AI coverage focuses on two opposite ends: the end of every job due to AI, or the overselling of AI as a complicated program that fails to replace human tasks. These extremes are detrimental to the thoughtful discussion AI deserves, and are not helpful when figuring out technical and social implications of AI. The following will provide a better perspective to the current AI landscape.
What’s New
While ML is not a new concept, AI has recently shifted to LLMs or Large Language Models. LLMs are text-based models that can generate vast amounts of written text, code, and even images, based on prompts. These models are what resulted in the rapid proliferation of AI because, for the first time ever, the interface is as simple as typing.
Currently, AI tools have shifted from something only data scientists work with, to technology embedded in everyday programs. From email programs drafting messages, to programming tools that auto-fill code, to design software that creates layouts, people are more directly interacting with the technology.
How these models actually get built
There are a few intricate details I will skip on to give you the basics. First, one must feed a language model reams of text. Then it can begin to create patterns pertaining to how language is structured on a statistical level. It learns what text often follows what text and the order in which ideas generally follow. It doesn’t memorize facts like a person would. It’s predicting what the statistically most likely next piece of text will be. Because of this, the language models can appear to be certain of their answers when they are actually very wrong because although our patterns can help us solve problems, we usually fact check.
People using AI tools in professional situations need to take this into account. AI tools are very good in specific contexts where drafting is required, summarizing, and completing repetitive tasks where there is a lot of structure. You should not rely on these tools in a professional context for truth or for anything requiring a fair amount of judgement.
Where AI is actually helpful
Some areas show more promise than others. Coding assistants help developers eliminate the time wasted on repetitive input and coding errors. Image generation and other content creation tools permit smaller teams to create backlogs for content that, prior to AI, would be less accessible. Finally, there are considerable benefits to using AI to abstract data and perform repetitive analytical tasks where humans are unproductive and failed by the AI.
Where it’s not great: creativity and AI do not mesh, nor does novelty problem-solving, time-critical decisions without sufficient context, and high-stakes tasks. Regarding AI-generated materials, it should be regarded as an intermediate/draft answer for validation, not the final material.
What to watch
Application assistants, coding assistants, and the various other generative tools (images, audio, and video) have made incredible advancements recently. Rather than chasing every sub-category tool release, it’s better to make a solid foundation with one or two tools in each of those categories.
The barriers that matter
There’s a lot of focus on the capabilities of AI tools, but not enough on what they can’t do. For instance, many AI models lack awareness of the world around them. Language models can answer most questions, but will give incorrect responses to questions about the news or other time-sensitive information if they do not have access to up-to-date information. Each AI model has a limit to how much context it can consider. That’s why longer, more complicated tasks tend to require more oversight, and not less, as the task grows.
There are also real costs associated that aren’t discussed as much. Using large AI models required extensive computing power, which results in elevated subscription costs and increased environmental impacts from the infrastructure used to run it. Considering these factors is justified as it makes an AI model less appealing for certain tasks that aren’t required.
What to expect going forward…
AI is no longer its own separate tech. Similar to how cloud computing became quietly integrated infrastructure as opposed to a separate service, AI is also starting to integrate as a layer within programs. Instead of asking “will AI replace X” people must ask “how is X going to evolve because AI is now a part of it.” This thinking should be applied during the evaluation of how AI alters journalism instead of answering the increasingly sensationalized and polarizing questions the headlines present.
The accompanying articles will look into specific offerings and what they’re really like to use.