Your curated collection of saved posts and media
Megakernel for Laguna XS + GRPO training on a single RTX 4090: ~180 tok/s @ batch 1 ~499 tok/s @ batch 8 In-place, memory-efficient forward & backward wiith custom kernels. @poolsideai @eric_alcaide https://t.co/8zSwUK87LS https://t.co/qd7pgKzsBI
Cognitive offloading and the speedup illusion in human-AI interaction Sunny Yu, Myra Cheng, Ahmad Jabbar, Ilia Sucholutsky, Katherine M. Collins, Dan Jurafsky, Robert D. Hawkins https://t.co/9pvuG5Kt2N [ππ.π²π ππ.π·π²] https://t.co/MP2VFqSqul
Uberβs COO has said that itβs getting βharder to justifyβ its AI costs because there was no way to show a link between AI spend and any meaningful increase in useful features. This is the first time Iβve seen a company say this directly. https://t.co/xUhZvtpwah https://t.co/nDj9GIXssV
Massive output uptick due to agentic AI. Complete flat adoption. https://t.co/s6ubPsy0SL
A French engineer who lives quietly in Paris has spent 30 years writing software that the entire internet now runs on without knowing his name. He wrote the code that streams every YouTube video, every Netflix show, every TikTok clip. He wrote the code that runs the virtual servers underneath AWS, Google Cloud, and Microsoft Azure. He calculated more digits of pi than anyone in history. He has no Twitter. He has no marketing. He just keeps shipping. His name is Fabrice Bellard. Here is the story, because almost nobody outside the systems programming world knows what one man has built. Fabrice was born in 1972 in Grenoble, France. He studied at Γcole Polytechnique, the top French engineering school. He never went to Silicon Valley. He never built a startup empire. He just wrote code. In 2000 he started a project called FFmpeg, an open-source multimedia framework for encoding, decoding, and streaming video. He was 28. The project did one thing nobody else had done well. It handled every video and audio format that existed, in one library, on every operating system. He led it himself for years. Today FFmpeg is the invisible engine of the internet. YouTube uses it. Netflix uses it. VLC uses it. Chrome and Firefox use parts of it. Every Android phone, every iPhone, every smart TV, every video editing tool you have ever touched runs FFmpeg somewhere underneath. If you have watched a video on a screen in the last 20 years, Fabrice's code processed it. He was not done. In 2003 he started QEMU, a machine emulator and virtualizer. He wrote it solo until version 0.7.1 in 2005. QEMU lets you run any operating system on any other operating system. It became the foundation of modern virtualization. KVM, the Linux kernel hypervisor, runs on top of QEMU. Every major cloud provider, AWS, Google Cloud, Microsoft Azure, IBM Cloud, runs virtual machines on infrastructure built around it. The Quick Emulator is the most cited piece of cloud infrastructure code on Earth. He kept going. In 2001 he won the International Obfuscated C Code Contest with a small C compiler that grew into TCC, the Tiny C Compiler. TCC can compile and boot a Linux kernel from source in under 15 seconds. In 2004 he calculated the most digits of pi ever computed at the time, using a personal desktop computer and an algorithm he derived himself called Bellard's formula. In 2011 he wrote a complete PC emulator in pure JavaScript that runs Linux in your browser, a project called JSLinux that engineers still cannot believe is real. In 2019 he released QuickJS, a small but complete JavaScript engine that fits where V8 cannot. In 2021 he released NNCP, a neural network based lossless data compressor that immediately took the lead on the Large Text Compression Benchmark. Then he turned his attention to large language models. He built TextSynth Server, a web server with a REST API for running LLMs locally. He released ts_zip and ts_sms, compression utilities that use language models to compress text and short messages at ratios traditional algorithms cannot reach. He released TSAC, a very low bitrate audio compression system. In December 2025 he released Micro QuickJS, a new JavaScript engine for microcontrollers, separate from QuickJS, designed for environments with almost no memory. Fabrice co-founded a telecom company called Amarisoft in 2012, where he serves as CTO. Amarisoft builds 4G and 5G base station software used by carriers and labs around the world. He has been running it for over a decade while continuing to ship personal projects from his own home page at bellard dot org He has no Twitter. He has no Instagram. He gives almost no interviews. His personal website is a flat list of projects with no styling, no fonts, no marketing copy. Just titles and links. A quiet French engineer who never moved to Silicon Valley wrote the code that quietly runs the internet. He is still shipping.
