Last week, Google launched its new AI, or rather its new big language model, dubbed Gemini. The Gemini 1.0 model is available in three versions: Gemini Nano is supposed to be best suited for tasks on a specific device, Gemini Pro is supposed to be the best option for a wider range of tasks, and Gemini Ultra is Google's largest language model that will handle the most complex tasks you can give it. Something that Google was keen to highlight at the launch of Gemini Ultra was that the language model outperformed the latest version of OpenAI's GPT-4 in 30 of the 32 most commonly used tests to measure the capabilities of language models. The tests cover everything from reading comprehension and various math questions to writing code for Python and image analysis. In some of the tests, the difference between the two AI models was only a few tenths of a percentage point, while in others it was up to ten percentage points.

Technology companies are racing to develop artificial intelligence that can run "unplugged" from the internet, providing users with a more personalized and private experience. During this year's Intel Innovation summit, company CEO Pat Gelsinger unveiled new "AI PCs" that will increase the use of AI on the devices themselves and not depend on the cloud, according to a report from Spectrum. The company is not alone in its quest to optimize its devices to run artificial intelligence "at the edge," unplugged from the internet and run on local hardware. Apple and Qualcomm have also been involved in the race, the report noted, leading a drive toward AI meant to act more as a personalized assistant for the end user. Most AI tools today rely heavily on data centers that require a stable internet connection, at times overburdening servers attempting to keep up with the growing demand.

If 2023 was the year artificial intelligence became a household topic of conversation, it's in many ways because of Sam Altman, CEO of the artificial intelligence research organization OpenAI. Altman, who was named TIME's 2023 "CEO of the Year" spoke candidly about his November ousting--and reinstatement--at OpenAI, how AI threatens to contribute to disinformation, and the rapidly advancing technology's future potential in a wide-ranging conversation with TIME Editor-in-Chief Sam Jacobs as part of TIME's "A Year in TIME" event on Tuesday. Altman shared that his mid-November sudden removal from OpenAI proved a learning experience--both for him and the company at large. "We always said that some moment like this would come," said Altman. "I didn't think it was going to come so soon, but I think we are stronger for having gone through it."

Recent publications explore AI biases in detecting objects and people in the environment. However, there is no research tackling how AI examines nature. This case study presents a pioneering exploration into the AI attitudes (ecocentric, anthropocentric and antipathetic) toward nature. Experiments with a Large Language Model (LLM) and an image captioning algorithm demonstrate the presence of anthropocentric biases in AI. Moreover, to delve deeper into these biases and Human-Nature-AI interaction, we conducted a real-life experiment in which participants underwent an immersive de-anthropocentric experience in a forest and subsequently engaged with ChatGPT to co-create narratives. By creating fictional AI chatbot characters with ecocentric attributes, emotions and views, we successfully amplified ecocentric exchanges. We encountered some difficulties, mainly that participants deviated from narrative co-creation to short dialogues and questions and answers, possibly due to the novelty of interacting with LLMs. To solve this problem, we recommend providing preliminary guidelines on interacting with LLMs and allowing participants to get familiar with the technology. We plan to repeat this experiment in various countries and forests to expand our corpus of ecocentric materials.

Multi-modality promises to unlock further uses for large language models. Recently, the state-of-the-art language model GPT-4 was enhanced with vision capabilities. We carry out a prompting evaluation of GPT-4V and five other baselines on structured reasoning tasks, such as mathematical reasoning, visual data analysis, and code generation. We show that visual Chain-of-Thought, an extension of Chain-of-Thought to multi-modal LLMs, yields significant improvements over the vanilla model. We also present a categorized analysis of scenarios where these models perform well and where they struggle, highlighting challenges associated with coherent multimodal reasoning.

When solving challenging problems, language models (LMs) are able to identify relevant information from long and complicated contexts. To study how LMs solve retrieval tasks in diverse situations, we introduce ORION, a collection of structured retrieval tasks spanning six domains, from text understanding to coding. Each task in ORION can be represented abstractly by a request (e.g. a question) that retrieves an attribute (e.g. the character name) from a context (e.g. a story). We apply causal analysis on 18 open-source language models with sizes ranging from 125 million to 70 billion parameters. We find that LMs internally decompose retrieval tasks in a modular way: middle layers at the last token position process the request, while late layers retrieve the correct entity from the context. After causally enforcing this decomposition, models are still able to solve the original task, preserving 70% of the original correct token probability in 98 of the 106 studied model-task pairs. We connect our macroscopic decomposition with a microscopic description by performing a fine-grained case study of a question-answering task on Pythia-2.8b. Building on our high-level understanding, we demonstrate a proof of concept application for scalable internal oversight of LMs to mitigate prompt-injection while requiring human supervision on only a single input. Our solution improves accuracy drastically (from 15.5% to 97.5% on Pythia-12b). This work presents evidence of a universal emergent modular processing of tasks across varied domains and models and is a pioneering effort in applying interpretability for scalable internal oversight of LMs.

