Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Models are usually trained separately on specific tasks (e.g., language vs image generation). The newly proposed joint autoregressive mixture (JAM) algorithm combines disparate models with a clever interleaved cross attention and gentle fine-tuning achieves dominant performance in many multimodal tasks.
Sponsored
The MLOps Community’s LLMs in Production conference
Joint embedding models combine two data types into a single space. CLIP is a popular one for images and text. Recently, Google researchers proposed a Sigmoid CLIP model which performed very well and was Vision Transformer based. They've now released more information about the models and updated their code and paper.
InternLM-XComposer is a cutting-edge tool that can create articles with both text and images that fit perfectly together.
For technical leaders, embracing generative AI in the workplace brings a wealth of opportunity to drive productivity and transformation. It also means leaders need to consider how to design meaningful career pathways for team members. Workera’s guide shows leaders how they can do it.
The White House may mandate cloud companies to report large-scale computing purchases, targeting potential foreign AI threats. Critics cite surveillance concerns and changing compute costs. Entities like Microsoft and OpenAI support similar oversight measures.
Anthropic has partnered with Boston Consulting Group (BCG) to offer their AI assistant, Claude, to BCG customers globally. This collaboration focuses on responsible AI deployment, with use cases including market research, fraud detection, and business analysis. Both entities aim to establish a new standard for ethically deployed enterprise AI.
Google introduced a tool, Google-Extended, which enables website publishers to prevent their data from being used in Google's AI training while still appearing in Google Search results.
Sponsored
The semantic layer is key to delivering the ‘last mile’ of analytics — making data available in tools such as Power BI, Tableau, and Excel. But implementing it in the data-intensive insurance industry can get tricky. Watch this tech talk
After a huge seed round and building a team, ML stack, and data pipeline, the Mistral team has released an extremely performant 7B model. It was initially released via a magnet link (torrent) along with a minimal GitHub Repo and permissive license.
Meta is launching its own AI chatbot assistant along with various AI characters across its platforms. The chatbot is designed as a general-purpose assistant, uses Microsoft's Bing for real-time web results, and can generate images similar to OpenAI's DALL-E.
OpenAI has reintroduced a feature in ChatGPT that enables it to access current web information using the "Browse with Bing" functionality. It is initially available for Plus and Enterprise subscribers.
Llava is a method that inserts vision capabilities into language models. It has now been instruction-tuned and run through an RLHF process. Llava improves instruction following and reasoning abilities.
Large models like ChatGPT have been used to improve search results, but many of these tests are hard to repeat because they're hidden or secret. Researchers have introduced RankVicuna, an open-source tool that can reorder search results using language models.
Large language models are powerful, but often too big for some devices due to their computational demands. This new method, QA-LoRA, makes models more efficient by cleverly changing how they're built and fine-tuned.
Sponsored
Want to know what users are doing inside your app? Why not just ask? TLDR members get EXCLUSIVE 3 months free*
This project introduces a new tool, UnitedHuman, that uses images from many different sources and combines them, even if they're in different sizes or focused on different parts.
Tuning language models with RL is becoming a standard step in the model deployment pipeline. This will now be easier with diffusion models and also with the release of DRLX, which allows users to specify a reward model or use a predefined one.
This project made a new tool that helps improve the quality of pictures in a way that's easier to understand than other methods out there. The tool, called HMTF-Net, looks at different parts of a photo and figures out the best way to enhance each part.
A Reddit user's AI-generated spiral art using Stable Diffusion and ControlNet AI tech went viral, signaling changing perceptions of AI-generated art.
The dust is settling around the excitement around generative AI, which means people are settling in to figure out how it can be useful in the enterprise. Short story: it’s hugely beneficial. Long story: it’s going to take time.
Sponsored
Luda's AI Agents are interactive, 3d-rendered, and simulated with real-world physics. They're fast to train, cheap to run, and accessible from any browser. Watch an AI agent in training.
Apple’s former chief design officer Jony Ive and OpenAI CEO Sam Altman are allegedly exploring opportunities to collaborate on building an unspecified artificial intelligence hardware device.
OpenAI is in discussions to possibly sell shares in a move that would boost the company’s valuation from $29 billion to somewhere between $80 billion and $90 billion.
Powered by Nvidia, Getty launches a commercial image generation service with strong commercial guarantees.
YouTube unveiled new AI tools for creators, including AI-generated photo and video backgrounds, topic suggestions, and music recommendations. A highlight is Dream Screen for AI-generated backgrounds in YouTube Shorts. This shift may reshape how creators design content, with AI-driven insights influencing content focus and an increased prevalence of AI-generated material on the platform.
Neural radiance fields (NeRF) are great for creating virtual views but struggle with shiny or see-through objects. This new method, called refractive-reflective field, fixes this by better handling light changes and producing clearer images of such tricky objects.
Most current tools for semantic image editing, where images are changed based on their meaning, struggle with blending new objects smoothly into scenes. This new method improves this by smartly understanding both fully seen and partially hidden objects.
Over the last ten years, there have been big improvements in fixing damaged videos using special technologies. However, many of these tools aren't perfect at fixing common video issues seen in things like live streaming. This study created BSCV, a new collection of tools and videos to help fix these problems even better.
Sponsored
The MLOps Community
Category theory is the abstract study of relationships. It is typically used as an additional layer of abstraction in mathematics. That gives it powerful tools to study machine learning as well.
OpenLM is a library and framework for training medium-sized language models up to 7B parameters. It is designed with GPU speed and hackability in mind. Models trained on 1.25T tokens are available as validation of convergence.
A great resource for annotated papers you may find useful.
The surge in Generative AI, exemplified by Nvidia's success, underscores a vast demand for GPUs and AI training, leading to significant investments in the sector. As major tech companies drive this growth, the challenge for startups is to shift focus from infrastructure to creating tangible end-customer value.
Although OpenAI has made strides in preventing GPT-4 with vision from breaking CAPTCHAS and estimating people’s age based on age and race, it still sometimes struggles to make inferences and is still prone to hallucinating.
AI giants are now hiring poets and writers in a bid to boost the literary quality of generative writing tools.
The CIA and other US intelligence agencies are developing an AI chatbot to assist agents in sifting through vast amounts of information. The chatbot, which will be trained on publicly available data, aims to provide sources with its answers, but concerns arise over what counts as public data and potential privacy issues.
Guest author Dan Kraemer for Crunchbase News goes into his methodology around leveraging AI to brainstorm the next big idea.
Sponsored
Ring 一 Acquired by Amazon for $1.2B Nest 一 Acquired by Google for $3.2B
Amazon is investing up to $4B in Anthropic with a set of trades in IP and compute options on the table. Anthropic’s safety governance remains unchanged.
Researchers have created a new model called PVLFF that can look at 3D scenes and not only identify what's in them, but also distinguish between different instances of the same object—all by using text descriptions.
LayoutLM is a model that is good at reading business documents. It was made even smarter with two new training exercises—one to help it understand complicated layouts and another to get numeric values right.
NeuralLabeling is a tool for detailed scene annotations using 3D techniques.
This study introduces a method that simulates active stereo in passive cameras without a physical pattern projector. By virtually overlaying patterns using depth data from external sensors, the researchers overcome the limitations of traditional setups. This approach enhances both stereo algorithms and deep learning networks on various datasets.
DeepEval provides a Pythonic way to run offline evaluations on your LLM pipelines so you can launch comfortably into production.
Generate knowledge graphs from your data with GPT-4.