Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Meta has a great new paper about using a mixture of judges models to properly perform multi task RLHF in post training. It dramatically improves final performance of the models on a number of benchmarks.
This survey examines the honesty of LLMs, a key factor in aligning AI with human values. It reviews challenges such as models confidently presenting wrong answers and difficulties in distinguishing known from unknown knowledge.
has your back—TRY IT FOR JUST $1 IF YOU SIGN UP IN THE NEXT 12 HOURS [10]. That means; Hawaii from $174, Italy from $291, or France from $295 round trip. Join over 1 million members saving on travel this
Sponsored
LexEval is a benchmark designed to evaluate large language models (LLMs) in the legal domain. Given the high demands for accuracy, reliability, and fairness in legal applications, LexEval aims to help assess the strengths and weaknesses of LLMs in legal contexts.
PerCo (SD) is a new perceptual image compression method based on Stable Diffusion v2.1 aimed at ultra-low bit ranges.
This project highlights the need to evaluate large language models (LLMs) on their combined abilities, not just individual skills. While most models are trained on specialized datasets targeting specific capabilities, real-world tasks often require a blend of expertise across multiple areas, referred to as cross-capabilities.
Nvidia did a really great ablation on different ways to include images into a language model. It found that the LlaVa concatenation approach performs the best.
AI is rapidly expanding access to traditionally expensive and inaccessible services like education, healthcare, and personal styling. Generative AI, like ChatGPT, provides affordable, personalized assistance, acting as tutors, healthcare advisors, and stylists, reducing the reliance on costly human professionals. This shift democratizes access, making high-end services available to a wider audience at a fraction of the cost.
OpenAI CEO Sam Altman has proposed a global initiative to build data centers and chip factories to support advanced AI development. Altman initially sought trillions but is now targeting hundreds of billions. The plan involves collaboration with global tech giants and governments but faces regulatory and logistical challenges. Despite initial skepticism, discussions continue with potential expansions in the US, Europe, and Asia to bolster computing power for AI.
Inbox Zero is an open source AI personal assistant for email.
OpenAI plans to restructure into a for-profit benefit corporation to attract investors - its non-profit will retain a minority stake.
While LLMs can boost productivity, they lack critical thinking and self-awareness.
Sponsored
While rule-based code reviewers have been around for some time, AI code reviews [4] are new and getting tons of attention. Many believe them to not only be a step-function change over predecessors like SonarQube, but a valid replacement of human code reviewers.
SoftBank plans to invest $500 million in OpenAI's funding round, boosting its valuation to potentially $150 billion. Microsoft is also participating in the round, which reflects OpenAI's 1,700% revenue growth despite a projected $5 billion in losses.
OpenAI's recent financial documents reveal $300 million in monthly revenue, a 1,700% increase since early 2023, with annual sales projected to reach $3.7 billion. Despite this growth, the company expects to lose approximately $5 billion this year due to high operational costs. OpenAI is seeking $7 billion in a new funding round, which will value the company at $150 billion.
OpenAI CEO Sam Altman is urging the Biden administration to build AI data centers consuming up to five gigawatts of power to maintain US technological leadership over China. The plan details constructing multiple data centers in the US. Other tech giants, like Microsoft and Amazon, are also securing nuclear power agreements to support their AI operations.
Emu 3 is a next token prediction model that outperforms SDXL on image synthesis, LlaVa-1.6 on image understanding, and OpenSora 2 on Video generation. It is a 9B parameter model trained on all these tasks in an interleaved manner, similar to Gemini.
Using a pretrained diffusion model for tasks like depth estimation is extremely popular and powerful. This work shows how some of the previous methods were slightly wrong and improves performance while dramatically simplifying the modeling process.
SegVLAD is an approach for visual place recognition that focuses on image segments rather than entire images.
A fork of CleanRL that has been optimized to use PyTorch's newest performance and stability features. It is dramatically faster while also being simpler to understand and extend.
MaskLLM is a pruning method that reduces computational overhead in large language models by applying learnable sparsity.
E.T. Bench is a new benchmark designed to evaluate video language models on fine-grained, event-level tasks. Unlike previous benchmarks that focus on video-level questions, E.T. Bench covers a range of time-sensitive tasks across multiple domains.
