Gary Marcus: Two Models of AI Oversight -- and How Things Could Go Deeply Wrong https://bit.ly/3Qnxd9A June 12, 2023 Originally published on The Road to AI We Can Trust (http://bit.ly/3juuD3j) The Senate hearing that I participated in a few weeks ago (https://bit.ly/44QxHt1) I was thrilled by what I saw of the Senate that day: genuine interest and genuine humility. Senators acknowledged that they were too slow to figure out what do about social media, that the moves were made then, and that there was now a sense of urgency.

Battles between human and artificial intelligence are no longer science fiction. The strikes in Hollywood led by the united guilds of actors and screenwriters have a common, intangible enemy: the algorithms and computer-generated imagery that are increasingly programmed by studios to render them redundant. In New York last week, a new front in that stand-off was opened by a group of American novelists – including John Grisham, Jodi Picoult and Jonathan Franzen – who are suing OpenAI, the creators of the ChatGPT program. The legal case may help to define and protect those increasingly porous boundaries between human creativity and the robots that mimic it. In the meantime, Amazon, these days flooded by self-published books written by AI, has taken its first half-hearted steps to curtail that practice.

In all the frenzied discourse about large language models (LLMs) such as GPT-4 there is one point on which everyone seems to agree: these models are essentially stochastic parrots – namely, machines that are good at generating convincing sentences, but do not actually understand the meaning of the language they are processing. They have somehow "read" (that is, ingested) everything ever published in machine-readable form and create sentences word by word, at each point making a statistical guess of "what one might expect someone to write after seeing what people have written on billions of webpages, etc". Ever since ChatGPT arrived last November, people have been astonished by the capabilities of these parrots – how humanlike they seem to be and so on. But consolation was drawn initially from the thought that since the models were drawing only on what already resided in their capacious memories, then they couldn't be genuinely original: they would just regurgitate the conventional wisdom embedded in their training data. That comforting thought didn't last long, though, as experimenters kept finding startling and unpredictable behaviours of LLMs – facets now labelled "emergent abilities".

Figure 1 | Schematic overview of the data regimes for which we show experimental results. These regimes are characterized by how much data from the target environment is available to the agent, and how much (potentially generalizable) experience has been collected in other environments. The experiments both demonstrate that TDMs are capable single-environment models (marked purple) and generalize across environments (marked yellow). If sufficient data from the target environment is available, we can learn a single-environment specialist model (section 5.1). If there are only small amounts of data from the target environment, but more data from other environments, a generalist model can be pre-trained and then fine-tuned on the target environment (section 5.2.1). Finally, if we are able to train a generalist model on large amounts of data from different environments, we can zero-shot apply this model to our target environment without fine-tuning (section 5.2.2). We also show an example for unsuccessful generalization (no color) in section E.

For both public and private firms, comparable companies' analysis is widely used as a method for company valuation. In particular, the method is of great value for valuation of private equity companies. The several approaches to the comparable companies' method usually rely on a qualitative approach to identifying similar peer companies, which tend to use established industry classification schemes and/or analyst intuition and knowledge. However, more quantitative methods have started being used in the literature and in the private equity industry, in particular, machine learning clustering, and natural language processing (NLP). For NLP methods, the process consists of extracting product entities from e.g., the company's website or company descriptions from some financial database system and then to perform similarity analysis. Here, using companies' descriptions/summaries from publicly available companies' Wikipedia websites, we show that using large language models (LLMs), such as GPT from OpenAI, has a much higher precision and success rate than using the standard named entity recognition (NER) methods which use manual annotation. We demonstrate quantitatively a higher precision rate, and show that, qualitatively, it can be used to create appropriate comparable companies peer groups which could then be used for equity valuation.

