Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
With the rise of fake content made by AI, there's a need to spot the specific parts of videos that are altered. The new tool "UMMAFormer" can identify these fake segments in videos more effectively than older methods.
While big models like MiniGPT-4 are good at understanding images, they struggle with detailed anomalies in industrial settings. This solution, AnomalyGPT, trains these models on simulated anomaly images and their descriptions, and can detect industrial faults without manual adjustments.
Researchers have figured out a way to speed up the process of turning text into images using a technique called Rectified Flow.
An inference library for running local LLMs on modern consumer GPUs.
This research shows that if a model is trained to understand one programming language, it gets better at understanding others too.
Moderna CEO Stéphane Bancel reveals the company's extensive use of AI, which predates the ChatGPT era, in drug development, emphasizing its role in creating the COVID-19 vaccine, inventing enzymes, and interacting with regulators.
OpenAI has responded to lawsuits from authors who allege ChatGPT was trained on pirated books. Arguing transformative use, the firm seeks dismissal of most claims, comparing the situation to a Google Books copyright case and refuting Digital Millennium Copyright Act violations.
Sponsored
A survey of 1000 enterprise AI leaders
Amazon is launching a new AI tool that generates product listings.
Sponsored
Smart-doorbell pioneer, RING, pitched on the hit show SHARK TANK at a valuation of just $7 MILLION — the sharks shut them down. 5 YEARS LATER, the company sold to Amazon for more than $1 BILLION. Turning the 10% stake initially offered to the sharks from $700K into $100M!
Google's philanthropic arm is investing $20 million in the Digital Futures Project to promote responsible AI development. The initiative supports global institutions, addressing AI's societal impacts. Google.org stresses collaboration for AI's potential benefit.
Adobe, IBM, Nvidia, and others have committed to the White House’s accord on AI safety and trustworthiness, echoing earlier pledges by Meta and Google. These voluntary agreements stress pre-release testing and risk-sharing. AI regulation remains a step behind innovation under the Biden administration.
Optical character recognition (OCR) is the process of extracting text from images. It can fail on documents with lots of jargon or special characters like mathematics. This work from Facebook research showcases strong performance across academic domains, enabling the digitization of many old texts.
Researchers were able to train a model to accurately recognize smells by training a graph neural network on a custom data set of thousands of hand-labeled molecules.
This study looked at whether advanced AI writing assistants like InstructGPT were making people's writing sound too similar and potentially making public conversations less diverse.
Researchers have introduced Scenimefy, a new tool to transform everyday photos into detailed anime-style images.
Arcus scales RAG to planet scale by using a hierarchical retriever. After clustering documents into groups based on their semantic content, you can progressively filter over these groups to narrow the search space. This results in much more relevant context being retrieved, fewer hallucinations, and better reliability over planet-scale corpuses of data.
A drop-in replacement for the OpenAI Whisper API that offers 20x faster transcription written in Rust.
Roblox introduced an AI assistant to help creators design virtual experiences at the 2023 RDC. Set for release late this year or early next, the tool will enable advanced gameplay and 3D model generation.
The desire to encourage AI development while limiting potential harm, combined with the breakneck pace of innovation, makes regulating AI an especially tricky situation.
The idea of personalized services beyond what we have around shopping is really interesting. This post breaks down the opportunities and some of the companies already playing in this space.
MLCommons has unveiled a new benchmark for how quickly top-of-the-line hardware can run AI models, with Nvidia coming in first and Intel coming in second.
Sponsored
We’ve all tried every project management tool under the sun… And we’ve all given up on most of them at one point or another.
In collaboration with Microsoft, thousands of open-domain books have been transcribed with state-of-the-art text-to-speech. They're available on most streaming platforms. Information about the process is included in the link.
Meta has been snapping up AI training chips and building out data centers in order to create a more powerful new chatbot it hopes will be as sophisticated as OpenAI’s GPT-4.
Language model inference is generally slow because these models are heavily memory-bound. So people have introduced speculative decoding using a smaller draft model to “run ahead” and offer suggestions to the large model. This works well, but it’s complicated to implement, and finding a good draft model is tough. Instead, you can create a Medusa model that acts as its own draft by splitting off activations from various internal stages. This speeds up Llama 2x without compromising performance.