Ora, lege, relege... https://t.co/MC9uKHxrrn
Singer-songwriter Matthew Sweet suffered a stroke while on tour opening for Hanson, and management reps have set up a GoFundMe for his health care. https://t.co/4FyU7SL3As
The Good Omens final season is now a single 90-minute episode https://t.co/axoWsgxpms https://t.co/0WL66diE5k
Well, that last comment says it all, doesn't it? https://t.co/SmPohdSWb1
Solar growth is exponential. That's why coal jobs are NEVER coming back. https://t.co/ZeQeyypzJy
#truth: If Youβre Going to Do a SaaS Start-Up β¦ You Have to Give it 24 Months https://t.co/yaKi91xEPt via @saastr
Congrats to @JoshuaRosenthal and the whole team at @RowdMap! What a great win. https://t.co/gTrHCzzk3z
Louisville's @RowdMap acquired in $70M cash transaction: https://t.co/74PRMGKRHw Measuring healthcare provider performance & efficiency
I'm so proud to be the owner of the very first Chevy Bolt sold in Kentucky. The future is carbon free and EVs will help make it happen. https://t.co/swMJiU6CmP
I like the shape of this Model 3 better than my Bolt, but this Chevy is a lot more fun to drive than this Tesla. Thanks @qz https://t.co/le2CQyIwgA
Fascinating how blockchains could socially disintermediate corporations: The Blockchain Man https://t.co/B10lXMskg6 via @ribbonfarm
At its 258th Commencement on Sunday, Brown conferred honorary doctorates on six candidates who have achieved great distinction in a variety of fields, including economics, civil rights, artificial intelligence, national security, business and community leadership. https://t.co/6t7jUCboy8
Itβs a real honor to receive an honorary doctorate of science from @BrownUniversity . π https://t.co/86SWE882xY
At its 258th Commencement on Sunday, Brown conferred honorary doctorates on six candidates who have achieved great distinction in a variety of fields, including economics, civil rights, artificial intelligence, national security, business and community leadership. https://t.co/6t
1/ Introducing GPIC: a Giant Permissive Image Corpus and benchmark for visual generation! π100M VLM-captioned image-text pairs for training π1M image-text pairs for benchmarking πΌοΈ~28 trillion pixels π€Centrally Hosted β Fully permissive for research + commercial use Dataset, benchmark and modelsπ§΅π Co-led with @KyleSargentAI
The CoRL 2026 keynote lineup is here! πΉ Russ Tedrake β MIT; stealth startup @RussTedrake πΉ Fei-Fei Li β Stanford; World Labs @drfeifei πΉ Wolfram Burgard β UT Nuremberg @wolfram_burgard Join us in Austin this November. https://t.co/uiOkizDNIc https://t.co/dR4AY5UZK4
We turned dreams into worlds. Then filled them with history's greatest minds. Not a video. A world, running directly in your browser. Step inside β https://t.co/W62gfXuEZO
π€Unclear on word and sentence embeddings? Check out this awesome summary with clear visual explanations on the @CohereAI website by @luis_likes_math π° Post: https://t.co/Lklh77Jf9D https://t.co/MizcNNPWig
π€ Unclear on word and sentence embeddings? Check out this awesome summary with clear visual explanations on the @CohereAI website by @luis_likes_math https://t.co/auvUgsUgTz
Gemma 4 12B can now run locally on just 8GB RAM via Dynamic GGUFs. Google's new model, Gemma 4 12B Unified supports image, audio and 256K context. You can run and train the model via Unsloth Studio. GGUF: https://t.co/8cL321pVDh Guide: https://t.co/odRo9WjRpA https://t.co/Ax09ZTXFF3
Meet Gemma 4 12B! A unified, encoder-free multimodal model designed to bring high-performance intelligence directly to your laptop, and released under an Apache 2.0 license. Bridging the gap between edge efficiency and advanced reasoning. Here is whatβs new with Gemma 4 12B: π
