The Gradient of Generative AI Release: Methods and Considerations

Abstract

As increasingly powerful generative AI systems are developed, the release method greatly varies. We propose a framework to assess six levels of access to generative AI systems: fully closed; gradual or staged access; hosted access; cloud-based or API access; downloadable access; and fully open. Each level, from fully closed to fully open, can be viewed as an option along a gradient. We outline key considerations across this gradient: release methods come with tradeoffs, especially around the tension between concentrating power and mitigating risks. Diverse and multidisciplinary perspectives are needed to examine and mitigate risk in generative AI systems from conception to deployment. We show trends in generative system release over time, noting closedness among large companies for powerful systems and openness among organizations founded on principles of openness. We also enumerate safety controls and guardrails for generative systems and necessary investments to improve future releases.Preprint. Under review.

The AI community building the future? A quantitative analysis of development activity on Hugging Face Hub

“…Following years of debate about the safety of openly releasing AI models [17,18,36,37], recent years have seen the emergence and proliferation of "open" models, which individuals and organisations have shared on an open access basis on platforms such as the HF Hub [4]. Prior to this, AI models, in particular large language models (LLMs), were principally developed and maintained behind closed doors, albeit with open science practices, such as the sharing of publications on arXiv and code on platforms like GitHub.…”

“…The proliferation of open models, especially foundation models, has ignited heated debate about their potential benefits and risks [16][17][18][19][20]43]. On the one hand, open models are said to promise benefits for research, innovation, and competition by lowering entry barriers and widening access to state-of-the-art AI [44].…”

“…While the benefits and risks of open source AI have been widely debated [16][17][18][19][20], the practices and processes involved in open model development have received relatively little attention. To date, only a handful of scholars have explored various aspects of open model development, including user contributions to grassroots initiatives [12,14], commercial participation in model development [14,21], model maintenance practices [22], and the processes and tools used by open data engineering communities [23].…”

Microarchitectural Side-Channel Threats, Weaknesses and Mitigations: A Systematic Mapping Study

“…We observe that overall 4% (5 out of 104) of primary studies belong to this theme spanning across years 2015-2022 and we anticipate this trend to accelerate. Factors like generative AI [37] will certainly be among the forefront of discovering new microarchitecture vulnerabilities and perhaps generate after-fixes on-the-fly with automated reasoning. With sustained miniaturization in chip fabrication and advances in lithographic techniques, we anticipate newer generation of chips with affordable onboard AI hardware being readily available, which could be programmed for any applicationspecific purpose including security.…”

Transductive Learning for Textual Few-Shot Classification in API-based Embedding Models

“…Despite the success of the scaling paradigm, significant challenges still exist especially when the many practical constraints of real-world scenarios have to be met: labeled data can be severely limited (i.e., few-shot scenario (Song et al, 2022;Ye et al, 2021)), data privacy is critical for many industries and has become the subject of increasingly many regulatory pieces (Commission, 2020(Commission, , 2016, compute costs need to be optimized (Strubell et al, 2019). Furthermore, these challenges are made even more complex as stronger foundation models are now available only through APIs (e.g., Ope-nAI's GPT-3, GPT-4 or ChatGPT, Anthropic's Claude or Google's PaLM (Chowdhery et al, 2022)) which has led to some of their parameters being concealed, presenting new challenges for model adaptation (Solaiman, 2023). This paper is centered on the fundamental task of fewshot text classification, specifically focusing on cloud-based/API access.…”