Qualcomm's Snapdragon X Elite and X Plus chips are turning Windows on Arm into a viable platform, competing closely with Apple's M3 chip, Intel Core Ultra, and AMD Ryzen processors in performance and battery life. Despite not leading in GPU performance, the Snapdragon chips excel in multi-core benchmarks and power efficiency. The new wave of Snapdragon-equipped laptops are cost-effective alternatives to MacBooks and traditional Intel or AMD-based systems.
PaliGemma is a strong vision language model built on SigLIP and Gemma 2B. This technical report shows many of the decisions made in architecture choices and data collection.
FlashAttention is the foundation of modern fast language models. This new version uses 75% of H100 capacity (up from 35% before). There are a number of important system improvements that led to this capability gain.
OV-DINO is a new method for open-vocabulary detection that tackles the challenges of integrating diverse data sources and leveraging language-aware capabilities.
Demystify the “Agentic AI” hype with the experts at OctoAI and watch this builder-focused session [10] on the foundations of function calling, the first layer in developing AI agents.
OVFormer is a novel method for Open-Vocabulary Video Instance Segmentation (VIS) that addresses key issues in the field. It improves embedding alignment and leverages video-based training to enhance temporal consistency.
This study addresses semantic change detection using satellite image time series (SITS-SCD) by integrating both change detection and semantic segmentation.
The PosFormer model significantly improves Handwritten Mathematical Expression Recognition (HMER) by addressing the limitations of sequence-based models.
This article discusses bridging the gap between techno-optimists and pessimists on AI and advocates for a pragmatic approach to using AI as an augment to human intelligence. It covers AI's impact on creativity and art, the ethical use of AI, and AI's role in education. Suggesting the term "Augmented Intelligence" as a more accurate descriptor, the article emphasizes that AI is not a threat, but a tool that enhances human capabilities.
Anthropic's new features in Claude allow developers to automate prompt engineering, improving AI application development by generating, testing, and refining prompts with quick feedback.
Microsoft and Apple have opted not to join OpenAI's board due to regulatory scrutiny, with Microsoft stepping down from its observer role and Apple reversing plans to become an observer.
Anole is an open autoregressive multimodal model built on Meta's Chameleon model. This work adds image generation capabilities back into the model with targeted fine-tuning.
From the creator of ControlNet, IC-Light, and many other image generation systems, Paints UNDO is a system where a model creates strokes that are used to recreate an image. Interestingly, unlike previous stroke systems, this model can undo strokes and often rethinks its approach entirely part way through - very similarly to a human artist.
Scalable Rectified Flow Transformers are used, in part, for Stable Diffusion 3. This repository is a minimal implementation of this that also includes muP hyperparameter sweeps.
One popular area of VC investing is building the “picks and shovels” for AI (e.g. model fine-tuning, observability, and inference). VCs are betting that companies will not want to build things in-house as they add AI to their products. The TAM for AI inference is quite constrained. VCs need to believe in massive TAM expansion for their investments to pay off. While AI inference platforms are a boon for startups in the short term, they are detrimental to startups over longer periods.
Cyclists in San Francisco have reported close encounters and safety concerns with Waymo's and Cruise's autonomous vehicles. Nearly 200 complaints have been made to the California DMV highlighting erratic driving and near-misses involving these self-driving cars. While the companies claim their vehicles have enhanced safety features, the incidents raise questions about their readiness for broader deployment amidst increased regulatory scrutiny.
OpenAI and Los Alamos National Laboratory (LANL) are collaborating to explore safe applications for AI in bioscientific research, with a focus on using models like GPT-4o to assist in laboratory tasks through multimodal capabilities such as vision and voice.
Selling AI to enterprise means getting serious about security and compliance — usually involving features like single sign-on (SSO), Directory Sync (SCIM), RBAC, and Audit Logs.
xAI has ended its deal with Oracle. It will build its own datacenter after Grok 2 training finishes. The company originally contracted for 24k H100s from Oracle.
VC firm Andreessen Horowitz has secured thousands of AI chips, including Nvidia H100 GPUs, to dole out to its AI portfolio companies in exchange for equity.
Quora's Poe has introduced a feature called Previews, enabling users to create interactive apps such as data visualizations and games directly through chatbot interactions.
System 2 models are those that use something like Chain of Thought to spend more test time compute and improve reasoning. It turns out that you can distill this behavior into a System 1 model that runs faster with similar accuracy.
A new RNN variant that outperforms Mamba on a number of tasks. Importantly, the update function is itself an ML model - which enables long contexts and in-context learning.
A universal semantic layer makes it much easier to run both BI and AI workloads [12] on your data warehouse. AtScale [12], now available as a Snowflake Native App in the Snowflake Marketplace, allows you to define and consume a universal semantic layer directly within your Snowflake account.
Tinygrad is a toolkit for training deep learning models. This set of notes is a great introduction for AI compilers and prepares one for a deep dive into Tinygrad internals.
Goldman Sachs released a critical 31-page report titled "Gen AI: Too Much Spend, Too Little Benefit?", arguing that generative AI's productivity benefits and returns are significantly limited and that its power demands will drastically increase utility spending. The report highlights doubts about AI's ability to transform industries, pointing out high costs, power grid challenges, and lack of clear productivity gains or significant revenue generation. It suggests a potentially bleak future for the technology without major breakthroughs.
The rapid adoption of ChatGPT, which reached 100 million users in two months, stands in contrast to historical tech developments like the iPhone and e-commerce, which took years to gain traction. Despite the initial hype, many users have not found long-term utility with ChatGPT and enterprise deployment of large language models remains limited, indicating a need for further development to achieve meaningful product-market fit and sustained value.
This project investigates the journey towards Artificial General Intelligence (AGI), which aims to perform diverse real-world tasks with human-like efficiency.