Microsoft is promising an AI-driven future for Windows 11--but initially, only select PC owners could try Copilot features as they rolled out in preview. If you didn't own a laptop with a Snapdragon X Elite or X Plus processor, you had to wait. That changes today, with Microsoft's announcement of two Copilot features becoming available to compatible Intel and AMD PCs: Microsoft Recall and Click to Do. Of the two, Microsoft Recall has greater notoriety--it captures your activity as you use your computer, taking screenshots every few seconds. Windows then leans on AI to sift through the accumulated data when you're seeking specific information like a recipe, conversation, or even document you previously worked on.

Meta's aggressive push to make its AI assistant a ubiquitous presence in its apps continues to pay off. Meta AI is on the verge of passing its next major milestone, with "nearly" 600 million monthly users, Mark Zuckerberg shared in an update. Meta AI, which debuted last fall, passed 500 million users back in October. The update came alongside the release of Meta's latest Llama 3.3 70B model. According to Meta, the latest text model has "similar performance to the Llama 3.1 405B model," but comes "at a fraction of the cost."

LLMs are created by performing machine learning (ML) on large amounts of data. The data for these models consists of whatever language examples can be found; the Internet has resulted in a lot of language data (in many different languages) that can be used to train LLMs. Radically simplifying, the training results in a set of probabilities that can be used to tell, for any word or string of words, which word or words are the most likely to be associated with those words. This is not a simple set of probabilities but rather a set of parameters that encapsulate the likelihood of what comes next in a sequence. Models are often described by the size of the training set and the number of parameters used to build the probability model.

OpenAI released its most advanced AI model yet, called o1, for paying users on Thursday. The launch kicked off the company's "12 Days of OpenAI" event--a dozen consecutive releases to celebrate the holiday season. OpenAI has touted o1's "complex reasoning" capabilities, and announced on Thursday that unlimited access to the model would cost 200 per month. In the video the company released to show the model's strengths, a user uploads a picture of a wooden birdhouse and asks the model for advice on how to build a similar one. The model "thinks" for a short period and then spits out what on the surface appears to be a comprehensive set of instructions. Close examination reveals the instructions to be almost useless.

With multiple AI platforms and bots competing against each other--there's Copilot, Gemini, ChatGPT, Claude, and Perplexity, to name just a few--we're seeing new updates and upgrades appear on a frequent basis. One of the newest additions OpenAI has pushed out to ChatGPT is called Canvas, and it's a little bit like an AI-powered Google Docs. OpenAI describes it as "a new way of working with ChatGPT to write and code," and it means you're essentially collaborating with the AI on a text document or on program code. You can already do this in the main chat interface of course, but with Canvas it's a bit more like having an AI coworker with you. Right now, you have to be a ChatGPT Enterprise, ChatGPT Pro, or ChatGPT Plus user (from 20 a month) to access the Canvas model.

Open-source multimodal large language models (MLLMs) have shown significant potential in a broad range of multimodal tasks. However, their reasoning capabilities remain constrained by existing instruction-tuning datasets, which were predominately repurposed from academic datasets such as VQA, AI2D, and ChartQA. These datasets target simplistic tasks, and only provide phrase-level answers without any intermediate rationales. To address these challenges, we introduce a scalable and cost-effective method to construct a large-scale multimodal instruction-tuning dataset with rich intermediate rationales designed to elicit CoT reasoning. Using only open models, we create a dataset containing 12M instruction-response pairs to cover diverse, reasoning-intensive tasks with detailed and faithful rationales. Experiments demonstrate that training MLLMs on this dataset significantly improves reasoning capabilities, achieving state-of-the-art performance on benchmarks such as MathVerse (+8.1%), MMMU-Pro (+7%), and MuirBench (+13.3%). Additionally, the model demonstrates notable improvements of up to 4% on non-reasoning-based benchmarks. Ablation studies further highlight the importance of key components, such as rewriting and self-filtering, in the dataset construction process.

This manuscript gives a big-picture, up-to-date overview of the field of (deep) reinforcement learning and sequential decision making, covering value-based RL, policy-gradient methods, model-based methods, and various other topics (including a very brief discussion of RL+LLMs).

Few-shot learning and parameter-efficient fine-tuning (PEFT) are crucial to overcome the challenges of data scarcity and ever growing language model sizes. This applies in particular to specialized scientific domains, where researchers might lack expertise and resources to fine-tune high-performing language models to nuanced tasks. We propose PETapter, a novel method that effectively combines PEFT methods with PET-style classification heads to boost few-shot learning capabilities without the significant computational overhead typically associated with full model training. We validate our approach on three established NLP benchmark datasets and one real-world dataset from communication research. We show that PETapter not only achieves comparable performance to full few-shot fine-tuning using pattern-exploiting training (PET), but also provides greater reliability and higher parameter efficiency while enabling higher modularity and easy sharing of the trained modules, which enables more researchers to utilize high-performing NLP-methods in their research.

The rise of large language models (LLMs) has highlighted the importance of prompt engineering as a crucial technique for optimizing model outputs. While experimentation with various prompting methods, such as Few-shot, Chain-of-Thought, and role-based techniques, has yielded promising results, these advancements remain fragmented across academic papers, blog posts and anecdotal experimentation. The lack of a single, unified resource to consolidate the field's knowledge impedes the progress of both research and practical application. This paper argues for the creation of an overarching framework that synthesizes existing methodologies into a cohesive overview for practitioners. Using a design-based research approach, we present the Prompt Canvas, a structured framework resulting from an extensive literature review on prompt engineering that captures current knowledge and expertise. By combining the conceptual foundations and practical strategies identified in prompt engineering, the Prompt Canvas provides a practical approach for leveraging the potential of Large Language Models. It is primarily designed as a learning resource for pupils, students and employees, offering a structured introduction to prompt engineering. This work aims to contribute to the growing discourse on prompt engineering by establishing a unified methodology for researchers and providing guidance for practitioners.

Large Language Models (LLMs) based on transformers achieve cutting-edge results on a variety of applications. However, their enormous size and processing requirements make deployment on devices with constrained resources extremely difficult. Among various efficiency considerations, model binarization and Early Exit (EE) are common effective solutions. However, binarization may lead to performance loss due to reduced precision affecting gradient estimation and parameter updates. Besides, the present early-exit mechanisms are still in the nascent stages of research. To ameliorate these issues, we propose Binarized Early Exit Transformer (BEExformer), the first-ever selective learning transformer architecture to combine early exit with binarization for textual inference. It improves the binarization process through a differentiable second-order approximation to the impulse function. This enables gradient computation concerning both the sign as well as the magnitude of the weights. In contrast to absolute threshold-based EE, the proposed EE mechanism hinges on fractional reduction in entropy among intermediate transformer blocks with soft-routing loss estimation. While binarization results in 18.44 times reduction in model size, early exit reduces the FLOPs during inference by 54.85% and even improves accuracy by 5.98% through resolving the "overthinking" problem inherent in deep networks. Moreover, the proposed BEExformer simplifies training by not requiring knowledge distillation from a full-precision LLM. Extensive evaluation on the GLUE dataset and comparison with the SOTA works showcase its pareto-optimal performance-efficiency trade-off.