GB BF16 weights 🔓 MIT licensed 📄 text + tables + formulas + layout ⚙️ Transformers + SGLang Can we squeeze this onto a 16GB GPU? Links in ALT

Applied the 3090 mods on my 3 GPUs

If Flávio loses, Jair Bolsonaro’s decision to back his son over Tarcísio de Freitas will be widely seen as a decisive error. Many on the right will argue that Tarcísio, who polled better against Lula and was the centrão’s favorite, would have had a better shot. Jair, aware of his son’s weaknesses, chose to keep power in the family.

Ouça o que o Renan tem a dizer, pelo menos uma vez seja humilde e aceite o erro. A vitoria do Lula foi construída em cima da arrogância da família Bolsonaro. Os deputados do PL ganharam na dancinha, mas quem dançou foi você.

PREDICTION - Google overtakes OpenAI and potentially Anthropic by year end Swapping all our Sol and Sonnet workloads with Argon. It’s just better Their next release Carbon will beat Fable and Astra. It almost seems inevitable 😲

This feels horribly like 8 Jan 2023, when something that shouldn't have taken anyone at all by surprise, took everyone by surprise.

Deepseek Flash is very hard to beat on easy tasks - ship updates every month - 10x cheaper than sonnet - handles multi modal outputs - ultra fast options will soon be available Their performance and price combination will almost become impossible to match

Modders have unlocked hidden AI compute performance on NVIDIA’s RTX 20, 30 and 40-series GPUs. The modifications remove software-level limits on certain Tensor Core operations, without changing the hardware or flashing a new VBIOS. - RTX 4090: 165.6 → 331.6 TFLOPS FP16 Tensor Core throughput -RTX 3090 Ti: 87.5 → 157.4 TFLOPS BF16 performance - RTX 2080 Ti: 59.9 → 111 TFLOPS FP16→FP32 performance - RTX 3090: 73.0 → 122.8 TFLOPS in a PyTorch GEMM test The gains are mainly relevant to AI and compute workloads, not gaming.

I love that David Bowie's and Brian May's reactions were captured when Annie Lennox gives it everything, and that George Michael singing along in the back! Fantastic video and the very best music.