Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Predicting more than one token at a time is an interesting paradigm of active research. If successful, it would dramatically improve generation time for many large language models. The approach in this post, which mirrors consistency models from image synthetics, attempts to use a parallel decoding strategy on fine-tuned LLMs to speed up generation. Early results match speculative decoding performance of 3x.
The YOCO architecture is a decoder-decoder model that reduces GPU memory demands while retaining global attention capabilities. It consists of a self-decoder and cross-decoder, allowing for efficient caching and reuse of key-value pairs. YOCO achieves favorable performance compared to traditional Transformers, with significant improvements in inference memory, latency, and throughput, making it suitable for large language models and long context lengths.
Sponsored
Click _capture_ when you start and _stop_ when you're done - then sit back and let Guidde's generative AI [10] do the rest. Guidde will add a step by step description of your workflow [10], include a voiceover in the language of your choice (100+ voices available), and make it
Buzz is a novel dataset that includes preference data in the pretraining mix. Its researchers have also released several models that were trained using this data. They found that the models perform well on a number of human preference tasks.
This new post-processing algorithm addresses model bias by applying a "fairness cost" to recalibrate output scores, ensuring compliance with various group fairness criteria such as statistical parity, equal opportunity, and equalized odds.
DiffMatch is a novel semi-supervised change detection method that leverages visual language models to synthesize pseudo labels for unlabeled data, providing additional supervision signals.
Discussion of the various ways to extend context for language models. It doesn't provide much in the way of evaluations, but it's a fascinating discussion of where the field is exploring.
A comprehensive survey that explores Mamba's applications across various visual tasks and its evolving impact. Stay updated on new findings and advancements on the Mamba project.
Sponsored
The next-generation Neat Bar brings beautifully simple yet powerful audio and video quality to Zoom and Microsoft Teams meetings. Learn
Paris-based Mistral AI is raising funds at a valuation of $6 billion, tripling its previous valuation from December.
OpenAI's Preferred Publishers Program, detailed in a pitch deck, offers financial incentives and enhanced visibility within ChatGPT for select news publishers.
Alibaba Cloud has launched the latest version of its large language model, Tongyi Qianwen Qwen2.5, marking significant improvements in reasoning, code comprehension, and textual understanding over Qwen2.0.
(normally $1,690).
Sponsored
Exclusive for TLDR readers – sign up in the next 12 hours to get
Google DeepMind and Isomorphic Labs have developed the 3rd generation of AlphaFold, a powerful protein folding prediction model. It is launching AlphaFold Server, which is a free way to interact with the model. AlphaFold 3 is 50% more accurate than previous generations. It correctly predicted the folded structure of the spike protein on Coronavirus OC43.
Microsoft is developing a new AI model named MAI-1, which reportedly boasts about 500 billion parameters, aiming to surpass other major AI models by Google and OpenAI.
Apple has announced its next-generation Apple Silicon chip, the M4, the company's first chip designed specifically for AI. Featuring a 3-nanometer architecture and debuting in the 2024 iPad Pro, the new chip boasts a 10-core CPU that is 50% faster than the M2 and a new Neural Engine capable of processing 38 trillion operations per second. It enhances applications that leverage AI, although specific AI features for iPadOS haven't been detailed yet.
A novel labeling method that leveraged two passes of VLMs resulted in an extremely detailed pairing of images and text. The captions have far greater detail than any previous dataset and can aid in training strong multimodal models.
A document released by OpenAI outlining how it wants its models to behave in various scenarios.
A new diffusion model introduces a unidirectional block attention mechanism to overcome memory limitations in generating ultra-high-resolution images.
DocRes is a new model that simplifies document image restoration by handling five tasks: dewarping, deshadowing, appearance enhancement, deblurring, and binarization within a single system.
QoQ, a novel quantization algorithm, accelerates large language model inference by using 4-bit weights, 8-bit activations, and 4-bit KV cache.
ChemFlow is a new framework designed to enhance molecular science by efficiently navigating chemical space using deep generative models.
