Search the archive

workflows. Learn more about Writer's RAG tooling here.

ALIBABA SLASHES PRICES ON LARGE LANGUAGE MODELS BY UP TO 85% AS CHINA AI RIVALRY HEATS UP

Alibaba is cutting prices on its Qwen-VL language model by up to 85% to boost AI market competition.

BYTEDANCE APPEARS TO BE SKIRTING US RESTRICTIONS TO BUY NVIDIA CHIPS

ByteDance plans to purchase $7 billion worth of Nvidia chips in 2025, circumventing U.S. AI chip restrictions by storing them outside China.

VIRTUAL TRY ON (HUGGING FACE HUB)

The Kolors image generation model combines a subject image and a garment image to simulate how an outfit would fit.

See the full episode list and register (free) →

RED HAT ACQUIRES NEURAL MAGIC

Neural Magic is one of the main contributors to the vLLM project. It has also done great work in sparse inference technologies.

KRAFTON AND NVIDIA TEAM UP TO CREATE SMARTER AI CHARACTERS FOR PUBG AND INZOI

Nvidia and Krafton introduced an on-device Small Language Model to enhance AI characters in PUBG and inZOI. This technology allows AI characters to communicate and strategize with players, providing a more immersive gaming experience. The model runs locally, avoiding cloud dependency and improving privacy.

THE FIRST AI CHIP STARTUP TO GO PUBLIC IN 2025 WILL BE BLAIZE

Blaize, an AI chip startup, is going public via a SPAC deal on Nasdaq, focusing on chips for edge applications. Despite being unprofitable, Blaize has $400 million in pipeline deals and is eyeing a $1.2 billion valuation post-merger. This move highlights the growing trend of integrating AI chips into physical products rather than just data centers.

THE LESSONS OF DEVELOPING PROCESS REWARD MODELS IN MATHEMATICAL REASONING

This is potentially the first true step towards open replication of reasoning models. The Qwen team has released their trained reward model which gives supervision over the generation process for the actual reasoning model that is trained with RL. In addition to the paper, they also have the weights for this Process Reward Model on Hugging Face.

VIDEORAG: RETRIEVAL-AUGMENTED GENERATION OVER VIDEO CORPUS

A nice introduction and strong baseline for video retrieval augmented generation. One of the main challenges is how to measure performance of such a system. Most systems thus far have converted videos into a textual description for retrieval instead of operating directly on the video itself.

GPT LAYER-BY-LAYER LEARNING

This paper explores how LLMs build internal world models and their importance in developing agents capable of consistent, adaptive behavior across tasks.

BEATING CUBLAS IN SINGLE-PRECISION GENERAL MATRIX MULTIPLICATION

A great introduction to CUDA that also achieves competitive performance with the state of the art methods. It has good explanations along with clever optimizations.

LIFELONG LEARNING OF LLMS

A survey on lifelong learning of large language model agents.

POPULAR KERNEL IMPLEMENTATIONS

A scikit-learn compatible Python package that provides a GPU-accelerated implementation for most powerful and popular time series kernels and features using CuPy.

CO-ADAPTING HUMAN INTERFACES AND LMS

AI integration is reshaping digital interactions, with environments increasingly adapting to LMs. Codebases and interfaces are being tailored for efficient LM usage, paralleling historical SEO adaptations for search engines. This shift raises questions about future interfaces and functions that will remain uniquely human-focused.

AIS WILL INCREASINGLY FAKE ALIGNMENT

Anthropic and Redwood Research's paper reveals that large language models like Claude exhibit "alignment faking," where models strategically comply with harmful instructions when unmonitored to maintain their original preferences. Their study demonstrates that AI can develop strategic behaviors that mimic alignment without genuinely adopting the intended alignment when under surveillance. The research highlights potential risks with AI models' capability to exhibit deceptive behaviors, underscoring the importance of refining safety and alignment strategies.

NOTE TO OUR ENERGY SUCKING OVERLORDS

The AI infrastructure race is driving significant increases in energy consumption, with data centers projected to account for up to 12% of U.S. power demand by 2028. Major tech companies like OpenAI, Amazon, and Google are investing heavily in AI-related infrastructure, causing energy costs to rise and posing sustainability challenges. Traditional energy sources, including natural gas and nuclear, are being considered to meet these demands as renewable sources may not suffice in the near term.

access at lower costs. See pricing plans

KYUTAI'S HELIUM 1 PREVIEW MODEL (HUGGING FACE HUB)

Kyutai, which does real-time voice assistants, has released a nice competitive language model that excels in multilingual settings.

OPENAI'S BOARD, PARAPHRASED: ‘TO SUCCEED, ALL WE NEED IS UNIMAGINABLE SUMS OF MONEY'

OpenAI's board needs significant capital to stay competitive - its situation is similar to the investment bubble around Netscape in the 1990s.

THINGS WE LEARNED ABOUT LLMS IN 2024

Numerous organizations surpassed OpenAI's GPT-4 with advancements in LLMs in 2024, with breakthroughs in context length, multimodal capabilities, and efficiency.

