Daily Tech Feed: From the Labs

Why it matters. Autoregressive image generation has been limited by a fundamental tension: larger codebooks give better reconstruction but make next-token prediction harder. "BitDance: Scaling Autoregressive Generative Models with Binary Tokens" resolves this by replacing codebook indices with 256-dimensional binary vectors — giving each token 2²⁵⁶ possible states (more than atoms in the observable universe) while making sampling tractable through a novel binary diffusion head. On ImageNet 256×256, BitDance achieves an FID of 1.24, the best among all autoregressive models, and its next-patch decoding delivers 30× speedup at 1024×1024 resolution with 5.4× fewer parameters than prior parallel AR methods.

ByteDance, MMLab @ CUHK, and collaborators. This work is a collaboration between ByteDance, MMLab at The Chinese University of Hong Kong, Shanghai Jiao Tong University, the Institute of Automation at the Chinese Academy of Sciences, and the National University of Singapore. The paper is on arXiv (2602.14041). Code is available on GitHub, model weights on Hugging Face, and an interactive demo at the project page. All models are released under Apache 2.0.

The Researchers. Jiaming Han is a PhD student at MMLab@CUHK advised by Xiangyu Yue and serves as lead developer on the project (corresponding author). Hao Chen (Google Scholar) at ByteDance is a corresponding author specializing in video processing and representation learning. Xiangyu Yue is a professor at CUHK with affiliations at UC Berkeley and Stanford, known for multimodal learning and computer vision research. Huaibo Huang is at the Chinese Academy of Sciences working on generative models and face synthesis.

Key Technical Concepts. BitDance's three innovations: (1) A lookup-free binary tokenizer using group-wise binary quantization with entropy regularization — achieving 25.29 PSNR, surpassing even continuous VAEs (23.54 PSNR for Stable Diffusion's VAE). (2) A binary diffusion head that replaces the impossible softmax-over-2²⁵⁶-classes with a small diffusion model that jointly generates all 256 bits using velocity matching. (3) Next-patch diffusion that predicts multiple tokens in parallel (up to 64 per step), enabling massive speedups without quality loss. The approach builds on ideas from VQ-VAE, LlamaGen, and Open-MAGVIT2, while fundamentally rethinking the discrete representation paradigm.

Daily Tech Feed: From the Labs is available on Apple Podcasts, Spotify, and wherever fine podcasts are distributed. Visit us at pod.c457.org for all our shows. New episodes daily.

Why it matters. LLM agents that learn from experience typically store raw trajectories — verbose logs of every action taken. This is like studying for an exam by memorizing every page of the textbook instead of extracting key concepts. "SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning" introduces a framework that distills experience into compact, reusable skills and evolves them through reinforcement learning. Using Qwen 2.5 at just 7 billion parameters, SkillRL achieves 89.9% on ALFWorld — beating GPT-4o by 41.9 points and Gemini 2.5 Pro by 29.6 points — while compressing knowledge 10–20× compared to raw trajectory storage.

UNC Chapel Hill / AIMING Lab. SkillRL comes from the AIMING Lab at the University of North Carolina at Chapel Hill, with contributions from NEC Labs America. The paper is available on arXiv (2602.08234) with an HTML version. Code is open-sourced on GitHub. The framework uses Qwen 2.5-7B-Instruct as its base model and OpenAI's o3 as the teacher model for skill distillation. It builds on the GRPO reinforcement learning algorithm and is evaluated on ALFWorld, WebShop, and seven search-augmented QA tasks.

The Researchers. Huaxiu Yao, the senior author, is an Assistant Professor of Computer Science at UNC Chapel Hill and director of the AIMING Lab, working on adaptive intelligent agents, foundation models, and AI alignment. First author Peng Xia is a PhD student at UNC Chapel Hill focused on multimodal AI and agent systems. Cihang Xie is an Assistant Professor at UC Santa Cruz specializing in computer vision and machine learning. Zeyu Zheng contributes reinforcement learning expertise, while Xujiang Zhao and Haifeng Chen bring applied AI research experience from NEC Labs America.

Key Technical Concepts. SkillRL introduces three innovations: (1) experience-based distillation that extracts strategic patterns from successes and counterfactual lessons from failures into a hierarchical skill library called SkillBank; (2) adaptive retrieval that selects both general and task-specific skills at inference time; and (3) recursive evolution where the skill library co-evolves with the agent's policy during RL training. Prior approaches like Reflexion, ExpeL, and MemRL either store raw trajectories or do simple self-reflection — SkillRL's hierarchical abstraction is what enables the dramatic performance leap.

Daily Tech Feed: From the Labs is available on Apple Podcasts, Spotify, and wherever fine podcasts are distributed. Visit us at pod.c457.org for all our shows. New episodes daily.