Accelerated timelines, competition, and a lack of concern from superiors about real-world effects are themes common across a broad spectrum of the biggest effect companies. Engineers are increasingly saying that a large part of their jobs is focused on satisfying investors and not falling behind in the competition rather than solving actual problems for users. Some have been switched over to AI teams to help support fast-paced rollouts without adequate training, even if they are new to the technology.
GitHub Copilot Workspace, now in a limited technical preview, offers a powerful programming tool that lets users code in plain English, from planning to implementation, within the browser. The AI integrates seamlessly with codebases, suggesting block-by-block code execution and handling complex tasks with less active user input compared to ChatGPT.
Teenage users on Character.AI are turning to AI chatbots for companionship and emotional support, which they find helpful but potentially addictive. There are concerns about the appropriateness of AI in mental health, especially when bots give problematic advice or increase reliance on virtual interactions.
The gpt-2-chatbot, which appeared recently in the LMSYS arena, has been confirmed as a new OpenAI model following a 429 rate limit error that revealed details from OpenAI's API.
OpenAI is developing a search engine for ChatGPT to allow the AI to search the web for answers.
The data curation platform adds to its $11m seed round from September, aiming to expand its team and continue building for enterprise.
Sponsored
AE Studio's team is stacked with elite software creators [4] who can turn any AI/ML concept into reality for your business. Automated reports, NLP, custom chatbots - if you can imagine it, AE Studio will build it [4].
Stack Overflow, a popular programming website, and OpenAI are partnering to provide a data API for OpenAI customers to retrieve real time and vetted data.
Elon Musk plans to enhance X's AI, Grok, to merge live news with social media commentary to provide updates and citations in real time. Grok will generate news summaries from user discussions on X, focusing on engagement and accuracy. The project faces challenges with proper citation and legal concerns.
Video from the founder of Unsloth on how its team uses PyTorch, writes their kernels, and designs their API surface. Unsloth's framework and library are extremely powerful and easy to use.
SlotGAT is a new approach that improves heterogeneous graph neural networks by addressing the semantic mixing issue in traditional message passing.
This new method detects deepfakes by focusing on masked image modeling, especially in the frequency domain. The approach differs from traditional methods and shows significant improvement in identifying synthetic images, even from new AI generative techniques.
The team at Hugging Face has released a new project that nicely packages common tools needed for robotics development.
Researchers have developed "Morph-Tokens" to improve AI's visual understanding and image generation capabilities. These tokens transform abstract concepts used for comprehension into detailed visuals for image creation, leveraging the advanced processing power of the MLLM framework.
Vibe-Eval is a newly launched benchmark designed to test multimodal chat models with 269 visual understanding prompts, including 100 particularly challenging ones.
DeepSeek has released a 200B+ parameter model with 21B active parameters. It performs extremely well on code and reasoning. It's not clear if it is overall better than Llama 3 70B, but it is a welcome addition to the open model ecosystem.
Automating prompt optimization for AI models suggests a future where manual prompt engineering may become obsolete, pointing towards more efficient, model-driven methods of generating effective prompts.
Vision-Language Models like GPT-4V are advancing rapidly in understanding and interacting with images and text. A recent study uncovers their significant limitations in visual deductive reasoning. Researchers tested these models using complex visual puzzles, like those found in IQ tests, and discovered that they struggle with multi-step reasoning and recognizing abstract patterns.
Microsoft's Responsible AI Transparency Report highlights its advancement in deploying AI responsibly in 2023, including the creation of 30 AI tools and implementing safety measures.
A new report by Stanford tracking worldwide AI trends.
Command R and R+ are state-of-the-art LLMs optimized for long context tasks and excel at key business-critical capabilities.
(normally $1,690). Join over 1 million happy travelers today.
Sponsored
Exclusive for TLDR readers – sign up in the next 12 hours to get
Lamini, an Enterprise AI platform, makes it possible for software teams within enterprises to develop new LLM capabilities that reduce hallucinations on proprietary data, run their LLMs securely from cloud VPCs to on-premise, and scale their infrastructure with model evaluations that prioritize ROI and business outcomes over hype. It raised a $25M Series A led by Amplify Partners.