In her keynote at Devoxx UK, Mairi Aitken, Ethics Fellow at the Alan Turing Institute, spoke about the limitations of AI in trying to tackle the complexity of human language, giving her name an AI of Gaelic origin. I gave an example of mispronunciation by . . She began her presentation by explaining that her own work at Allan’s Turing Institute focuses on predicting the social and ethical risks of large-scale language models in society. bottom. Beyond that, she seeks to understand how, when responsibly designed and developed, those risks can be minimized and the value that data and her AI can provide to society at large. increase.
She then gave a quick overview of what ChatGPT is for the very few who have never known it before.
This is a fairly advanced and sophisticated version of the predictive text program. It’s like predictive text on my phone that can’t spell my name correctly.Based on large language models […] It is built on a huge dataset of human languages. They are trained to recognize patterns in the language, mimic human language, and predict what word combinations will result in a persuasive response…
Essentially, the entire internet was used to train ChatGPT. And now, news about AI breakthroughs happens multiple times a day instead of multiple times a year like before the November announcement. From advising on customer support channels, providing legal and medical advice, to haiku and poetry on dating apps, and even writing witty and catchy lines, he has demonstrated his LLM advantages in a variety of industries. The movement to introduce it is progressing rapidly. Another amazing ability is the manipulation of code to generate code or analyze it to detect errors.
Regardless of the industry in which the content is generated, she said, regardless of whether the content is compelling or not, “I have absolutely no idea what those words and structures mean.” […] There is absolutely no way to understand the context or meaning of those words. In other words, it’s all about style, not content. ”
ChatGPT is the foundation for building new systems. LLM is a foundational model for AI. In other words, LLMs are trained on huge data sets, not for a specific task or function, but for general purposes that can be applied to different contexts. After all, a user working with software built on top of a model may not even know what’s under the hood, at its base. She stresses the importance of inspecting foundations like a construction inspector to see if what is being built is fit for purpose before building on top of it.
What construction materials are used?
In AI, models trained on biased or incomplete data produce biased or incomplete results. The same is true for LLMs, ChatGPT was trained on the internet, so we need to ask ourselves how much the internet is a mirror of society. According to Aitken, the Internet’s perspective is a highly distorted view of the world, dominated by powerful economic or political voices. Minority groups are usually even less represented.
Is this model fit for purpose? Is it designed for these specific conditions? Will it withstand extreme events?
Aitken rhetorically asks whether general-purpose foundations can support housing, hospitals, and prisons in the same way as construction. LLM has not been considered for a specific domain, so different sensitivities and risks need to be considered for different domains where LLM is applied. We have to ask ourselves whether it is suitable for the purposes of the particular industry we are interested in. For construction, you need to understand if it can withstand earthquakes, hurricanes, and fires. For LLMs, you have to withstand cyberattacks and malicious actors.
Built using ethical labor practices?
ChatGPT relies on the Moderation API to prevent harmful or dangerous consequences to some extent. However, according to Time magazine, Kenyan workers were exposed to the most harmful content imaginable. These workers worked long hours and unprotected for ‘small pay’ that was labeled as ‘really frightening and really harmful’ to workers.
What is the impact of the model on the environment?
Although invisible, LLM’s impact on the environment is significant. Aitken cites estimates that the amount of CO2 generated by training ChatGPT is equivalent to driving a car to the moon and back. It is training only and does not include modification, maintenance or operation. She encourages us to consider whether it’s really worth it that we’re using it. Or maybe it’s worth the cost we pay just for what we use it for.
AI systems always do what they are programmed to do. We program to do more at a higher level. […] But ChatGPT cannot replace human creativity and problem-solving skills
Towards the end of her keynote, she expressed her fears, proportionate to the enthusiasm around her, that it would make multiple jobs obsolete and, in some cases, even destroy the world. said there is also.Her reaction to both no AI is limited by the context it is given, so even if it produces more impressive content, it is due to the fact that the data on which the AI was trained was impressive in terms of both content and quality. AI cannot replace a person’s problem-solving ability, creativity, or emotional intelligence. Nevertheless, she argues that developers not only have a high responsibility to consider the practical and technical limitations of underlying models, but also ethical and social considerations when taking a responsible approach to engaging with AI. Emphasizes that matters must also be followed. They are supposed to inspect the foundation before construction!