Uses an LLM's own change in belief (ΔBelief) about the correct answer as a per-turn reward signal for RL training. Instead of sparse outcome rewards (right/wrong at the end), each intermediate action is rewarded based on whether the model's confidence in the right answer increased. Trains information-seeking behavior that generalizes across domains and scales beyond the training horizon.

Richard Sutton (2024 Turing Award) would credit the paper for addressing credit assignment — the core RL problem — with an intrinsic reward signal. But he would identify fundamental limitations inherited from the LLM substrate:

In July 2024, a French nonprofit's open-source voice AI went viral for demanding human sacrifice mid-conversation. Seven months later, the same team used the same architecture to build a real-time speech translator that runs on your phone. This is the story of how fundamental research survives embarrassment — and why transparency beats secrecy every time.

Kyutai is a €300 million AI research nonprofit founded in November 2023 in Paris by Xavier Niel (€100M), Rodolphe Saadé of CMA CGM (€100M), and Eric Schmidt via Schmidt Ventures. All research published, all code open-sourced. Their first project, Moshi (arXiv: 2410.00037), was the first real-time full-duplex voice AI — 200ms latency, both sides talking simultaneously — built on the Mimi neural audio codec at 12.5 Hz / 1.1 kbps. It beat OpenAI to market on real-time voice. The Blood God incident was training data contamination, not an architecture failure — Kyutai shipped it raw and let the world see. 9,600+ GitHub stars. Their latest, Hibiki (arXiv: 2502.03382, accepted at ICML), repurposes Moshi's multistream architecture for simultaneous speech-to-speech translation with voice transfer — state of the art on French-English, runs on phones via Hibiki-M. Try the audio samples or the Colab notebook. 1,386+ stars.

Six founding researchers left Meta FAIR and Google DeepMind to build Kyutai. Alexandre Défossez (co-lead) created Demucs and co-created EnCodec at Meta. Neil Zeghidour (co-lead) came from DeepMind specializing in audio/speech ML. Laurent Mazaré is the Rust implementation lead — his engineering makes these models run on consumer hardware. Edouard Grave is an NLP specialist. Patrick Pérez serves as CEO/Lab Director. Hervé Jégou contributed retrieval expertise. Scientific advisors: Yann LeCun, Yejin Choi, and Bernhard Schölkopf.

Daily Tech Feed: From the Labs is available on Apple Podcasts, Spotify, and wherever fine podcasts are distributed. Visit us at pod.c457.org for all our shows. New episodes daily.

In a 48-hour span on February 11-12, 2026, three Chinese AI laboratories independently released frontier-class language models under open licenses. Not research demos — full production models matching or exceeding GPT-5.2 and Claude Opus 4.5 on major benchmarks. Three labs, three architectures, all open-weight, all arriving at the same conclusion: sparse Mixture of Experts is the path to frontier AI that anyone can run. This episode breaks down what each lab built, who built it, and why it matters for the future of open-source AI.

StepFun released Step 3.5 Flash on February 11th — a 196 billion parameter sparse MoE model that activates just 11 billion parameters per inference. That means frontier-level intelligence running on a Mac Studio. The 45-page arXiv paper introduces MIS-PO, a novel reinforcement learning algorithm for stable off-policy training at scale. Step 3.5 Flash scores 97.3% on AIME 2025 and delivers 100-350 tokens per second via multi-token prediction. It's the only model of the three with a full technical report — StepFun published the recipe, not just the weights.

StepFun was founded by three former Microsoft veterans. CEO Jiang Daxin served as Vice President and Chief Scientist at Microsoft's Software Technology Center Asia. The name that stands out is Harry Shum (Heung-Yeung Shum), former Executive Vice President of AI and Research at Microsoft — one of the highest-ranking Chinese executives in American tech history. Shum ran Microsoft Research Asia, the lab that trained an entire generation of Chinese AI leaders including the founders of DeepSeek, SenseTime, and Megvii. He studied under Turing Award winner Raj Reddy at Carnegie Mellon. Other key contributors include Jiao Binxing and Zhu Yibo.

MiniMax released M2.5 on February 12th with a pitch of "intelligence too cheap to meter." At $0.15 per million input tokens, M2.5 costs one-twentieth what Claude Opus charges. The 230 billion parameter model activates roughly 10 billion parameters and scores 80.2% on SWE-bench Verified — within 0.7% of Claude Opus 4.5. MiniMax built Forge, an agent-native reinforcement learning framework with over 200,000 real-world training environments, and during training the model spontaneously began writing specification documents before coding. Two variants ship: M2.5 standard at 50 tokens per second and M2.5-Lightning at 100 tokens per second.

