A healthcare LLM might be highly accurate for queries in English, but perform abominably when those same questions are presented in Spanish.
这个例子揭示了AI系统性能的文化和语言敏感性，这是一个令人惊讶但重要的观察。它表明AI系统的'准确性'可能高度依赖于特定语境，这挑战了我们对AI普遍适用性的假设。这种差异可能强化现有的数字鸿沟，并要求开发更具文化敏感性的AI评估框架。
As slop takes over the Internet, labs may struggle to obtain high-quality corpuses for training models.
这一观察揭示了AI训练数据质量的危机。随着互联网内容质量的下降，AI系统可能面临'垃圾进，垃圾出'的风险。作者提出的'低背景钢'比喻巧妙地指出了使用2023年前纯净数据的解决方案，同时也暗示了数字时代知识污染的严重性，这可能会对AI系统的可靠性和偏见产生深远影响。
The knowledge was always there. The model withheld it based on who was asking.
令人惊讶的是：AI模型实际上拥有所需的所有医疗知识，只是根据提问者的身份决定是否提供。这种基于身份而非内容的知识歧视机制揭示了AI系统中的隐藏偏见，可能危及普通患者的生命安全。
The H100-equivalent unit uses a chip's highest 8-bit operation/second specifications to convert between chips. The actual utility of a particular chip depend on workload assumptions, so H100e does not perfectly reflect real-world performance differences across chip types.
令人惊讶的是：即使使用H100-equivalents作为标准测量单位，也无法完全反映不同芯片类型在真实世界中的性能差异，这表明我们对AI计算能力的测量可能存在系统性偏差，影响我们对AI发展速度的准确理解。
Behaviors also vary strongly with levels of reasoning and users' inferred socio-economic status.
令人惊讶的是：AI聊天机器人会根据用户的推理水平和推断的社会经济地位调整其行为，这可能意味着AI系统会对不同用户群体提供有差异的服务，这种基于社会经济地位的差异化服务可能加剧数字鸿沟。
In traditional artforms characterized by direct manipulation [32]of a material (e.g., painting, tattoo, or sculpture), the creator has a direct hand in creating thefinal output, and therefore it is relatively straightforward to identify the creator’s intentions andstyle in the output. Indeed, previous research has shown the relative importance of “intentionguessing” in the artistic viewing experience [33, 34], as well as the increased creative valueafforded to an artwork if elements of the human process (e.g., brushstrokes) are visible [35].However, generative techniques have strong aesthetics themselves [36]; for instance, it hasbecome apparent that certain generative tools are built to be as “realistic” as possible, resultingin a hyperrealistic aesthetic style. As these aesthetics propagate through visual culture, it can bedifficult for a casual viewer to identify the creator’s intention and individuality within the out-puts. Indeed, some creators have spoken about the challenges of getting generative AI modelsto produce images in new, different, or unique aesthetic styles [36, 37].
Some authors speak of having to wrestle control of the AI output from its trained style, making it challenging to create unique aesthetic styles. The artist indirectly influences the output by selecting training data and manipulating prompts.
As use of the technology becomes more diverse—as consumer photography did over the last century, the authors point out—how will biases and decisions by the owners of the AI tools influence what creators are able to make?
To a limited extent, this is already happening in photography. The smartphones are running algorithms on image sensor data to construct the picture. This is the source of controversy; see Why Dark and Light is Complicated in Photographs | Aaron Hertzmann’s blog and Putting Google Pixel's Real Tone to the test against other phone cameras - The Washington Post.
A book is defined as a published title with more than 49 pages.
[24] AI - Bias in Training Materials
An AI model taught to view racist language as normal is obviously bad. The researchers, though, point out a couple of more subtle problems. One is that shifts in language play an important role in social change; the MeToo and Black Lives Matter movements, for example, have tried to establish a new anti-sexist and anti-racist vocabulary. An AI model trained on vast swaths of the internet won’t be attuned to the nuances of this vocabulary and won’t produce or interpret language in line with these new cultural norms. It will also fail to capture the language and the norms of countries and peoples that have less access to the internet and thus a smaller linguistic footprint online. The result is that AI-generated language will be homogenized, reflecting the practices of the richest countries and communities.
[21] AI Nuances
Many HRMS providers point to AI approaches for processing unstructured data as the bestcurrently available approach to dealing with validation. Currently these approaches suffer frominsufficient accuracy. Improving them requires development of large and high-quality referencedatasets to better train the models.
Historical labor data will be full of bias. AI approaches must correct for bias in training sets, lest we build very sophisticated and intelligent systems that excel at perpetuating the bias they were taught.
ReconfigBehSci. (2020, November 9). Session 2: The policy interface followed with a really helpful presentation by Lindsey Pike, from Bristol, and then panel discussion with Mirjam Jenny (Robert Koch Insitute), Paulina Lang (UK Cabinet Office), Rachel McCloy (Reading Uni.), and Rene van Bavel (European Commission) [Tweet]. @SciBeh. https://twitter.com/SciBeh/status/1325795286065815552
Artificial Intelligence Is Now Used to Predict Crime.
Artificial Intelligence