Due to the powerful capabilities demonstrated by large language model (LLM), there has been a recent surge in efforts to integrate them with AI agents to enhance their performance. In this paper, we have explored the core differences and characteristics between LLM-based AI agents and traditional AI agents. Specifically, we first compare the fundamental characteristics of these two types of agents, clarifying the significant advantages of LLM-based agents in handling natural language, knowledge storage, and reasoning capabilities. Subsequently, we conducted an in-depth analysis of the key components of AI agents, including planning, memory, and tool use. Particularly, for the crucial component of memory, this paper introduced an innovative classification scheme, not only departing from traditional classification methods but also providing a fresh perspective on the design of an AI agent's memory system. We firmly believe that in-depth research and understanding of these core components will lay a solid foundation for the future advancement of AI agent technology. At the end of the paper, we provide directional suggestions for further research in this field, with the hope of offering valuable insights to scholars and researchers in the field.

Vision-language models (VLMs) have shown to be effective at image retrieval based on simple text queries, but text-image retrieval based on conversational input remains a challenge. Consequently, if we want to use VLMs for reference resolution in visually-grounded dialogue, the discourse processing capabilities of these models need to be augmented. To address this issue, we propose fine-tuning a causal large language model (LLM) to generate definite descriptions that summarize coreferential information found in the linguistic context of references. We then use a pretrained VLM to identify referents based on the generated descriptions, zero-shot. We evaluate our approach on a manually annotated dataset of visually-grounded dialogues and achieve results that, on average, exceed the performance of the baselines we compare against. Furthermore, we find that using referent descriptions based on larger context windows has the potential to yield higher returns.

Large language models (LLMs) have achieved dramatic proficiency over NLP tasks with normal length. Recently, multiple studies have committed to extending the context length and enhancing the long text modeling capabilities of LLMs. To comprehensively evaluate the long context ability of LLMs, we propose BAMBOO, a multi-task long context benchmark. BAMBOO has been designed with four principles: comprehensive capacity evaluation, avoidance of data contamination, accurate automatic evaluation, and different length levels. It consists of 10 datasets from 5 different long text understanding tasks, i.e. question answering, hallucination detection, text sorting, language modeling, and code completion, to cover core capacities and various domains of LLMs. We conduct experiments with five long context models on BAMBOO and further discuss four key research questions of long text. We also qualitatively analyze current long context models and point out future directions for enhancing long text modeling capacities. We release our data, prompts, and code at https://github.com/RUCAIBox/BAMBOO.

Recent advancements in large language models have showcased their remarkable generalizability across various domains. However, their reasoning abilities still have significant room for improvement, especially when confronted with scenarios requiring multi-step reasoning. Although large language models possess extensive knowledge, their behavior, particularly in terms of reasoning, often fails to effectively utilize this knowledge to establish a coherent thinking paradigm. Generative language models sometimes show hallucinations as their reasoning procedures are unconstrained by logical principles. Aiming to improve the zero-shot chain-of-thought reasoning ability of large language models, we propose Logical Chain-of-Thought (LogiCoT), a neurosymbolic framework which leverages principles from symbolic logic to verify and revise the reasoning processes accordingly. Experimental evaluations conducted on language tasks in diverse domains, including arithmetic, commonsense, symbolic, causal inference, and social problems, demonstrate the efficacy of the enhanced reasoning paradigm by logic.

The widespread use of Large Language Models (LLMs), celebrated for their ability to generate human-like text, has raised concerns about misinformation and ethical implications. Addressing these concerns necessitates the development of robust methods to detect and attribute text generated by LLMs. This paper investigates "Cross-Model Detection," evaluating whether a classifier trained to distinguish between source LLM-generated and human-written text can also detect text from a target LLM without further training. The study comprehensively explores various LLM sizes and families, and assesses the impact of conversational fine-tuning techniques on classifier generalization. The research also delves into Model Attribution, encompassing source model identification, model family classification, and model size classification. Our results reveal several key findings: a clear inverse relationship between classifier effectiveness and model size, with larger LLMs being more challenging to detect, especially when the classifier is trained on data from smaller models. Training on data from similarly sized LLMs can improve detection performance from larger models but may lead to decreased performance when dealing with smaller models. Additionally, model attribution experiments show promising results in identifying source models and model families, highlighting detectable signatures in LLM-generated text. Overall, our study contributes valuable insights into the interplay of model size, family, and training data in LLM detection and attribution.