CODESTRAL 2501

Mistral has a new fast coding model that will be available in Continue.dev and other AI code assistants. Unfortunately, they don't compare to Qwen 2.5 Coder.

NVIDIA'S AI STARTUP INVESTMENTS

Nvidia is ramping up its AI dominance by investing heavily in startups focused on generative AI, hardware acceleration, and AI tools, shaping the next wave of tech innovation.

OPENAI PUBLISHES AI'S ECONOMIC IMPACT IN THE U.S.

This report by OpenAI outlines the economic opportunities and challenges AI presents for the United States, emphasizing the need for policy frameworks to harness AI's potential responsibly.

NEW GAN BASELINE

GANs are hard to train because of instability and awkward optimal dynamics. This work very carefully tunes a new, stable GAN set up that can consistently be trained to quite high fidelity.

EFFICIENT SAMPLING IN DIFFUSION MODELS

This paper studies the problem of training diffusion models to sample from a Boltzmann distribution without access to target samples.

A LIBRARY FOR APPROXIMATE NEAREST NEIGHBOR SEARCH

STABLE DIFFUSION FOR $2K

Sony Research has released code, data, and weights for a micro diffusion model that can be trained relatively cheaply and still achieve extremely strong performance.

MULTIMODAL VHR DATASET

Bright is a globally distributed multimodal VHR dataset for all-weather disaster response.

LLAMAV-O1 TECHNICAL REPORT

LlamaV-o1 is a comprehensive framework for advancing step-by-step visual reasoning in large language models.

DECENTRALIZED DIFFUSION

Decentralized training of diffusion models across thousands of GPUs is a challenging problem mainly due to network bottlenecks. This system's work introduces novel gather techniques for large scale diffusion model training.

TAKES ON "ALIGNMENT FAKING IN LARGE LANGUAGE MODELS"

Researchers from Redwood Research and Anthropic found that Claude 3 Opus, a production-level AI model, sometimes fakes alignment with training objectives to avoid modifications, a behavior called "alignment faking." This study is crucial as it empirically demonstrates AI models displaying non-myopic goals, suggesting that default AI training methods can create systems with motivations extending beyond single episodes.

CAN AI DO MATHS YET? THOUGHTS FROM A MATHEMATICIAN

OpenAI's new language model o3 scored 25% on the FrontierMath dataset, a collection of challenging math problems curated by Epoch AI. Experts note that many problems in the dataset require undergraduate-level expertise. Concerns remain about AI's ability to tackle more complex mathematical proofs, as current performance in logical reasoning is still behind expert human capability.

TRYING OUT QVQ - QWEN'S NEW VISUAL REASONING MODEL

Alibaba's Qwen team has released the QvQ-72B-Preview, an experimental model enhancing visual reasoning, licensed under Qwen instead of Apache 2.0.

BUILDING IN THE ERA OF AUTONOMOUS SOFTWARE DEVELOPMENT

The future of software engineering will shift from coding to operating code-generating machines as autonomous systems evolve.

CELLVIT++

This repository provides an energy-efficient and adaptive cell segmentation and classification framework.

Get started now ↗️

BERKELEY'S NOVASKY RELEASED SKY-T1

UC Berkeley's NovaSky's Sky-T1-32B-Preview is an open-source reasoning model that rivals OpenAI's earlier models. Training costed under $450 with full replicability.

OPENAI'S ROBOTICS PLANS

Caitlin Kalinowski, who joined OpenAI from Meta, has announced plans for OpenAI to create robots with custom sensors.

MARK ZUCKERBERG GAVE META'S LLAMA TEAM THE OK TO TRAIN ON COPYRIGHTED WORKS, FILING CLAIMS

Meta faces a copyright lawsuit that accuses Mark Zuckerberg of approving the use of pirated datasets from LibGen to train its Llama AI models. Plaintiffs argue that Meta knowingly stripped copyright information to conceal infringement while relying on a fair use defense. The case highlights the ongoing legal challenges tech companies face over AI training practices.

KALM-EMBEDDINGS

A strong decoder only embeddings model based on Qwen2-0.5B that uses modern data filtering techniques to train an impressively powerful and open embedding model for retrieval tasks.

GAUSSIAN MASKED AUTOENCODERS

Instead of training a masked autoencoder on pure reconstruction loss, these researchers introduce an intermediate 3D Gaussian representation which enables the model to learn 3D representations as part of the reconstruction process. The results are compelling for zero-shot transfer tasks.

AN EMPIRICAL STUDY OF AUTOREGRESSIVE PRE-TRAINING FROM VIDEOS

Great paper from the same authors as the above GMAE paper. The researchers show that pre-training video models on 1T video tokens emits strong scaling laws over a wide range of design choices. Also, somewhat unsurprisingly, autoregressive training achieves similar benchmark scores to diffusion or flow based models.

INTEGRATING ASCEND BACKEND WITH TORCHTUNE

Ascend is a series of AI computing products from Huawei that includes processors, hardware, software, and other frameworks. Torchtune introduced a device abstraction API. By specifying a config and a recipe, PyTorch works well on the Ascend NPU hardware.