Second big release from us today: Nemotron-3.5-ASR-Streaming! π40 languages β‘οΈ80ms - 1s controllable latency π₯240 - 2400 concurrent streams on 1xH100 π§±FastConformer Cache-Aware RNN-T architecture https://t.co/lxmcAnKeOl
Introducing Harness-1, a 20B search agent trained with a state-externalizing harness. > frontier-level long-horizon search, rivaling Opus-4.6 and outperforming GPT-5.4 > Context-1-level cost and latency > externalizes candidates, evidence, verification, and search history > open-source
π mlx-audio v0.4.4 is out β our biggest model drop yet. 15+ new TTS, ASR & VAD models, faster long-form transcription, and an expanded OpenAI-compatible audio server. All running local on Apple Silicon. π€ New TTS β’ VoxCPM2 β 2B, 48kHz, 30 languages β’ MOSS-TTS / TTSD / 1.5 β’ Higgs Audio v3 β’ Miso, Dramabox, Irodori-TTS v3 VoiceDesign π New STT/ASR β’ Mega-ASR (Qwen3-ASR-1.7B + LoRA routing) β’ Nemotron 3.5 ASR (streaming) β’ granite-speech-4.1-2b-nar, Fun-ASR-Nano β’ Cohere ASR β 1.7Γ faster long-form π VAD & codecs: Silero VAD, FSMN-VAD, Step-Audio 2 βοΈ Server: OpenAI-compatible response_format, /v1/audio/voices, word timestamps, realtime server-side VAD turns h/t @lllucas Huge thanks to all the contributors π > uv pip install -U mlx-audio https://t.co/muDYzy10FA
Before the week ends, let's acknowledge one of the most INSANE week ever for open AI, with 25+ notable open-weight drops across every modality: π§ LLMs β NVIDIA Nemotron 3 Ultra: 550B hybrid Mamba-MoE, only 55B active, 1M context, MMLU 89.1. NVFP4 variant claims ~5x throughput on Blackwell. First openly-weighted 550B hybrid Mamba-Transformer, closing the gap with frontier closed models. β Google Gemma 4 12B: fully open dense any-to-any (text/image/audio/video), 256k context, encoder-free, 140+ languages, AIME 2026 at 77.5. Shipped with a 23-checkpoint QAT wave (mobile ONNX + MLX). Most deployable model of the week. β StepFun Step-3.7-Flash: 198B sparse MoE VLM, ~11B active, SWE-Bench PRO 56.3. Apache 2.0. β Liquid AI LFM2.5-8B-A1B: edge MoE, just 1.5B active, 128k ctx, MATH500 88.8, MLX-ready. Best on-device option this week. β JetBrains Mellum2-12B-A2.5B-Thinking: their first open MoE, near-Qwen3-14B coding at 2.5B active. Apache 2.0. π¨ Image gen (the surprise of the week) β Ideogram 4: their FIRST-EVER open weights. 9.3B flow-matching DiT trained from scratch. #2 overall behind GPT Image 2, top open-weight model on Design Arena + LMArena. Strongest open checkpoint for text-rich images, full stop. It has taste. Still can't believe this is open weights. π Audio & Speech (a breakout week for open TTS, 4 labs shipped) β Boson Higgs Audio v3 4B: 102 languages, 21 emotions, singing/whispering/shouting, sub-second TTFA. β RedNote dots.tts: the only fully continuous (no codec) open TTS pipeline, Apache 2.0. β Google Magenta RealTime 2: real-time music gen, <200ms latency, text+audio+MIDI. multimodalart ported it to PyTorch within hours with live ZeroGPU demos. β NVIDIA Nemotron-3.5 ASR: 600M streaming, 17x more concurrent streams vs Parakeet RNNT 1.1B. ποΈ Vision & VLMs β PaddleOCR-VL-1.6: SOTA document parsing at 1B params, Apache 2.0. β Baidu NAVA: 6.3B joint audio-video gen, best-in-class A/V sync, Apache 2.0. π¬ Video, 3D & World Models β NVIDIA Cosmos3-Super: 64B omnimodal world model coupling action trajectories with video+audio gen, for Physical AI. β JD JoyAI-Echo: up to 5-min multi-shot text-to-video on LTX-2.3. β ByteDance Bernini-R + VAST TripoSplat (single-image-to-3D Gaussian splats, MIT).