Recent advancements in large language models (LLMs) have notably propelled natural language processing (NLP) capabilities, demonstrating significant potential in safety engineering applications. Despite these advancements, LLMs face constraints in processing specialized tasks, attributed to factors such as corpus size, input processing limitations, and privacy concerns. Obtaining useful information from reliable sources in a limited time is crucial for LLM. Addressing this, our study introduces an LLM-based Q&A system for safety engineering, enhancing the comprehension and response accuracy of the model. We employed prompt engineering to incorporate external knowledge databases, thus enriching the LLM with up-to-date and reliable information. The system analyzes historical incident reports through statistical methods, utilizes vector embedding to construct a vector database, and offers an efficient similarity-based search functionality. Our findings indicate that the integration of external knowledge significantly augments the capabilities of LLM for in-depth problem analysis and autonomous task assignment. It effectively summarizes accident reports and provides pertinent recommendations. This integration approach not only expands LLM applications in safety engineering but also sets a precedent for future developments towards automation and intelligent systems.

This paper introduces a novel approach to enhance the capabilities of Large Language Models (LLMs) in processing and understanding extensive text sequences, a critical aspect in applications requiring deep comprehension and synthesis of large volumes of information. Recognizing the inherent challenges in extending the context window for LLMs, primarily built on Transformer architecture, we propose a new model architecture, referred to as Zebra. This architecture efficiently manages the quadratic time and memory complexity issues associated with full attention in the Transformer by employing grouped local-global attention layers. Our model, akin to a zebra's alternating stripes, balances local and global attention layers, significantly reducing computational requirements and memory consumption. Comprehensive experiments, including pretraining from scratch, continuation of long context adaptation training, and long instruction tuning, are conducted to evaluate the Zebra's performance. The results show that Zebra achieves comparable or superior performance on both short and long sequence benchmarks, while also enhancing training and inference efficiency.

The advent of Foundation Models is transforming machine learning across many modalities (e.g., language, images, videos) with prompt engineering replacing training in many settings. Recent work on tabular data (e.g., TabPFN) hints at a similar opportunity to build Foundation Models for classification for numerical data. In this paper, we go one step further and propose a hypernetwork architecture that we call MotherNet, trained on millions of classification tasks, that, once prompted with a never-seen-before training set generates the weights of a trained ``child'' neural-network. Like other Foundation Models, MotherNet replaces training on specific datasets with in-context learning through a single forward pass. In contrast to existing hypernetworks that were either task-specific or trained for relatively constraint multi-task settings, MotherNet is trained to generate networks to perform multiclass classification on arbitrary tabular datasets without any dataset specific gradient descent. The child network generated by MotherNet using in-context learning outperforms neural networks trained using gradient descent on small datasets, and is competitive with predictions by TabPFN and standard ML methods like Gradient Boosting. Unlike a direct application of transformer models like TabPFN, MotherNet generated networks are highly efficient at inference time. This methodology opens up a new approach to building predictive models on tabular data that is both efficient and robust, without any dataset-specific training.

With the rapid development of large language models (LLMs), it is highly demanded that LLMs can be adopted to make decisions to enable the artificial general intelligence. Most approaches leverage manually crafted examples to prompt the LLMs to imitate the decision process of human. However, designing optimal prompts is difficult and the patterned prompts can hardly be generalized to more complex environments. In this paper, we propose a novel model named Large Decision Model with Memory (LDM$^2$), which leverages a dynamic memory mechanism to construct dynamic prompts, guiding the LLMs in making proper decisions according to the faced state. LDM$^2$ consists of two stages: memory formation and memory refinement. In the former stage, human behaviors are decomposed into state-action tuples utilizing the powerful summarizing ability of LLMs. Then, these tuples are stored in the memory, whose indices are generated by the LLMs, to facilitate the retrieval of the most relevant subset of memorized tuples based on the current state. In the latter stage, our LDM$^2$ employs tree exploration to discover more suitable decision processes and enrich the memory by adding valuable state-action tuples. The dynamic circle of exploration and memory enhancement provides LDM$^2$ a better understanding of the global environment. Extensive experiments conducted in two interactive environments have shown that our LDM$^2$ outperforms the baselines in terms of both score and success rate, which demonstrates its effectiveness.