Building a stack

You don’t start with the software. Instead, I would suggest that anyone, seeking to enhance their workflows or creative endeavours with AI applications, should start by writing out (in longhand) a list of tasks they want an AI application to assist with. This could be writing blog posts, or Linkedin posts, learned articles, or drawing up presentations. It could involve needing powerful tools, to undertake large amounts of detailed legal or historical research, or simply using the image generation facility of an LLM, to create colourful pictures or figures, to enliven or draw attention to content. This in turn means that once the job list has been drafted, a decision can be made as to what form of AI “stack” will be best suited, to assist with the work.

An AI “stack” is no more and no less, than a collection of applications that at any one time, you choose to subscribe to, and use, for whatever purposes have brought you within the thickets of the AI jungle. Whilst I suppose it is possible to have a “stack” of just one AI application, that presupposes that you can achieve everything that you want to achieve, to the level that you want to achieve it with a single application. In reality, all the major large language models, whilst similar in concept, each have their strengths and weaknesses. In addition, there are many hundreds of other AI applications, which are narrower in focus, but perhaps greater in depth for particular tasks that you might want to undertake, and so you may add one of more these to the stack.

In the past, I have experimented with the Google suite of applications, and Perplexity, but at the current time, I have the following stack. ChatGPT has pride of place, for a number of reasons. First, it is a well known generalist application. I also note that it is the first AI application I knowingly used, and in a sense I can trace my knowledge and familiarity with AI applications to the day that I started using it. Like all AI applications, it has been progressively improved and in particular its ability to generate images, is, at the current time, very strong. It does not permit the creation of videos, as OpenAI, its maker, has recently discontinued Sora, which was a video maker. But for most of my day to day use, ChatGPT is where I usually start.

Second in the stack, is, perhaps a little controversially, Grok. For those who do not know, Grok is part of Elon Musk’s empire. It has also featured heavily in adverse headlines this year, for its image editing and video creation features, which have been used by bad actors, to create sexualised images of real people. These features were not initially constrained, or given “guardrails” because Grok positions itself, to reflect Musk’s particular take on “free speech” and aligns itself with the X application, and its editorial policy. But the application is also fast, considerably faster, I would say than ChatGPT, has a more irreverent tone, and sharper discernment. It is also manifestly agentic: give it a task, and you can watch in real time as 4 agents get to work on your instructions. The video feature can be used to either make videos de novo, or to animate existing pictures, or even images created by ChatGPT.

Third, I have selected an application called Gamma. Imagine Powerpoint on steroids: it enables you to design and format slides, or write presentations, from prompts or through importing other material, and using prompts to turn it into a presentation. It is fast, powerful and makes it incredibly easy to create beautiful slides, with images or graphical content, in a way that avoids pain. The application is being developed to permit a wider use of its design qualities, expanding them to facilitate the creation of websites and other digital documents. Once created a document such as a presentation can be exported as a PDF or Powerpoint or other file. This application genuinely saves me a great deal of time.

Fourthly, I have not given up on Microsoft Copilot. Although its interface is clunky, and nowhere near as intuitive, or as easy to deploy as other LLMs, Copilot has the massive advantage that it lives within Microsoft Office, and so can be deployed within Word or other programmes, to be used in the context of that particular Office based application. Thus having drafted a skeleton argument or other document, it is possible to use Copilot, not only for spellings and grammar checks, but also to detect and correct internal inconsistencies, to reduce prolix prose into leaner, sharper content, and to improve my drafts.

Lastly, intrigued by the hype/press on Claude, and its Opus model, I have taken out a subscription to Claude. Claude is meant to excell at drafting really long documents, and I have a couple of those in mind. At the moment, Claude is something of a blank canvass. I now have to set it up. Setting up an LLM such as Claude, ChatGPT or Grok, means devising and installing within it, bespoke and specific instructions/style guides as to how I want it to approach particular tasks, how I want it to carry out its work, how I want it to produce text, and a myriad of other possible refinements. The instructions are “always on” and enable the LLM to approach fresh tasks, consistently with other, earlier tasks I have given to it, and to maintain continuity across my work.

Subscriptions to any particular AI application are cheap. Usually hovering about the £20/$20 mark for a monthly subscription. It follows that there is limited downside to taking out an application’s subscription, which unlocks the real abilities of the software, and just playing around with it, or trying to break it, according to your point of view. But the key remains, to determine what tasks you want AI to assist with, and then build an appropriate stack.