Can AI truly edit audio, not just generate it? π§ Tencent Hy, in collaboration with SJTU, SII, NTU, TJU, ZODA, PKU, FDU, and other collaborators, introduces MMAE. MMAE--A Massive Multitask Audio Editing Benchmark, is the first comprehensive evaluation benchmark for speech and audio "Bananaπ" Instead of simply requiring the AI to "generate" audio, it demands that the AI understand an existing audio clip and precisely modify it according to natural language instructionsβaltering what needs to be changed while leaving the rest untouched. Current models show an Exact Match Rate (EMR) below 5%, revealing a major gap in reliable audio editing. MMAE includes: β 2,000 high-fidelity samples from real-world scenarios β 17,741 fine-grained rubric evaluation items β 7 modality settings across sound, music, speech and their mixtures β 6 task complexity from basic modifications to multi-hop reasoning and multi-round editing β 8 operation types across local and global granularities How to use: arXiv: https://t.co/TM81ahH7PZ GitHub: https://t.co/UR1dRUKqMD HuggingFace: https://t.co/1MHR1n3LJn Demo: https://t.co/tz2TVHaCk8
Gemma 4 MTP just got officially merged into llama.cpp This means you can use Gemma 4 QAT + MTP for a lightweight + super fast setup. Excited to see what the community builds with it https://t.co/1te7tgdi2H
Best models for your hardware this week. 8-12GB - https://t.co/5SYi6D56FR incredible model, so fast, so small 16-32GB - latest Google model, Gemma 12B: https://t.co/TLm2x2l3lk really solid performance up neck and neck with a model 2x its size from a month ago. Jetbrains new model, best in class on livecode bench 32-96gb - Nex-N2-Mini GPT style postrain of Qwen-35B it seems to be its class leader caveman style reasoning https://t.co/EL1ePzwI58 - Jackrongβs Qwopus is the #1 overall Q4 of Qwen3.6-27B on our benchmark suite of 5 agent + coding benchmarks (1200 samples total) https://t.co/P1gypZwufi 192gb - Step-3.7-Flash is hard to beat, high scores, really fast inference, vision capable, later cutoff dates https://t.co/oaVf5wMILx 384gb - Nex-N2-Pro GPT style post train of Qwen-3.5-397B incredibly strong and #1 on deepswe if their claims are right https://t.co/LsGXZRl6nh 768gb - very promising post-train of GLM-5.1 that wins out on 8 benchmarks https://t.co/25KElLHEos
Best models for your hardware this week. 8-12GB - https://t.co/5SYi6D56FR incredible model, so fast, so small 16-32GB - latest Google model, Gemma 12B: https://t.co/TLm2x2l3lk really solid performance up neck and neck with a model 2x its size from a month ago. Jetbrains new model, best in class on livecode bench 32-96gb - Nex-N2-Mini GPT style postrain of Qwen-35B it seems to be its class leader caveman style reasoning https://t.co/EL1ePzwI58 - Jackrongβs Qwopus is the #1 overall Q4 of Qwen3.6-27B on our benchmark suite of 5 agent + coding benchmarks (1200 samples total) https://t.co/P1gypZwufi 192gb - Step-3.7-Flash is hard to beat, high scores, really fast inference, vision capable, later cutoff dates https://t.co/oaVf5wMILx 384gb - Nex-N2-Pro GPT style post train of Qwen-3.5-397B incredibly strong and #1 on deepswe if their claims are right https://t.co/LsGXZRl6nh 768gb - very promising post-train of GLM-5.1 that wins out on 8 benchmarks https://t.co/25KElLHEos
Today weβre releasing Quasar-Preview! Our first public proof that the Quasar architecture works at real scale. [ 18B MoE - 2B active / 5M context ] Built with Loop Transformer + Quasar attention Trained on Bittensor through decentralized infrastructure π https://t.co/TN4QZsqCNJ