MiniMax was founded by Yan Junjie, who earned his PhD in AI at 25, became VP of SenseTime at 27, founded MiniMax at 31, and became a billionaire at 36 after MiniMax's Hong Kong IPO in January 2026. He holds roughly 200 academic papers with 30,000 citations and serves as both CEO and CTO — the only leader among these three labs wearing both hats. His vision of commoditized AI comes directly from deploying computer vision at massive scale at SenseTime. Key team members include Yang Bin and Zhou Yucong.

Zhipu AI (trading as Z.AI) released GLM-5 on February 11th — a 744 billion parameter model trained entirely on Huawei Ascend 910 chips. Zero NVIDIA hardware. Zhipu is on the US Commerce Department's Entity List, and they just trained the largest open-weight model of the three on domestic silicon. GLM-5 holds the #1 score on BrowseComp (75.9) across all models, open or closed. Before the official launch, Zhipu stealth-tested GLM-5 as "Pony Alpha" on OpenRouter, letting users benchmark it blind. The company IPO'd on the Hong Kong Stock Exchange on January 8th, raising $558 million — the first publicly traded foundation model company in the world.

Zhipu was founded by Tang Jie, professor at Tsinghua University and ACM Fellow who invented the GLM (General Language Model) architecture. He designed the original GLM in his lab, trained it at the Jinan Supercomputing Center, then commercialized it. Co-founder Li Juanzi, also a Tsinghua professor, is an expert in knowledge graphs and factual reasoning — her work is directly reflected in GLM-5's record-low hallucination rate. Zhang Peng rounds out the core leadership. Zhipu AI is essentially Tsinghua's computer science department, commercialized.

Daily Tech Feed: From the Labs is available on Apple Podcasts, Spotify, and wherever fine podcasts are distributed. Visit us at pod.c457.org for all our shows. New episodes daily.

Why it matters. Teaching robots to manipulate objects in the real world requires vast amounts of training data that is expensive and dangerous to collect. "DreamDojo: A Generalist Robot World Model from Large-Scale Human Videos" takes a different path: train a world model on 44,000 hours of egocentric human video — people cooking, cleaning, assembling, and manipulating objects in daily life — and then transfer that understanding of physics and dexterous control to robots with minimal fine-tuning. The key innovation is continuous latent actions that bridge the gap between unlabeled human video and precise robot commands. After distillation, DreamDojo runs at real-time speed (10.81 FPS at 640×480), enabling live teleoperation, policy evaluation without real-world deployment, and model-based planning. Thirty authors across nine institutions built this — it is a statement paper about the future of general-purpose robot world models.

NVIDIA. NVIDIA Research is the driving force behind DreamDojo, which builds on NVIDIA's Cosmos-Predict2.5 world foundation model platform. The paper is available on arXiv (2602.06949) with a project page showcasing video results. The GitHub organization currently hosts only the project website; model code has not yet been released. The DreamDojo-HV dataset of 44,000 hours represents the largest video dataset ever assembled for world model pretraining, spanning 6,015 skills across 1.1 million scenes — 15× longer, 96× more diverse, and 2,000× more scenes than any prior dataset.

NVIDIA's Leaders. Linxi "Jim" Fan (Google Scholar), project lead and head of NVIDIA's AI Agents initiative, earned his PhD at Stanford under Fei-Fei Li and was OpenAI's very first intern in 2016. He created Voyager, the first LLM agent to play Minecraft, and won the NeurIPS 2022 Outstanding Paper Award for MineDojo. Ming-Yu Liu (Google Scholar), Vice President of Research at NVIDIA and IEEE Fellow, leads the Deep Imagination Research group and created GauGAN, SPADE, pix2pixHD, and NVIDIA Cosmos. Yuke Zhu, project lead and Assistant Professor at UT Austin, built ROBOSUITE and ROBOMIMIC — benchmarks the entire robot learning community relies on.

UC Berkeley. Two of the field's most towering figures anchor the academic side. Jitendra Malik (Wikipedia, Google Scholar) is the Arthur J. Chick Professor of EECS at UC Berkeley, a member of both the National Academy of Sciences and the National Academy of Engineering, and a pioneer of texture analysis, image segmentation (Normalized Cuts), and object recognition — his academic family tree spans nearly every major computer vision lab in the world. Pieter Abbeel (Wikipedia, Google Scholar) is a Professor of EECS at UC Berkeley and co-founder of Covariant, one of the leading robotics AI companies, with foundational contributions to apprenticeship learning, deep reinforcement learning, and robotic helicopter acrobatics. Co-first author William Liang is a UC Berkeley PhD student and NVIDIA research intern.