The authors of this study have created a way to test how well vision-language models can also think logically about what they see.
This project introduces NuPrompt, a new set of instructions in everyday language that helps self-driving cars understand and track objects around them better. It works really well!
RestGPT aims to construct a large language model-based autonomous agent to control real-world applications, such as movie databases and music players.
PointLLM is a tool that helps language models understand 3D objects, not just text and 2D images.
The debate surrounding large language models concerns whether they possess human-like reasoning abilities or rely on memorization and pattern-matching, with studies showing mixed results. Evaluating LLMs for true abstract reasoning is challenging due to their tendency to excel in tasks related to their training data, necessitating further research for clarification.
This article shares seven practical patterns for integrating large language models into systems: measuring performance, adding recent knowledge, fine-tuning for tasks, caching to save time/cost, ensuring quality output, managing errors, and harnessing user feedback.
OpenAI faces a lawsuit from writers alleging illegal use of their copyrighted works to train ChatGPT. They argue that ChatGPT's capabilities to summarize and analyze their content indicate direct training on their works, constituting copyright infringement. This follows similar suits from other authors against OpenAI.
OpenAI has confirmed that AI writing detectors are ineffective.
Training costs dominated inference when the current AI race began. Inference costs can quickly balloon as more people deploy language models to production. TensorRT has long been the go-to for speed. A version specifically for language models on H100s is now available.
The NERSC is offering a limited-time discount on its supercomputer for GPU-based workloads.
Imbue, formerly known as Generally Intelligent, has secured $200M in Series B funding, achieving a valuation of over $1 billion. The funding aims to accelerate the development of AI systems capable of reasoning and coding. Imbue's mission is to produce truly personal computers that amplify user agency, freedom, and dignity, with a vision for AI agents that understand and autonomously act upon user goals.
If you use a diffusion model to generate 2D normal maps, which contain distance and shape information, and then use a 3D reconstruction algorithm, you can generate realistic 3D human avatars using an underlying body model. This is useful for synthetic avatars and character creation.
This study looks at how well ChatGPT can suggest movies or books you might like.
The authors of this study developed a new way to accurately figure out where a basketball is in 3D space using just one photo, which is a big deal for things like sports analytics and robotics.
Sponsored
Can AI get you to 80% automated test coverage within 4 months? Maybe one day. Until then, there’s QA Wolf
Create and share easy-to-make, built-to-last, innovative, and customizable experiences by leveraging XenonJs’ rapidly growing ecosystem of components in a no-code environment. Seamlessly add your own components or models to the ecosystem.
As the technology has matured, WebAssembly hype has died down. WebAssembly is now good enough to be used in many production scenarios, including browser-based model deployment.
The Segment Anything Model (SAM) works great for regular images, but not so well for medical ones. SAM-Med2D was created by training SAM on a huge set of medical images with various kinds of input information. It performs much better on medical images.
OpenAI and CEO Sam Altman aim to develop artificial general intelligence (AGI) safely, but the pursuit of funding and products has shifted the company’s culture. Releasing ChatGPT acclimated the public to AI. Critics say that OpenAI’s commercial focus distracts from mitigating AI risks, but OpenAI insists that its mission remains unchanged.
Investors are shifting focus from backing AI enablers like NVIDIA to supporting founders building transformative, AI-first products that redefine work and life. Three major AI-first opportunities include AI-first networks and marketplaces transitioning from human-driven to algorithmic content creation, redefining enterprise software categories where AI becomes the main feature, and AI "co-pilot" augmenting services, especially in areas like wealth management. The potential extends beyond software, tapping into vast service markets.
Meta has introduced an opt-out tool allowing Facebook users to delete personal data used in training generative AI models. The "Generative AI Data Subject Rights" form focuses on third-party information, which includes publicly available data or licensed sources. The move comes as data privacy concerns rise, with international agencies urging tech companies to adhere to global data protection and privacy laws.
Microsoft is teaming up with digital pathology provider Paige to build the world’s largest image-based artificial intelligence model for identifying cancer.
It is estimated that ChatGPT uses up to 500 milliliters of water (about equal to what’s in a 16-ounce bottle) every time it is asked a series of between 5 to 50 prompts or questions.
Sponsored