Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Anthropic has announced a new batch processing API that enables developers to send batches of up to 10,000 queries per batch. Each batch is processed in less than 24 hours and costs 50% less than standard API calls. This makes processing non-time-sensitive tasks more efficient and cost-effective.
A deep dive post mortem on what happened with the much hyped Reflection-70B model. It had bugs in its benchmark code. Its team has fixed them and performance is still very impressive, but not quite to the level they claimed.
Model merging is a way to combine two models and get the performance of each. However, it doesn't always scale properly with model size. This paper explores what is needed to make it work well with a really large model.
Normalizing the internal representation of language models is the target of a lot of resource activity. This work shows that if you put every internal vector on a hypersphere, it dramatically decreases convergence time for reasonably sized models.
Language models are getting so good at writing code that many benchmarks are saturating. This paper suggests a benchmark that is harder and aims to determine how models perform on reasoning and code generation tasks.
Sponsored
on October 16 to learn how to boost developer productivity and improve code security in your SDLC. You'll explore how to harness the full potential of AI coding assistants with Sonar's latest capabilities -
The intensify package is a nice way to show logprobs or other real number modulated text based on the color of the terminal.
JEDi is a new metric based on Joint Embedding Predictive Architecture (JEPA). It improves evaluation accuracy while needing fewer samples and better aligning with human assessments, offering a stronger alternative to FVD.
PRFusion and PRFusion++ are multimodal models designed to improve place recognition in robotics and computer vision.
ETH Zurich used reinforcement learning to enable the ANYMal robot to climb ladders with 90% success rate. Enhancing quadrupedal robots' capabilities in industrial settings, the system improved climbing speed by 232x compared to existing methods. This research highlights the integration of robot morphology and advanced control policies.
This article explores the pitfalls of data leakage in machine learning, highlighting two scenarios where improper data handling led to misleading model performance. One company made the mistake of filtering data based on an upper price limit before modeling, while another faced issues by not adhering to a strict chronological split. Key takeaways include the importance of detecting leakage and understanding its impact on model accuracy.
AI development faces potential challenges from a "data wall" as language models approach training on all available text. This article argues against relying on human analogies for overcoming data limitations, highlighting the vast data and evolutionary processes that contribute to human intelligence. While human learning strategies might not directly apply to AI, this doesn't rule out other modalities or algorithmic progress for advancing AI capabilities.
Sponsored
Better code = better software. Sonar's tools and code assurance workflows instill confidence and create accountability for AI-generated code, helping you produce software of unparalleled
Braintrust, which helps Airtable, Brex, Notion, and Stripe build AI products, has raised $36M in a Series A led by a16z.
Hamming AI brings trust to AI voice agents.
Clout Kitchen raised $4.45 million in seed funding to develop AI-powered "Backseat AI" buddies for League of Legends that offer gameplay tips and commentary in the voices of popular creators.
Sponsored
Private credit—i.e., loans and debt financing from non-bank lenders—is a resilient asset class that's averaged a 12% return globally (and 14.5% on Percent [5]). It's been popular with institutional investors for a while, especially as it has outperformed both stocks and bonds in the last 3 market downturns.
Kobold Materials, a mining startup that uses AI to identify and extract metals, has raised a substantial new round of funding.
OpenAI has secured over $10 billion in liquidity, reaching a $157 billion valuation after its latest funding round. The company raised $6.6 billion from major investors, including Microsoft and Nvidia, but faces high operational costs, primarily due to the need for more GPUs to support LLM training. OpenAI is exploring restructuring options to optimize financial growth and sustainability in the AI ecosystem.
Black Forest Labs has launched a beta API providing its Flux image generation models, including the new Flux1.1 Pro, which offers faster generation and enhanced image quality. The startup, backed by Andreessen Horowitz, aims for market growth with ongoing developments despite controversy over its data usage and competition from major players like OpenAI and Google.
The Dynamic Diffusion Transformer (DyDiT) enhances the efficiency of diffusion models used in image generation. It improves upon the existing Diffusion Transformer (DiT) by dynamically adjusting computational efforts during different timesteps and spatial regions, reducing redundancy.
The Frame-Aware Video Diffusion Model (FVDM) improves video generation by addressing limitations in current models. Unlike traditional approaches that use a single timestep for the whole clip, the FVDM uses a vectorized timestep variable, allowing each frame to follow its own noise schedule.
This study reveals significant flaws in current membership inference attacks used to detect unauthorized data usage in diffusion models. CopyMark is a more realistic benchmark for evaluating MIAs on pre-trained models that offers unbiased datasets and fair evaluation methods.
Sponsored
On October 10th, Intercom is bringing together customer service leaders to explore the impact and opportunity of AI. Hear success stories from businesses like yours, learn strategies to integrate AI into your support operations, and see the launch of Fin 2, the first
ImageFolder is a semantic tokenizer designed to address the trade-off between image reconstruction and generation quality in visual generative models.
Grounded-VideoLLM is a new Video-Large Language Model designed to improve fine-grained understanding of specific video moments. By adding a temporal stream and discrete temporal tokens, the model better captures relationships between frames and timestamps.
Chunking Causal Transformer (CCT) is a new autoregressive architecture designed for robotic manipulation tasks.
FacePoke is a tool for quickly editing faces in videos and images.
LLMs rely on tremendous amounts of unstructured data for training, but this data must be cleaned and processed before it's useful. Existing data processing systems based on CPUs and today's software architectures weren't designed to handle the scale or complexity of this data, making data cleaning and processing a slow and expensive process that limits AI development. DataPelago has launched a Universal Data Processing Engine built to overcome existing data processing performance, cost, and scalability limitations that make AI development faster and cheaper.
Google will show ads in its AI Overviews to counter alternatives like ChatGPT while enhancing user engagement. Despite generating controversies over misinformation, AI Overviews have boosted search engagement, especially in younger demographics. Google is also rolling out AI-organized search results in the U.S. to diversify content presentation without including ads.
OpenAI's rushed development and safety testing of its latest models, including GPT-4o and o1, have caused internal friction and senior staff departures.
This professor quit teaching due to the pervasive use of LLMs like ChatGPT among students.
A great 200 line example of parallelism from scratch from a large model training lead at Hugging Face.
Sponsored
Would you spend one dollar if it meant you could travel the world?
Meta has trained a state-of-the-art movie model. It has 30B parameters and took 6,144 H100s to train. The model was trained on 1B images and 100m highly curated videos. Interestingly, it is a Temporal Autoencoder and Flow matching Llama. Meta released a 92 page paper, one of the most detailed research reports on the topic.
The Open FinLLM Leaderboard offers a specialized evaluation platform tailored for financial language models. It focuses on tasks crucial for finance, such as stock movement prediction, sentiment analysis, and information extraction from financial reports.
Nvidia's RTX-Acceleration combined with Ollama allows for running local models in the browser.
Large language models have limitations due to next token prediction. However, OpenAI's o1 model was trained with a new objective over reasoning traces. Despite this change, o1 still shows signs of limitations due to next token prediction.
This paper describes a method that is almost like a neutral TF/IDF in that it takes information from the entire corpus instead of just a single embedding document. It essentially encodes contextual information about neighboring documents. This method is state-of-the-art on MTEB.
This paper shows a novel design architecture for using encoder-only transformers as next token prediction models. This has been done in other ways before, but the researchers find a great way to prevent cheating. The models have hints of strong performance at scale and also are capable of in-context learning.
Sponsored
In advertising, timing is everything. Reach software developers, AI/ML engineers, executives and other tech professionals as they actively read TLDR. With 10 interest-based newsletters, TLDR helps you reach your target audience. Learn more [11] about launching your first campaign with us.
New model testing is limited in the image domain by poor-quality small datasets like CIFAR10. This GitHub repository contains a tool that creates infinite complicated fractals as images or videos for testing.
An immensely viral repository that uses language models to automate the job application process and add an additional layer of personalization to the process.
This project proposes a new technique called Pairwise Relevance Distillation (PairDistill) to improve the accuracy of dense retrieval methods.
This study develops high-performance graph algorithms that help solve complex, large-scale problems like detecting fraud in financial networks or finding efficient delivery routes.