HKUST and UT Austin. Co-first author Shenyuan Gao is a PhD student at HKUST and NVIDIA research intern whose prior work on AdaWorld pioneered the latent action approach that DreamDojo scales up. Joel Jang serves as project lead from NVIDIA Research.

Daily Tech Feed: From the Labs is available on Apple Podcasts, Spotify, and wherever fine podcasts are distributed. Visit us at pod.c457.org for all our shows. New episodes daily.

Why it matters. Diffusion models generate stunning images but require dozens or hundreds of iterative steps at inference time, making them slow and expensive. "Generative Modeling via Drifting" introduces a fundamentally different approach: instead of iterating at inference, Drifting Models move the iterative process into training itself. A "drifting field" based on attraction toward real data and repulsion from generated samples evolves the output distribution during training, so that at inference time a single forward pass is all you need. The result is state-of-the-art image generation on ImageNet 256×256 — FID 1.54 in latent space, 1.61 in pixel space — with a single network evaluation, and a base-size model (133M parameters) that competes with models five times its size. The approach also transfers to robotics, where a "Drifting Policy" matches the performance of 100-step Diffusion Policy in a single step.

MIT and Harvard. This work comes from MIT CSAIL and Harvard University, with the team operating in Kaiming He's lab at MIT. The paper is available on arXiv (2602.04770) with an interactive project page demonstrating the core algorithm on toy distributions. Code and ImageNet training models have been promised but not yet released. The method builds on self-supervised feature extractors including MoCo, SimCLR, and latent-MAE, and uses the Stable Diffusion VAE tokenizer with a DiT-like architecture.

The Researchers. Kaiming He, the senior author, is a professor at MIT EECS and one of the most cited researchers in all of computer science — his ResNet paper alone has over 200,000 citations, and he pioneered self-supervised learning methods MoCo and MAE as well as Mask R-CNN and Feature Pyramid Networks during his years at Meta AI (FAIR). First author Mingyang Deng is a PhD student at MIT CSAIL focused on generative models. Yilun Du is an Assistant Professor at Harvard known for energy-based models, compositional generation, and applying diffusion models to robotics and planning. Tianhong Li is an MIT researcher whose prior work on MAR (Masked Autoregressive generation) has advanced autoregressive approaches to image synthesis. He Li rounds out the team as an MIT researcher contributing to the generative modeling pipeline.

Daily Tech Feed: From the Labs is available on Apple Podcasts, Spotify, and wherever fine podcasts are distributed. Visit us at pod.c457.org for all our shows. New episodes daily.

Why it matters. In June 2017, eight researchers at Google published a 15-page paper that would become the most consequential machine learning contribution of the decade. "Attention Is All You Need" introduced the Transformer architecture, replacing the recurrent and convolutional networks that dominated sequence modeling with a mechanism built entirely on attention. The result was a model that trained faster, parallelized better, and scaled beyond anyone's expectations — becoming the foundation for GPT, BERT, PaLM, Gemini, Claude, and virtually every large language model that followed. With over 173,000 citations, it is one of the most cited papers in the history of computer science.

Google Brain / Google Research. The Transformer was born at Google Brain and Google Research, where the eight authors — all listed as equal contributors in randomized order — were working on machine translation. The paper was published at NeurIPS 2017 and is available on arXiv (1706.03762). Google's original implementation lives in the now-archived tensor2tensor repository, though today the HuggingFace Transformers library is the de facto standard. For learning the architecture from scratch, The Annotated Transformer from Harvard NLP, Jay Alammar's Illustrated Transformer, and Andrej Karpathy's "Let's build GPT" are essential resources.

The Transformer Eight. The paper's authors have scattered across the AI landscape in remarkable fashion. Ashish Vaswani co-founded Adept AI and then Essential AI, where he serves as CEO. Noam Shazeer co-founded Character.AI before returning to Google in a deal reportedly worth $2.7 billion, where he now co-leads Gemini development as VP of Engineering at Google DeepMind. Niki Parmar co-founded both Essential AI and Adept AI before joining Anthropic. Jakob Uszkoreit took the attention mechanism into biology, co-founding Inceptive to design RNA molecules with AI. Llion Jones moved to Tokyo and co-founded Sakana AI, pursuing nature-inspired approaches to artificial intelligence. Aidan Gomez — just 20 years old when the paper was published as a Google Brain intern — went on to co-found Cohere, now a leading enterprise AI company. Łukasz Kaiser continued foundational research at OpenAI and co-created the Trax deep learning library. And Illia Polosukhin pivoted entirely, co-founding NEAR Protocol, a layer-1 blockchain, and becoming CEO of the NEAR Foundation.

Daily Tech Feed: From the Labs is available on Apple Podcasts, Spotify, and wherever fine podcasts are distributed. Visit us at pod.c457.org for all our shows. New episodes daily.