Why it matters: Minecraft may not sound like an important tool that supports advanced AI research. After all, what could possibly be so important about teaching a machine to play a sandbox game released more than a decade ago? Based on OpenAI's recent efforts, a well-trained Minecraft bot is more relevant to AI advancement than most people might realize. OpenAI has always focused on artificial intelligence (AI) and machine learning advances that benefit humanity. Recently, the company successfully trained a bot to play Minecraft using more than 70,000 hours of gameplay videos. The achievement is far more than just a bot playing a game.

The semiconductor chip manufacturing and design work is in full swing with large players working on building process nodes as low as 3nm. But, there is a limit to how many transistors can be infused on a single chip. Even with the introduction of multi-core processors, in which multiple single-core processors could be attached to increase power, concerns over whether this is enough to sustain in the long run looms. At this point, it seems as though we have reached saturation levels, and chants of Moore's law--which states that every 12-18 months, the processing power doubles--slowing down or nearing an end have been restored. However, a new ray of light--the cloud--has been powering Moore's law and will continue to do so at least for the next decade or two, propelling the most cutting-edge innovation.

I recently caught up with Ian Hogarth and Nathan Benaich, who each year produce The State of AI Report, a must-read snapshot of how commercial applications of A.I. are evolving. Benaich is the founder of Air Street Capital, a solo venture capital fund that is one of the savviest early-stage investors in A.I.-based startups I know. Hogarth is the former co-founder of concert discovery app Songkick and has since go on to become a prominent angel investor as well one of the founders behind the founder-lead European venture capital platform Plural. There's always a lot to digest in their report. But one of the key takeaways from this year's State of AI is that concerns established tech giants and their affiliated A.I. research labs would monopolize the development of A.I. have been proven, if not exactly wrong, then at least premature. While it is true that Alphabet (which has both Google Brain and Deepmind in its stable), Meta, Microsoft, and OpenAI (which is closely partnered now with Microsoft) are building large "foundational models" for natural language processing and image and video generation, they are hardly the only players in the game.

Please bear me and point out if I said something wrong and glad to hear your voice). In the past five years, a series of Transformer-based models has been created and relevant works have been done. Pre-trained large language models with few-shot prompting becomes the new paradigm for tackling a broad range of NLP-related tasks. This is amazing and really useful for NLP applications. But no significant improvement of model architecture (algorithm side) has been done; everything is still transformer-based.

1989 … 2021 [gMLP] … 2022 [ConvNeXt] [PVTv2] [ViT-G] [AS-MLP] [ResTv2] [CSWin Transformer] [Pale Transformer] 1991 … 2020 [ALBERT] [GPT-3] [T5] [Pre-LN Transformer] [MobileBERT] [TinyBERT] [BART]…

Machine Learning algorithms have had a profound impact on the field of computer science over the past few decades. These algorithms performance is greatly influenced by the representations that are derived from the data in the learning process. The representations learned in a successful learning process should be concise, discrete, meaningful, and able to be applied across a variety of tasks. A recent effort has been directed toward developing Deep Learning models, which have proven to be particularly effective at capturing high-dimensional, non-linear, and multi-modal characteristics. In this work, we discuss the principles and developments that have been made in the process of learning representations, and converting them into desirable applications. In addition, for each framework or model, the key issues and open challenges, as well as the advantages, are examined.

Handling and digesting a huge amount of information in an efficient manner has been a long-term demand in modern society. Some solutions to map key points (short textual summaries capturing essential information and filtering redundancies) to a large number of arguments/opinions have been provided recently (Bar-Haim et al., 2020). To complement the full picture of the argument-to-keypoint mapping task, we mainly propose two approaches in this paper. The first approach is to incorporate prompt engineering for fine-tuning the pre-trained language models (PLMs). The second approach utilizes prompt-based learning in PLMs to generate intermediary texts, which are then combined with the original argument-keypoint pairs and fed as inputs to a classifier, thereby mapping them. Furthermore, we extend the experiments to cross/in-domain to conduct an in-depth analysis. In our evaluation, we find that i) using prompt engineering in a more direct way (Approach 1) can yield promising results and improve the performance; ii) Approach 2 performs considerably worse than Approach 1 due to the negation issue of the PLM.

Recent work in large language modeling (LLMs) has used fine-tuning to align outputs with the preferences of a prototypical user. This work assumes that human preferences are static and homogeneous across individuals, so that aligning to a a single "generic" user will confer more general alignment. Here, we embrace the heterogeneity of human preferences to consider a different challenge: how might a machine help people with diverse views find agreement? We fine-tune a 70 billion parameter LLM to generate statements that maximize the expected approval for a group of people with potentially diverse opinions. Human participants provide written opinions on thousands of questions touching on moral and political issues (e.g., "should we raise taxes on the rich?"), and rate the LLM's generated candidate consensus statements for agreement and quality. A reward model is then trained to predict individual preferences, enabling it to quantify and rank consensus statements in terms of their appeal to the overall group, defined according to different aggregation (social welfare) functions. The model produces consensus statements that are preferred by human users over those from prompted LLMs (>70%) and significantly outperforms a tight fine-tuned baseline that lacks the final ranking step. Further, our best model's consensus statements are preferred over the best human-generated opinions (>65%). We find that when we silently constructed consensus statements from only a subset of group members, those who were excluded were more likely to dissent, revealing the sensitivity of the consensus to individual contributions. These results highlight the potential to use LLMs to help groups of humans align their values with one another.

It has become possible to distill complex linguistic information of text into multidimensional dense numeric vectors with the use of the distributional hypothesis. As a consequence, text representation methods have been evolving at such a quick pace that the research community is struggling to retain knowledge of the methods and their interrelations. We contribute threefold to this lack of compilation, composition, and systematization by providing a survey of current approaches, by arranging them in a genealogy, and by conceptualizing a taxonomy of text representation methods to examine and explain the state-of-the-art. Our research is a valuable guide and reference for artificial intelligence researchers and practitioners interested in natural language processing applications such as recommender systems, chatbots, and sentiment analysis.

It has been a long-standing dream to design artificial agents that explore their environment efficiently via intrinsic motivation, similar to how children perform curious free play. Despite recent advances in intrinsically motivated reinforcement learning (RL), sample-efficient exploration in object manipulation scenarios remains a significant challenge as most of the relevant information lies in the sparse agent-object and object-object interactions. In this paper, we propose to use structured world models to incorporate relational inductive biases in the control loop to achieve sample-efficient and interaction-rich exploration in compositional multi-object environments. By planning for future novelty inside structured world models, our method generates free-play behavior that starts to interact with objects early on and develops more complex behavior over time. Instead of using models only to compute intrinsic rewards, as commonly done, our method showcases that the self-reinforcing cycle between good models and good exploration also opens up another avenue: zero-shot generalization to downstream tasks via model-based planning. After the entirely intrinsic task-agnostic exploration phase, our method solves challenging downstream tasks such as stacking, flipping, pick & place, and throwing that generalizes to unseen numbers and arrangements of objects without any additional training.