Researchers explore how AI detects "illusory" faces—seeing faces in inanimate objects—using a new dataset.
This article highlights how large language models (LLMs) are revolutionizing table extraction from complex documents, overcoming the limitations of traditional methods like OCR, rule-based systems, and machine learning. LLMs demonstrate flexibility and contextual understanding, notably enhancing accuracy in diverse and intricate table structures. Despite challenges like hallucination and high resource demands, combining traditional techniques with LLMs is currently the most effective strategy for automated table extraction.
Microsoft considered reallocating its US-based server power to GPUs for AI but ultimately scrapped the plan. Big Tech companies, including Microsoft, Google, and Amazon, heavily invest in AI but primarily see underwhelming returns in generative AI applications. The industry's reliance on SaaS and the integration of AI tools, which often add little genuine utility while incurring high costs, highlights a growing desperation to maintain growth amidst a slowing market.
Wispr Flow is an AI dictation app that lets you speak naturally and write in your style across every application.
Samsung has unveiled the Galaxy Tab S10 series, which features AI-enhanced functionalities.
Tesla's FSD exhibited dangerous behavior requiring human intervention every 13 miles during a 1,000-mile evaluation by AMCI Testing.
AI is poised to amplify human capabilities, enabling tasks that seem unimaginable today. Through deep learning and increased computational resources, AI will offer advancements like personal assistants, educational tutors, and healthcare advisors. The focus must remain on making AI accessible and mitigating its risks, paving the way for widespread prosperity in the Intelligence Age.
Apple has opted out of investing in OpenAI's $6.5 billion funding round, leaving Microsoft and Nvidia as potential investors. Microsoft is reportedly expanding its investment in OpenAI by another $1 billion. Despite not investing, Apple and OpenAI are partnering to integrate ChatGPT into iOS 18.
AlphaChip is a reinforcement learning model that significantly accelerates and optimizes chip design processes. It generates superhuman chip layouts, currently used in Google's TPUs, in hours rather than weeks. This AI-driven method has broad applications, influencing both Google's hardware and external organizations like MediaTek.
The Parallel Gated Network (PGN) is a novel architecture designed to overcome the limitations of traditional RNNs in handling long-term dependencies. It reduces the information propagation path and uses gated mechanisms to effectively capture historical and current time step information.
DoSSR is a diffusion-based super-resolution model that enhances both performance and efficiency by leveraging pretrained diffusion models and starting the process with low-resolution images.
FlowTurbo is a technique designed to speed up sampling in flow-based models without sacrificing quality.
Sponsored
Paragon's Growth Marketing Manager mentions “TLDR allows us to connect with engineers who aren't on social media…Our CTO reads TLDR but doesn't scroll LinkedIn. ”
This repository introduces the Prototype-based Masked Audio Model, which improves sound event detection by making better use of unlabeled data. This technique generates pseudo labels using a Gaussian mixture model to guide the training of a Transformer-based audio model.
Vector Post-Training Quantization is a method designed to enable extremely low-bit quantization of large language models, optimizing memory and storage use during deployment.
LightAvatar is a head avatar model that improves rendering speed and efficiency using neural light fields (NeLFs).
Aider has been able to dramatically improve the performance of general-purpose code editing by using o1 as an architect and DeepSeek as the writer.
This model was trained on army handbooks and has integrated deep knowledge that is rare for fine-tuned models.
Ichiban Prolog was embedded into Hellabot, a Go-based IRC bot, to avoid recompilation for new triggers. This integration allows for dynamic Prolog code execution, enabling users to modify bot logic on-the-fly. Future improvements could reduce interpreter setup impact and move more logic into Prolog.
Arcade AI, a generative product company that launched this week, has announced funding from notable investors as it pushes to build its “prompt to product” system, which allows for the creation of products that are immediately purchasable.
Elecrow is suspected of using AI to clone a voice for promotional videos without consent.
Anthropic, a generative AI startup backed by Amazon and other tech firms, is in talks to secure funding that could value the company at $40 billion.
Sponsored
AI isn't the future of customer support — it's the present: