Designing Knowledge Futures: Investigating the impact of Generative AI on the future of Knowledge Work.

生成性人工智能(GAI)及其对知识工作的影响是一个令人着迷的、令人难以置信的及时的主题，我对此非常感兴趣。GAI已经在深刻地影响和增强知识工作。当早期的人工智能工具影响到与日常任务相关的人类劳动领域时，我们开始看到模型体系结构中的新技术发展，例如通用预先训练的变形金刚和扩散模型，以及用户体验设计中，如Dialogic(ChatGPT，Bard等)。和基于文本提示的界面(DALL-E、中途等)，正在深刻地影响非常规的问题解决、认知和创造性劳动。也许其最直接和最明显的影响是增加程序任务，如产生技术或行政文本。然而，人工智能也影响着聚合和发散思维，为问题提供了方便的潜在解决方案，并根据特定标准帮助评估和优化这些解决方案。随着基于简单文本提示生成新奇图像、音乐和创造性文本的工具的出现，人工智能正在极大地影响创造性工作。对于那些与新兴技术动态合作的创造性从业者来说，这些工具展示了创造性过程中令人兴奋的新发展。然而，对于那些与生产技术具有更静态关系的做法来说，这些工具似乎威胁到它们的经济可持续性。有两条线特别有兴趣进行探索。第一种是应用设计研究的正式方法，以及迄今为止通过我的实践展示的方法来评估旨在改善GAI可能对创造性实践者产生的一些负面影响的技术建议。第二，我感兴趣的是我们如何在知识和创造性工作的背景下理解和产生新颖性，以及人工智能系统如何阻碍或增强这一点。我之前曾通过一个名为《未来假阳性》的多渠道视频作品探索过这些问题。这项工作试图说明机器学习系统引导的交互中人类参与减少的含义，以及这种系统所表现出的内在历史偏见--Berardi(2009)和Fisher(2014)将这一概念概念化为“对未来的缓慢取消”，Pasquelli将其称为“过去的独裁”(2020)。这部电影是使用在自动驾驶汽车镜头的多摄像头数据集上训练的下一帧预测算法制作的。这一输出随后通过自动驾驶汽车目标检测算法，在视频之上进行分层分析。最终结果是从训练数据中提取的半连贯视觉形式的组合，并用显示模型置信度分数的边界框进行注释。通过将预测和分类机器学习算法(即试图产生和理解的模型)链接在一起，我试图将人类从循环中移除的效果可视化。

The subject of Generative AI (GAI) and its impact on knowledge work is a fascinating and incredibly timely subject of considerable interest to me.GAI is already profoundly impacting and augmenting knowledge work. Where earlier tranches of AI tools impacted the areas of human labour concerned with routine tasks, we're beginning to see how new technical developments in model architectures, such as the General Pre-trained Transformers and Diffusion Models, and in user experience design, such as dialogic (ChatGPT, Bard, etc.) and text-prompt based interfaces (DALL-E, Midjourney, etc.), are profoundly impacting non-routine problem-solving, cognitive, and creative labour.Perhaps its most immediate and visible impact is in augmenting procedural tasks, such as the production of technical or administrative text. However, AI is also impacting both convergent and divergent thinking, by offering a convenient supply of potential solutions to problems, as well as aiding in the evaluation and optimisation of these solutions based on particular criteria.With the advent of tools to generate novel images, music, and creative text based on simple text prompts, AI is dramatically influencing creative work. For those creative practitioners that work dynamically with emerging technology, these tools present exciting new developments in the creative process. However, for those practices characterised by a more static relationship to production technologies, these tools appear as threats to their economic sustainability.There are two lines of inquiry particularly interested in exploring. The first is applying the formal methods of Research through Design and those exhibited through my practice to date towards the evaluation of technical proposals intended to ameliorate some of the negative impacts that GAI might have on creative practitioners. Systems such as those that identify GAI-produced media (e.g., AI watermarking technology) or those that automatically assign attribution to samples in training datasets based on computed visual similarity scores.Second, I'm interested in how we understand and produce novelty in the context of knowledge and creative work, and how AI systems impede or augment this.I have previously explored these issues through a multi-channel video work entitled "Future False Positive". The work attempts to illustrate the implications of reduced human involvement in interactions guided by machine learning systems, as well as the intrinsic historical bias exhibited by such systems - a concept that Berardi (2009) and Fisher (2014) conceptualise as the "slow cancellation of the future" and Pasquinelli refers to as the "dictatorship of the past" (2020). The film was produced using a Next Frame Prediction algorithm trained on a multi-camera dataset of self-driving car footage. The output of this was then passed through a self-driving car object detection algorithm, layering analysis on top of the video. The final result was a composition of semi-coherent visual forms drawn from the training data, annotated with bounding boxes displaying the confidence scores of the model. By chaining together prediction and classification machine learning algorithms, that is, models that attempt to produce and understand, I am attempting to visualise the effect of removing humans from the loop.